Paper Session 17: Digital Health

Poor diets are a leading cause of morbidity and mortality. Exposure to low-quality food environments, such as ‘food deserts’ with limited access to grocery stores and ‘food swamps’ saturated with fast food outlets (FFO), is hypothesized to negatively impact diet and related disease. However, research linking such exposure to diet and health outcomes has generated mixed findings and led to unsuccessful policy interventions. A major research limitation has been a predominant focus on static food environments around the home and workplace, and sparse availability of information on the dynamic food environments people are exposed to and the food outlets they visit as they move throughout the day. We leverage population-scale mobility data – anonymized GPS trajectories from people’s smartphones, collected from approximately 15 million U.S. adult residents – to examine peoples’ visits to food outlets and FFO in and beyond their home neighborhoods. First, we show that visits to FFO provide strong and significant indicators of FF intake and diet-related disease. Second, we develop a semi-causal framework using several natural experiments to evaluate how food choice is influenced by features of food environments people are exposed to in their daily routines vs. individual preference. We find that individuals in large urban areas habitually travel considerable distances to food outlets (median: ~8km), consume most fast food away from home during lunch on workdays, and experience greater dietary and health impact from FEs outside their home neighborhood. These findings reveal a critical research gap in understanding how at-risk populations interact with FEs beyond their home neighborhoods in the context of their lifestyle routines, and the causal effects of these FE exposures on food choices — insights pivotal to designing food environment interventions that meet people when and where they are needed. Through simulation, we explore example intervention strategies to illustrate the value of accounting for dynamic food environments. In ongoing work, we have partnered with the Los Angeles County Department of Regional Planning and Department of Public Health to use these insights to directly guide the design of implementable interventions to food environments that optimally increase equitable access to healthy foods.

Presenter - Abigail L. Horn, PhD, PhD,
University of Southern California

Co-Author - Kayla De La Haye, PhD, PhD,
University of Southern California

Co-Author - Esteban Moro, PhD,
Northeastern University

Co-Author - Bernardo Garcia Bulle Bueno, MS,
MIT

Co-Author - Brooke M. Bell, PhD, PhD,
Tufts University

Despite the growing potential of large language models (LLMs) in mental health services, their application in diagnostic processes remains unexplored. This study described the development and evaluation of CapyEngine, an LLM-powered diagnostic tool designed to assist in mental disorder diagnosis.

We developed and evaluated CapyEngine through three phases. In Phase 1, we created a symptom database using the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition, Text Revision (DSM-5-TR). We then developed CapyEngine's architecture using LLMs, embedding models, and vector searches. In Phase 2, we conducted interviews and usability tests with mental health professionals (n = 7) to identify challenges in traditional diagnostic practices and potential areas for CapyEngine's application. In Phase 3, we compared CapyEngine's diagnostic accuracy against ChatGPT-4 and clinicians using 35 standardized case scenarios test questions from psychiatry and clinical psychology board exams. Questions were input into CapyEngine and the top 10 recommended diagnoses were obtained. ChatGPT-4 was prompted to provide the top ten potential diagnoses for each question. Clinicians (n = 3) received similar instruction to generate at least 10 potential diagnoses for each question. Responses were then analyzed to determine accuracy within the top 10, top 5, and top 1 diagnoses.

CapyEngine achieved 62.86% accuracy for identifying correct diagnoses within the top 10 options, and 48.57% accuracy for top diagnosis. ChatGPT-4 showed 100% accuracy within the top 10 and top 5 options, but only 31.43% for the top diagnosis. Clinicians outperformed both AI models with 82.86% accuracy within the top 10 and 57.14% for top diagnosis.

CapyEngine shows promise in augmenting the mental health diagnostic process. Future enhancements will focus on incorporating non-symptom-based diagnostic factors, developing specialized embedding models, and addressing cultural sensitivity. Further research is needed to assess the risks and benefits of integrating AI tools like CapyEngine into clinical workflows and to address barriers to adoption.

Author - Liying Wang,
Florida State University

Co-Author - Yunzhang Jiang,
Nexcuria Labs

Digital health interventions (DHIs) can offer personalized health behavior support, at scale, and on demand. Yet, the uptake of DHIs remains poor amongst vulnerable or minority groups and attrition is a problem that threatens their effectiveness. Taking a systems thinking approach, this qualitative study aimed to (1) better understand the mechanisms of action that underpin DHI use by a multiethnic and diverse group of adults, and (2) suggest evidence-based actions that can be taken to address the issue of uptake and sustained use of DHIs by everyone in society.

