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.

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%.

Consumer hearable technologies such as earbuds are increasingly embedding physiological sensors, including photoplethysmography (PPG) and inertial measurements. They create unique opportunities to passively monitor stress and deliver digital interventions such as music. However, PPG signals recorded from ear canals are often very noisy due to head movement and fit issues. This work proposes algorithms to estimate heart rate variability (HRV) features from noisy PPG signals recorded using earbuds. We have used template matching to determine the signal quality for dynamic fine filtering around the estimated heart rate. We have also improved the inter-beat interval (IBI) outlier detection and removal algorithm using the global-local context of the input PPG signal. The mean absolute error of estimating RMSSD decreased from 70.83 milliseconds (ms) to 24.88 ms, and SDNN decreased from 46.89 ms to 16.60 ms.

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.

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.

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.

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.

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%.