Contact-free sensing gained much traction in the past decade. While remote monitoring of some parameters (heart rate) is approaching clinical levels of precision, others remain challenging (blood pressure). We investigated the feasibility of estimating blood pressure (BP) via pulse transit time (PTT) in a novel remote single-site manner, using a modified RGB camera. A narrow-band triple band-pass filter allowed us to measure the PTT between different skin layers, harvesting information from green and near-infrared wavelengths. The filter minimizes the inter-channel influence and band overlap, however, some overlap remains within the filter bands. We further resolve this using a color-channel model and a novel channel-separation method. Using the proposed setup and algorithm, we obtained multi-wavelength (MW) PTTs in an experiment inducing BP changes to 9 subjects. The results showed good absolute Pearson’s correlation coefficient between both MW PTT and systolic BP (R = 0.61, p = 0.08) as well as diastolic BP (R = 0.54, p = 0.05), pointing to feasibility of the proposed novel remote MW BP estimation via PTT. This was further confirmed in a leave-one-subject-out experiment, where a simple Random Forest regression model achieved mean absolute errors of 3.59 and 2.63 mmHg for systolic and diastolic BP respectively.

Virtual Reality (VR) has become a mainstream opportunity to develop immersive environments where people are exposed to relevant stimuli in a systematic but safe way. Besides the controlled exposure to stimuli, VR systems enable a more efficient and unbiased gathering of information with practical applications in diagnosis, intervention, and health monitoring processes. Nonetheless, scientific literature has demonstrated that interacting with VR technology and its peripherals can be challenging, especially in populations less familiar with technology (e.g., older adults) or with clinical conditions that are not adequately accommodated by the technology, such as cognitive (e.g., dementia) or physical impairments (e.g., blindness). This workshop will focus on systems, technological approaches, and theoretical frameworks to develop user/patient-centred VR tools for diagnosis, intervention, and monitoring of health and wellbeing outcomes. Moreover, we welcome work focused on accessibility issues, aiming to promote equal access and opportunity to use VR tools by people with diverse abilities.

An important problem in the field of context recognition is preserving the battery life of the sensing device. In this work we introduce cost-sensitive learning that takes into account the cost of having the sensors active and tries to balance it against the classification error. The benefits can be amplified by adapting the classifier to each context. This approach was tested on two real-life datasets where we achieved 78% and 80% reduction in energy consumption in exchange for almost no accuracy loss.

The recognition of the user's context with wearable sensing systems is a common problem in ubiquitous computing. However, the typically small battery of such systems often makes continuous recognition impractical. An efficient method to reduce the strain on the battery is to employ duty-cycling – periodically turning sensors on and off. Its benefits increase if the duration of those periods is tailored to each context. In this work we present a general mathematical model to predict the effect of different duty-cycle parameters on the system and discuss ways of selecting suitable ones. The methodology was tested on a real-life dataset where it accurately predicted the system performance and found duty-cycle parameters that were better than those found with expert knowledge alone.

We present the Fit4Work system for monitoring and management of physical, mental and environmental stress at the workplace. The system was designed specifically for older workers who are subject to sedentary stressful work in an office environment. It uses commercially available devices and intelligent methods, which utilize machine-learning models to monitor the three aspects of the users' lifestyle, and provide recommendations for improving them. The results show that the system can adequately recognize the user's physical activities, estimate energy expenditure and detect mental stress, as well as recognize and reason about unhealthy environment. The system provides recommendations according to the monitoring results.

Abstract Refrigeration systems have been a vital component of our lives for more than a century. Apart from storing food, they are used to store sensitive goods such as pharmaceutical products or reactive chemicals. The deterioration of the refrigeration system performance due to aging or malfunction directly affects the quality of stored goods. Therefore, an early detection of deviation in performance is an important task. This paper presents a system that monitors operation of refrigeration devices and alerts the user to possible irregularities in the operation run. The emphasis is on recognition of gradual changes of performance that indicate upcoming hardware problems. The system consists of 2 modules: human‐defined expert rules and machine learning. The machine‐learning module learns to recognize abnormal behaviour of devices automatically. Furthermore, it can distinguish between different abnormal events and allow the user to classify some of the types as normal, so that they not longer raise an alarm. The machine learning was evaluated by comparing its recognition of abnormal events and classification accuracy of such events to the performance of a human operator. The system can in principle be adapted to any electronic device that periodically applies some system for sustaining a predefined quality (e.g., temperature).

