Predicting the Self-Initiation of Behavior Change from Self-Reflection of Real-Life Stressors

Presentation Type: Paper - Regular

Click to View Presentation

Presentation Time: 03:30 PM - 03:42 PM

Abstract Keywords: Human-computer Interaction, Mobile Health, User-centered Design Methods, Quantitative Methods, Real-World Evidence Generation
Programmatic Theme: Consumer Health Informatics

Understanding which individuals are ready to initiate stress-related behavior change is critical for developing adaptive consumer health technologies. We present a feasibility study that models the likelihood of self-initiated behavior change using survival analysis applied to longitudinal, participant-generated stressor data. Drawing on a 14-week field study with 99 participants who logged wearable-triggered stressors and received weekly reflective visualizations, we operationalize behavior change using affirmative self-reports of behavioral change. Using cumulative features derived from stress frequency, stressor intensity, and engagement patterns, we evaluate three survival models. Our results show that stress frequency, intensity, and engagement significantly predict which individuals are more likely to initiate behavioral change. These findings demonstrate the feasibility of leveraging longitudinal stress data to enable consumer health systems to estimate when behavioral change readiness peaks.

Speaker(s):
Sameer Neupane, PhD
University of California San Francisco
Author(s):
Sameer Neupane, Postdoc Scholar - University of California San Francisco; Alexander Moreno, PhD - MBZUAI; Mithun Saha, Student - University of Memphis; Peter Washington, PhD - University of California, San Francisco; Santosh Kumar, PhD - University of Memphis;

---

Optimizing AI-Driven Messaging for Type 2 Diabetes Management: Insights from Patient Preference Elicitation

Presentation Type: Paper - Student

Click to View Presentation

Presentation Time: 03:42 PM - 03:54 PM

Abstract Keywords: Patient Engagement and Preferences, Large Language Models (LLMs), Human-computer Interaction, Chronic Care Management
Programmatic Theme: Consumer Health Informatics

Generative AI (GenAI) allows for improved user experience within conversational agents for diabetes management by supporting dynamic, context-aware conversations. In this study, we elicited patient preferences for the communication style of a GenAI-based conversational agent (uMatter) developed to support diabetes management. We conducted an online survey with 125 individuals with type 2 diabetes. The survey included a discrete choice experiment to evaluate participant preferences for different types of messaging attributes. The survey also elicited participant perceptions and feedback on the messages from uMatter. We found significant preference heterogeneity for the inclusion of emojis within the messages. Additionally, qualitative findings indicated that participants had different desired personas and communication styles for the conversational agent. We propose strategies from recent human-computer interaction and natural language processing research that can be used to design GenAI-based conversational agents that align with the communication preferences of patients.

Speaker(s):
Angela Mastrianni, PhD
NYU Grossman School of Medicine
Author(s):
Angela Mastrianni, PhD - NYU Grossman School of Medicine; Defne Levine, MPH - NYU Grossman School of Medicine; Katerina Andreadis, MS - NYU Grossman School of Medicine; Lynn Xu, MPH - NYU Grossman School of Medicine; Priscilla D’Antico, PhD - NYU Grossman School of Medicine; Antionette Schoenthaler, EdD - NYU Grossman School of Medicine; Devin Mann, MD - NYU Grossman School of Medicine;

---

Latent Patterns of Reliability in mHealth Nutrition Data

Presentation Type: Paper - Student

Click to View Presentation

Presentation Time: 03:54 PM - 04:06 PM

Abstract Keywords: Mobile Health, Machine Learning, Data Standards, Qualitative Methods, Healthcare Quality, Artificial Intelligence, Public Health
Programmatic Theme: Consumer Health Informatics

Mobile Health applications offer a cost-effective solution for weight management, yet the prevalence of self-reporting bias—such as underreporting and inconsistent logging—challenges data reliability. While small-scale validation studies exist, large-scale evaluations of reporting quality using machine learning remain scarce. This study utilizes a dataset of 549,382 MyFitnessPal users to develop a Hidden Markov Model (HMM) with Gaussian emissions, characterizing latent behavioral states based on the alignment between caloric surplus and weight change. By analyzing transitions across weight loss, gain, and maintenance cohorts, we identify distinct phenotypes of data trustworthiness. Our results reveal a significant longitudinal trend of "reporting fatigue," where even consistent users frequently migrate toward underreporting or dropout states. This work establishes a quality-evaluation framework, demonstrating that HMMs can effectively detect shifts in data integrity at scale. These insights enable mHealth platforms to implement timely interventions, reinforcing positive tracking behaviors to improve both data reliability and clinical outcomes.

