Mass casualty incidents (MCIs) are a growing concern, characterized by complexity and uncertainty that demand adaptive decision-making strategies. The victim tagging step in the emergency medical response must be completed quickly and is crucial for providing information to guide subsequent time-constrained response actions. In this paper, we present a mathematical formulation of multi-agent victim tagging to minimize the time it takes for responders to tag all victims. Five distributed heuristics are formulated and evaluated with simulation experiments. The heuristics are on-the go, practical solutions that represent varying levels of situational uncertainty in the form of global or local communication capabilities, demonstrating practical constraints. We further investigate the performance of a multi-agent reinforcement learning (MARL) strategy, factorized deep Q-network (FDQN), to minimize victim tagging time as compared to baseline heuristics. Extensive simulations demonstrate that between the heuristics, methods with local communication are more efficient for adaptive victim tagging, specifically choosing the nearest victim with the option to replan. Analyzing all experiments, we find that the FDQN approach outperforms heuristics in smaller-scale scenarios, while heuristics excel in more complex scenarios. Our experiments contain diverse complexities and agent configurations that explore the upper limits of MARL capabilities for real-world applications and reveal key insights.

Policy distillation is a widely used class of deep reinforcement learning transfer approaches designed to minimize the divergence between student and expert policies. These methods significantly improve learning efficiency and effectiveness but still require many agent-environment interactions. In this work, we explore parameter transfer as an alternative approach that is not subject to this limitation. Through empirical evaluation, we show that parameter transfer reliably enables positive transfer across various tasks, including those where state-of-the-art methods exhibit limited effectiveness. We also show that parameter transfer is robust to the choice of expert and that it remains effective even after the student network is pruned. Finally, we demonstrate that parameter transfer solves an open problem in constrained reinforcement learning. Our work demonstrates that parameter transfer is not only simple to understand and implement, but also practically useful.

Offline reinforcement learning in high-dimensional, discrete action spaces is challenging due to the exponential scaling of the joint action space with the number of sub-actions and the complexity of modeling sub-action dependencies. Existing methods either exhaustively evaluate the action space, making them computationally infeasible, or factorize Q-values, failing to represent joint sub-action effects. We propose \textbf{Bra}nch \textbf{V}alue \textbf{E}stimation (BraVE), a value-based method that uses tree-structured action traversal to evaluate a linear number of joint actions while preserving dependency structure. BraVE outperforms prior offline RL methods by up to $20\times$ in environments with over four million actions.

Humans tend to project their own emotions onto others, perceiving others’ emotional states as more similar to their own. Yet, no study has investigated this emotional projection in robots, despite being frequently anthropomorphized into having emotions. In this work, we demonstrate how participants’ emotional state influences their perception of a robot’s emotions. We conducted a user study in which participants felt varying emotions and stress levels. In each state, they observed a robot perform an emotionally ambiguous action and indicated the emotion they believed the robot was communicating. Our results indicate that our participants projected their own emotional state onto the robot. Increased stress manifested in participants as increased uncertainty about the robots’ emotions. As participants became more familiar with the robot, their felt and perceived emotions remained partially aligned; however, the effect of stress diminished. Our findings suggest future social robotic studies must consider participants’ emotional states in their evaluations.

Nutrients are critical to the functioning of the human body and their imbalance can result in detrimental health concerns. The majority of nutritional literature focuses on macronutrients, often ignoring the more critical nuances of micronutrient balance, which require more precise regulation. Currently, micronutrient status is routinely assessed via complex methods that are arduous for both the patient and the clinician. To address the global burden of micronutrient imbalance, innovations in assessment must be accessible and noninvasive. In support of this task, this article synthesizes useful background information on micronutrients themselves, reviews the state of biofluid and physiological analyses for their assessment, and presents actionable opportunities to push the field forward. By taking a unique, clinical perspective that is absent from technological research on the topic, we find that the state of the art suffers from limited clinical relevance, a lack of overlap between biofluid and physiological approaches, and highly invasive and inaccessible solutions. We present opportunities for future work to maximize the impact of a novel assessment method by incorporating clinical relevance, the holistic nature of micronutrition, and prioritizing accessible and noninvasive systems.

Micronutrient imbalance is a global issue, and its detection is invasive and expensive. Despite being largely preventable, imbalances of one or more micronutrients is pervasive and has major downstream health effects [1]. Women, children, and underserved populations in particular bear the greatest burden of micronutrient imbalance [4--6]. However, the true scope of this issue is often unseen and unaddressed because of the barriers to accessible micronutrient status assessment [2]. Status assessment of micronutrients is often done via indirect, subjective dietary logs, in-person clinical examinations, or complex analyses on blood (e.g. liquid chromatography-coupled mass spectrometry). While valuable, these assessments are expensive, flawed, and burdensome on the patient and the clinician.