In this paper we propose a method for continuous stress monitoring using data provided by a commercial wrist device equipped with common physiological sensors and an accelerometer. The method consists of three machine-learning components: a laboratory stress-detector that detects short-term stress every 2 minutes; an activity recognizer that continuously recognizes user's activity and thus provides context information; and a context-based stress detector that first aggregates the predictions of the laboratory detector, and then exploits the user's context in order to provide the final decision in a 20 minute interval. The method was trained on 21 subjects in a laboratory setting and tested on 5 subjects in a real-life setting. The accuracy on 55 days of real-life data was 92%. The method is currently being implemented as a smartphone application, which will be demonstrated at the conference.

We present the "e-Gibalec" system, which was designed to stimulate and motivate schoolchildren to be physically active and to assist physical education (PE) teachers and parents to track their progress. The system consists of a smartphone application which uses built-in sensors to monitor activities, and a web application for PE teachers and parents. We present an overview of the system and its functionality. We also provide an initial evaluation of the accuracy of the monitoring obtained on 10 schoolchildren, and their subjective evaluation of the mobile application, which is quite positive.

BACKGROUND: Understanding the interactions between antibodies and the linear epitopes that they recognize is an important task in the study of immunological diseases. We present a novel computational method for the design of linear epitopes of specified binding affinity to Intravenous Immunoglobulin (IVIg). RESULTS: We show that the method, called Pythia-design can accurately design peptides with both high-binding affinity and low binding affinity to IVIg. To show this, we experimentally constructed and tested the computationally constructed designs. We further show experimentally that these designed peptides are more accurate that those produced by a recent method for the same task. Pythia-design is based on combining random walks with an ensemble of probabilistic support vector machines (SVM) classifiers, and we show that it produces a diverse set of designed peptides, an important property to develop robust sets of candidates for construction. We show that by combining Pythia-design and the method of (PloS ONE 6(8):23616, 2011), we are able to produce an even more accurate collection of designed peptides. Analysis of the experimental validation of Pythia-design peptides indicates that binding of IVIg is favored by epitopes that contain trypthophan and cysteine. CONCLUSIONS: Our method, Pythia-design, is able to generate a diverse set of binding and non-binding peptides, and its designs have been experimentally shown to be accurate.

This paper presents an approach to designing a method for the estimation of human energy expenditure (EE). The approach first evaluates different sensors and their combinations. After that, multiple regression models are trained utilizing data from different sensors. The EE estimation method designed in this way was evaluated on a dataset containing a wide range of activities. It was compared against three competing state-of-the-art approaches, including the BodyMedia Fit armband, the leading consumer EE estimation device. The results show that the proposed method outperforms the competition by up to 10.2 percentage points.

This paper presents an approach to detecting perceived stress in students using data collected with smartphones. The goal is to develop a machine-learning model that can unobtrusively detect the stress level in students using data from several smartphone sources: accelerometers, audio recorder, GPS, Wi-Fi, call log and light sensor. From these, features were constructed describing the students' deviation from usual behaviour. As ground truth, we used the data obtained from stress level questionnaires with three possible stress levels: "Not stressed", "Slightly stressed" and "Stressed". Several machine learning approaches were tested: a general models for all the students, models for cluster of similar students, and student-specific models. Our findings show that the perceived stress is highly subjective and that only person-specific models are substantially better than the baseline.

Energy demand in a smart grid is directly related to energy consumption, as defined by user needs and comfort experience. This article presents a multi-agent architecture for smart control of space heating and cooling processes, in an attempt to enable flexible ways of monitoring and adjusting energy supply and demand. In this proposed system, control agents are implemented in order to perform temperature set-point delegation for heating and cooling systems in a building, offering a means to observe and learn from both the environment and the occupant. Operation of the proposed algorithms is compared with traditional algorithms utilized for room heating, using a simulated model of a residential building and real data about user behaviour. The results show (i) the performance of machine learning for the occupancy forecasting problem and for the problem of calculating the time to heat or cool a room; and (ii) the performance of the control algorithms, with respect to energy consumption and occupant comfort. The proposed control agents make it possible to significantly improve an occupant comfort with a relatively small increase in energy consumption, compared with simple control strategies that always maintain predefined temperatures. The findings enable the smart grid to anticipate the energy needs of the building.

