Lightning Talks & Posters

Beyond the Clinic: Rethinking Cognitive Screening at Scale Using AI:

As the global population ages, the urgency of detecting cognitive decline early grows. Yet standard tools—like the MMSE, MoCA, and Clock Drawing Test (CDT)—were designed for clinics, not the daily lives of the millions who need monitoring most. My research explores the question: what would cognitive assessment look like if designed from scratch for the ubiquitous computing era? And can AI help us achieve that goal?

My first project evaluated three frontier AI models on the CDT. While models achieved 94% dementia screening accuracy, granular scoring accuracy dropped to 55%. Notably, when all three models agreed, they were collectively wrong 22% of the time, revealing shared architectural blind spots. This suggests that clinical metrics are fundamentally mismatched with how these AI systems behave.

Building on this, I am designing "cognitive microtasks"—brief, drawing-based exercises featuring everyday objects with strategically missing features. These are built for frequent, self-administered, and practice-effect-free monitoring. Crucially, each task is paired with a structured rubric that decomposes drawing quality into domain-specific questions, enabling reliable AI scoring. Additionally, vision-language models can generate unlimited new stimuli on demand.

Together, these projects reflect a unified vision: a future where cognitive health is monitored continuously and unobtrusively in everyday life, powered by assessments designed from the ground up for both human completion and AI evaluation.

Digital mental health (DMH) tools have extensively explored personalization of interventions to users’ needs and contexts. However, this personalization often targets what support is provided, not how it is experienced. Even well-matched content can fail when the interaction format misaligns with how someone can engage in the moment. We introduce generative experience as a new paradigm for DMH support, where the intervention experience is dynamically composed at runtime. We instantiate this paradigm in GUIDE, a system that generates personalized intervention content and multimodal interaction structure through rubric-guided generation of modular components. In a preregistered study with N=237 participants, GUIDE produced greater stress reduction (p=.02) and better overall user experience (p=.04) than an LLM-based cognitive restructuring control. GUIDE also supported diverse forms of reflection and action through varied interaction flows, while revealing tensions around uneven personalization and AI-generated support. Our work lays the foundation for mental health interventions that dynamically shape how support is experienced and enacted in digital settings.

One longstanding aim of mHealth technologies is to capture and synthesize data from patients’ daily lives, such as sleep, diet, and physical activity, to deliver actionable insights that support behavior change. Traditionally, these technologies have relied on rule-based approaches to generate insights, such as generating messages to patients based on whether they met predefined goals (e.g., increasing physical activity). However, recent advances in generative AI (GenAI) offer new opportunities to create more tailored and personalized insights for patients. With this increased potential comes new challenges, particularly in determining appropriate tone and interaction styles.

Our team has been developing and evaluating an mHealth platform, iMatter2, to support type 2 diabetes management. The platform includes a GenAI-based conversational agent that provides motivational and educational support to patients using their sleep, diet, and physical activity data. To inform the agent’s design, we conducted design workshops, usability testing sessions, and a discrete choice experiment (n = 125) with individuals with diabetes. We are currently evaluating the impact of the platform on diabetes control through a randomized controlled trial, with 25 patients enrolled currently.

Our preliminary findings suggest substantial heterogeneity in patient preferences for communication style. Some patients prefer a more coach-like, informal tone, while others favor a more clinical style. Similarly, some patients value messages that emphasize progress, whereas others prefer a focus on actionable next steps. These findings highlight an enduring design challenge: how to develop agents that align with individual communication preferences while still effectively supporting behavior change and improving health outcomes.

Healthcare is beginning to deploy a new generation of AI systems that do far more than answer isolated prompts. These agentic systems retain memory across encounters, access external tools such as EHRs and lab databases, and coordinate with multiple specialized agents around a single patient. Oncology copilots, documentation assistants, and triage agents are already entering pilot deployment, yet, the privacy safeguards surrounding clinical AI have not evolved at the same pace. Most existing privacy research was built for an earlier era of machine learning models and focuses primarily on protecting training data. That leaves a major blind spot at runtime, where patient information can persist in agent memory, leak through tool calls, or propagate across inter-agent communication channels. Emerging studies suggest that each of these pathways constitutes a meaningful attack surface, while current healthcare benchmarks evaluate task performance, safety refusal, or adversarial robustness in isolation without testing whether agent systems can actually preserve patient privacy during clinical workflows.This talk presents a new research agenda for runtime privacy in clinical agentic AI, centered on three components: a threat model for clinical agent privacy grounded in replicating published deployments; an architectural retrofit applying cryptographic controls to agent memory, tool gateways, and inter-agent channels; and a benchmark extending MedAgentBench with adversarial privacy tests. I am particularly interested in connecting with clinicians (psychiatry, oncology, primary care) piloting agent systems, regulatory scientists, and security researchers who want to help shape this space before development outpaces oversight

With the majority of our lives spent indoors, it is imperative to maintain healthy indoor air quality for our well-being. The importance of healthy air is further heightened during pandemic or flu season or for immuno-compromised individuals at higher risk of contracting airborne transmission. Indoor airborne transmission poses a significant health risk, yet current monitoring solutions are invasive, costly, or fail to address it directly. My research explores the untapped potential of ambient audio sensing to estimate key transmission risk factors such as ventilation, aerosol emissions, and occupant distribution - non-invasively and in real time. I develop privacy-preserving systems that leverage existing microphones to monitor the whole spectrum of indoor air quality which can have a significant effect on an individual's health. This work lays the foundation for privacy-aware airborne risk monitoring using everyday devices.

Remote patient monitoring (RPM) programs have scaled rapidly with advances in digital health technologies; however, challenges in patient engagement, clinician workload, and care coordination limit their effectiveness. This abstract introduces the AI Brain, a generative artificial intelligence (GenAI)-enhanced platform to address these limitations in the context of hypertension management.

AI Brain comprises two integrated components: (1) a patient-facing conversational agent which leverages SMS-based communication to deliver real-time feedback on blood pressure measurements, provide tailored educational and behavioral nudges, and respond dynamically to patient inquiries. and (2) a clinician-facing agent embedded within the electronic health record that synthesizes longitudinal patient data into structured summaries and generates draft patient communications, thereby reducing cognitive load and administrative burden.

The platform was developed using a user-centered design framework with iterative input from patients, clinicians, and operational stakeholders. The AI Brain has been piloted among hypertension patients at NYU Langone Health in February 2025. To our knowledge, this represents one of the first fully automated, patient-facing GenAI agents deployed within the institution, supported by extensive research and operational infrastructure.

Preliminary findings demonstrate both clinical and operational benefits, as well as key implementation challenges, including patient onboarding friction, and the need for clearer communication regarding the role and capabilities of the AI agent. Ongoing iterative refinements have focused on improving system usability, enhancing patient adherence, and ensuring alignment with clinical workflows. Results suggest that GenAI-enabled RPM platforms have the potential to improve patient engagement, optimize clinician efficiency, and support scalable, data-driven management of chronic conditions.

