The mental health impacts of adolescent social media use are complex, with evidence linking use to increased anxiety, depression, loneliness, and self-harm, while also offering opportunities for connection among isolated youth. A central sociotechnical challenge presented is youth loneliness: online platforms may support well-being when they strengthen meaningful relationships, but may be detrimental when they displace in-person interaction. This work addresses this challenge through the co-design of a digital intervention aimed at promoting face-to-face social engagement among adolescents. Building on a youth-generated concept developed during a hackathon, we apply systems analysis and user-centered design processes—including goal definition, alternative generation and evaluation, and iterative prototyping—to translate an early-stage idea into a minimum viable system. A targeted review of prior digital and behavioral interventions informs the design, grounding it in behavioral activation, self-determination theory, and social support theory, and identifying key behavioral targets for sustaining friendships: initiation, follow-through, and relationship deepening. These insights guide the development of core system features centered on a "friendship garden" metaphor, in which peer connections are represented as entities that grow through shared, in-person activities. The prototype is iteratively refined through wireframing and feedback from a Youth Advisory Board (YAB) and is designed to support intentional interaction, shared accountability, and sustained relationship maintenance. Ongoing work includes a feasibility and user experience study with adolescents (target n=10) to evaluate usability, engagement, and potential for reducing loneliness. This work contributes a theoretically grounded, co-designed digital intervention and provides design insights for promoting offline social connection through technology.

Personal AI tools can now be generated from natural-language requests, but they often remain isolated after creation. We present PSI, a shared-state architecture that turns independently generated modules into coherent instruments: persistent, connected, and chat-complementary artifacts accessible through both GUIs and a generic chat agent. By publishing current state and write-back affordances to a shared personal-context bus, modules enable cross-module reasoning and synchronized actions across interfaces. We study PSI through a three-week autobiographical deployment in a self-developed personal AI environment and show that later-generated instruments can be integrated automatically through the same contract. PSI identifies shared state as the missing systems layer that transforms AI-generated personal software from isolated apps into coherent personal computing environments.

Social interactions are fundamental to well-being, yet automatically detecting them in daily life—particularly using wearables—remains underexplored. Most existing systems are evaluated in controlled settings, focus primarily on in-person interactions, or rely on restrictive assumptions (e.g., requiring multiple speakers within fixed temporal windows), limiting generalizability to real-world use. We present an on-watch interaction detection system designed to capture diverse interactions in naturalistic settings. A core component is a foreground speech detector trained on a public dataset. Evaluated on over 100,000 labeled foreground speech and background sound instances, the detector achieves a balanced accuracy of 85.51%, outperforming prior work by 5.11%. We evaluated the system in a real-world deployment (N=38), with over 900 hours of total smartwatch wear time. The system detected 1,691 interactions, 77.28% were confirmed via participant self-report, with durations ranging from under one minute to over one hour. Among correct detections, 81.45% were in-person, 15.7% virtual, and 1.85% hybrid. We further developed a 15-second window-level audio-only model that enables faster interaction prediction, achieving a balanced accuracy of 90.39% and a sensitivity of 91.01% on 33,698 labeled windows. These results demonstrate the feasibility of real-world interaction sensing and open the door to adaptive, context-aware systems responding to users’ dynamic social environments.

Digital mental health interventions (DMHIs), such as cognitive bias modification for interpretations (CBM-I), offer promise for increasing access to anxiety treatment among underserved adolescents, but data regarding their efficacy are mixed. Paraprofessionals and other caring adults in youth's lives, such as non-parental adult mentors, may be able to support the use of DMHIs and increase teen engagement. The present mixed methods evaluation of a pilot open trial tested the feasibility, acceptability, and preliminary efficacy of implementing MindTrails Teen (an app-based, youth-adapted version of the web-based MindTrails CBM-I intervention) within mentor/mentee dyads. Thirty participants (composed of 15 dyads) participated in remote data collection for 5 weeks. A subset of participants (n = 7 mentors; n = 7 mentees) also provided qualitative feedback. Intervention outcomes (change in anxiety symptoms, and positive and negative interpretation bias), feasibility, and acceptability were assessed via a mix of qualitative interviews, quantitative change in questionnaire scores, and program completion and fidelity metrics. Outcomes were compared to pre-registered benchmarks. Large effect sizes were observed for changes in anxiety among youth. Small to medium effects were observed for change in positive interpretation bias, and no change was found for negative interpretation bias. Intervention outcomes should be considered with caution given very low internal consistency of the interpretation bias measure and the lack of a control comparison group. Acceptability of the intervention was rated positively by mentors and youth. Feasibility benchmarks were met for mentors but not for youth. Qualitative feedback indicated mentors perceived the app as helpful to their mentees, found that it either improved or did not affect their relationship, but also identified implementation challenges. Youth overall perceived the app as helpful but identified barriers to engagement.

