Parkinson's disease (PD) significantly affects patients’ quality of life through debilitating motor symptoms, such as Freezing of Gait (FoG). Continuous, in-home monitoring of FoG is essential for timely clinical intervention but remains challenging due to high power consumption, annotation cost, and the controlled environments required by current wearables. We introduce LIFT-PD 1 , a novel self-supervised learning (SSL) framework for real-time, patient-independent FoG detection that uniquely utilizes a single waist-worn accelerometer—an approach traditionally considered less optimal due to weaker gait signatures. LIFT-PD leverages SSL on unlabeled data collected from uncontrolled, real-world settings and employs a novel Differential Hopping Windowing Technique (DHWT) to address gait variability and dataset imbalance. Additionally, an opportunistic inference module selectively activates the deep learning model only during patient movement, significantly reducing power consumption and enabling continuous monitoring ( \(>\) 48 hours). Experimental results show that LIFT-PD achieves a 7.25% increase in precision and 4.4% improvement in accuracy compared to supervised and semi-supervised baseline models while requiring approximately 40% fewer labeled training samples. Evaluations across diverse patient characteristics-including severity, medication state, age, and gender-confirm the model's robustness and clinical applicability, positioning LIFT-PD as a practical, energy-efficient, and scalable solution for continuous real-world FoG monitoring in PD.

FuSE-MET addresses critical challenges in deploying human activity recognition (HAR) systems in uncontrolled environments by effectively managing noisy labels, sparse data, and undefined activity vocabularies. By integrating BERT-based word embeddings with domain-specific knowledge (i.e., MET values), FuSE-MET optimizes label merging, reducing label complexity and improving classification accuracy. Our approach outperforms the state-of-the-art techniques, including ChatGPT-4, by balancing semantic meaning and physical intensity.

Coping with stress is one of the most frequently cited reasons for chronic cannabis use. Therefore, it is hypothesized that cannabis users exhibit distinct physiological stress responses compared to non-users, and that these differences may be especially pronounced during moments of cannabis consumption. However, there is a scarcity of publicly available datasets that allow such hypotheses to be tested under real-world conditions. This paper introduces a dataset named CAN-STRESS, collected using Empatica E4 wristbands. The dataset includes multimodal physiological measurements (such as skin conductance, heart rate, and skin temperature) from 82 participants (39 cannabis users and 43 non-users) as they went about their daily routines. In addition to sensor data, participants provided self-reported survey responses that included perceived stress ratings and timestamps of key daily events such as cannabis use, physical activity, and sleep. To demonstrate the utility of the dataset for downstream applications, we present a preliminary machine learning task aimed at classifying cannabis users versus nonusers based on physiological features. Our model achieves a classification accuracy of approximately 96% and an f1-score of around 98%. An analysis of feature importance using SHAP values revealed that electrodermal activity and heart rate metrics were the most influential predictors, consistent with their established roles in stress detection. We publicly release the CANSTRESS dataset, which we believe serves as a reliable and rich resource for studying the physiological correlates of cannabis use and stress in naturalistic settings.

High-quality real-world datasets are essential for advancing data-driven approaches in type 1 diabetes (T1D) management, including personalized therapy design, digital twin systems, and glucose prediction models. However, progress in this area has been limited by the scarcity of publicly available datasets that offer detailed and comprehensive patient data. To address this gap, we present$A Z T 1 D$, a dataset containing data collected from 25 individuals with T1D on automated insulin delivery (AID) systems. AZT1D includes continuous glucose monitoring (CGM) data, insulin pump and insulin administration data, carbohydrate intake, and device mode (regular, sleep, and exercise) obtained over 6-8 weeks for each patient. Notably, the dataset provides granular details on bolus insulin delivery (i.e., total dose, bolus type, correction-specific amounts) features that are rarely found in existing datasets. By offering rich, naturalistic data, AZT1D supports a wide range of artificial intelligence and machine learning applications aimed at improving clinical decision-making and individualized care in T1D.

Accurate blood glucose forecasting enables proactive management of metabolic health, particularly when leveraging data from wearable sensors that capture data about physiological and behavioral health. However, existing models struggle with integrating multimodal time-series data with inconsistent sampling rates. This paper proposes a novel forecasting framework that incorporates a time-aware cross-attention mechanism with an LSTM architecture to predict blood glucose levels using continuous glucose monitoring (CGM) data alongside physiological and behavioral signals, such as heart rate (HR), electrodermal activity, accelerometry, and dietary intake. The proposed method dynamically encodes temporal features without the need for preprocessing and employs gated multi-head cross-attention layers to fuse sensor modalities effectively. We evaluate our approach on a newly constructed dataset involving$\mathbf{1 2}$participants. Our method outperforms the baseline and state-of-the-art GlySim models across multiple prediction horizons ranging from 5 minutes to 90 minutes, achieving up to 17.8% improvement in Root Mean Squared Error (RMSE) values.

Counterfactual explanations (CFs) offer human-centric insights into machine learning predictions by highlighting minimal changes required to alter an outcome. Therefore, CFs can be used as (i) interventions for abnormality prevention and (ii) augmented data for training robust models. In this work, we explore large language models (LLMs), specifically GPT-4o-mini, for generating CFs in a zero-shot and three-shot setting. We evaluate our approach on two datasets: the AI-Readi flagship dataset for stress prediction and a public dataset for heart disease detection. Compared to traditional methods such as DiCE, CFNOW, and NICE, our few-shot LLM-based approach achieves high plausibility (up to 99%), strong validity (up to 0.99), and competitive sparsity. Moreover, using LLM-generated CFs as augmented samples improves downstream classifier performance (an average accuracy gain of 5%), especially in low-data regimes. This demonstrates the potential of prompt-based generative techniques to enhance explainability and robustness in clinical and physiological prediction tasks. Code base: github.com/shovito66/SenseCF.

Sensor channel selection is an important optimization problem in resource-constrained wearable systems with the goal of identifying an optimal set of input sensors for efficient machine learning. We introduce a framework for this optimization problem, mathematically formulate the minimum-cost channel selection (MCCS), and propose two novel algorithms to solve the problem. Branch and bound channel selection finds a globally optimal channel subset and the greedy channel selection finds the best intermediate subset based on our proposed penalty function. These proposed channel selection algorithms are conditioned with both performance and the cost of the channel subset. We evaluate both algorithms on two publicly available time series datasets for activity recognition and mental task classification. Branch and bound channel selection achieve a cost saving between 92.6% and 95.7%, and the greedy approach reduces the cost between 51.8% and 91.4, % for performance thresholds of 50% and 70%.

Wearable EEG (electroencephalogram) systems have demonstrated potential in epilepsy monitoring, sleep assessment, and determining cognitive workload to improve human decision-making. However, analyzing EEG signals is challenging due to their non-stationary nature and susceptibility to noise. In particular, achieving high accuracy in machine learning tasks requires large amounts of labeled data, which is difficult to obtain due to the time-consuming and labor-intensive nature of data labeling. To address these challenges, we propose a self-supervised learning (SSL) approach for cognitive workload classification using wavelet-based augmentations of EEG signals. First, two augmentations per channel are generated, and their wavelets are computed. The visual representations of these wavelets are then fed to the SSL pretext phase as contrastive pairs to pre-train the model. Finally, the pre-trained model is fine-tuned for workload classification using small amounts of labeled EEG data. Experimental results on the EEG During Mental Arithmetic Tasks (EEGMAT) dataset show that our method outperforms the state-of-the-art supervised models. Notably, our model achieves an accuracy of 99.5% with only 50% of the labeled data, demonstrating the effectiveness of our approach in scenarios with limited labeled data availability. Furthermore, the proposed approach achieves an accuracy of 98.6% in the leave-one-subject-out analysis.

Inter-beat interval (IBI) measurement enables estimation of heart-tare variability (HRV) which, in turn, can provide early indication of potential cardiovascular diseases (CVDs). However, extracting IBIs from noisy signals is challenging since the morphology of the signal gets distorted in the presence of noise. Electrocardiogram (ECG) of a person in heavy motion is highly corrupted with noise, known as motion-artifact, and IBI extracted from it is inaccurate. As a part of remote health monitoring and wearable system development, denoising ECG signals and estimating IBIs correctly from them have become an emerging topic among signal-processing researchers. Apart from conventional methods, deep-learning techniques have been successfully used in signal denoising recently, and diagnosis process has become easier, leading to accuracy levels that were previously unachievable. We propose a deep-learning approach leveraging tiramisu autoencoder model to suppress motion-artifact noise and make the R-peaks of the ECG signal prominent even in the presence of high-intensity motion. After denoising, IBIs are estimated more accurately expediting diagnosis tasks. Results illustrate that our method enables IBI estimation from noisy ECG signals with SNR up to -30 dB with average root mean square error (RMSE) of 13 milliseconds for estimated IBIs. At this noise level, our error percentage remains below 8% and outperforms other state-of-the-art techniques.

Machine learning algorithms are increasingly used for inference and decision-making in embedded systems. Data from sensors are used to train machine learning models for various smart functions of embedded and cyber-physical systems ranging from applications in healthcare, autonomous vehicles, and national security. However, recent studies have shown that machine learning models can be fooled by adding adversarial noise to their inputs. The perturbed inputs are called adversarial examples. Furthermore, adversarial examples designed to fool one machine learning system are also often effective against another system. This property of adversarial examples is called adversarial transferability and has not been explored in wearable systems to date. In this work, we take the first stride in studying adversarial transferability in wearable sensor systems from four viewpoints: (1) transferability between machine learning models; (2) transferability across users/subjects of the embedded system; (3) transferability across sensor body locations; and (4) transferability across datasets used for model training. We present a set of carefully designed experiments to investigate these transferability scenarios. We also propose a threat model describing the interactions of an adversary with the source and target sensor systems in different transferability settings. In most cases, we found high untargeted transferability, whereas targeted transferability success scores varied from 0% to 80%. The transferability of adversarial examples depends on many factors such as the inclusion of data from all subjects, sensor body position, number of samples in the dataset, type of learning algorithm, and the distribution of source and target system dataset. The transferability of adversarial examples decreased sharply when the data distribution of the source and target system became more distinct. We also provide guidelines and suggestions for the community for designing robust sensor systems. Code and dataset used in our analysis is publicly available here. 1

Coping with stress is reportedly one of the main reasons for chronic cannabis use. Developing a real-time system that offers cannabis users alternative methods to cope with stress is of interest in medical applications. To develop such a system, it is necessary to design a reliable mechanism for identifying cannabis use sessions in uncontrolled environments using physiological markers captured with wearable sensors. Therefore, the primary objective of this study is to design a system that can identify sessions of cannabis consumption by utilizing one of the most significant biomarkers of stress, Electrodermal Activity (EDA). We conducted a user study to collect physiological sensor data in real-life setting. We then model the cannabis use detection as a supervised learning problem and train a neural network model. To improve the performance of the proposed model for a specific subject, transfer learning techniques were used to retrain the base model on the new user data. Trained model achieved average f1-score of 0.68 and accuracy of 71.58% on the test data from Leave One Subject Out (LOSO) analysis. After applying transfer learning, the retrained model achieved average f1-score of 0.8 and accuracy of 83.61% when detecting the cannabis consumption period for the same subjects.