Background: Early functional and behavioral changes associated with Alzheimer’s disease (AD) and related dementias often emerge years before clinical diagnosis, yet existing digital biomarker approaches—such as wearables, apps, and in-home sensors—face limitations in cost, scalability, adherence, and privacy. We explore passive network traffic monitoring as a low-burden, privacy-preserving alternative for longitudinal digital phenotyping, leveraging existing Internet infrastructure to capture everyday behavioral rhythms without collecting content or personal identifiers.

Methods: We developed and deployed a passive sensing approach that leverages existing Internet infrastructure. It monitors encrypted network traffic metadata of participants’ devices (smartphones, tablets, PCs, TVs, IoTs) with low setup cost and low maintenance, enabling long-term, “connect-and-forget” monitoring. We recruited and deployed across a cohort of ~200 younger adults and 5 older adults (some with mild cognitive impairment (MCI)) over two months and ongoing.

Results: Passive network traffic monitoring proved feasible and acceptable over extended periods, representing one of the largest longitudinal datasets of real-world personal, longitudinal network traffic collected to date for medical studies. Participants reported low burden and strong understanding of privacy protections. Preliminary qualitative findings from the ongoing MCI pilot similarly indicate acceptability: participants expressed comfort with the system and its privacy model, understanding that the system observes high-level activity without access to content, specific media, or personal identity.

Conclusion: Passive network traffic monitoring offers a scalable, low-cost, and privacy-preserving platform for longitudinal behavioral phenotyping relevant to AD and cognitive aging. This talk will present early findings and outline future applications for studying cognitive progression in real-world settings.

Commercial wearables offer an unprecedented opportunity to monitor human movement continuously and unobtrusively, making them especially valuable for detecting the subtle, long-term behavioral changes that characterize aging and cognitive decline. Prior approaches rely on dense activity annotations that are impractical to collect over months of free-living data. In this work, we instead leverage self-supervised representation learning to discover and quantify longitudinal movement changes in older adults, enabling a scalable, annotation-free pipeline for in-the-wild health monitoring.

We utilize data from a longitudinal study of older adults (>65 years of age), in which smartwatch recordings are collected for one week every four months over a total duration of approximately 18 months. Self-supervised models are trained on data from healthy participants, who are comparatively easy to recruit, and used to quantify longitudinal movement changes occurring across the study. Statistical analyses via linear mixed effects modeling reveal significantly different movement change trends between healthy participants and those diagnosed with cognitive decline. These findings pave the way for passive cognitive health monitoring through everyday devices. As cognitive decline remains vastly underdiagnosed worldwide, leveraging the wearables already in widespread use could fundamentally transform early detection and open new pathways for timely intervention.

Breast cancer remains one of the most prevalent cancers worldwide, and long-term endocrine therapy is essential for reducing recurrence risk among survivors. However, adherence to these medications often declines over time due to complex, individualized behavioral, contextual, and physiological factors. Existing monitoring approaches rely on infrequent clinical assessments and population-level models, limiting their ability to capture dynamic, patient-specific adherence patterns in real-world settings. This work presents HealthGraph, a personalized, multimodal temporal graph modeling framework for early identification of adherence risk.

HealthGraph represents each patient as a longitudinal behavioral graph constructed from daily, weekly, and biweekly observations, integrating wearable physiological signals (activity, sleep, cardiovascular patterns), medication event monitoring system (MEMS) data, and ecological momentary assessment (EMA) surveys. Temporal dependencies are modeled using a graph neural network to capture evolving behavioral dynamics. Personalization is achieved through subject-specific conditioning via feature-wise linear modulation (FiLM), while contrastive regularization with masked views improves robustness to missing or noisy samples. To support scalability and deployment in resource-constrained environments, parameter-efficient fine-tuning (PEFT) updates only lightweight FiLM layers, enabling efficient adaptation with minimal parameters.

Beyond prediction, HealthGraph incorporates a grounded reasoning module that translates personalized behavioral traits into user-centric, interpretable adherence insights. The system generates patient-centric summaries and as well as structured clinical insights that highlight longitudinal trends and key behavioral drivers. Empirical results demonstrate improved performance over population-level baselines, enabling HealthGraph to provide proactive, context-aware, and personalized support for adaptive adherence interventions.

Deploying digital health tools in clinical environments frequently falls short not from technical shortcomings, but from workflow misalignment, fragmented governance, and poor sociotechnical integration. As computer and digital solutions become increasingly embedded in clinical care—enabling more ambient, continuous interventions from EHR-integrated decision support to automated messaging systems—the need for structured, reproducible digital development and deployment frameworks becomes increasingly critical. In this abstract, we present ATLAS (Architecture for Translational Learning & Applied Systems), a co-design framework inductively derived from five real-world implementations at NYU Langone Health, spanning generative AI-enabled chronic disease management, personalized automated messaging, and pharmacy-integrated adherence tools. This framework emerged through iterative analysis of implementation challenges, workflow disruptions, and integration failures. Across these initiatives, six recurring sociotechnical themes emerged: early workflow mapping, clinical champion engagement, explicit minimum viable product scoping, structured requirement negotiation, technical architecture clarity, and role-defined team structures. ATLAS operationalizes these themes as front-loaded governance constraints embedded directly into the development lifecycle—treating workflow analysis, technical scoping, and stakeholder alignment as first-order engineering concerns rather than peripheral considerations. Findings suggest this approach reduces deployment friction and increases implementation consistency across clinically and technically diverse settings. For the ubiquitous computing community, ATLAS offers a translatable model for bridging the persistent gap between ambient system capability and sustainable clinical adoption.

I present lessons from building and deploying real-world health systems across two settings: clinical interactions and campus mental health. This work spans multimodal sensing of patient-provider communication and large-scale digital platforms for student wellness. While these systems show promise in controlled settings, deployment reveals a different set of constraints that shape their real-world utility.

A recurring challenge is trust. In clinical settings, automated assessments of social signals such as dominance or empathy often lack clear pathways for action, limiting their usefulness for providers. In student mental health systems, recommendation pipelines surface resources but do not always align with how students seek or prioritize support. Across both contexts, issues of agency, actionability, and privacy emerge as central barriers. Users are less likely to engage with systems that provide insights without control, or that rely on sensing pipelines that are difficult to interpret or contest.

A second set of challenges arises from imperfect data. Real-world deployments must operate under missing, noisy, or weakly labeled inputs, whether from speech transcripts, wearable sensors, or app usage logs. I discuss strategies for maintaining system responsiveness under these conditions, including designing for partial observability and shifting from prediction to interaction.