Speaker(s):
Missy Bridgwater, Masters of Science
University of California
Author(s):
Jeffrey Feng, BASc - UCLA; Henry Zheng, PhD Candidate - UCLA; Alexander Bui, PhD - UCLA Medical Informatics;

---

When Mean Absolute Error Misleads: Label Range Confounds Per-Participant Evaluation in Wearable Stress Detection

Presentation Type: Paper - Student

Click to View Presentation

Presentation Time: 04:06 PM - 04:18 PM

Abstract Keywords: Mobile Health, Patient-/Person-Generated Health Data, Machine Learning
Programmatic Theme: Consumer Health Informatics

Wearable models that predict stress could enable just-in-time mental health interventions. However, model performance is typically reported using aggregate metrics that obscure participant-level differences and potential sources of error. Using the K-EmoPhone dataset, consisting of seven days of wearable data and stress reports for 70 participants who passed our inclusion criteria, we examined 24 factors related to label structure, response behavior, and physiological signal quality to identify systematic sources of prediction error. Across model architectures and evaluation paradigms, the primary source of error was the structure of individual stress ratings rather than the modeling strategy. Label ambiguity factors explained up to 70% of error variance. These findings suggest that stress predictability depends strongly on how individuals report stress. Improving self-report reliability and label quality may therefore yield greater gains than further advances in modeling alone.

Speaker(s):
Tanvir Islam, PhD
University of Hawai'i at Manoa
Author(s):
Tanvir Islam, PhD - University of Hawai'i at Manoa; Ali Kargarandehkordi, PhD - University of Hawaii at Manoa; Alexander Kostrinsky-Thomas, DO - UCSF; Haopeng Zhang, PhD - University of Hawaii at Manoa; Finale Doshi-Velez, PhD - Harvard; Xuhai Xu, PhD - Columbia University / Google; Kristina Phillips, PhD - Kaiser Permanente; Roberto Benzo, PhD - The Ohio State University; Sameer Neupane, PhD - University of California San Francisco; Nhung Nguyen, PhD, PharmD - University of California, San Francisco; Adrian Aguilera, PhD - University of California, Berkeley; Peter Washington, PhD - University of California, San Francisco;

---

Cross-Cohort Surrogate Modeling Links Conversational Language Features to Multimodal Biomarkers

Presentation Type: Paper - Regular

Click to View Presentation

2026 Symposium LIEAF Presentation

Presentation Time: 04:18 PM - 04:30 PM

Abstract Keywords: Machine Learning, Patient-/Person-Generated Health Data, Quantitative Methods
Programmatic Theme: Academic Informatics / LIEAF

Conversational language has emerged as a promising digital marker for cognitive decline and neurodegeneration, yet its neurobiological basis remains poorly understood. A key barrier is that existing datasets rarely combine speech recordings with multimodal biomarker measurements. To bridge this gap, we propose a cross-cohort surrogate biomarker framework leveraging shared clinical variables to transfer neurobiological information across datasets. We trained predictive models on clinical variables common to both NACC and I-CONECT to estimate over 300 biomarkers, including CSF, structural MRI, and PET measures, then applied these models to I-CONECT participants lacking direct biomarker data. Spatial mapping revealed regional variation in MRI biomarker predictability, and correlation analyses uncovered structured associations between conversational language features and multiple surrogate biomarkers. These findings suggest that conversational language reflects underlying neurobiological variation relevant to neurodegeneration, and that surrogate modeling can enable language–biology research in cohorts where direct biomarker collection is unavailable.

Speaker(s):
Yijiang Pang, PhD Student
Michigan State University
Author(s):
Alexander Hart, MD - University of Michigan Medical School; Hiroko Dodge, PhD - Massachusetts General Hospital, Harvard Medical School; Jiayu Zhou, Ph.D. - University of Michigan;

---

Estimation-Theoretic Bias Reduction for Oscillometric Blood Pressure Readings

Presentation Type: Paper - Student
Presentation Time: 04:30 PM - 04:42 PM

Abstract Keywords: Healthcare Quality, Information Visualization, Interoperability and Health Information Exchange, Data Mining, Qualitative Methods
Programmatic Theme: Clinical Research Informatics

Oscillometry is the standard method for non-invasive, cuff-based blood pressure (BP) measurement, but it introduces systematic errors that may impact clinical accuracy. This study investigates the sources of these errors—primarily the limitations of oscillometry itself and respiration-induced fluctuations—using BP waveform data from the MIMIC database. Oscillometry tends to underestimate systolic BP and overestimate diastolic BP, while respiration introduces cyclical variations that further degrade measurement precision. To mitigate these effects, we propose an estimation-theoretic framework employing least squares (LS) and maximum likelihood (ML) methods for correcting both single and multiple BP measurements. LS estimation corresponds to conventional multi-measurement averaging protocols, whereas the ML approach incorporates prior knowledge of measurement errors, offering improved performance. Our results demonstrate that averaging repeated BP measurements and incorporating prior knowledge of measurement error statistics can improve the accuracy of non-invasive BP estimation.

Speaker(s):
Masoud Nateghi, PhD Student
Emory University
Author(s):
Reza Sameni, PhD - Emory University;

---