Electronic fetal monitoring (EFM) is designed for the early detection of fetal risks and the prevention of serious neurological impairment but suffers from high false positive rates. The Fetal Reserve Index (FRI) is an expert-based system that combines EFM with maternal, obstetrical, and fetal risk factors and displays superior performance in risk detection than EFM alone. Towards translating the FRI into an automated risk prediction system that can make recommendations to clinicians in real-time, we here develop machine learning classifiers that calculate feature importance based on historical data from labor cases and predict the risk of developing neurological impairment. We train random forest and multilayer perceptron (MLP) models to classify abnormal and normal delivery cases and to assess the model performance using a dataset of 1462 labor cases. The random forest classifier achieves a macro average f-1 score of 0.82 with an abnormal case recall of 0.59. Alternatively, MLP classifiers provide higher abnormal case recall at a cost of lower accuracy and macro average f-1 score. Future work will aim to optimize weightings and trade-offs of statistical performance to achieve further improvements for clinical practice.

Regulating blood glucose concentration is crucial for every individual, particularly for patients with diabetes or prediabetes to manage their metabolic health. Poor glucose control results in dysglycemia. Frequent dysglycemia exposure increases the risk of cardiovascular disease, seizures, loss of consciousness, and potentially death. Patients often struggle with glucose control due to a multitude of interrelated behavioral, physiological, and biological factors such as food, insulin intake, and metabolism rate. There is a need for a solution that can accurately predict future adverse dysglycemic events and important parameters such as the area under the glucose curve (AUC). However, current research uses limited input parameters, lacks potential meal-based predictions, is data-hungry and computationally expensive, and predicts a single health outcome. In this research, GlucoseAssist1, a novel, personalized, AI-driven system was developed to predict glucose response and area under the glucose curve in real-time and identify dysglycemic events based on diet, health, and medication data. Importantly, the devised tiered architecture uses a multimodal convolutional neural network and random forest classifier with time series data from a clinical dataset with 20,040 Continuous Glucose Monitor (CGM) records. GlucoseAssist accurately predicts blood glucose response for the next 30 minutes with a Root Mean Squared Error of 1.23, Mean Absolute Error of 0.920, and an accuracy of 97.07% for the identification of dysglycemic events.

Stress monitoring is an important area of research with significant implications for individuals' physical and mental health. We present a data-driven approach for stress detection based on convolutional neural networks while addressing the problems of the best sensor channel and the lack of knowledge about stress episodes. Our work is the first to present an analysis of stress-related sensor data collected in real-world conditions from individuals diagnosed with Alcohol Use Disorder (AUD) and undergoing treatment to abstain from alcohol. We developed polynomial-time sensor channel selection algorithms to determine the best sensor modality for a machine learning task. We model the time variation in stress labels expressed by the participants as the subjective effects of stress. We addressed the subjective nature of stress by determining the optimal input length around stress events with an iterative search algorithm. We found the skin conductance modality to be most indicative of stress, and the segment length of 60 seconds around user-reported stress labels resulted in top stress detection performance. We used both majority undersampling and minority oversampling to balance our dataset. With majority undersampling, the binary stress classification model achieved an average accuracy of 99% and an f1-score of 0.99 on the training and test sets after 5-fold cross-validation. With minority oversampling, the performance on the test set dropped to an average accuracy of 76.25% and an f1-score of 0.68, highlighting the challenges of working with real-world datasets.

Recent years have witnessed a growing body of research on autonomous activity recognition models for use in deployment of mobile systems in new settings such as when a wearable system is adopted by a new user. Current research, however, lacks comprehensive frameworks for transfer learning. Specifically, it lacks the ability to deal with partially available data in new settings. To address these limitations, we proposeOptiMapper, a novel uninformed cross-subject transfer learning framework for activity recognition. OptiMapper is a combinatorial optimization framework that extracts abstract knowledge across subjects and utilizes this knowledge for developing a personalized and accurate activity recognition model in new subjects. To this end, a novel community-detection-based clustering of unlabeled data is proposed that uses the target user data to construct a network of unannotated sensor observations. The clusters of these target observations are then mapped onto the source clusters using a complete bipartite graph model. In the next step, the mapped labels are conditionally fused with the prediction of a base learner to create a personalized and labeled training dataset for the target user. We present two instantiations of OptiMapper. The first instantiation, which is applicable for transfer learning across domains with identical activity labels, performs a one-to-one bipartite mapping between clusters of the source and target users. The second instantiation performs optimal many-to-one mapping between the source clusters and those of the target. The many-to-one mapping allows us to find an optimal mapping even when the target dataset does not contain sufficient instances of all activity classes. We show that this type of cross-domain mapping can be formulated as a transportation problem and solved optimally. We evaluate our transfer learning techniques on several activity recognition datasets. Our results show that the proposed community detection approach can achieve, on average, 69 percent utilization of the datasets for clustering with an overall clustering accuracy of 87.5 percent. Our results also suggest that the proposed transfer learning algorithms can achieve up to 22.5 percent improvement in the activity recognition accuracy, compared to the state-of-the-art techniques. The experimental results also demonstrate high and sustained performance even in presence of partial data.

Postprandial hyperglycemia (PPHG) is detrimental to health and increases risk of cardiovascular diseases, reduced eyesight, and life-threatening conditions like cancer. Detecting PPHG events before they occur can potentially help with providing early interventions. Prior research suggests that PPHG events can be predicted based on information about diet. However, such computational approaches (1) are data hungry requiring significant amounts of data for algorithm training; and (2) work as a black-box and lack interpretability, thus limiting the adoption of these technologies for use in clinical interventions. Motivated by these shortcomings, we propose, DietNudge1, a machine learning based framework that integrates multi-modal data about diet, insulin, and blood glucose to predict PPHG events before they occur. Using data from patients with diabetes, we demonstrate that our model can predict PPHG events with up to 90% classification accuracy and an average F1 score of 0.93. The proposed decision-tree-based approach also identifies modifiable factors that contribute to an impending PPHG event while providing personalized thresholds to prevent such events. Our results suggest that we can develop simple, yet effective, computational algorithms that can be used as preventative mechanisms for diabetes and obesity management.

Physical activity is a cornerstone of chronic conditions and one of the most critical factors in reducing the risks of cardiovascular diseases, the leading cause of death in the United States. App-based lifestyle interventions have been utilized to promote physical activity in people with or at risk for chronic conditions. However, these mHealth tools have remained largely static and do not adapt to the changing behavior of the user. In a step toward designing adaptive interventions, we propose BeWell24Plus, a framework for monitoring activity and user engagement and developing computational models for outcome prediction and intervention design. In particular, we focus on devising algorithms that combine data about physical activity and engagement with the app to predict future physical activity performance. Knowing in advance how active a person is going to be in the next day can help with designing adaptive interventions that help individuals achieve their physical activity goals. Our technique combines the recent history of a person’s physical activity with app engagement metrics such as when, how often, and for how long the app was used to forecast the near future’s activity. We formulate the problem of multimodal activity forecasting and propose an LSTM-based realization of our proposed model architecture, which estimates physical activity outcomes in advance by examining the history of app usage and physical activity of the user. We demonstrate the effectiveness of our forecasting approach using data collected with 58 prediabetic people in a 9-month user study. We show that our multimodal forecasting approach outperforms single-modality forecasting by 2.2% to 11.1% in mean-absolute-error.

Stress detection and classification from wearable sensor data is an emerging area of research with significant implications for individuals’ physical and mental health. In this work, we introduce a new dataset, ADARP, which contains physiological data and self-report outcomes collected in real-world ambulatory settings involving individuals diagnosed with alcohol use disorders. We describe the user study, present details of the dataset, establish the significant correlation between physiological data and self-reported outcomes, demonstrate stress classification, and make our dataset public to facilitate research.

Effective diagnosis of neuro-degenerative diseases is critical to providing early treatments, which in turn can lead to substantial savings in medical costs. Machine learning models can help with the diagnosis of such diseases like Parkinson’s and aid in assessing disease symptoms. This work introduces a novel system that integrates pervasive computing, mobile sensing, and machine learning to classify hand-drawn images and provide diagnostic insights for the screening of Parkinson’s disease patients. We designed a computational framework that combines data augmentation techniques with optimized convolutional neural network design for on-device and real-time image classification. We assess the performance of the proposed system using two datasets of images of Archimedean spirals drawn by hand and demonstrate that our approach achieves 76% and 83% accuracy respectively. Thanks to 4x memory reduction via integer quantization, our system can run fast on an Android smartphone. Our study demonstrates that pervasive computing may offer an inexpensive and effective tool for early diagnosis of Parkinson’s disease1.

Detecting stress from wearable sensor data enables those struggling with unhealthy stress coping mechanisms to better manage their stress. Previous studies have investigated how mechanisms for detecting stress from sensor data can be optimized, comparing alternative algorithms and approaches to find the best possible outcome. One strategy to make these mechanisms more accessible is to reduce the number of sensors that wearable devices must support. Reducing the number of sensors will enable wearable devices to be a smaller size, require less battery, and last longer, making use of these wearable devices more accessible. To progress towards this more convenient stress detection mechanism, we investigate how learning algorithms perform on singular modalities and compare the outcome with results from multiple modalities. We found that singular modalities performed comparably or better than combined modalities on two stress-detection datasets, suggesting that there is promise for detecting stress with fewer sensor requirements. From the four modalities we tested, acceleration, blood volume pulse, and electrodermal activity, we saw acceleration and electrodermal activity to stand out in a few cases, but all modalities showed potential. Our results are acquired from testing with random holdout and leave-one-subject-out validation, using several machine learning techniques. Our results can inspire work on optimizing stress detection with singular modalities to make the benefits of these detection mechanisms more convenient.