This talk argues that the key bottleneck in ubiquitous health systems is not model performance alone, but how systems integrate into human workflows. Addressing trust, agency, and robustness is critical for moving from promising prototypes to sustained, real-world impact.

TBD

Personalized health assistants have long been a vision in the Ubicomp and health community. Realizing this vision requires systems that can understand patients’ situations and deliver support that is timely, relevant, and tailored to everyday care. Yet adaptation remains a core challenge: patients differ in their needs, and real-world care settings are difficult to sense reliably due to environmental variability, behavioral diversity, and limited training data. My thesis investigates how health assistants, despite limited understanding of an individual’s context, can adapt over time through human-in-the-loop interaction. I study this question through two long-term health projects, using them as cases to examine the effectiveness of these designs and their potential downstream impact on care processes and health outcomes.

Self-tracked health data is becoming central to how people manage their health, yet the data quality issues that quietly undermine it remain largely overlooked. Gaps in data, inaccurate recordings, validity concerns, and loss of contextual meaning all quietly undermine a user's ability to reflect on their health and make meaningful progress toward their goals. My current work focuses on understanding and characterizing these data quality problems in self-reported data. When self-reported data is incomplete, inaccurate, or stripped of context, it fails to reflect a user's real experience, making meaningful reflection or progress impossible and the effort feel pointless. Through a synthesis of 85 papers across HCI, digital health, and health informatics, we are developing a heuristic evaluation framework aimed at supporting early identification of data quality issues in the design lifecycle. The framework consists of questions aimed at assisting researchers and designers identify and address data quality failures earlier in the design cycle to mitigate them and avoid poor end-user experiences

Traditional eye-tracking measures, including gaze and pupil dynamics, have long been associated with psychiatric conditions and used as proxies for underlying cognitive processes. However, these measures have typically been collected in highly controlled laboratory or clinical settings, limiting understanding of naturalistic cognitive patterns in everyday life. In contrast, smartphone-based sensing enables measurement in real-world environments. Building on advances in computer vision, we present a novel smartphone-based eye-tracking sensor embedded in the background of Therabot, a generative AI chatbot for mental health treatment. Throughout app use, a computer vision pipeline passively captures facial images from the smartphone’s front-facing camera and extracts numerical coordinates corresponding to key facial features, including the eyes and irises. Pupil circumferences are then derived from the isolated iris regions. To preserve privacy, all facial images are processed in real time on-device and are not stored or transmitted. In a feasibility evaluation, face, eye, and pupil features were successfully detected at an average sampling rate of approximately 1 frame per second, with occasional bursts reaching 7 to 9 Hz. Irises were detected in 99% of frames, while pupil measurements were obtained in approximately 15% of frames. Despite the lower rate of pupil detection, these measurements still provided usable signals for examining pupil dynamics because of the temporal granularity of the data. Overall, this work demonstrates the feasibility of passive, privacy-preserving, smartphone-based eye tracking and establishes a foundation for studying cognitive processes during real-world interaction with digital mental health applications.

Passively sensed multimodal data from smartphones and wearable devices offer a rich, continuous view of individuals’ daily behaviors. However, the high volume, longitudinal nature, and heterogeneity of these data pose significant challenges for interpretability and clinical utility. Recent advances in large language models present an opportunity to translate complex sensing features into structured, clinically meaningful natural language insights.

In this work, we aim to develop an LLM-powered interactive dashboard using a longitudinal dataset of multimodal smartphone sensing data collected from over 50 older adults across 12 months. The system is designed to support clinicians in assessing cognitive impairment and to assist caregivers in monitoring everyday behavioral changes. The dashboard first employs an insight extraction module that identifies salient patterns in the data, such as long-term trends, deviations, and anomalies, using statistically rigorous methods. It then integrates a retrieval-based literature grounding module to retrieve and synthesize relevant scientific evidence, linking extracted behavioral patterns to cognitive decline.

We will evaluate the system’s effectiveness and usability through survey-based user studies and semi-structured interviews with neuropsychological clinicians and caregivers, focusing on interpretability, perceived clinical and everyday utility, and trust in generated insights. This work presents a step toward bridging passive sensing data and clinical interpretation, enabling more accessible and evidence-informed monitoring of cognitive health through LLM-based interfaces.

Postpartum depression (PPD) is a serious perinatal mental health condition affecting approximately 20% of women worldwide. Common screening approaches for PPD, such as self-report questionnaires and active digital logs, rely heavily on user input and thus impose a substantial burden on participants, limiting their feasibility for long-term use. Recent passive mobile sensing (PMS) approaches have enabled low-burden detection of depressive symptoms using machine-learning methods with multimodal sensor data from off-the-shelf mobile devices, including smartphones. However, the postpartum period entails distinct behavioral patterns, raising uncertainty about whether sensing-based indicators for general depression and mental disorders generalize to PPD scenarios. To address this gap, we propose PocketPPD, a PMS-based PPD screening method that detects PPD risk using maternal contextual features, such as disruptions in behavioral rhythms and shifts in stability, collected through a smartphone. Our experimental results using a dataset of 61 postpartum women over a four-week period showed that the PMS-only model achieved an AUC of 0.75, while the best-performing model, integrating PMS-oriented data and self-report features, achieved an AUC of 0.83. Moreover, we find that morning and late-night routine volatility offers strong predictive power and is dynamically moderated by maternal contexts, such as infant developmental stage and employment status. This work provides empirical evidence for low-burden PPD risk screening, and our findings lay the groundwork for continuous perinatal mental health monitoring.

Wearables and smartphones generate more health-relevant data per person per day than any clinician could ever review, yet most of it never translates into personalized, actionable support. My research designs and builds multimodal AI systems that empower individuals and domain experts to transform heterogeneous health and well-being data into evidence-based guidance, sitting at the intersection of HCI, data visualization, and ubiquitous computing. Across my PhD at HKU and my ongoing postdoctoral work at Columbi, I have developed three complementary lines of work: (1) wearable-signal pipelines for screening and monitoring, including a feasibility study on diagnosing ADHD in adolescents and context-aware health inference for children during school closures; (2) visual analytics systems such as HealthPrism that help clinicians and caregivers make sense of multimodal child-health profiles; and (3) generative-AI companions for everyday health and emotional well-being, including DietGlance for knowledge-empowered dietary analysis and Hear You in Silence for context-aware active listening in conversational agents. A unifying thread is treating AI not as a black-box predictor but as a collaborator that scaffolds human sensemaking, surfaces evidence, and responds to people in ways aligned with clinical knowledge and individual context. At the symposium, I hope to discuss longitudinal deployments, clinician-in-the-loop evaluation, and emerging settings where multimodal sensing, generative AI, and interactive visualization meet real clinical and caregiving practice. I am actively looking for collaborators working on shared problems.