Forecasting behavioral patterns can help humans comprehend their habits and tendencies, allowing them to inter-vene before problems arise. Among the numerous techniques of modeling human behavior, cyclic modeling can better identify recurring patterns, providing regularity and predictability in human behavior. However, existing approaches to cyclic modeling ignore the effects of human characteristics, such as resilience and coping abilities, which substantially influence the stability of behavior. To explore the value of adding such information in behavior modeling and prediction, we introduce a transformer-based architecture with an advanced attention mechanism and parallel operation that models regularity in human behavior from multidimensional sensing signals and predicts future behavior patterns. The architecture further transforms static human characteristics metadata into dynamic time series that conform to behavioral patterns and serve as covariates to assist prediction. Our experiments with a wearable dataset indicate that our architecture 1) is more accurate in forecasting human behavior patterns than current time-series models and 2) enables us to investigate the impact of human characteristics on behavior patterns. Specifically, we analyze the influence of resilience and coping strategies on behavioral regularity and predictability. We show that supplementing bio-behavioral wearable data with resilience and coping scores in the forecasting model increases the convergence speed of the model and decreases prediction loss.

Humanoid robots provide a unique opportunity for personalized interaction using emotion recognition. However, emotion recognition performed by humanoid robots in complex social interactions is limited in the flexibility of interaction as well as personalization and adaptation in the responses. We designed an adaptive learning system for real-time emotion recognition that elicits its own ground-truth data and updates individualized models to improve performance over time. A Convolutional Neural Network based on off-the-shelf ResNet50 and Inception v3 are assembled to form an ensemble model which is used for real-time emotion recognition through facial expression. Two sets of robot behaviors, general and personalized, are developed to evoke different emotion responses. The personalized behaviors are adapted based on user preferences collected through a pre-test survey. The performance of the proposed system is verified through a 2-stage user study and tested for the accuracy of the self-supervised retraining. We also evaluate the effectiveness of the personalized behavior of the robot in evoking intended emotions between stages using trust, empathy and engagement scales. The participants are divided into two groups based on their familiarity and previous interactions with the robot. The results of emotion recognition indicate a 12% increase in the F1 score for 7 emotions in stage 2 compared to pre-trained model. Higher mean scores for trust, engagement, and empathy are observed in both participant groups. The average similarity score for both stages was 82% and the average success rate of eliciting the intended emotion increased by 8.28% between stages, despite their differences in familiarity thus offering a way to mitigate novelty effect patterns among user interactions.

Continuous passive sensing of daily behavior from mobile devices has the potential to identify behavioral patterns associated with different aspects of human characteristics. This paper presents novel analytic approaches to extract and understand these behavioral patterns and their impact on predicting adaptive and maladaptive personality traits. Our machine learning analysis extends previous research by showing that both adaptive and maladaptive traits are associated with passively sensed behavior providing initial evidence for the utility of this type of data to study personality and its pathology. The analysis also suggests directions for future confirmatory studies into the underlying behavior patterns that link adaptive and maladaptive variants consistent with contemporary models of personality pathology.

This paper presents a computational framework for modeling biobehavioral rhythms - the repeating cycles of physiological, psychological, social, and environmental events - from mobile and wearable data streams. The framework incorporates four main components: mobile data processing, rhythm discovery, rhythm modeling, and machine learning. We evaluate the framework with two case studies using datasets of smartphone, Fitbit, and OURA smart ring to evaluate the framework’s ability to (1) detect cyclic biobehavior, (2) model commonality and differences in rhythms of human participants in the sample datasets, and (3) predict their health and readiness status using models of biobehavioral rhythms. Our evaluation demonstrates the framework’s ability to generate new knowledge and findings through rigorous micro- and macro-level modeling of human rhythms from mobile and wearable data streams collected in the wild and using them to assess and predict different life and health outcomes.

We present a machine learning approach that uses data from smartphones and fitness trackers of 138 college students to identify students that experienced depressive symptoms at the end of the semester and students whose depressive symptoms worsened over the semester. Our novel approach is a feature extraction technique that allows us to select meaningful features indicative of depressive symptoms from longitudinal data. It allows us to detect the presence of post-semester depressive symptoms with an accuracy of 85.7% and change in symptom severity with an accuracy of 85.4%. It also predicts these outcomes with an accuracy of >80%, 11–15 weeks before the end of the semester, allowing ample time for pre-emptive interventions. Our work has significant implications for the detection of health outcomes using longitudinal behavioral data and limited ground truth. By detecting change and predicting symptoms several weeks before their onset, our work also has implications for preventing depression.