We propose a novel active learning framework for activity recognition using wearable sensors. Our work is unique in that it takes limitations of the oracle into account when selecting sensor data for annotation by the oracle. Our approach is inspired by human-beings’ limited capacity to respond to prompts on their mobile device. This capacity constraint is manifested not only in the number of queries that a person can respond to in a given time-frame but also in the time lag between the query issuance and the oracle response. We introduce the notion ofmindful active learningand propose a computational framework, calledEMMA, to maximize the active learning performance taking informativeness of sensor data, query budget, and human memory into account. We formulate this optimization problem, propose an approach to model memory retention, discuss the complexity of the problem, and propose a greedy heuristic to solve the optimization problem. Additionally, we design an approach to perform mindful active learning in batch where multiple sensor observations are selected simultaneously for querying the oracle. We demonstrate the effectiveness of our approach using three publicly available activity datasets and by simulating oracles with various memory strengths. We show that the activity recognition accuracy ranges from 21 to 97 percent depending on memory strength, query budget, and difficulty of the machine learning task. Our results also indicate that EMMA achieves an accuracy level that is, on average, 13.5 percent higher than the case when only informativeness of the sensor data is considered for active learning. Moreover, we show that the performance of our approach is at most 20 percent less than the experimental upper-bound and up to 80 percent higher than the experimental lower-bound. To evaluate the performance of EMMA for batch active learning, we design two instantiations of EMMA to perform active learning in batch mode. We show that these algorithms improve the algorithm training time at the cost of a reduced accuracy in performance. Another finding in our work is that integrating clustering into the process of selecting sensor observations for batch active learning improves the activity learning performance by 11.1 percent on average, mainly due to reducing the redundancy among the selected sensor observations. We observe that mindful active learning is most beneficial when the query budget is small and/or the oracle’s memory is weak. This observation emphasizes advantages of utilizing mindful active learning strategies in mobile health settings that involve interaction with older adults and other populations with cognitive impairments.

Wearables are poised to transform health and wellness through automation of cost-effective, objective, and real-time health monitoring. However, machine learning models for these systems are designed based on labeled data collected, and feature representations engineered, in controlled environments. This approach has limited scalability of wearables because (i) collecting and labeling sufficiently large amounts of sensor data is a labor-intensive and expensive process; and (ii) wearables are deployed in highly dynamic environments of the end-users whose context undergoes consistent changes. We introduce TransNet , a deep learning framework that minimizes the costly process of data labeling, feature engineering, and algorithm retraining by constructing a scalable computational approach. TransNet learns general and reusable features in lower layers of the framework and quickly reconfigures the underlying models from a small number of labeled instances in a new domain, such as when the system is adopted by a new user or when a previously unseen event is to be added to event vocabulary of the system. Utilizing TransNet on four activity datasets, TransNet achieves an average accuracy of 88.1% in cross-subject learning scenarios using only one labeled instance for each activity class. This performance improves to an accuracy of 92.7% with five labeled instances.

Despite the fact that deep neural networks are powerful models and achieve appealing results on many tasks, they are too large to be deployed on edge devices like smartphones or embedded sensor nodes. There have been efforts to compress these networks, and a popular method is knowledge distillation, where a large (teacher) pre-trained network is used to train a smaller (student) network. However, in this paper, we show that the student network performance degrades when the gap between student and teacher is large. Given a fixed student network, one cannot employ an arbitrarily large teacher, or in other words, a teacher can effectively transfer its knowledge to students up to a certain size, not smaller. To alleviate this shortcoming, we introduce multi-step knowledge distillation, which employs an intermediate-sized network (teacher assistant) to bridge the gap between the student and the teacher. Moreover, we study the effect of teacher assistant size and extend the framework to multi-step distillation. Theoretical analysis and extensive experiments on CIFAR-10,100 and ImageNet datasets and on CNN and ResNet architectures substantiate the effectiveness of our proposed approach.

With the recent advances in both machine learning and embedded systems research, the demand to deploy computational models for real-time execution on edge devices has increased substantially. Without deploying computational models on edge devices, the frequent transmission of sensor data to the cloud results in rapid battery draining due to the energy consumption of wireless data transmission. This rapid power dissipation leads to a considerable reduction in the battery lifetime of the system, therefore jeopardizing the real-world utility of smart devices. It is well-established that for difficult machine learning tasks, models with higher performance often require more computation power and thus are not power-efficient choices for deployment on edge devices. However, the trade-offs between performance and power consumption are not well studied. While numerous methods (e.g., model compression) have been developed to obtain an optimal model, these methods focus on improving the efficiency of a "single" model. In an entirely new direction, we introduce an effective method to find a combination of "multiple" models that are optimal in terms of power-efficiency and performance by solving an optimization problem in which both performance and power consumption are taken into account. Experimental results demonstrate that on the ImageNet dataset, we can achieve a 20% energy reduction with only 0.3% accuracy drop compared to Squeeze-and-Excitation Networks. Compared to a pruned convolutional neural network for human activity recognition, while consuming 1.7% less energy, our proposed policy achieves 1.3% higher accuracy.

Catastrophic forgetting affects the training of neural networks, limiting their ability to learn multiple tasks sequentially. From the perspective of the well established plasticity-stability dilemma, neural networks tend to be overly plastic, lacking the stability necessary to prevent the forgetting of previous knowledge, which means that as learning progresses, networks tend to forget previously seen tasks. This phenomenon coined in the continual learning literature, has attracted much attention lately, and several families of approaches have been proposed with different degrees of success. However, there has been limited prior work extensively analyzing the impact that different training regimes -- learning rate, batch size, regularization method-- can have on forgetting. In this work, we depart from the typical approach of altering the learning algorithm to improve stability. Instead, we hypothesize that the geometrical properties of the local minima found for each task play an important role in the overall degree of forgetting. In particular, we study the effect of dropout, learning rate decay, and batch size on forming training regimes that widen the tasks' local minima and consequently, on helping it not to forget catastrophically. Our study provides practical insights to improve stability via simple yet effective techniques that outperform alternative baselines.

We introduce a novel approach for robust estimation of physiological parameters such as interbeat interval (IBI) and heart rate variability (HRV) from cardiac signals captured with wearable sensors in the presence of motion artifacts. Motion artifact due to physical exercise is known as a major source of noise that contributes to a significant decline in the performance of IBI and HRV estimation techniques for cardiac monitoring in free-living environments. Therefore, developing robust estimation algorithms is essential for utilization of wearable sensors in daily life situations. The proposed approach includes two algorithmic components. First, we propose a combinatorial technique to select characteristic points that define heartbeats in noisy signals in time domain. The heartbeat detection problem is defined as a shortest path search problem on a direct acyclic graph that leverages morphological features of the cardiac signals by taking advantage of the time-continuity of heartbeats - each heartbeat ends with the starting point of the next heartbeat. The graph is constructed with vertices and edges representing candidate morphological features and IBIs, respectively. Second, we propose a fusion technique to combine physiological parameters estimated from different morphological features using the shortest path algorithm to obtain more accurate IBI/HRV estimations. We evaluate our techniques on motion-corrupted photoplethysmogram and electrocardiogram signals. Our results indicate that the estimated IBIs are highly correlated with the ground truth (r = 0.89) and detected HRV parameters indicate high correlation with the true HRV parameters. Furthermore, our findings demonstrate that the developed fusion technique, which utilizes different morphological features, achieves a correlation coefficient that is at least 3% higher than that obtained using single physiological characteristic.

Per-core Dynamic Voltage and Frequency (V/F) Scaling (DVFS) is a well-known methodology for achieving energy efficiency in multicore systems. Heuristic DVFS techniques provide fast, suboptimal V/F predictions while Dynamic Programming (DP) methods solve smaller sub-problems iteratively and use their outcomes to evaluate V/F levels globally, but at the cost of overhead delays. We propose an efficient DP framework using the Viterbi algorithm, which uses the Energy-Delay Product (EDP) as an objective function to predict the best V/F levels using applications' profiled information, to minimize energy consumption and execution time. Experimental results show that our framework outperforms heuristics using the EDP criteria and provides near-optimal solutions when maximizing energy saving is as, or more, important than minimizing execution time penalty. In fact, across several benchmarks, our proposed algorithm provides from a 12 to 75 percent improvement in EDP compared to heuristic methods. Furthermore, using a Pareto frontier to evaluate solutions of the algorithms under study, we demonstrate that our framework's energy-time solution is on average only 9 percent worse than the optimal solution. In addition, we show that our dynamic programming solution is 3 to 18 percent closer to a theoretical lower-bound when compared to the studied heuristic methods.

Activity recognition is central to many motion analysis applications ranging from health assessment to gaming. However, the need for obtaining sufficiently large amounts of labeled data has limited the development of personalized activity recognition models. Semi-supervised learning has traditionally been a promising approach in many application domains to alleviate reliance on large amounts of labeled data by learning the label information from a small set of seed labels. Nonetheless, existing approaches perform poorly in highly dynamic settings, such as wearable systems, because some algorithms rely on predefined hyper-parameters or distribution models that needs to be tuned for each user or context. To address these challenges, we introduce LabelForest 1, a novel non-parametric semi-supervised learning framework for activity recognition. LabelForest has two algorithms at its core: (1) a spanning forest algorithm for sample selection and label inference; and (2) a silhouette-based filtering method to finalize label augmentation for machine learning model training. Our thorough analysis on three human activity datasets demonstrate that LabelForest achieves a labeling accuracy of 90.1% in presence of a skewed label distribution in the seed data. Compared to self-training and other sequential learning algorithms, LabelForest achieves up to 56.9% and 175.3% improvement in the accuracy on balanced and unbalanced seed data, respectively.

Advances in embedded systems have enabled integration of many lightweight sensory devices within our daily life. In particular, this trend has given rise to continuous expansion of wearable sensors in a broad range of applications from health and fitness monitoring to social networking and military surveillance. Wearables leverage machine learning techniques to profile behavioral routine of their end-users through activity recognition algorithms. Current research assumes that such machine learning algorithms are trained offline. In reality, however, wearables demand continuous reconfiguration of their computational algorithms due to their highly dynamic operation. Developing a personalized and adaptive machine learning model requires real-time reconfiguration of the model. Due to stringent computation and memory constraints of these embedded sensors, the training/re-training of the computational algorithms need to be memory- and computation-efficient. In this paper, we propose a framework, based on the notion of online learning, for real-time and on-device machine learning training. We propose to transform the activity recognition problem from a multi-class classification problem to a hierarchical model of binary decisions using cascading online binary classifiers. Our results, based on Pegasos online learning, demonstrate that the proposed approach achieves 97% accuracy in detecting activities of varying intensities using a limited memory while power usages of the system is reduced by more than 40%.