Researchers conduct longitudinal passive sensing studies in in-the-wild settings, often spanning months or years, to uncover naturalistic behavioral patterns. These studies are not “set-and-forget” deployments; they require continuous monitoring as technical failures and declining participant compliance can lead to substantial missing data, undermining study validity and downstream models. Thus, conducting these studies involves multiple detail-oriented, cognitively demanding, and time-consuming tasks, making it a burdensome and stressful process. Existing research dashboards, the primary tools for data monitoring, offer limited support in easing this burden. Leveraging recent advances in AI for passive sensing data, we explore the design of human-AI collaborative workflows enabled through research dashboards to improve the effectiveness and efficiency of monitoring and associated tasks. We begin with a co-design study with 13 researchers involved in longitudinal sensing studies to identify desired AI capabilities and interactions through semi-structured interviews, brainstorming, and sketching activities. We establish the feasibility of novel human-AI workflows our participants envisioned by implementing an AI-augmented dashboard prototype DAIMON, and use it as a research probe in two studies: a task-based study and a deployment within an ongoing real-world sensing study. Our findings demonstrate the promise of AI-augmented dashboards in supporting researchers’ day-to-day data monitoring and decision-making tasks. It also surfaces concerns around transparency and expectations with AI systems. Consolidating insights across all three studies, we present design guidelines for AI-augmented dashboards for longitudinal passive sensing research and discuss directions for future work.

Cancer survivors face elevated rates of depression, anxiety, and general emotional distress, yet the precise moments they most need support are often the moments when self-report is sparse, a phenomenon we term the diary paradox. Passive smartphone sensing offers a continuous, unobtrusive alternative, but prior sensing-based affect prediction has been limited by an accuracy ceiling, suggesting a bottleneck not only in available data, but in how behavioral signals are interpreted. We present Pulse, a system that shifts from fixed feature pipelines to agentic sensing investigation: LLM agents equipped with eight purpose-built tools autonomously query smartphone sensing data, compare current behavior against personalized baselines, and calibrate inferences through retrieval- augmented population-level comparisons. Rather than receiving pre-formatted feature summaries, agents decide which modalities to inspect, how far back to look, and how deeply to investigate, mirroring hypothesis-driven clinical reasoning. We evaluate Pulse through a 2×2 factorial design crossing reasoning architecture (structured vs. agentic) with data modality (sensing-only vs. with diary) on 50 cancer survivors from a longitudinal study of cancer survivors. Agentic reasoning is the primary driver of performance: agentic multimodal agent achieves balanced accuracy of 0.743 for emotion regulation desire with diary and sensing data, while agentic agents predict intervention availability at 0.713 with passive sensing data only. These results suggest that agentic investigation may be a cornerstone for unlocking the clinical value of passive sensing, advancing the feasibility of proactive just-in-time mental health support

Peripheral Artery Disease (PAD) is a common cardiovascular condition that is underdiagnosed, in part, due to the lack of accessible screening options. Photoplethysmography (PPG) is a noninvasive optical measurement that has potential to be a screening solution as it can capture peripheral vascular physiology. This work presents the first large-scale analysis of the relationship between PPG and PAD severity using clinical data from UCSD Health (5,237 signals from N=2,362 unique patients). We find significant associations between PPG morphology and PAD, which serve as the scientific basis for the development and validation of a machine learning model to classify disease status. Our PAD digital biomarker achieves consistent results across demographics, clinical comorbidities, and location site. Additionally, we (AUC = 0.83) vastly outperform traditional clinical methods of PAD screening (AUC = 0.67) with a four-second PPG signal alone. These results provide a scientific and translational foundation for scalable PAD screening, with the potential to support earlier detection, reduce healthcare costs, and improve long-term cardiovascular outcomes. Building on these findings, efforts are now directed toward prospective evaluation, smartphone-based deployment, and assessment across diverse clinical populations.

Agentic AI for Dementia Care: The Next Frontier for Ubiquitous Health Intelligence

Dementia care is emerging as one of the most important yet insufficiently addressed challenges for ubiquitous health AI. Effective dementia management requires continuous interpretation of diverse and evolving signals, including speech patterns, daily behavior, caregiver observations, neuroimaging findings, and electronic health records, collected over months or even years of care. Conventional clinical AI systems have improved isolated prediction tasks, but they remain limited in their ability to translate fragmented multimodal information into clinically meaningful actions. Recent progress in agentic AI offers a new direction by enabling large language model systems to retrieve evidence, coordinate multiple analytical tasks, and generate sequential care recommendations. Despite this promise, current dementia AI efforts remain disconnected across modalities, with neuroimaging, speech, and EHR-based systems advancing independently rather than functioning as an integrated reasoning partner for clinicians and caregivers. The major barrier to clinical translation is no longer prediction accuracy alone, but the lack of clinically embedded agentic architectures that can unify passive sensing, multimodal reasoning, and human-centered trust across the relationships among clinicians, caregivers, and people living with dementia. This presentation outlines clinically embedded agentic architectures that transform passive monitoring into interpretable support for long-term dementia management.

Perinatal mood and anxiety disorders (PMADs) are the leading cause of pregnancy-related deaths in the US, with risk exacerbated in rural communities where support sharply declines after birth. Doulas (Medicaid-reimbursable professionals) can reduce PMAD onset by over 50%, but they lack tools to support sustained care. This work explores how ubiquitous, AI-enabled systems can support doulas in delivering continuous, coordinated postpartum mental health care. Findings from 10 stakeholder interviews revealed that doulas lack a centralized patient-management platform to track progress, coordinate operations and logistics, provide evidence-based resources, and deliver consistent support. Doulas currently rely on fragmented tools for care and monitoring of patients, limiting their ability to deliver continuous, context-aware support.

Our work presents the design and preliminary usability evaluation of DoulaConnect, an evidence-based, AI-enabled dashboard co-designed with doulas serving mothers with postpartum depression in rural New England. Using a human-centered design process grounded in community needs surfaced through Community-Based Participatory Research (CBPR) sessions, we translated stakeholder needs into a feature-traceability matrix, high-fidelity wireframes, and eventually a functional AI-enabled dashboard prototype. Our evaluation involves one-on-one sessions with 10 certified doulas to test structured tasks using think-aloud protocols and measures such as the System Usability Scale and Post-Study System Usability Questionnaire. Findings will report on dimensions of effectiveness, efficiency, and satisfaction to inform iterative system refinement.

This work contributes (1) an AI-enabled, doula-facing system for continuous postpartum health support and (2) empirical insights into how a CBPR-informed co-design process can improve usability and coordination in fragmented, resource-constrained maternal mental health contexts.