Correctly identifying an individual's social context from passively worn sensors holds promise for delivering just-in-time adaptive interventions (JITAIs) to treat social anxiety disorder. In this study, we present results using passively collected data from a within-subject experiment that assessed physiological response across different social contexts (i.e, alone vs. with others), social phases (i.e., pre- and post-interaction vs. during an interaction), social interaction sizes (i.e., dyadic vs. group interactions), and levels of social threat (i.e., implicit vs. explicit social evaluation). Participants in the study ($N=46$) reported moderate to severe social anxiety symptoms as assessed by the Social Interaction Anxiety Scale ($\geq$34 out of 80). Univariate paired difference tests, multivariate random forest models, and follow-up cluster analyses were used to explore physiological response patterns across different social and non-social contexts. Our results suggest that social context is more reliably distinguishable than social phase, group size, or level of social threat, but that there is considerable variability in physiological response patterns even among these distinguishable contexts. Implications for real-world context detection and deployment of JITAIs are discussed.

Mobile sensing is ubiquitous and offers opportunities to gain insight into state mental health functioning. Detecting state elevations in social anxiety would be especially useful given this phenomenon is highly prevalent and impairing, but often not disclosed. In the present work, we explore the feasibility of detecting fluctuations in state social anxiety among N = 46 undergraduate students with elevated symptoms of trait social anxiety. Participants engaged in two dyadic and two group social interactions via Zoom. We evaluated participants' state anxiety levels as they anticipated, immediately after experiencing, and upon reflecting on each social interaction, spanning a time frame of 2-6 minutes. We collected biobehavioral features (i.e., PPG, EDA, skin temperature, and accelerometer) via Empatica E4 devices as they participated in the varied social contexts (e.g., dyadic vs. group; anticipating vs. experiencing the interaction; experiencing varying levels of social evaluation). We additionally measured their trait mental health functioning. Mixed-effect logistic regression and leave-one-subject-out machine learning modeling indicated biobehavioral features significantly predict state fluctuations in anxiety, though balanced accuracy tended to be modest (59%). However, our capacity to identify instances of heightened versus low state anxiety significantly increased (with balanced accuracy ranging from 69% to 84% across different operationalizations of state anxiety) when we integrated contextual data alongside trait mental health functioning into our predictive models. We discuss these and other findings in the context of the broader anxiety detection literature.

Rates of stress and anxiety are alarmingly high in university communities, but most people do not receive treatment. Mobile health (mHealth) interventions show promise to improve psychological symptoms and increase access to interventions, but little is known about their effects in the moment. The present study evaluated the short-term impact of brief mHealth sessions to determine which intervention features are associated with the greatest momentary self-reported improvements. Participants (N = 100 undergraduate students, graduate students, and university staff members) completed brief training sessions 1–2 times daily of MASK, a new mobile application for the university community that uses Cognitive Bias Modification for Interpretations (CBM-I) to shift anxious thinking patterns. Training sessions varied based on stressor domain/topic selected and writing requirements, among other features. Linear mixed effects models were used to test whether stressor domain or writing requirements predict post-training: (1) momentary affect, (2) reappraisal self-efficacy, and (3) emotion regulation self-efficacy. Self-reported improvement in state affect, reappraisal self-efficacy, and emotion regulation self-efficacy occurred for six out of eight stressor domains. Additionally, training sessions requiring less (vs. more) writing were associated with greater positive changes in affect, but not reappraisal or emotion regulation self-efficacy. Stressor domain and writing requirements are associated with different in-the-moment cognitive and affective outcomes, pointing to the need to tailor mHealth programs to users’ specific needs and current stressors.