Recent advances in machine learning and deep neural networks have led to the realization of many important applications in the area of personalized medicine. Whether it is detecting activities of daily living or analyzing images for cancerous cells, machine learning algorithms have become the dominant choice for such emerging applications. In particular, the state-of-the-art algorithms used for human activity recognition (HAR) using wearable inertial sensors utilize machine learning algorithms to detect health events and to make predictions from sensor data. Currently, however, there remains a gap in research on whether or not and how activity recognition algorithms may become the subject of adversarial attacks. In this paper, we take the first strides on (1) investigating methods of generating adversarial example in the context of HAR systems; (2) studying the vulnerability of activity recognition models to adversarial examples in feature and signal domain; and (3) investigating the effects of adversarial training on HAR systems. We introduce Adar11Software code and experimental data for Adar are available online at https://github.com/rameshKrSah/Adar., a novel computational framework for optimization-driven creation of adversarial examples in sensor-based activity recognition systems. Through extensive analysis based on real sensor data collected with human subjects, we found that simple evasion attacks are able to decrease the accuracy of a deep neural network from 95.1% to 3.4% and from 93.1% to 16.8% in the case of a convolutional neural network. With adversarial training, the robustness of the deep neural network increased on the adversarial examples by 49.1% in the worst case while the accuracy on clean samples decreased by 13.2%.

Datacenters provide high performance and flexibility for users and cost efficiency for operators. Hyperscale datacenters are harnessing massively scalable computer resources for large-scale data analysis. However, cloud/datacenter infrastructure does not scale as fast as the input data volume and computational requirements of big data and analytics technologies. Thus, more applications need to share CPU at the node level that could have large impact on performance and operational cost. To address this challenge, in this paper we show that, concurrently fine-tune parameters at the application, microarchitecture, and system levels are creating opportunities to co-locate applications at the node level and improve energy-efficiency of the server while maintaining performance. Co-locating and self-tuning of unknown applications are challenging problems, especially when co-locating multiple big data applications concurrently with many tuning knobs, potentially requiring exhaustive brute-force search to find the right settings. This research challenge upsurges an imminent need to develop a technique that co-locates applications at a node level and predict the optimal system, architecture and application level configure parameters to achieve the maximum energy efficiency. It promotes the scale-down of computational nodes by presenting the Energy-Efficient Co-Locating and Self-Tuning (ECoST) technique for data intensive applications. ECoST proof of concept was successfully tested on MapReduce platform. ECoST can also be deployed on other data-intensive frameworks where there are several parameters for power and performance tuning optimizations. ECoST collects run-time hardware performance counter data and implements various machine learning models from as simple as a lookup table or decision tree based to as complex as neural network based to predict the energy-efficiency of co-located applications. Experimental data show energy efficiency is achieved within 4% of the upper bound results when co-locating multiple applications at a node level. ECoST is also scalable, being within 8% of upper bound on an 8-node server.

We propose a novel active learning framework for activity recognition using wearable sensors. Our work is unique in that it takes physical and cognitive limitations of the oracle into account when selecting sensor data to be annotated by the oracle. Our approach is inspired by human-beings' limited capacity to respond to external stimulus such as responding to a prompt on their mobile devices. This capacity constraint is manifested not only in the number of queries that a person can respond to in a given time-frame but also in the lag between the time that a query is made and when it is responded to. We introduce the notion of mindful active learning and propose a computational framework, called EMMA, to maximize the active learning performance taking informativeness of sensor data, query budget, and human memory into account. We formulate this optimization problem, propose an approach to model memory retention, discuss complexity of the problem, and propose a greedy heuristic to solve the problem. We demonstrate the effectiveness of our approach on three publicly available datasets and by simulating oracles with various memory strengths. We show that the activity recognition accuracy ranges from 21% to 97% depending on memory strength, query budget, and difficulty of the machine learning task. Our results also indicate that EMMA achieves an accuracy level that is, on average, 13.5% higher than the case when only informativeness of the sensor data is considered for active learning. Additionally, we show that the performance of our approach is at most 20% less than experimental upper-bound and up to 80% higher than experimental lower-bound. We observe that mindful active learning is most beneficial when query budget is small and/or oracle's memory is weak, thus emphasizing contributions of our work in human-centered mobile health settings and for elderly with cognitive impairments.

We propose two optimization techniques to minimize memory usage and computation while meeting system timing constraints for real-time classification in wearable systems. Our method derives a hierarchical classifier structure for Support Vector Machine (SVM) in order to reduce the amount of computations, based on the probability distribution of output classes occurrences. Also, we propose a memory optimization technique based on SVM parameters, which results in storing fewer support vectors and as a result requiring less memory. To demonstrate the efficiency of our proposed techniques, we performed an activity recognition experiment and were able to save up to 35% and 56% in memory storage when classifying 14 and 6 different activities, respectively. In addition, we demonstrated that there is a trade-off between accuracy of classification and memory savings, which can be controlled based on application requirements.

Lifestyle interventions with the focus on diet are crucial in self-management and prevention of many chronic conditions, such as obesity, cardiovascular disease, diabetes, and cancer. Such interventions require a diet monitoring approach to estimate overall dietary composition and energy intake. Although wearable sensors have been used to estimate eating context (e.g., food type and eating time), accurate monitoring of dietary intake has remained a challenging problem. In particular, because monitoring dietary intake is a self-administered task that involves the end-user to record or report their nutrition intake, current diet monitoring technologies are prone to measurement errors related to challenges of human memory, estimation, and bias. New approaches based on mobile devices have been proposed to facilitate the process of dietary intake recording. These technologies require individuals to use mobile devices such as smartphones to record nutrition intake by either entering text or taking images of the food. Such approaches, however, suffer from errors due to low adherence to technology adoption and time sensitivity to the dietary intake context. In this article, we introduce EZNutriPal , 1 an interactive diet monitoring system that operates on unstructured mobile data such as speech and free-text to facilitate dietary recording, real-time prompting, and personalized nutrition monitoring. EZNutriPal features a natural language processing unit that learns incrementally to add user-specific nutrition data and rules to the system. To prevent missing data that are required for dietary monitoring (e.g., calorie intake estimation), EZNutriPal devises an interactive operating mode that prompts the end-user to complete missing data in real-time. Additionally, we propose a combinatorial optimization approach to identify the most appropriate pairs of food names and food quantities in complex input sentences. We evaluate the performance of EZNutriPal using real data collected from 23 human subjects who participated in two user studies conducted in 13 days each. The results demonstrate that EZNutriPal achieves an accuracy of 89.7% in calorie intake estimation. We also assess the impacts of the incremental training and interactive prompting technologies on the accuracy of nutrient intake estimation and show that incremental training and interactive prompting improve the performance of diet monitoring by 49.6% and 29.1%, respectively, compared to a system without such computing units.

Activity recognition, as an important component of behavioral monitoring and intervention, has attracted enormous attention, especially in Mobile Cloud Computing (MCC) and Remote Health Monitoring (RHM) paradigms. While recently resource constrained wearable devices have been gaining popularity, their battery life is limited and constrained by the frequent wireless transmission of data to more computationally powerful back-ends. This paper proposes an ultra-low power activity recognition system using a novel adaptive compressed sensing technique that aims to minimize transmission costs. Coarse-grained on-body sensor localization and unsupervised clustering modules are devised to autonomously reconfigure the compressed sensing module for further power saving. We perform a thorough heuristic optimization using Grammatical Evolution (GE) to ensure minimal computation overhead of the proposed methodology. Our evaluation on a real-world dataset and a low power wearable sensing node demonstrates that our approach can reduce the energy consumption of the wireless data transmission up to 81.2 and 61.5 percent, with up to 60.6 and 35.0 percent overall power savings in comparison with baseline and a naive state-of-the-art approaches, respectively. These solutions lead to an average activity recognition accuracy of 89.0 percent-only 4.8 percent less than the baseline accuracy-while having a negligible energy overhead of on-node computation.

Wearable sensors utilize machine learning algorithms to infer important events such as the behavioral routine and health status of their end users from time-series sensor data. A major obstacle in large-scale utilization of these systems is that the machine learning algorithms cannot be shared among users or reused in contexts different from the setting in which the training data are collected. As a result, the algorithms need to be retrained from scratch in new sensor contexts, such as when the on-body location of the wearable sensor changes or when the system is utilized by a new user. The retraining process places significant burden on end users and system designers to collect and label large amounts of training sensor data. In this article, we challenge the current algorithm training paradigm and introduce Share-n-Learn to automatically detect and learn physical sensor contexts from a repository of shared expert models without collecting any new labeled training data. Share-n-Learn enables system designers and end users to seamlessly share and reuse machine learning algorithms that are trained under different contexts and data collection settings. We develop algorithms to autonomously identify sensor contexts and propose a gating function to automatically activate the most accurate machine learning model among the set of shared expert models. We assess the performance of Share-n-Learn for activity recognition when a dynamic sensor constantly migrates from one body location to another. Our analysis based on real data collected with human subjects on three datasets demonstrates that Share-n-Learn achieves, on average, 68.4% accuracy in detecting physical activities with context-varying wearables. This accuracy performance is about 19% more than ‘majority voting,’ 10% more than the state-of-the-art transfer learning, and only 8% less than the experimental upper bound.

A major barrier to the personalized Human Activity Recognition using wearable sensors is that the performance of the recognition model drops significantly upon adoption of the system by new users or changes in physical/behavioral status of users. Therefore, the model needs to be retrained by collecting new labeled data in the new context. In this study, we develop a transfer learning framework using convolutional neural networks to build a personalized activity recognition model with minimal user supervision.

With the flourishing development of body sensor networks, a variety of head-worn sensor-based devices have emerged in many domains, to facilitate applications involving head movements. This paper explores the potential of using head-mounted sensors coupled with computational algorithms, to assess visual field defects through analyzing head motion in reading activities. Visual field defects, such as homonymous hemianopia, is a common disorder that occurs after stroke, injury, or vascular brain damage. A customized reading experiment is conducted on 17 participants, while Google Glass is used for head motion monitoring and visual field defect simulation. The results show a 6%-10% drop in reading performance with the simulated condition. Several machine learnig algorithms demonstrate the distinguishability of head motion in reading activities for visual field defect, with an average accuracy of 91%. Furthermore, experiment results suggest that the difference in head motion between normal and impaired visual field is less significant under extreme reading conditions.