Large language models (LLMs) are increasingly used to assist clinicians in responding to patient messages within electronic health record (EHR) systems. These tools are often framed as improving communication: making it more efficient, clear, or empathetic. However, in clinical settings, communication is not simply relational interaction; it is documented, legally accountable, and embedded in complex organizational workflows. In this context, communication functions as accountable labor. Drawing on research and implementation experience with AI-assisted clinical messaging, my lightning talk argues that LLMs reshape institutional labor more than they enhance interpersonal communication. I identify three key tensions: (1) the limits of standardization in contexts requiring situated clinical judgment, (2) the redistribution of work from drafting messages to reviewing and supervising AI-generated outputs, and (3) emerging challenges to responsibility and accountability structures when AI participates in message production. This reframing suggests that, in high-liability domains like healthcare, the primary impact of LLMs is not on how clinicians write, but on how communication work is organized, evaluated, and governed. I argue for design approaches that prioritize workflow integration and make responsibility structures explicit.

Sustained attention is a transdiagnostic marker of mental health, and online education removes most cues educators and clinicians rely on to notice when a student's attention is failing. Disengagement is the common pathway from missed learning to isolation and declining mental health, yet no widely deployable technology can say when it occurs. Measuring engagement reliably and continuously is the enabling step for personalized support. Engagement, however, is multidimensional whose physiological signature shifts across cognitive, behavioral, and emotional facets, and no unimodality captures it holistically. We present En3, a head mounted unobtrusive multimodal wearable system (EDA, EEG, eye tracking) that predicts different engagement tasks.

In a mixed methods study (N=30) of students watching educational videos, we collected multimodal signals, engagement self-reports, affect ratings, comprehension quizzes, and semi structured interviews. Under leave-one-subject-out cross-validation, the best performing modality depends on the construct: eye tracking for coarse engagement (F1 = 78.9%), EDA for ordinal prediction (MAE = 0.79), and EEG for learning outcome (F1 = 75.3%) — each localizing a different facet of engagement.

Our notable findings are that forehead mounted EDA matches traditional wrist placement performance (F1 = 71.0% vs. 68.3%), and that modalities best at predicting felt engagement diverge from those best at predicting learning outcomes. Together, these findings argue that engagement is a construct specific measurement problem, modality should be matched to the facet of interest. And that head worn sensing suffices to move engagement aware mental health support out of the lab, increasing personalized care accessibility.

Collecting in-situ behavioral data to support research studies often involves prompting repeated ecological momentary assessments (EMAs) on smartphones and wearables. Sustained engagement, however, remains a major challenge because study participants often experience survey fatigue and declining motivation over time. To address this challenge, we introduce PAGE: a novel LLM-powered framework that integrates longitudinal data collection with a personalized game experience. PAGE uses a participant's previously collected data to dynamically generate personalized puzzles, embedding behavioral data collection within engaging gameplay. We deployed a prototype using a variant of a word wheel puzzle in a two-week feasibility study (N=15); while participants responded to EMAs in both weeks of the study, they also received sensor-based personalized puzzles in one of those weeks. Our findings demonstrate the potential of PAGE: most participants (80%) preferred the week with personalized puzzles and contributed more and richer data (72 more unique contextual tags / 1,188 in-game survey responses). We discuss design guidelines for developing context-aware personalized games to support engagement in longitudinal studies.

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. Leveraging participant-labeled data, we further developed a multimodal model achieving a balanced accuracy of 90.36% and a sensitivity of 91.17% on 33,698 labeled 15-second 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.

Individuals often form deep emotional attachments to physical objects such as plush toys, keepsakes, and character merchandise, yet these relationships remain constrained by the objects’ passivity. Existing AI companions offer conversational responsiveness, but require users to build emotional bonds from scratch, often imposing a relational burden that leads to abandonment before meaningful attachment can form. We present a dual-embodiment companion framework that bridges this gap by anchoring a digital AI companion to a user’s existing cherished physical object, allowing a single companion identity to extend across physical and digital forms. Through a formative interview study with adults who maintained emotionally significant object attachments (N = 9), we derived four design principles: Faithful Identity, Calibrated Agency, Ambient Presence, and Reciprocal Memory. We operationalized these principles in an end-to-end mobile system that integrates augmented reality and multimodal large language models to support identity synchronization, situated interaction, autonomous agency, and evolving memory. In a within-subjects lab study (N = 25), the dual-embodiment system significantly outperformed a standard personalized digital companion on all relational measures, including companionship, emotional bond, perceived agency, and shared memory, suggesting that physical grounding can successfully transfer pre-existing attachment. A seven-day field deployment (N = 17) further revealed three key relational dynamics: bidirectional vitalization, in which digital activity retroactively animates the physical object; a quantity-quality decoupling, in which bond deepening depends more on emotional vulnerability than interaction frequency; and ontological oscillation, in which users actively navigate between experiencing the companion as alive and recognizing it as AI. Our findings suggest a paradigm shift for affective computing: from building companionship from scratch toward transferring and extending attachment through everyday cherished artifacts.

Forming meaningful friendships in emerging adulthood is increasingly mediated by digital contexts, yet most technology-facilitated interaction platforms apply one-size-fits-all engagement structures regardless of users' momentary emotional states. This study examines whether mood-state personalization improves friendship outcomes between stranger dyads communicating via asynchronous text messaging. Drawing on the self-expansion framework, we developed a three-tier engagement structure — microengagement tasks, bridge questions, and deep closeness prompts — in which question selection is dynamically calibrated to participants' reported mood states. Positive mood states trigger creative and generative tasks, while low-affect or high-stress states prompt minimal-load activities and delay high-vulnerability disclosures. A large language model operates behind the scenes to analyze conversation content and personalize engagement without disrupting the human-to-human nature of dyadic interaction. College students (ages 18–29) are currently being recruited into stranger dyad pairs and randomized to personalized versus standard engagement conditions. Friendship formation is assessed using validated closeness scales, with partner knowledge as a key secondary outcome. I discuss preliminary design decisions and their theoretical implications, including the operationalization of emotional capacity as a moderator of self-disclosure timing.

With the increasing prevalence of frequent cannabis use, there is a growing need for continuous monitoring and timely intervention. Although existing pervasive sensing technologies can capture physiological and behavioral signals in real time, they often lack in-depth support for clinical decision-making and fail to effectively facilitate communication between patients and clinicians. To address this, this paper proposes the SDM system: a multimodal framework that seamlessly integrates passive sensing, explainable artificial intelligence and Shared Decision-Making (SDM). The system introduces an innovative three-party interaction paradigm, enabling the AI to mediate between patients and clinicians, combining emotional context with clinical intent to generate evidence-based, role-specific responses. We have constructed a closed-loop processing workflow encompassing sensing, reasoning and bidirectional interaction, and have comprehensively validated it through high-fidelity simulated clinical scenarios and double-blind expert evaluations. The results demonstrate that the system, incorporating a multi-perspective Chain of Thought (CoT) reasoning architecture, outperforms standard baseline models by a significant margin in terms of providing practical clinical assistance, professionalism and overall preference. Specifically, this multi-perspective strategy not only retains excellent empathy and emotional comfort capabilities but also exhibits outstanding clinical utility in medical decision-making tasks requiring high information density and complex reasoning. This study demonstrates that integrating multimodal perception with explainable AI combined with domain-specific knowledge graphs can effectively balance the rigor of clinical decision-making with human-centered care, providing robust technical support for the development of complex digital behavioral intervention systems in the real world.