Social anxiety is a common mental health condition linked to significant challenges in academic, social, and occupational functioning. A core feature is elevated momentary (state) anxiety in social situations, yet little prior work has measured or predicted fluctuations in this anxiety throughout the day. Capturing these intra-day dynamics is critical for designing real-time, personalized interventions such as Just-In-Time Adaptive Interventions (JITAIs). To address this gap, we conducted a study with socially anxious college students (N=91; 72 after exclusions) using our custom smartwatch-based system over an average of 9.03 days (SD = 2.95). Participants received seven ecological momentary assessments (EMAs) per day to report state anxiety. We developed a base model on over 10,000 days of external heart rate data, transferred its representations to our dataset, and fine-tuned it to generate probabilistic predictions. These were combined with trait-level measures in a meta-learner. Our pipeline achieved 60.4

Social interactions are a fundamental part of daily life and play a critical role in well-being. As emerging technologies offer opportunities to unobtrusively monitor behavior, there is growing interest in using them to better understand social experiences. However, automatically detecting interactions-particularly via wearable devices-remains underexplored. Existing systems are often limited to controlled environments, constrained to in-person interactions, and rely on rigid assumptions such as the presence of two speakers within a fixed time window. These limitations reduce their generalizability to capture diverse real-world interactions. To address these challenges, we developed a real-time, on-watch system capable of detecting both in-person and virtual interactions. The system leverages transfer learning to detect foreground speech (FS) and infers interaction boundaries based upon FS and conversational cues like whispering. In a real-world evaluation involving 11 participants over a total of 38 days (Mean = 3.45 days, SD = 2.73), the system achieved an interaction detection accuracy of 73.18%. Follow-up with six participants indicated perfect recall for detecting interactions. These preliminary findings demonstrate the potential of our system to capture interactions in daily life-providing a foundation for applications such as personalized interventions targeting social anxiety.

Anxiety is highly prevalent among college communities, with significant numbers of students, faculty, and staff experiencing severe anxiety symptoms. Digital mental health interventions (DMHIs), including Cognitive Bias Modification for Interpretation (CBM-I), offer promising solutions to enhance access to mental health care, yet there is a critical need to evaluate user experience and acceptability of DMHIs. CBM-I training targets cognitive biases in threat perception, aiming to increase cognitive flexibility by reducing rigid negative thought patterns and encouraging more benign interpretations of ambiguous situations. This study used questionnaire and interview data to gather feedback from users of a mobile application called "Hoos Think Calmly" (HTC), which offers brief CBM-I training doses in response to stressors commonly experienced by students, faculty, and staff at a large public university. Mixed methods were used for triangulation to enhance the validity of the findings. Qualitative data was collected through semi-structured interviews from a subset of participants (n = 22) and analyzed thematically using an inductive framework, revealing five main themes: Effectiveness of the Training Program; Feedback on Training Sessions; Barriers to Using the App; Use Patterns; and Suggestions for Improvement. Additionally, biweekly user experience questionnaires sent to all participants in the active treatment condition (n = 134) during the parent trial showed the most commonly endorsed response (by 43.30% of participants) was that the program was somewhat helpful in reducing or managing their anxiety or stress. There was overall agreement between the quantitative and qualitative findings, indicating that graduate students found it the most effective and relatable, with results being moderately positive but somewhat more mixed for undergraduate students and staff, and least positive for faculty. Findings point to clear avenues to enhance the relatability and acceptability of DMHIs across diverse demographics through increased customization and personalization, which may help guide development of future DMHIs.