Emerging big data applications heavily rely on machine learning algorithms which are computationally intensive. To meet computational requirements, and power and scalability challenges, FPGA based Hardware accelerators have found their way in data centers and cloud infrastructures. Recent efforts on HW acceleration of big data mainly attempt to accelerate a particular application and deploy it on a specific architecture that fits well its performance and power requirements. Given the diversity of architectures and ML applications, the important research question is which architecture is better suited to meet the performance, power and energy-efficiency requirements of a diverse range of ML-based analytics applications. In this work, we answer this question by investigating how the type of FPGA (low-end vs. high-end), and its integration with the CPU (on-chip vs. off-chip) along with the choice of CPU (high performance big vs. low power little servers) affects the speedup yield and power reduction in a CPU+FPGA architecture for machine learning applications implemented in MapReduce. We show that among the three architectural parameters, the type of CPU is the most dominant factor in determining the execution time and power in a CPU+FPGA architecture for MapReduce applications. The integration technology and FPGA type comes next, with the power and performance least sensitive to the FPGA type.

Diet and physical activity are known as important lifestyle factors in self-management and prevention of many chronic diseases. Mobile sensors such as accelerometers have been used to measure physical activity or detect eating time. In many intervention studies, however, stringent monitoring of overall dietary composition and energy intake is needed. Currently, such a monitoring relies on self-reported data by either entering text or taking an image that represents food intake. These approaches suffer from limitations such as low adherence in technology adoption and time sensitivity to the diet intake context. In order to address these limitations, we introduce development and validation of Speech2Health, a voice-based mobile nutrition monitoring system that devises speech processing, natural language processing (NLP), and text mining techniques in a unified platform to facilitate nutrition monitoring. After converting the spoken data to text, nutrition-specific data are identified within the text using an NLP-based approach that combines standard NLP with our introduced pattern mapping technique. We then develop a tiered matching algorithm to search the food name in our nutrition database and accurately compute calorie intake values. We evaluate Speech2Health using real data collected with 30 participants. Our experimental results show that Speech2Health achieves an accuracy of 92.2% in computing calorie intake. Furthermore, our user study demonstrates that Speech2Health achieves significantly higher scores on technology adoption metrics compared to text-based and image-based nutrition monitoring. Our research demonstrates that new sensor modalities such as voice can be used either standalone or as a complementary source of information to existing modalities to improve the accuracy and acceptability of mobile health technologies for dietary composition monitoring.

Wearable technologies play a central role in human-centered Internet-of-Things applications. Wearables leverage machine learning algorithms to detect events of interest such as physical activities and medical complications. A major obstacle in large-scale utilization of current wearables is that their computational algorithms need to be re-built from scratch upon any changes in the configuration. Retraining of these algorithms requires significant amount of labeled training data, a process that is labor-intensive and time-consuming. We propose an approach for automatic retraining of the machine learning algorithms in real-time without need for any labeled training data. We measure the inherent correlation between observations made by an old sensor view for which trained algorithms exist and the new sensor view for which an algorithm needs to be developed. Our multi-view learning approach can be used in both online and batch modes. By applying the autonomous multi-view learning in the batch mode, we achieve an accuracy of 83.7 percent in activity recognition which is an improvement of 9.3 percent due to the automatic labeling of the data in the new sensor node. In addition to gain the less computation advantage of incremental training, the online learning algorithm results in an accuracy of 82.2 percent in activity recognition.

Power consumption is identified as one of the main complications in designing practical wearable systems, mainly due to their stringent resource limitations. When designing wearable technologies, several system-level design choices, which directly contribute to the energy consumption of these systems, must be considered. In this article, we propose a computationally lightweight system optimization framework that trades off power consumption and performance in connected wearable motion sensors. While existing approaches exclusively focus on one or a few hand-picked design variables, our framework holistically finds the optimal power-performance solution with respect to the specified application need. Our design tackles a multi-variant non-convex optimization problem that is theoretically hard to solve. To decrease the complexity, we propose a smoothing function that reduces this optimization to a convex problem. The reduced optimization is then solved in linear time using a devised derivative-free optimization approach, namely cyclic coordinate search. We evaluate our framework against several holistic optimization baselines using a real-world wearable activity recognition dataset. We minimize the energy consumption for various activity-recognition performance thresholds ranging from 40% to 80% and demonstrate up to 64% energy savings.

With the proliferation of wearable devices in the Internet-of-Things applications, designing highly power-efficient solutions for continuous operation of these technologies in life-critical settings emerges. We propose a novel ultra-low power framework for adaptive compressed sensing in activity recognition. The proposed design uses a coarse-grained activity recognition module to adaptively tune the compressed sensing module for minimized sensing/transmission costs. We pose an optimization problem to minimize activity-specific sensing rates and introduce a polynomial time approximation algorithm using a novel heuristic dynamic optimization tree. Our evaluations on real-world data shows that the proposed autonomous framework is capable of generating feedback with -80% confidence and improves power reduction performance of the state-of-the-art approach by a factor of two.

As the second leading cause of blindness worldwide, glaucoma impacts a large population of individuals over 40. Although visual acuity often remains unaffected in early stages of the disease, visual field loss, expressed by tunnel vision condition, gradually increases. Glaucoma often remains undetected until it has moved into advanced stages. In this paper, we introduce a wearable system for automatic tunnel vision detection using head-mounted sensors and machine learning techniques. We develop several tasks, including reading and observation, and estimate visual field loss by analyzing user's head movements while performing the tasks. An integrated computational module takes sensor signals as input, passes the data through several automatic data processing phases, and returns a final result by merging task-level predictions. For validation purposes, a series of experiments is conducted with 10 participants using tunnel vision simulators. Our results demonstrate that the proposed system can detect mild and moderate tunnel visions with an accuracy of 93.3% using a leave-one-subject-out analysis.

Wearable monitoring devices for ubiquitous health care are becoming a reality that has to deal with limited battery autonomy. Several researchers focus their efforts in reducing the energy consumption of these motes: from efficient micro-architectures, to on-node data processing techniques. In this paper we focus in the optimization of the energy consumption of monitoring devices for the prediction of symptomatic events in chronic diseases in real time. To do this, we have developed an optimization methodology that incorporates information of several sources of energy consumption: the running code for prediction, and the sensors for data acquisition. As a result of our methodology, we are able to improve the energy consumption of the computing process up to 90% with a minimal impact on accuracy. The proposed optimization methodology can be applied to any prediction modeling scheme to introduce the concept of energy efficiency. In this work we test the framework using Grammatical Evolutionary algorithms in the prediction of chronic migraines.

Developing machine learning algorithms for applications of Internet-of-Things requires collecting a large amount of labeled training data, which is an expensive and labor-intensive process. Upon a minor change in the context, for example utilization by a new user, the model will need re-training to maintain the initial performance. To address this problem, we propose a graph model and an unsupervised label transfer algorithm (learn-on-the-go) which exploits the relations between source and target user data to develop a highly-accurate and scalable machine learning model. Our analysis on real-world data demonstrates 54% and 22% performance improvement against baseline and state-of-the-art solutions, respectively.

One of major challenges in chronic disease self-management is the lack of medication adherence. Despite the proliferation of mobile technologies, the potential of using pervasive computing solutions for improved medication management has remained almost unexplored. In this paper, we present a smart-phone based system capable of delivering adaptive activity-aware medication reminders by learning the user's activity of daily living and detecting the most appropriate and effective timing for medication reminders centered around the initial user-specified schedule.

We present a novel approach for monitoring beverage intake. Our system is composed of an ultrasonic sensor, an RGB color sensor, and machine learning algorithms. The system not only measures beverage volume but also detects beverage types. The sensor unit is lightweight that can be mounted on the lid of any drinking bottle. Our experimental results demonstrate that the proposed approach achieves more than 97% accuracy in beverage type classification. Furthermore, our regression-based volume measurement has a nominal error of 3%.

Wearable technologies play a central role in human-centered Internet-of-Things applications. Wearables leverage machine learning algorithms to detect events of interest such as physical activities and medical complications. These algorithms, however, need to be retrained upon any changes in configuration of the system, such as addition/ removal of a sensor to/ from the network or displacement/ misplacement/ mis-orientation of the physical sensors on the body. We challenge this retraining model by stimulating the vision of autonomous learning with the goal of eliminating the labor-intensive, time-consuming, and highly expensive process of collecting labeled training data in dynamic environments. We propose an approach for autonomous retraining of the machine learning algorithms in real-time without need for any new labeled training data. We focus on a dynamic setting where new sensors are added to the system and worn on various body locations. We capture the inherent correlation between observations made by a static sensor view for which trained algorithms exist and the new dynamic sensor views for which an algorithm needs to be developed. By applying our real-time dynamic-view autonomous learning approach, we achieve an average accuracy of 81.1% in activity recognition using three experimental datasets. This amount of accuracy represents more than 13.8% improvement in the accuracy due to the automatic labeling of the sensor data in the newly added sensor. This performance is only 11.2% lower than the experimental upper bound where labeled training data are collected with the new sensor.

Designing remote health monitoring systems requires a multi-faceted perspective that takes into account requirements and contexts imposed by the medical application, technology, and end-user. We study such a design perspective in the context of remote and real-time edema monitoring. Edema (accumulation of fluid in certain soft-tissues) is regarded as one of the most important symptoms for systematic diseases such as heart failure. Monitoring edema allows patients and caregivers to understand the state of sickness and effectiveness of the treatments. This article proposes a novel low-power context-aware and real-time wearable platform capable of continuous assessment of ankle edema in remote settings. Our system keeps track of changes in subject's ankle circumference as well as current body posture. An examination of our system with 15 subjects demonstrates the effectiveness and reliability of the proposed force-sensitive-resistor-based edema sensor (with an R2of 0.87 for our regression model and intraclass correlation of 0.97) as well as an over 96 percent accuracy in activity monitoring that provide the means to perform reliable data validation on ankle circumference measurements in a continuous manner. Furthermore, we devise a novel derivative-free power optimization approach to maximize the battery lifetime resulting in improvement in battery lifetime by a factor of 2.13.