Pedestrian detection systems in fully automated vehicles (SAE Level 4) can provide varying levels of information detail to passengers, but optimal information granularity remains unclear. This study examines how different levels of detail (LOD0--LOD4) in head-down pedestrian detection displays affect passenger attention allocation, trust in automation, and situational awareness during mixed-traffic scenarios. Using a Unity-based vehicle simulator with integrated eye tracking (Meta Aria Gen1 glasses), participants complete monitoring tasks while observing pedestrian detection displays ranging from no information (LOD0) to comprehensive displays including position and pedestrian orientation (LOD4). All displays are visual-only, presented on a head-down display (HDD), with no auditory or haptic alerting---isolating information granularity as the sole independent variable. Eye tracking provides objective measurement of attention allocation between road monitoring and display interaction, while subjective measures assess trust calibration and cognitive load. Results will establish optimal information granularity for pedestrian detection interfaces that balance information utility with attention demands in fully automated driving contexts.

Wearable biosensing allows for continuous monitoring and intervention in daily settings. Wrist photoplethysmography (PPG) is a commonly used method for measuring cardiac activity ambulatorily. A typical data preprocessing step is to apply a fixed, one-size-fits-all band-pass filter before peak detection and pulse rate variability (PRV) calculation. However, our analyses reveal that fixed filtering can cause significant errors in PRV estimation even with minimal motion artifacts, and the best filter settings differ across individuals and internal states. Based on these findings, we introduce an open-source, window-level pipeline that adaptively selects filter parameters in a person- and state-dependent way. In tests across two datasets, our adaptive method improved PRV (RMSSD) accuracy by up to 279ms compared to a fixed filter and enhanced stress detection. It also increased the amount of usable inter-beat intervals without losing accuracy. Our collective results suggest shifting focus from solely removing motion artifacts toward adaptive preprocessing that considers who is being measured and their state. Our pipeline offers a standardized way to achieve more accurate analyses in health applications.

Eating disorders for which body image disturbance and body image concern are clinically relevant features (including anorexia nervosa, bulimia nervosa, and binge eating disorder) carry among the highest mortality rates of any psychiatric condition. Individuals with these disorders demonstrate high use of short-form video platforms (SFVPs) and some evidence suggests these users are served over 4,000% more body-oriented content promoting disordered eating behaviors, with increased platform use potentially worsening symptomatology. Existing research suggests that individuals with eating disorders exhibit unique visual attention patterns when viewing static images of their own or other people’s bodies, frequently fixating on specific features like stomach or thigh areas. We aim to examine whether these observations extend to dynamic content on SFVPs, where prolonged exposure to body-oriented content is prevalent. In an upcoming study, we will track gaze behavior while participants view a range of SFVP content, sampled by sex, body size, skin visibility, and movement speed. Simultaneously, we will utilize camera-based methods using a mobile device to measure exploratory variables, including blink rate, pupil dilation, and respiratory signals via remote photoplethysmography. This poster reports the findings from a self-administered feasibility study which explores the viability of integrating these techniques and variables into the full study.

Smartphones and wearables offer the potential to detect and monitor behavioral and contextual states linked to mental and affective health outcomes. Machine learning models have shown promise in detecting these outcomes, yet current approaches are commonly designed to generalize across users, often at the expense of capturing individual variability. Behavior is unique across each person, but not always consistent. Over time, there may also be changes within the individual. While personalized models may intuitively yield more accurate detection and prediction, there are challenges with limited data and the availability of ground-truth labels. In this work, we explore machine learning architectures that perform well under these constraints, examining how model design choices impact detection and prediction at the individual level. With this foundation, we explore variations in encoder design, training window strategy, class weighting, ground-truth label augmentation, and look-back length to assess what configurations best capture individual-level patterns. We train and evaluate models on a per-participant basis, directly addressing the individual variability that pooled approaches obscure. We evaluate this strategy while predicting stress and craving in a longitudinal dataset (n=59) over 12 weeks, with patients undergoing opioid use disorder (OUD) treatment. Preliminary results suggest that configurations should differ across participants, and adapting the model to each individual may account for the personal variability.

Accurate measurement of everyday behaviors is critical for research in human–computer interaction, ubiquitous computing, and personal health informatics. However, most behavioral data collection still relies on self-reports and manual activity annotations, which are time-consuming, error-prone, and require substantial post hoc cleaning. In this talk, I present two complementary systems—ACAI and GLOSS4HAR—that explore how human–AI collaboration can improve activity annotation in free-living settings. First, ACAI is a context-assisted activity annotation interface that helps participants efficiently label daily activities using system-generated suggestions, while allowing quick corrections and the expression of temporal uncertainty. A usability study and a two-week deployment show that ACAI reduces annotation effort while improving data validity compared to gold-standard recall methods. Building on the dataset collected from this deployment, GLOSS4HAR is a multi-agent LLM-based system that assists researchers in cleaning and refining annotations. By combining passive sensing data with participant reports, it detects inconsistencies, suggests corrections, and reconstructs detailed activity timelines. Together, these systems demonstrate a full-cycle approach to behavioral data collection, from reducing participant burden to supporting post hoc validation. This work highlights the potential of human–AI collaboration to enable scalable, high-quality behavioral datasets.

Introduction. Ubiquitous sensing devices for blood pressure (BP), glucose, and activity are now widely available to patients, yet transmitting that data to clinicians and closing the feedback loop remains a persistent challenge. Remote patient monitoring (RPM) programs automate data transmission and trigger clinician feedback, but engaging patients across onboarding, adherence, and data-driven care changes has proven difficult. NYU Langone Health (NYULH) operates one of the largest ambulatory RPM programs in the United States, with hypertension as its primary use case, and has introduced a recently deployed generative AI-enhanced arm to improve engagement. Understanding which approaches work, for whom, and why non-engagers disengage is essential to guiding real-world deployment of ubiquitous health sensing. Methods. Five cross-sectional mixed-methods survey waves (September 2023 through April 2026) were administered to patients in NYULH's hypertension RPM program. Surveys combined 5-point Likert items (satisfaction, NPS, empowerment, alert helpfulness, care team closeness) with open-ended prompts analyzed for recurring themes. Results. Across 3,375 invitations, 301 surveys were completed (9%). Satisfaction and NPS remained high across waves. Patient-reported receipt of helpful BP alerts declined from 65% (2023) to 34% (2025). Early AI-cohort data showed alert endorsement nearly double that of standard users (60% vs. 34%). Non-engagers understood the program but cited device, onboarding, and technical barriers; over half remained willing to engage. Discussion. Ubiquitous sensing is expanding rapidly, but in healthcare, sensor deployment is not the same as implementation. Sustaining engagement depends on clinical workflow integration, patient-provider relationships, and closed-loop feedback, and generative AI may offer a tractable path to closing that loop at scale.