Activity recognition with a single accelerometer placed on the torso is fairly common task, but distinguishing standing from sitting in this way is very difficult because the torso is oriented the same way during both activities, and the transition between the two is very hard to classify into going down or up. We propose a novel approach based on the Multiple Contexts Ensemble (MCE) algorithm which classifies the activity with an ensemble of classifiers, each of which considers the problem in the context of a single feature. The improvement stems from using multiple viewpoints, based on accelerometer data only, designed specifically to distinguish standing from sitting. This approach improves the accuracy on the two activities by 24 percentage points compared to regular machine learning.

We propose a multi-objective machine learning approach guaranteed to find the Pareto optimal set of hybrid classification models consisting of comprehensible and incomprehensible submodels. The algorithm run-times are below 1 s for typical applications despite the exponential worst-case time complexity. The user chooses the model with the best comprehensibility-accuracy trade-off from the Pareto front which enables a well informed decision or repeats finding new Pareto fronts with modified seeds. For a classification trees as the comprehensible seed, the hybrids include single black-box model, invoked in hybrid leaves. The comprehensibility of such hybrid classifiers is measured with the proportion of examples classified by the regular leaves. We propose one simple and one computationally efficient algorithm for finding the Pareto optimal hybrid trees, starting from an initial classification tree and a black-box classifier. We evaluate the proposed algorithms empirically, comparing them to the baseline solution set, showing that they often provide valuable improvements. Furthermore, we show that the efficient algorithm outperforms the NSGA-II algorithm in terms of quality of the result set and efficiency (for this optimisation problem). Finally we show that the algorithm returns hybrid classifiers that reflect the expert's knowledge on activity recognition problem well.

Abstract This paper presents a method for combining domain knowledge and machine learning (CDKML) for classifier generation and online adaptation. The method exploits advantages in domain knowledge and machine learning as complementary information sources. Whereas machine learning may discover patterns in interest domains that are too subtle for humans to detect, domain knowledge may contain information on a domain not present in the available domain dataset. CDKML has three steps. First, prior domain knowledge is enriched with relevant patterns obtained by machine learning to create an initial classifier. Second, genetic algorithms refine the classifier. Third, the classifier is adapted online on the basis of user feedback using the Markov decision process. CDKML was applied in fall detection. Tests showed that the classifiers developed by CDKML have better performance than machine‐learning classifiers generated on a training dataset that does not adequately represent all real‐life cases of the learned concept. The accuracy of the initial classifier was 10 percentage points higher than the best machine‐learning classifier and the refinement added 3 percentage points. The online adaptation improved the accuracy of the refined classifier by an additional 15 percentage points.

Monitoring human energy expenditure is important in many health and sport applications, since the energy expenditure directly reflects the level of physical activity. The actual energy expenditure is unpractical to measure; hence, the field aims at estimating it by measuring the physical activity with accelerometers and other sensors. Current advanced estimators use a context-dependent approach in which a different regression model is invoked for different activities of the user. In this paper, we go a step further and use multiple contexts corresponding to multiple sensors, resulting in an ensemble of models for energy expenditure estimation. This provides a multi-view perspective, which leads to a better estimation of the energy. The proposed method was experimentally evaluated on a comprehensive set of activities where it outperformed the current state-of-the-art.

This paper presents a method for human activity recognition and energy expenditure estimation with two tri-axial accelerometers. Recognizing the activity of a person and measuring his/her energy expenditure is important for the management of several diseases. In the CHIRON project we aim to monitor congestive heart failure patients using wearable sensors and a smartphone. Our method uses a classifier for activity recognition constructed with machine learning. Attention was paid to the complexity of the attributes for machine learning, resulting in the omission of the most complex attributes in order to prolong the battery life. The recognized activity serves as an input to a classifier for energy expenditure estimation, which was also constructed with machine learning. The best-performing classifier turned out to be a composite of two activity-specific classifiers and a general classifier. Its mean absolute error was 0.91 metabolic equivalents of task (MET).

BACKGROUND: Alzheimer's disease has been known for more than 100 years and the underlying molecular mechanisms are not yet completely understood. The identification of genes involved in the processes in Alzheimer affected brain is an important step towards such an understanding. Genes differentially expressed in diseased and healthy brains are promising candidates. RESULTS: Based on microarray data we identify potential biomarkers as well as biomarker combinations using three feature selection methods: information gain, mean decrease accuracy of random forest and a wrapper of genetic algorithm and support vector machine (GA/SVM). Information gain and random forest are two commonly used methods. We compare their output to the results obtained from GA/SVM. GA/SVM is rarely used for the analysis of microarray data, but it is able to identify genes capable of classifying tissues into different classes at least as well as the two reference methods. CONCLUSION: Compared to the other methods, GA/SVM has the advantage of finding small, less redundant sets of genes that, in combination, show superior classification characteristics. The biological significance of the genes and gene pairs is discussed.