Wearables have emerged as a revolutionary technology in many application domains including healthcare and fitness. Machine learning algorithms, which form the core intelligence of wearables, traditionally deduce a computational model from a set of training examples to detect events of interest (e.g. activity type). However, in the dynamic environment in which wearables typically operate in, the accuracy of a computational model drops whenever changes in configuration of the system (such as device type and sensor orientation) occur. Therefore, there is a need to develop systems which can adapt to the new configuration autonomously. In this paper, using transfer learning as an organizing principle, we develop several algorithms for data mapping. The data mapping algorithms employ effective signal similarity methods and are used to adapt the system to the new configuration. We demonstrate the efficacy of the data mapping algorithms using a publicly available dataset on human activity recognition.

Wearable technologies play a central role in human-centered Internet-of-Things applications. Wearables leverage computational and machine learning algorithms to detect events of interest such as physical activities and medical complications. A major obstacle in large-scale utilization of current wearables is that their computational algorithms need to be re-built from scratch upon any changes in the configuration of the network. Retraining of these algorithms requires significant amount of labeled training data, a process that is labor-intensive, time-consuming, and infeasible. We propose an approach for automatic retraining of the machine learning algorithms in real-time without need for any labeled training data. We measure the inherent correlation between observations made by an old sensor view for which trained algorithms exist and the new sensor view for which an algorithm needs to be developed. By applying our real-time multi-view autonomous learning approach, we achieve an accuracy of 80.66% in activity recognition, which is an improvement of 15.96% in the accuracy due to the automatic labeling of the data in the new sensor node. This performance is only 7.96% lower than the experimental upper bound where labeled training data are collected with the new sensor.

Human-centered Internet-of-Things (IoT) applications utilize computational algorithms such as machine learning and signal processing techniques to infer knowledge about important events such as physical activities and medical complications. The inference is typically based on data collected with wearable sensors or those embedded in the environment. A major obstacle in large-scale utilization of these systems is that the computational algorithms cannot be shared between users or reused in contexts different than the setting in which the training data are collected. For example, an activity recognition algorithm trained for a wrist-band sensor cannot be used on a smartphone worn on the waist. We propose an approach for automatic detection of physical sensor-contexts (e.g., on-body sensor location) without need for collecting new labeled training data. Our techniques enable system designers and end-users to share and reuse computational algorithms that are trained under different contexts and data collection settings. We develop a framework to autonomously identify sensor-context. We propose a gating function to automatically activate the most accurate computational algorithm among a set of shared expert models. Our analysis based on real data collected with human subjects while performing 12 physical activities demonstrate that the accuracy of our multi-view learning is only 7.9% less than the experimental upper bound for activity recognition using a dynamic sensor constantly migrating from one on-body location to another. We also compare our approach with several mixture-of-experts models and transfer learning techniques and demonstrate that our approach outperforms algorithms in both categories.

Power consumption is a major obstacle in designing stringent resource constraint wearables. Several system-level design considerations contribute to energy consumption of these systems which must be taken into account while designing the system. We propose a power-performance optimization framework, namely CyHOP (Cyclic and Holistic Optimization framework), for connected wearable motion sensors. While existing work focus solely on one design parameter, our approach globally trades-off the performance of activity recognition and power consumption. CyHOP is capable of optimally adjusting the system to fulfill specific application needs. Using a smoothing technique, the initial multi-variate non-convex optimization problem is reduced to a convex problem and solved using our devised derivative-free optimization approach, namely, cyclic coordinate search. Our model performs a linear search by cycling through the system variables on each iteration until it converges to the global optimum. Using real-world data collected with wearable motion sensors during activity monitoring, we validate our approached with various performance thresholds ranging from 40% to 80%.

The utility of wearables is currently limited to lab experiments and controlled environments mainly because computational algorithms embedded in wearables fail to produce accurate measurements in uncontrolled, dynamically changing, and potentially harsh environments. With the exponentially growing adoption of these systems in human-centered Internet-of-Things (IoT) applications, development of resource-efficient solutions to enhance the accuracy of this systems remains a considerable research challenge. In this paper, we introduce an energy-efficient framework for uncertainty management of networked wearables. The core components of our framework are anomaly screening units for detecting anomalies that require handling, thus resulting in one order of magnitude less energy consumption compared to the conventional frameworks. Furthermore, our screening approach achieves 98.3% accuracy in detecting anomalies based on real data collected with wearable motion sensors.

Wearables are being widely utilized in health and wellness applications, primarily due to the recent advances in sensor and wireless communication, which enhance the promise of wearable systems in providing continuous and real-time monitoring and interventions. Wearables are generally composed of hardware/software components for collection, processing, and communication of physiological data. Practical implementation of wearable monitoring in real-life applications is currently limited due to notable obstacles. The wearability and form factor are dominated by the amount of energy needed for sensing, processing, and communication. In this article, we propose an ultra-low-power granular decision-making architecture, also called screening classifier, which can be viewed as a tiered wake-up circuitry, consuming three orders of magnitude-less power than the state-of-the-art low-power microcontrollers. This processing model operates based on computationally simple template matching modules, based on coarse- to fine-grained analysis of the signals with on-demand and gradually increasing the processing power consumption. Initial template matching rejects signals that are clearly not of interest from the signal processing chain, keeping the rest of processing blocks idle. If the signal is likely of interest, the sensitivity and the power of the template matching modules are gradually increased, and ultimately, the main processing unit is activated. We pose optimization techniques to efficiently split a full template into smaller bins, called mini-templates, and activate only a subset of bins during each classification decision. Our experimental results on real data show that this signal screening model reduces power consumption of the processing architecture by a factor of 70% while the sensitivity of detection remains at least 80%.

Metabolic equivalent of task (MET) indicates the intensity of physical activities. This measurement is used in providing physical activity intervention in many chronic illnesses such as coronary heart disease, type-2 diabetes, and cancer. Due to the small size, portability, low power consumption, and low cost, wearable motion sensors are widely used to estimate MET values. However, one major obstacle in widespread adoption of current wearable monitoring systems is that the sensors must be worn on predefined locations on the body. This imposes much discomfort for users as they are not allowed to wear the sensors on their own desired body locations. In addition, non-adherence to the predefined location of the sensors results in significant reduction in the accuracy of physical activity monitoring. In this paper, we propose a framework for sensor location-independent MET estimation. We introduce a sensor localization approach that allows users to wear the sensors on different body locations without having to adhere to a specific installation protocol. We study how such an algorithm impacts the performance of MET estimation algorithms. Using daily physical activity data, we demonstrate that an automatic sensor localization algorithm decreases the estimation error of the MET calculation by a factor of 2.3 compared to the case without sensor localization. Furthermore, our sensor localization algorithm achieves an accuracy of 90.8% in detecting on-body locations of wearable sensors. The integration of sensor localization and MET estimation achieves an accuracy of 80% in calculating the MET values of daily physical activities.

Biometric gait analysis using wearable sensors offers an objective and quantitative method for gait parameter extraction. However, current techniques are constrained to specific platform parameters, and hence significantly lack generality, scalability and sustainability. In this paper, we propose a platform-independent and self-adaptive approach for gait cycle detection and cadence estimation. Our algorithm utilizes physical kinematic properties and cyclic patterns of foot acceleration signals to automatically adjust internal parameters of the algorithm. As a result, the proposed approach is robust to noise and changes in sensor platform parameters such as sampling rate and sensor resolution. For the evaluation purpose, we use acceleration signals collected from 16 subjects in a clinical setting to examine the accuracy and robustness of the proposed algorithm. The results show that our approach achieves a precision above 98% and a recall above 95% in stride detection, and an average accuracy of 98% in cadence estimation under various uncertainty conditions such as noisy signals and changes in sampling frequency and sensor resolution.

Wireless health systems enable remote and continuous monitoring of individuals, with applications in elderly care support, chronic disease management, and preventive care. The underlying sensing platform provides constructs that consider the quality of information driven from the system and ensure the reliability/validity of the outcomes to support the decision-making processes. In this paper, we present an approach to integrate contextual information within the data processing flow in order to improve the quality of measurements. We focus on a pilot application that uses wearable motion sensors to calculate metabolic equivalent of task (MET) of exergaming movements. Exergames need to show energy expenditure values, often using accelerometer approximations applied to general activities. We focus on two contextual factors, namely “activity type” and “sensor location,” and demonstrate how these factors can be used to enhance the measured values, since allocating larger weights to more informative sensors can improve the final measurements. Further, designing regression models for each activity provides better results than any generalized model. Indeed, the averaged R2 value for the movements using simple sensor location improve from a general 0.71 to as high as 0.84 for an individual activity type. The different methods present a range of R2 value averages across activity type from 0.64 for sensor location to 0.89 for multidimensional regression, with an average game play MET value of 7.93. Finally, in a leaveone-subject-out cross validation, a mean absolute error of 2.231 METs is found when predicting the activity levels using the best models.

Remote health monitoring (RHM) has emerged as a solution to help reduce the cost burden of unhealthy lifestyles and aging populations. Enhancing compliance to prescribed medical regimens is an essential challenge to many systems, even those using smartphone technology. In this paper, we provide a technique to improve smartphone battery consumption and examine the effects of smartphone battery lifetime on compliance, in an attempt to enhance users' adherence to remote monitoring systems. We deploy WANDA-CVD, an RHM system for patients at risk of cardiovascular disease (CVD), using a wearable smartphone for detection of physical activity. We tested the battery optimization technique in an in-lab pilot study and validated its effects on compliance in the Women's Heart Health Study. The battery optimization technique enhanced the battery lifetime by 192% on average, resulting in a 53% increase in compliance in the study. A system like WANDA-CVD can help increase smartphone battery lifetime for RHM systems monitoring physical activity.

Wearable sensory devices are becoming the enabling technology for many applications in healthcare and well-being, where computational elements are tightly coupled with the human body to monitor specific events about their subjects. Classification algorithms are the most commonly used machine learning modules that detect events of interest in these systems. The use of accurate and resource-efficient classification algorithms is of key importance because wearable nodes operate on limited resources on one hand and intend to recognize critical events (e.g., falls) on the other hand. These algorithms are used to map statistical features extracted from physiological signals onto different states such as health status of a patient or type of activity performed by a subject. Conventionally selected features may lead to rapid battery depletion, mainly due to the absence of computing complexity criterion while selecting prominent features. In this paper, we introduce the notion of power-aware feature selection, which aims at minimizing energy consumption of the data processing for classification applications such as action recognition. Our approach takes into consideration the energy cost of individual features that are calculated in real-time. A graph model is introduced to represent correlation and computing complexity of the features. The problem is formulated using integer programming and a greedy approximation is presented to select the features in a power-efficient manner. Experimental results on thirty channels of activity data collected from real subjects demonstrate that our approach can significantly reduce energy consumption of the computing module, resulting in more than 30 percent energy savings while achieving 96.7 percent classification accuracy.