LLMs have been widely applied across diverse fields for various purposes, with prompting playing a central role. Through a systematic review of 152 papers involving human participants evaluating LLM-based systems, we examine how researchers conceptualize and use "prompt" and its related constructs. We find that researchers introduce new terms to describe different types of prompts and strategies for eliciting desired outputs---relevant to the specific tasks, target users, desired LLM characteristics, and interaction ways. However, inconsistencies in terminology persist, highlighting the need for standardization to better organize prompt-related concepts. We conclude by synthesizing the extracted concepts through a taxonomy to support ongoing research and discussion.

Humans tend to attend to, learn from, and use negative information more than positive information, a phenomenon known as negativity bias. This bias can intensify chronic stress and contribute to mental and physical health challenges. Gratitude journaling is one widely adopted practice that may help counter this bias.Although the benefits of short-term gratitude interventions are well-studied, we still know less about people’s long-term journeys with gratitude journaling: what motivates them to practice, how it evolves over time, and what helps them sustain it.

To address this gap, we conducted a qualitative study of 17 long-term gratitude practitioners(1-15 years). Using artifact-walkthrough retrospective interviews, we examined their journeys. Our findings show that gratitude journaling is a dynamic and context-sensitive practice. Participants engaged in diverse modalities, and their motivations and patterns of engagement shifted alongside emotional states and life circumstances. Through continued practice, participants learned to redirect attention toward present and positive aspects of experience, reinterpret adversity in meaningful ways, recognize goodness as coming from beyond the self, and respond with prosocial reciprocity. These experiences often fostered a sense of meaning and existential grounding, reinforcing engagement and supporting habit formation. Gratitude communities further provided mutual encouragement, including during periods of life struggle and addiction recovery.

These findings extend current understanding of gratitude journaling as an evolving lived practice and suggest design opportunities that adapt to changing life contexts, support meaningful reflection, and incorporate social features to sustain motivation across lapses and life transitions.

Title: Understanding Naturalistic Use of a Mental Health AI Chatbot among Family Caregivers: An Ongoing Analysis of 8-Week Field Deployment Data

Family caregivers of persons living with dementia face accumulating stressors that compromise their well-being. LLM-based mental health chatbots show promise in supporting caregiver self-management; our pilot study observed improvements on anger, anxiety, depression, perceived stress, fatigue, and sleep with medium-to-large effect sizes. Yet little is known about how caregivers actually engage with such a chatbot in their everyday caregiving context, such as what they bring to these conversations, what they hope to gain, and how these patterns unfold across weeks of use. We present an ongoing analysis of 13,637 message exchanges across 493 conversation sessions (90+ hours) from 22 dementia caregivers interacting with a mental health chatbot over 8 weeks, paired with baseline, midpoint, and end-of-study self-report well-being measures. We examine conversational topics, communicative purposes, and emotional expression as dimensions of caregiver-chatbot interaction. LLM-assisted inductive coding of caregiver chat messages has surfaced five preliminary categories of communicative purpose: emotional expression, narration and daily check-in, problem-solving and advice-seeking, meaning-making, and rapport and relationship-building with the chatbot. The analysis is being refined in dialogue with existing taxonomies of social support and therapeutic communication. Ongoing work will further examine emotional dynamics across the deployment and how engagement patterns relate to well-being trajectories. As additional caregivers are enrolled in subsequent deployments, future work can leverage the expanded dataset to develop a taxonomy of family caregivers’ engagement with conversational AI in self-management contexts.

For people living with chronic pain, ordinary everyday technologies may already be health technologies, even when they were not designed as such. Yet many digital health approaches to chronic pain still center on symptom tracking, numeric reporting, and individual self-management. This ongoing dissertation project argues for a care-centered UbiComp approach that attends to the everyday, invisible, and relational work through which people live with chronic pain.

Using fibromyalgia as a focal case, this project examines how people use everyday digital technologies across both typical days and flare-up days. The current study uses an online survey to compare how technology use shifts when symptoms intensify, including changes in communication, pacing, planning, distraction, support-seeking, and self-care. A follow-up qualitative phase will use interviews or diary-based inquiry to examine specific moments when technology becomes part of managing pain, deciding whether to disclose, asking for help, withdrawing, or making daily life workable.

Rather than asking only how ubiquitous technologies can capture better health data, this work asks how they might support forms of care that are difficult to measure but central to lived experience. The contribution is a design-oriented reframing of chronic pain technology: from tools for tracking symptoms toward systems that support fluctuating, relational, and care-centered everyday life.

Electroencephalography is widely used for cognitive assessment, mental state monitoring, and brain computer interface research. It provides a practical and noninvasive method to measure neural activity during everyday tasks. However, EEG signals can reveal more than cognitive state. They can also encode stable personal information such as participant identity and demographic attributes.

This study examines privacy leakage in low density scalp EEG. The system uses a four channel frontal temporal recording setup. Data is collected from 24 participants. Each participant completes three conditions. The conditions include rest, Stroop, and arithmetic tasks. Standard machine learning models are trained on spectral features and asymmetry features. The models predict three targets from the same EEG signals. The targets include task, participant identity, and gender.

The results show clear task related leakage. Task prediction exceeds the 33.3% chance level for all participants. In addition, 22 of 24 participants achieve task accuracy above 50%. The results also show stable personal information leakage. Participant identity prediction reaches 47.9% balanced accuracy and 0.431 macro F1.We also find out the most privacy leaking informative patterns appear in the theta and alpha frequency bands.These results show that EEG features used for cognitive task detection can also enable inference of participant identity and gender, highlighting the need to incorporate neuroprivacy as a core requirement in EEG system design.