Social anxiety, characterized by fear of social interactions and negative evaluation, is both pervasive and impairing. Passive sensing offers opportunities for timely detection and intervention, yet most prior work has emphasized trait-level anxiety, with limited success in predicting intra-day fluctuations in state anxiety. To address this gap, we studied 72 socially anxious students who used our smartwatch–smartphone system for an average of nine days. The smartwatch collected multimodal physiological and behavioral signals (e.g., heart rate, movement), while the smartphone delivered seven randomly timed ecological momentary assessments per day to capture state anxiety. These multimodal data formed the basis for anxiety modeling. We present a weakly supervised multimodal framework that trains a base model on a public dataset, then transfers learning and personalizes predictions via meta-learning. Our model outperformed baselines, achieving 72.1% balanced accuracy on our dataset and 66.9% on an external dataset, demonstrating potential to support just-in-time-adaptive interventions.

While audio data shows promise in addressing various health challenges, there is a lack of research on on-device audio processing for smartwatches. Privacy concerns make storing raw audio and performing post-hoc analysis undesirable for many users. Additionally, current on-device audio processing systems for smartwatches are limited in their feature extraction capabilities, restricting their potential for understanding user behavior and health. We developed a real-time system for on-device audio processing on smartwatches, which takes an average of 1.78 minutes (SD = 0.07 min) to extract 22 spectral and rhythmic features from a 1-minute audio sample, using a small window size of 25 milliseconds. Using these extracted audio features on a public dataset, we developed and incorporated models into a watch to classify foreground and background speech in real-time. Our Random Forest-based model classifies speech with a balanced accuracy of 80.3%.

Wearable devices with embedded sensors can provide personalized healthcare and wellness benefits in digital phenotyping and adaptive interventions. However, the collection, storage, and transmission of biometric data (including processed features rather than raw signals) from these devices pose significant privacy concerns. This quantitative, data-driven study examines the privacy risks associated with wearable-based digital phenotyping practices, with a focus on user reidentification (ReID), which is the process of identifying participants’ IDs from deidentified digital phenotyping datasets. We propose a machine-learning-based computational pipeline to evaluate and quantify model outcomes under various configurations, such as modality inclusion, window length, and feature type and format, to investigate the factors influencing ReID risks and their predictive trade-offs. This pipeline leverages features extracted from three wearable sensors, resulting in up to 68.43% accuracy in ReID risk for a sample size of N=45 socially anxious participants based on only descriptive features of 10-second observations. Additionally, we explore the trade-offs between privacy risks and predictive benefits by adjusting various settings (e.g., the ways to process extracted features). Our findings highlight the importance of privacy in digital phenotyping and suggest potential future directions.

For infinite action contextual bandits, smoothed regret and reduction to regression results in state-of-the-art online performance with computational cost independent of the action set: unfortunately, the resulting data exhaust does not have well-defined importance-weights. This frustrates the execution of downstream data science processes such as offline model selection. In this paper we describe an online algorithm with an equivalent smoothed regret guarantee, but which generates well-defined importance weights: in exchange, the online computational cost increases, but only to order smoothness (i.e., still independent of the action set). This removes a key obstacle to adoption of smoothed regret in production scenarios.

This work introduces the Eigen Memory Tree (EMT), a novel online memory model for sequential learning scenarios. EMTs store data at the leaves of a binary tree and route new samples through the structure using the principal components of previous experiences, facilitating efficient (logarithmic) access to relevant memories. We demonstrate that EMT outperforms existing online memory approaches, and provide a hybridized EMT-parametric algorithm that enjoys drastically improved performance over purely parametric methods with nearly no downsides. Our findings are validated using 206 datasets from the OpenML repository in both bounded and infinite memory budget situations.