Monitoring human movements using wireless wearable sensors finds applications in a variety of domains including healthcare and wellness. In these systems, sensory devices are tightly integrated with the human body and infer status of the user through signal and information processing. Typically, highly accurate observations can be made at the cost of deploying a sufficiently large number of sensors, which in turn results in increased energy consumption of the system and reduced adherence to using the system. Therefore, optimizing power consumption of the system while maintaining acceptable accuracy plays a crucial role in realizing these stringent resource constraint systems. In this paper, we present an activity monitoring approach that minimizes power consumption of the system subject to a lower bound on the classification accuracy. The system utilizes computationally simple template-matching blocks that perform classifications on individual sensor nodes. The system further employs a boosting approach to enhance accuracy of the distributed classifier by selecting a subset of sensors optimized in terms of power consumption and capable of achieving a given lower bound accuracy criterion. A proof-of-concept evaluation with three participants performing 14 transitional actions was conducted, where collected signals were segmented and labeled manually for each action. The results indicated that the proposed approach provides more than a 65% reduction in the power consumption of the signal processing, while maintaining 80% sensitivity in classifying human movements.

Exergaming is expanding as an option for sedentary behavior in childhood/adult obesity and for extra exercise for gamers. This paper presents the development process for a mobile active sports exergame with near-realistic motions through the usage of body-wearable sensors. The process begins by collecting a dataset specifically targeted to mapping real-world activities directly to the games, then, developing the recognition system in a fashion to produce an enjoyable game. The classification algorithm in this paper has precision and recall of 77% and 77% respectively, compared with 40% and 19% precision and recall on current activity monitoring algorithms intended for general daily living activities. Aside from classification, the user experience must be strong enough to be a successful system for adoption. Indeed, fast and intense activities as well as competitive, multiplayer environments make for a successful, enjoyable exergame. This enjoyment is evaluated through a 30 person user study. Multiple aspects of the exergaming user experience trials have been merged into a comprehensive survey, called ExerSurvey. All but one user thought the motions in the game were realistic and difficult to cheat. Ultimately, a game with near-realistic motions was shown to be an enjoyable, active video exergame for any environment.

Hyperexcitability in hand is a disorder characterized by exaggerated muscle movement, and is a common symptom associated with neuro-degenerative diseases and spinal cord injuries. Current assessment methods for hyperexcitability rely on subjective examination, or on methods that evaluate the overall hand grip performance without particularization in the excitation. This paper introduces a system that utilizes an inexpensive body sensor device combined with a series of signal processing units that extract information specifically related to physiological phenomena generated by hyperexcitability. A clinical cohort study has been conducted on nine patients with cervical spinal cord injuries (mean age 58.2 ± 13.5). The experimental results show that the proposed signal processing mechanism accurately detects and analyzes the body signal. The medical significance of the experimental results is also investigated. This opens up a new opportunity for patients and clinical professionals to obtain accurate feedback of patient's motor function in an economical and ubiquitous manner.

Advances in technology have led to the development of wearable sensing, computing, and communication devices that can be woven into the physical environment of our daily lives, enabling a large variety of new applications in several domains, including wellness and health care. Despite their tremendous potential to impact our lives, wearable health monitoring systems face a number of hurdles to become a reality. The enabling processors and architectures demand a large amount of energy, requiring sizable batteries. In this article, we propose a granular decision-making architecture for physical movement monitoring applications. The module can be viewed as a tiered wake-up circuitry. This decision-making module, in combination with a low-power microcontroller, allows for significant power saving through an ultra low-power processing architecture. The significant power saving is achieved by performing a preliminary ultra low-power signal processing, and hence, keeping the microcontroller off when the incoming signal is not of interest. The preliminary signal processing is performed by a set of special-purpose functional units, also called screening blocks, that implement template matching functions. We formulate and solve an optimization problem for selecting screening blocks such that the accuracy requirements of the signal processing are accommodated while the total power is minimized. Our experimental results on real data from wearable motion sensors show that the proposed algorithm achieves 63.2% energy saving while maintaining a sensitivity of 94.3% in recognizing transitional actions.

With the growing cost associated with the diagnosis and treatment of chronic neuro-degenerative diseases, the design and development of portable monitoring systems becomes essential. Such portable systems will allow for early diagnosis of motor function ability and provide new insight into the physical characteristics of ailment condition. This paper introduces a highly mobile and inexpensive monitoring system to quantify upper-limb performance for patients with movement disorders. With respect to the data analysis, we first present an approach to quantify general motor performance using the introduced sensing hardware. Next, we propose an ailment-based analysis which employs a significant-feature identification algorithm to perform cross-patient data analysis and classification. The efficacy of the proposed framework is demonstrated using real data collected through a clinical trial. The results show that the system can be utilized as a preliminary diagnostic tool to inspect the level of hand-movement performance. The ailment-based analysis performs an intergroup comparison of physiological signals for cerebral vascular accident (CVA) patients, chronic inflammatory demyelinating polyneuropathy (CIDP) patients, and healthy individuals. The system can classify each patient group with an accuracy of up to 95.00% and 91.42% for CVA and CIDP, respectively.

Many wearable embedded systems benefit from classification algorithms where statistical features extracted from physiological signals are mapped onto different user's states such as health status of a patient or type of activity performed by a subject. Conventionally selected features lead to rapid battery depletion in these battery-operated systems, mainly due to the absence of computing complexity criterion while selecting prominent features. In this paper, we introduce the notion of power-aware feature selection, which minimizes energy consumption of the signal processing for classification applications. Our approach takes into consideration the energy cost of individual features that are calculated in real-time. The problem is formulated using integer programming and a greedy approximation is presented to select the features in a power-efficient manner. Experimental results on thirty channels of activity data demonstrate that our approach can significantly reduce energy consumption of the computing module resulting in more than 30$% energy savings while achieving 96.7% classification accuracy.

Recent years have witnessed a large influx of applications in the field of cyber-physical systems. An important class of these systems is body sensor networks (BSNs) where lightweight embedded processors and communication systems are tightly coupled with the human body. BSNs can provide researchers, care providers and clinicians access to tremendously valuable information extracted from data that are collected in users' natural environment. With this information, one can monitor the progression of a disease, identify its early onset, or simply assess user's wellness. One major obstacle is managing repositories that store the large amount of sensing data. To address this issue, we propose a data mining approach inspired by the experience in the areas of text and natural language processing. We represent sensor readings with a sequence of characters, called motion transcripts. Transcripts reduce complexity of the data significantly while maintaining morphological and structural properties of the physiological signals. To further take advantage of the physiological signal's structure, our data mining technique focuses on the characteristic transitions in the signals. These transitions are efficiently captured using the concept ofn-grams. To facilitate a lightweight and fast mining approach, we reduce the overwhelmingly large number ofn-grams via information gain (IG) feature selection. We report the effectiveness of the proposed approach in terms of the speed of mining while maintaining an acceptable accuracy in terms of the F-score combining both precision and recall.

One important application of body sensor networks is action recognition. Action recognition often implicitly requires partitioning sensor data into intervals, then labeling the partitions according to the action that each represents or as a non-action. The temporal partitioning stage is called segmentation, and the labeling is called classification. While many effective methods exist for classification, segmentation remains problematic. We present a technique inspired by continuous speech recognition that combines segmentation and classification using hidden Markov models. This technique is distributed across several sensor nodes. We show the results of this technique and the bandwidth savings over full data transmission.

Cardiotocography is used to monitor the effect of contractions on fetal heart rate (FHR) and is used in the final trimester of pregnancy to detect fetal distress. This simple and efficient (patent-pending) architecture for Cardiotocography employs the Doppler technique to measure Fetal Heart Rate (FHR) and is scalable to any number of FHR monitors. The Doppler circuitry is designed for low cost and low power consumption to facilitate portability and compatibility with wireless data transmission. Measured results indicate performance consistent with conventional bedside systems, i.e. 8 bits of resolution with update rates ≥2 samples per second.

Advances in technology have led to development of wearable sensing, computing and communication devices that can be woven into the physical environment of our daily lives, enabling a large variety of new applications in several domains including wellness and health care. Despite their tremendous potential to impact our lives, wearable health monitoring systems face a number of hurdles to become a reality. The enabling processors and architectures demand a large amount of energy, requiring sizable batteries. In this paper, we propose a granular decision making architecture that can be viewed as a tiered wake up circuitry. This module, in combination with a low-power microcontroller, enables an ultra low-power architecture. The significant power saving is achieved by performing a preliminary ultra low-power signal processing and hence, keeping the microcontroller off when the incoming signal is not of interest. The preliminary signal processing is performed by a set of special purpose functional units, also called screening blocks, that implements template matching functions. We formulate and solve an optimization problem to select screening blocks such that the accuracy requirements of the signal processing are accommodated while the total power is minimized. Our experimental results on real data from wearable motion sensors show that the proposed algorithm achieves 65.2% energy saving while maintaining 92.7% sensitivity in recognizing human movements.

Monitoring human activities using wearable sensor nodes has the potential to enable many useful applications for everyday situations. Limited computation, battery lifetime and communication bandwidth make efficient use of these platforms crucial. In this paper, we introduce a novel classification model that identifies physical movements from body-worn inertial sensors while taking collaborative nature and limited resources of the system into consideration. Our action recognition model uses a decision tree structure to minimize the number of nodes involved in classification of each action. The decision tree is constructed based on the quality of action recognition in individual nodes. A clustering technique is employed to group similar actions and measure quality of per-node identifications. We pose an optimization problem for finding a minimal set of sensor nodes contributing to the action recognition. We then prove that this problem is NP-hard and provide fast greedy algorithms to approximate the solution. Finally, we demonstrate the effectiveness of our distributed algorithm on data collected from five healthy subjects. In particular, our system achieves a 72.4% reduction in the number of active nodes while maintaining 93.3% classification accuracy.