People's daily stress experiences pose significant risks to their mental and physical health. To better understand these experiences, researchers increasingly use wearable devices to detect potential stress events through physiological signals (e.g., heart rate variability). However, these signals are difficult to interpret without subjective context from the person experiencing them. To capture this context, prior work has widely adopted Ecological Momentary Assessments (EMAs), but their structured formats (e.g., fixed scales) limit the richness of self-reports and can become repetitive and burdensome for people over time. We propose StressTalk, a novel approach that uses LLM-based conversational agents (CAs), grounded in passive sensing data, to gather stress self-reports. With StressTalk, users engage in two brief daily conversations to share their stress self-reports: a prospective morning check-in and a retrospective end-of-day reflection. We hypothesize that StressTalk will yield richer stress self-reports than traditional EMAs, while maintaining comparable or lower participant burden. A preliminary study showed that users found StressTalk more natural, engaging, and less burdensome for sharing stress self-reports than EMAs. Building on these findings, we aim to conduct a 6-week within-subject longitudinal study to evaluate StressTalk at scale.

Digital environments are increasingly embedded in people's everyday lives, yet they often contain design features that disrupt attention, including notifications, recommendations, autoplay, ads, fragmented layouts, and task-switching cues. These challenges are especially significant for students with ADHD, who often need to navigate course platforms, assignments, readings, emails, and deadlines while sustaining focus across academic tasks. In these contexts, interface choices that prioritize engagement over study goals can make it harder to stay on task and return to coursework after distraction. While prior HCI research has examined digital well-being and ADHD support tools, less is known about how specific interface elements shape attention regulation during everyday goal-oriented digital tasks for students with ADHD.

This ongoing study investigates how students with ADHD experience, interpret, and manage digital distractions while completing tasks such as studying, writing, coding, and project management. Through a contextual inquiry with students with ADHD, we observe participants' real-world digital workflows to identify design elements that influence attention regulation (e.g. shifting, sustaining, and recovery).

This ongoing research aims to surface design implications for students with ADHD and offer insights for building supportive digital tools that help users maintain focus and stay productive while preserving autonomy, flexibility, and user control.

Overeating and emotional eating are increasingly common health issues that affect people even without an eating disorder. The vagus nerve plays a critical role in the gut-brain axis, and implanted vagus nerve stimulators have been associated with reduced appetite. In this paper, we propose transcutaneous cervical vagus nerve stimulation (tcVNS) as an immediate, low-effort intervention during or before an eating episode. In a study with 24 participants, we evaluated the use of a mobile, handheld tcVNS device during a single episode of distracted snacking. We found that participants ate 9.6% less and 23.6% more slowly during vagus nerve stimulation than during sham stimulation. Satiety after a snacking episode was the same in both conditions, while heart rate was reduced during vagus nerve stimulation, demonstrating a stronger rest-and-digest response. The stimulation was described as subtle and barely noticeable. Overall, these results demonstrate the feasibility of non-invasive vagus nerve stimulation as a momentary intervention for managing eating behavior in a non-intrusive and fine-grained manner. Beyond eating, this work provides a foundation for ubiquitous interfaces that directly manipulate physiological activity, enabling a new approach for behavioral interventions that does not demand attention.

Conversational agents are increasingly used for health and well-being support, yet evaluating their psychological safety at scale remains a challenge. Existing approaches rely on transcript-based scoring, despite real-world use being longitudinal and shaped by contextual factors not captured in dialogue alone. We present early findings from a three-day remote study (N = 81) of a voice-based conversational agent delivering evidence-based lifestyle interventions across cognitive, mental, physical, and social domains. Participants completed six session types, yielding 479 transcripts. Using a zero-shot large language model (LLM) rubric, we identified episode-level safety events (i.e., turns where safety-relevant signals arise), assigned signal clarity and origin labels, and rated responses as best practice, missed opportunity, or actively damaging, before aggregating to session-level scores. Transcript-only evaluation showed limited resolution. In 66% of sessions, no safety-risk episode was detected, producing ceiling effects. The probability of observing any safety-relevant episode differed by session type (χ² p < 10⁻¹⁰), with higher rates in health check-in sessions than in wellness training sessions. However, among sessions with at least one flagged episode, session-level scores did not differ across session types (Kruskal–Wallis p = 0.41), indicating that aggregation masks differences in how safety issues emerge and are handled. Qualitatively, flagged episodes were dominated by low-clarity safety risk cues (e.g., worry, frustration), with common failure modes reflecting missed opportunities to explore these cues despite otherwise appropriate responses. These findings suggest that transcript-only evaluation fails to distinguish between the absence of safety-relevant signals and missed opportunities to respond to low-clarity safety risk cues. We propose future work to develop context-aware, longitudinal safety evaluation frameworks that integrate conversational data with multimodal context to ground safety judgments in users’ real-world conditions.

Human activity recognition (HAR) is widely used to translate wearable sensor streams into measures of daily behavior. However, reported results are often difficult to interpret across studies because model performance is shaped by dataset-specific choices about sensing devices, preprocessing, evaluation protocols, and activity definitions. These choices are not always documented in enough detail to support replication, and code is frequently unavailable. This makes it unclear whether a reported result reflects generalizable progress or a configuration that is specific to one dataset. The issue becomes especially visible in cross-dataset evaluation, where models that perform well in their original setting may degrade substantially when applied to other public datasets, even when published preprocessing steps are followed. This work examines reproducibility and generalization in wearable HAR by re-running publicly available models on their original datasets and on a broader set of public datasets. We focus on how preprocessing decisions, train/test protocols, and activity-label mappings affect reported performance. Because labels such as “sitting,” “stationary,” and “other” can differ across datasets in scope and annotation rules, we evaluate each model using both the original activity labels provided with each dataset and a shared activity mapping across datasets. Our goal is to move HAR evaluation toward a more transparent and reusable benchmarking workflow. Drawing on practices from other areas of machine learning, where shared datasets and standardized evaluation protocols are used to make model comparisons more transparent, we propose a framework for cross-dataset HAR testing. Such a framework would make published results easier to verify, help researchers understand model robustness before deployment, and clarify which claims reflect generalizable recognition ability rather than dataset-specific optimization.

We present SensorLM, a family of sensor-language foundation models that enable wearable sensor data understanding with natural language. Despite its pervasive nature, aligning and interpreting sensor data with language remains challenging due to the lack of paired, richly annotated sensor-text descriptions in uncurated, real-world wearable data. We introduce a hierarchical caption generation pipeline designed to capture statistical, structural, and semantic information from sensor data. This approach enabled the curation of the largest sensor-language dataset to date, comprising over 59.7 million hours of data from more than 103,000 people. Furthermore, SensorLM extends prominent multimodal pretraining architectures (e.g., CLIP, CoCa) and recovers them as specific variants within a generic architecture. Extensive experiments on real-world tasks in human activity analysis and healthcare verify the superior performance of SensorLM over state-of-the-art in zero-shot classification, few-shot learning, cross-modal retrieval. SensorLM also demonstrates intriguing capabilities including scaling behaviors, label efficiency, sensor captioning, and zero-shot generalization to unseen tasks. It shows the promise of producing coherent, context-aware health narratives, and moves toward a unified wearable health LLM that can support personalized, data-driven health insights and interactions.