Most feature selection methods only perform well on datasets with relatively small set of features. In the case of large feature sets and small number of data points, almost none of the existing feature selection methods help in achieving high accuracy. This paper proposes a novel approach to optimize the feature selection process through Frequent Pattern Growth algorithm to find sets of features that appear frequently among the top features selected by the main feature selection methods. Our experimental evaluation on two datasets containing a small and very large number of features shows that our approach significantly improves the accuracy results of the dataset with a very large number of features.

As robots are increasingly endowed with social and communicative capabilities, they will interact with humans in more settings, both collaborative and competitive. We explore human-robot relationships in the context of a competitive Stackelberg Security Game. We vary humanoid robot expressive language (in the form of “encouraging” or “discouraging” verbal commentary) and measure the impact on participants' rationality, strategy prioritization, mood, and perceptions of the robot. We learn that a robot opponent that makes discouraging comments causes a human to play a game less rationally and to perceive the robot more negatively. We also contribute a simple open source Natural Language Processing framework for generating expressive sentences, which was used to generate the speech of our autonomous social robot.

Timebanking is a growing type of peer-to-peer service exchange, but is hampered by the effort of finding good transaction partners. We seek to reduce this effort by using a Matching Algorithm for Service Transactions (MAST). MAST matches transaction partners in terms of similarity of interests and complementarity of abilities and needs. We present an experiment involving data and participants from a real timebanking network, that evaluates the acceptability of MAST, and shows that such an algorithm can retrieve matches that are subjectively better than matches based on matching the category of people's historical offers or requests to the category of a current transaction request.

The rapid adoption of smartphones along with a growing habit for using these devices as alarm clocks presents an opportunity to use this device as a sleep detector. This adds value to UbiComp and personal informatics in terms of user context and new performance data to collect and visualize, and it benefits healthcare as sleep is correlated with many health issues. To assess this opportunity, we collected one month of phone sensor and sleep diary entries from 27 people who have a variety of sleep contexts. We used this data to construct models that detect sleep and wake states, daily sleep quality, and global sleep quality. Our system classifies sleep state with 93.06% accuracy, daily sleep quality with 83.97% accuracy, and overall sleep quality with 81.48% accuracy. Individual models performed better than generally trained models, where the individual models require 3 days of ground truth data and 3 weeks of ground truth data to perform well on detecting sleep and sleep quality, respectively. Finally, the features of noise and movement were useful to infer sleep quality.

There is a growing interest in personal health technologies that sample behavioral data from a patient and visualize this data back to the patient for increased health awareness. However, a core challenge for patients is often to understand the connection between specific behaviors and health, i.e. to go beyond health awareness to disease insight. This paper presents MONARCA 2.0, which records subjective and objective data from patients suffering from bipolar disorder, processes this, and informs both the patient and clinicians on the importance of the different data items according to the patient's mood. The goal is to provide patients with a increased insight into the parameters influencing the nature of their disease. The paper describes the user-centered design and the technical implementation of the system, as well as findings from an initial field deployment.

We present an approach to learning to recognize concurrent activities based on multiple data streams. One example is recognition of concurrent activities in hospital operating rooms based on multiple wearable and embedded sensors. This problem differs from standard time series classification in that there is no natural single target dimension, as multiple activities are performed at the same time. Hence, most existing approaches fail. The key innovations that allow us to tackle this problem is (1) learning to recognize base activities from raw sensor data, (2) creating artificial joint activities from base activities using frequent pattern mining and (3) handling temporal dependency using virtual evidence boosting.

This paper presents a recommender system for teams of medical professionals working collaboratively in hospital operating rooms. The system recommends relevant virtual actions, such as retrieval of information resources and initiation of communication with professionals outside the operating rooms. Recommendations are based on the current state of the ongoing operation as recognised from sensor data using machine learning techniques. The selection and non-selection of virtual actions during operations are interpreted as implicit feedback and used to update the weight matrices that guide recommendations. A pilot user study involving medical professionals indicates that the adaptation mechanism is effective and that the system provides adequate recommendations.

This paper presents "Activity Analysis" as a method for conducting and analyzing field studies based on Activity Theory. Two cases of activity analysis of work in a hospital ward and inside an operating room are presented. Guidelines for moving from Activity Analysis to systems design is presented and illustrated with the design of a context-aware system for hospitals.

In Ubiquitous Computing (Ubicomp) research, substantial work has been directed towards sensor-based detection and recognition of human activity. This research has, however, mainly been focused on activities of daily living of a single person. This paper presents a sensor platform and a machine learning approach to sense and detect phases of a surgical operation. Automatic detection of the progress of work inside an operating room has several important applications, including coordination, patient safety, and context-aware information retrieval. We verify the platform during a surgical simulation. Recognition of the main phases of an operation was done with a high degree of accuracy. Through further analysis, we were able to reveal which sensors provide the most significant input. This can be used in subsequent design of systems for use during real surgeries.