Mobile sensing is a ubiquitous and useful tool to make inferences about individuals' mental health based on physiology and behavior patterns. Along with sensing features directly associated with mental health, it can be valuable to detect different features of social contexts to learn about social interaction patterns over time and across different environments. This can provide insight into diverse communities' academic, work and social lives, and their social networks. We posit that passively detecting social contexts can be particularly useful for social anxiety research, as it may ultimately help identify changes in social anxiety status and patterns of social avoidance and withdrawal. To this end, we recruited a sample of highly socially anxious undergraduate students (N=46) to examine whether we could detect the presence of experimentally manipulated virtual social contexts via wristband sensors. Using a multitask machine learning pipeline, we leveraged passively sensed biobehavioral streams to detect contexts relevant to social anxiety, including (1) whether people were in a social situation, (2) size of the social group, (3) degree of social evaluation, and (4) phase of social situation (anticipating, actively experiencing, or had just participated in an experience). Results demonstrated the feasibility of detecting most virtual social contexts, with stronger predictive accuracy when detecting whether individuals were in a social situation or not and the phase of the situation, and weaker predictive accuracy when detecting the level of social evaluation. They also indicated that sensing streams are differentially important to prediction based on the context being predicted. Our findings also provide useful information regarding design elements relevant to passive context detection, including optimal sensing duration, the utility of different sensing modalities, and the need for personalization. We discuss implications of these findings for future work on context detection (e.g., just-in-time adaptive intervention development).

Individuals high in social anxiety symptoms often exhibit elevated state anxiety in social situations. Research has shown it is possible to detect state anxiety by leveraging digital biomarkers and machine learning techniques. However, most existing work trains models on an entire group of participants, failing to capture individual differences in their psychological and behavioral responses to social contexts. To address this concern, in Study 1, we collected linguistic data from N=35 high socially anxious participants in a variety of social contexts, finding that digital linguistic biomarkers significantly differ between evaluative vs. non-evaluative social contexts and between individuals having different trait psychological symptoms, suggesting the likely importance of personalized approaches to detect state anxiety. In Study 2, we used the same data and results from Study 1 to model a multilayer personalized machine learning pipeline to detect state anxiety that considers contextual and individual differences. This personalized model outperformed the baseline’s F1-score by 28.0%. Results suggest that state anxiety can be more accurately detected with personalized machine learning approaches, and that linguistic biomarkers hold promise for identifying periods of state anxiety in an unobtrusive way.

In an era of countless content offerings, recommender systems alleviate information overload by providing users with personalized content suggestions. Due to the scarcity of explicit user feedback, modern recommender systems typically optimize for the same fixed combination of implicit feedback signals across all users. However, this approach disregards a growing body of work highlighting that (i) implicit signals can be used by users in diverse ways, signaling anything from satisfaction to active dislike, and (ii) different users communicate preferences in different ways. We propose applying the recent Interaction Grounded Learning (IGL) paradigm to address the challenge of learning representations of diverse user communication modalities. Rather than requiring a fixed, human-designed reward function, IGL is able to learn personalized reward functions for different users and then optimize directly for the latent user satisfaction. We demonstrate the success of IGL with experiments using simulations as well as with real-world production traces.

In adversarial environments, one side could gain an advantage by identifying the opponent’s strategy. For example, in combat games, if an opponent’s strategy is identified as overly aggressive, one could lay a trap that exploits the opponent’s aggressive nature. However, an opponent’s strategy is not always apparent and may need to be estimated from observations of their actions. This paper proposes to use inverse reinforcement learning (IRL) to identify strategies in adversarial environments. Specifically, the contributions of this work are 1) the demonstration of this concept on gaming combat data generated from three pre-defined strategies and 2) the framework for using IRL to achieve strategy identification. The numerical experiments demonstrate that the recovered rewards can be identified using a variety of techniques including visual analysis, cluster analysis, and supervised classification.

This paper considers if a reward function learned via inverse reinforcement from a human expert can be used as a feedback intervention to alter future human performance as desired (i.e., human to human apprenticeship learning). To learn reward functions two new algorithms are developed: a kernel-based inverse reinforcement learning algorithm and a Monte Carlo reinforcement learning algorithm. The algorithms are benchmarked against well-known alternatives within their respective corpus and are shown to outperform in terms of efficiency and optimality. To test the feedback intervention two randomized experiments are performed with 3,256 human participants. The experimental results demonstrate with significance that the rewards learned from "expert" individuals are effective as feedback interventions. In addition to the algorithmic contributions and successful experiments, the paper also describes three reward function modifications to improve reward function feedback interventions for humans.