In practice, an incomplete heuristic search nearly always finds better solutions if it is allowed to search deeper, i.e. expand and heuristically evaluate more nodes in the search tree. On the rare occasions when searching deeper is not beneficial, a curious phenomenon called ‘search pathology’ occurs. In this article, we study the pathology and gain of a deeper search of the minimin algorithm in the 8-puzzle, a domain often used for evaluating single-agent search algorithms. We have analysed the influence of various properties of the search tree and the heuristic evaluation function on the gain and the pathology. In order to investigate a broad range of the properties, the original 8-puzzle was extended with diagonal moves, yielding a larger variety of search trees. It turned out that in the 8-puzzle, a substantial proportion of the solvable positions is pathological under various parameters. More importantly, the search parameters that enable the highest gains are quite consistent in pathological and non-pathological positions alike, thus pointing to potentially successful search strategies.

Due to the rapid aging of the population, many technical solutions for the care of the elderly are being developed, often involving fall detection with accelerometers. We present a novel approach to fall detection with location sensors. In our application, a user wears up to four tags on the body whose locations are detected with radio sensors. This makes it possible to recognize the user’s activity, including falling any lying afterwards, and the context in terms of the location in the apartment. We compared fall detection using location sensors, accelerometers and accelerometers combined with the context. A scenario consisting of events difficult to recognize as falls or non- falls was used for the comparison. The accuracy of the methods that utilized the context was almost 40 percentage points higher compared to the methods without the context. The accuracy of pure location-based methods was around 10 percentage points higher than the accuracy of accelerometers combined with the context.

This paper presents an approach to fall detection with accelerometers that exploits posture recognition to identify postures that may be the result of a fall. Posture recognition as a standalone task was also studied. Nine placements of up to four sensors were considered: on the waist, chest, thigh and ankle. The results are compared to the results of a system using ultra wide band location sensors on a scenario consisting of events difficult to recognize as falls or non-falls. Three accelerometers proved sufficient to correctly recognize all the events except one(a slow fall). The location-based system was comparable to two accelerometers, except that it was able to recognize the slow fall because it resulted in lying outside the bed, whose location was known to the system. One accelerometer was able to recognize only the most clear-cut fall. Two accelerometers achieved over 90% accuracy of posture recognition, which was better than the location-based system. Chest and waist accelerometers proved best at both tasks, with the chest accelerometer having a slight advantage in posture recognition.

Incomplete single-agent search methods are often better suited to real-time pathfinding tasks than complete methods (such as A*). Incomplete methods conduct a limited-depth lookahead search, i.e., expand a part of the space centered on the agent, and heuristically evaluate the distances from the frontier of the expanded space to the goal. Actions selected this way are not necessarily optimal, but it is generally believed that deeper lookahead increases the quality of decisions. However, in two-player games, where similar methods are used, it has long been known that this is not always the case [7, 1]. This phenomenon has been termed minimax pathology. More recently pathological behavior was discovered in single-agent search as well [3]. Some attempts to explain it have been made [5, 6], but the pathology in single-agent search is largely still not understood. In this paper we investigate lookahead pathology in real-time pathfinding on maps from commercial computer games. First, we present an empirical study showing a degree of pathology in over 90% of the problems considered. Second, we give four explanations for such wide-spread pathological behavior.

Real-time heuristic search is a challenging type of agent-centered search because the agent's planning time per action is bounded by a constant independent of problem size. A common problem that imposes such restrictions is pathfinding in modern computer games where a large number of units must plan their paths simultaneously over large maps. Common search algorithms (e.g., A*, IDA*, D*, ARA*, AD*) are inherently not real-time and may lose completeness when a constant bound is imposed on per-action planning time. Real-time search algorithms retain completeness but frequently produce unacceptably suboptimal solutions. In this paper, we extend classic and modern real-time search algorithms with an automated mechanism for dynamic depth and subgoal selection. The new algorithms remain real-time and complete. On large computer game maps, they find paths within 7% of optimal while on average expanding roughly a single state per action. This is nearly a three-fold improvement in suboptimality over the existing state-of-the-art algorithms and, at the same time, a 15-fold improvement in the amount of planning per action.