Body sensor networks are emerging as a promising platform for remote human monitoring. With the aim of extracting bio-kinematic parameters from distributed body-worn sensors, these systems require collaboration of sensor nodes to obtain relevant information from an overwhelmingly large volume of data. Clearly, efficient data reduction techniques and distributed signal processing algorithms are needed. In this paper, we present a data processing technique that constructs motion transcripts from inertial sensors and identifies human movements by taking collaboration between the nodes into consideration. Transcripts of basic motions, called primitives, are built to reduce the complexity of the sensor data. This model leads to a distributed algorithm for segmentation and action recognition. We demonstrate the effectiveness of our framework using data collected from five normal subjects performing ten transitional movements. The results clearly illustrate the effectiveness of our framework. In particular, we obtain a classification accuracy of 84.13% with only one sensor node involved in the classification process.

Body sensor networks (BSNs) have proved their viability to greatly improve quality of medical care by providing continuous and in-home monitoring solutions. Highly constrained nature of the platform demands a design that efficiently utilizes limited resources of the system. Energy optimization techniques are especially desirable as the system lifetime is constrained by small batteries that power sensor nodes in a BSN. In this paper, we introduce a novel data-centering routing model to minimize communication energy, taking collaborative nature of signal processing for healthcare applications into consideration. Transmission energy for a path is determined as a compromise between the path length and the amount of data being transmitted along the path. Data produced by different nodes are aggregated to form packets of large size that consume smaller energy per bit. We formulate the problem as a minimum concave cost multicommodity flow problem and propose two approaches to find both optimal and approximate solutions. We evaluate performance of our energy minimization techniques on a variety of synthesized signal processing task graphs, as well as a real application for evaluating human postural control system. The results show an average of 35% energy saving with our proposed routing against a simple shortest path approach.

Monitoring human movements using wireless sensory devices promises to revolutionize the delivery of healthcare services. Such platforms use inertial information of their subjects for motion analysis. Potentially, each action or disease can be discovered by collaborative processing of sensor data from multiple locations on the body. This functionality is provided by a Body Sensor Network (BSN), which consists of several wireless sensor nodes positioned on different parts of the body. In spite of the revolutionary potential of this platform, power requirements and wearability have limited the commercialization of these systems. In this paper, we present an energy-efficient communication model for BSN applications which uses buffers to limit communication to short bursts, decreasing power usage and simplifying the communication. We formulate an optimization problem to reduce transmissions among sensor nodes and present an ILP-based solution and a fast greedy heuristic algorithm. We show that despite the decreased transmission efficiency, our greedy algorithm can be adopted for fast allocation of buffers in real-time. We experimentally compare the performance of both of the proposed approaches to the performance of an unbuffered system. Our results demonstrate that ILP and greedy solutions can reduce the amount of transmissions by an average factor of 70 and 41, respectively.

The evaluation of the postural control system (PCS) has applications in rehabilitation, sports medicine, gait analysis, fall detection, and diagnosis of many diseases associated with a reduction in balance ability. Standing involves significant muscle use to maintain balance, making standing balance a good indicator of the health of the PCS. Inertial sensor systems have been used to quantify standing balance by assessing displacement of the center of mass, resulting in several standardized measures. Electromyogram (EMG) sensors directly measure the muscle control signals. Despite strong evidence of the potential of muscle activity for balance evaluation, less study has been done on extracting unique features from EMG data that express balance abnormalities. In this paper, we present machine learning and statistical techniques to extract parameters from EMG sensors placed on the tibialis anterior and gastrocnemius muscles, which show a strong correlation to the standard parameters extracted from accelerometer data. This novel interpretation of the neuromuscular system provides a unique method of assessing human balance based on EMG signals. In order to verify the effectiveness of the introduced features in measuring postural sway, we conduct several classification tests that operate on the EMG features and predict significance of different balance measures.

Mobile sensor-based systems are emerging as promising platforms for healthcare monitoring. An important goal of these systems is to extract physiological information about the subject wearing the network. Such information can be used for life logging, quality of life measures, fall detection, extraction of contextual information, and many other applications. Data collected by these sensor nodes are overwhelming, and hence, an efficient data processing technique is essential. In this paper, we present a system using inexpensive, off-the-shelf inertial sensor nodes that constructs motion transcripts from biomedical signals and identifies movements by taking collaboration between the nodes into consideration. Transcripts are built of motion primitives and aim to reduce the complexity of the original data. We then label each primitive with a unique symbol and generate a sequence of symbols, known as motion template, representing a particular action. This model leads to a distributed algorithm for action recognition using edit distance with respect to motion templates. The algorithm reduces the number of active nodes during every classification decision. We present our results using data collected from five normal subjects performing transitional movements. The results clearly illustrate the effectiveness of our framework. In particular, we obtain a classification accuracy of 84.13% with only one sensor node involved in the classification process.

Body sensor networks are emerging as a promising platform for healthcare monitoring. These systems are composed of battery-operated embedded devices which process physiological data. The reduction in the power consumption is an important factor to increase the lifetime for such systems and to enhance their wearability through reducing the size of the battery. In this paper, we develop an energy-efficient communication scheme that uses buffers to reduce the number of transmissions among the sensor nodes constrained to limited hardware resources. A direct acyclic graph is used to model the information flow. We define a communication optimization problem and solve it using convex optimization techniques. We present results that support the efficiency of the proposed technique.

Advances in technology have led to the development of various light-weight sensor devices that can be woven into the physical environment of our daily lives. Such systems enable on-body and mobile health-care monitoring. Our interest particularly lies in the area of movement-monitoring platforms that operate with inertial sensors. In this paper, we introduce the notion of compatibility graphs and describe how they can be utilized for power optimization. We first formulate an action coverage problem that will consider the sensing coverage from a collaborative signal processing perspective. Our solution is capable of eliminating redundant sensor nodes while maintaining the quality of service. The problem we outline can be transformed into an NP-hard problem. Therefore, we propose an ILP formulation to attain a lower bound on the solution and a fast greedy technique. Moreover, we present a system for dynamically activating and deactivating sensor nodes in real time. We then use our graph representation to develop an efficient formulation for maximum lifetime. This formulation provides sufficient information for finding activation duties for each sensor node. Finally, we demonstrate the effectiveness of our techniques on data collected from several subjects.

This article presents an efficient technique for placement and routing of sensors/actuators and processing units in a grid network. The driver application that we present is a medical jacket, which requires an extremely high level of robustness and fault tolerance. The power consumption of such jacket is another key technological constraint. Our proposed interconnection network is a mesh of wires. A jacket made of fabric and wires would be susceptible to accidental damage via tears. By modeling the tears, we evaluate the probability of having failures on every segment of wires in our mesh interconnection network. Then, we study two problems of placement and routing in the sensor networks such that the fault tolerance is maximized while the power consumption is minimized. We develop efficient integer linear programming (ILP) formulations to address these problems and perform both placement and routing, simultaneously. This ensures that the solution is a lower bound for both problems. We evaluate the effectiveness of our proposed techniques on a variety of benchmarks.

Human movement models often divide movements into parts. In walking, the stride can be segmented into four different parts, and in golf and other sports, the swing is divided into sections based on the primary direction of motion. These parts are often divided based on key events, also called temporal parameters. When analyzing a movement, it is important to correctly locate these key events, and so automated techniques are needed. There exist many methods for dividing specific actions using data from specific sensors, but for new sensors or sensing positions, new techniques must be developed. We introduce a generic method for temporal parameter extraction called the hidden Markov event model based on hidden Markov models. Our method constrains the state structure to facilitate precise location of key events. This method can be quickly adapted to new movements and new sensors/sensor placements. Furthermore, it generalizes well to subjects not used for training. A multiobjective optimization technique using genetic algorithms is applied to decrease error and increase cross-subject generalizability. Further, collaborative techniques are explored. We validate this method on a walking dataset by using inertial sensors placed on various locations on a human body. Our technique is designed to be computationally complex for training, but computationally simple at runtime to allow deployment on resource-constrained sensor nodes.

Advances in technology have led to the development of various light-weight sensory devices that can be woven into the physical environment of our daily lives. Such systems enable on-body and mobile health-care monitoring. Our interest particularly lies in the area of movement monitoring platforms that operate with inertial sensors. In this paper, we propose a power optimization technique that will consider the sensing coverage problem from a collaborative signal processing perspective. We introduce compatibility graphs and describe how they can be utilized for power optimization. The problem we outline can be transformed into an NP-hard problem. Therefore, we propose an ILP formulation to attain a lower bound on the solution and a fast greedy technique. Along side this, we introduce a system for dynamically activating and deactivating sensor nodes in real-time. Finally, we elucidate the effectiveness of our techniques on data collected from several subjects.

Body sensor networks (BSNs) create an enormous opportunity to revolutionalize the way we learn, work, entertain and live today. A particularly promising application of BSNs is in health monitoring. Research indicates that various disorders in aging ranging from mild cognitive impairment to dementia and Alzheimer’s could be diagnosed early based on the study of locomotion. Sensor platforms integrated into clothing provide the possibility of reliable locomotion monitoring. In this work, we demonstrate a real-time wireless sensor system that quantitatively measures some of the factors involved in locomotion.

Monitoring human activities using wearable wireless sensor nodes has the potential to enable many useful applications for everyday situations. The deployment of a compact and computationally efficient grammatical representation of actions reduces the complexities involved in the detection and recognition of human behaviors in a distributed system. In this paper, we introduce a road map to a linguistic framework for the symbolic representation of inertial information for physical movement monitoring. Our method for creating phonetic descriptions consists of constructing primitives across the network and assigning certain primitives to each movement. Our technique exploits the notion of a decision tree to identify atomic actions corresponding to every given movement. We pose an optimization problem for the fast identification of primitives. We then prove that this problem is NP-Complete and provide a fast greedy algorithm to approximate the solution. Finally, we demonstrate the effectiveness of our phonetic model on data collected from three subjects.

Transaction level modeling allows exploring several SoC design architectures leading to better performance and easier verification of the final product. In this paper, we present an approach for design and verification of transaction level models. Verification is integrated as part of the design-flow. In the proposed method, we first model the design in UML. Then, we translate it into the reactive objects language, Rebeca (Marjan Sirjani et al., 2004), which is an actor-based language with formal foundation. A model in Rebeca is a set of concurrently executed reactive objects (called rebecs) interacted by asynchronous message passing. After mapping UML to Rebeca, Rebeca code will be translated into Promela which is a language for formal verification. Checking the correctness of the design is performed on-the-fly with the LTL properties using the SPIN model checker. Finally, we translate the verified design to SystemC and map the properties to a set of assertions that can be re-used to validate the design at lower levels through simulation.