A qualitative study was conducted whereby lay facilitators approached adult patrons at seven public eateries (Hawker Centers) in residential areas of Singapore to engage in a short informal interview. A semi-structured topic guide was used to explore the daily motivations and barriers towards uptake and sustained use of DHIs. Data were analyzed using inductive thematic analysis followed by deductive mapping to behavior change theory, using the Theory and Techniques Tool, to identify the mechanisms of action (MoA) and effective behavior change techniques (BCTs) that influence DHI uptake and use.

Lay facilitators interacted with 118 individuals who discussed both positive and negative perceptions and experiences regarding the use of DHIs. Five themes that explain DHI usage were identified: (1) awareness, (2) appraisal of value, (3) accessibility, (4) trust, and (5) user experience. Themes were mapped to 15 MoAs and 29 corresponding BCTs that could inform strategies to improve uptake and use of DHIs. Community-based promotion of credible and accessible DHIs alongside behavioral cuing and digital literacy training could overcome perceived barriers towards DHI uptake. Brief counselling integrated within primary care and routine screening programs could assist individuals in their appraisal of DHIs and motivate their uptake and use. Variable rewards that link to individuals’ core values could motivate longer term DHI use. Finally, DHI designers should not underestimate the importance of simple and gamified user experiences, as well as features that support feedback processes and behavioral cueing, to ensure sustained use.

The design and wide-scale implementation of accessible, motivating, trustworthy, and user-friendly DHIs, that are promoted widely in community settings, will address barriers to uptake and sustained use by diverse and vulnerable communities and narrow the digital divide.

Co-Author - Jumana Hashim,
National University of Singapore

Co-Author - Sarah Edney,
National University of Singapore

Co-Author - Falk Mueller-Riemenschneider,
National University of Singapore

Co-Author - Tai E Shyong,
National University of Singapore

Co-Author - Jillian Ryan,
Painted Dog Research

Co-Author - Tobias Kowatsch,
University of St. Gallen

Background. Despite the efficacy of adjuvant endocrine therapy (AET) for improving survival for women with hormone receptor-positive early-stage breast cancer, adherence rates remain low, particularly among younger women who face more barriers. Our THRIVE study (NCT03592771) tested the effectiveness of a remote monitoring intervention on AET adherence and found no significant effects on 1-year AET adherence. We conducted a post hoc trial analysis to examine if age moderated the intervention's effectiveness.

Methods: This non-blinded randomized controlled trial included women with early-stage breast cancer prescribed AET at a large cancer center with 14 clinics. Participants used a pillbox to electronically monitor AET adherence for 1 year and completed surveys at enrollment. Consented participants were randomized into (1) “App”, receiving access to the study adherence and symptom monitoring app for 6 months, with increasing/severe symptoms and missed doses reported in the app triggering follow-ups from the oncology team; (2) “App+Feedback”, receiving additional weekly text messages about managing symptoms, adherence, and communication for 6 months; or (3) “Enhanced Usual Care (EUC).” The primary outcome was 1-year AET adherence captured with the pillbox (≥80% Proportion of Days Covered vs. <80%). We used multiple imputations with chained equations for missing outcomes due to loss of follow-up or missing responses. A linear probability model was then used to assess the interaction between the study arm and age on AET adherence. Marginal effects then calculated to estimate adherence by study arm and age.

Results. Among 304 women randomized (104 EUC, 98 App, and 102 App+Feedback), the 12-month retention rate was 87.5% (n=266); the median age was 60 (range 31 to 83). In adjusted analyses, the App arm was associated with a higher likelihood of adherence, which decreased with increasing age (P<0.01). For example, at age 35, 5.4% of EUC participants were adherent, vs. 48.2% of App, a 42.9 percentage point (ppt) difference (P=0.01). While at age 55, 48.3% of EUC vs. 52.0% of App participants were AET adherent; a 3.7 ppt difference (P=0.62). App+Feedback arm had a similar trend, but the interaction did not reach statistical significance (P=0.09).

Discussion. A remote symptom monitoring app significantly improved AET adherence among younger participants but not older participants. Targeting younger women, who face more adherence barriers and are more likely to benefit from remote symptom monitoring, could enhance adherence, mitigate disparities, and improve longer-term survival outcomes.

Co-Author - Rebecca A. Krukowski, PhD, FSBM, PhD, FSBM,
University of Virginia

Co-Author - Xin Hu, PhD,
Emory University

Co-Author - Edward J. Stepanski, PhD,
Stepanski Research Consulting

Co-Author - Gregory Vidal, MD, PhD,
West Cancer Center Research Institute

Co-Author - Lee S. Schwartzberg, MD,
Renown Institute for Cancer