Live Research Spotlight 7: AI and Digital Health

BACKGROUND: The long-term (i.e., ≥ 12 months) effects of commercial mHealth apps on physical activity (PA) have been scarcely studied. PURPOSE: To conduct a longitudinal examination of a ‘top tier’ commercial mHealth app on population-level PA. METHODS: A 24-month pre-post quasi-experimental study was conducted between December 2016 and June 2019 with Carrot Rewards app users in Ontario, Canada. Users with valid smartphone-assessed baseline step count data were included. Simple linear regression models analyzed weekly mean daily step counts. Post-hoc estimates examined differences in weekly mean daily step count from baseline. RESULTS: The total sample included 516,818 users (% female: 62.83; age [SD]: 33.46 [12.65] years; baseline daily step count [SD]: 6,035 [3,706]). Compared to baseline, weekly mean daily step counts were 464 (95% CIs: 453, 475) and 242 (95% CIs: 221, 264) steps/day higher at 12 and 24 months, respectively. CONCLUSION: This commercial mHealth app increased population-level PA over 24 months.

Author - Lisa Nguyen,
Western University

Co-Author - Marc Mitchell, PhD, PhD,
Western University

BACKGROUND: Memory and learning deficits are among the most impactful and longest-lasting symptoms after traumatic brain injury (TBI). Despite the persistence of memory deficits and their implications for community reintegration, memory rehabilitation is restricted to short-term care within structured therapy sessions. Technology may extend memory rehabilitation into daily life and increase the duration and contextual diversity of learning opportunities. Ecological momentary assessment and intervention frameworks leverage mobile phone technology to assess and support individuals’ behaviors across contexts and have shown benefit in other chronic conditions. However, few studies have used ecological momentary intervention for adults with chronic TBI, and none have addressed memory.

OBJECTIVES: To develop and test the usability of memory ecological momentary intervention (MEMI). MEMI is designed to introduce new information, cue retrieval, and assess learning across time and contexts.

METHODS: We developed MEMI by leveraging automated text messages using a REDCap/Twilio interface linked to the Gorilla online experimental platform. We recruited 14 adults with chronic, moderate-severe TBI to participate in 3 rounds of iterative usability testing: one ThinkAloud round using MEMI with an experimenter (n=4) and two real-world rounds using MEMI in their daily lives for a week (n=5/round). We analyzed engagement and quantitative and qualitative user feedback to assess MEMI’s usability and acceptability.

RESULTS: Participants were an average of 38.1 (± 12.3, range: 22-54) years old, with an average time since injury of 6.6 (± 8.1) years. Half the sample was female. Participants were highly engaged, completing an average of 11.8 out of 12 possible sessions (2 per day). They rated MEMI as highly usable, with scores on the System Usability Scale across all rounds equivalent to an A+ on a standardized scale. In semi-structured interviews, they stated that MEMI was simple and easy to use, that daily retrieval sessions were not burdensome, and that they perceived MEMI as helpful for memory. We identified a few small issues (e.g., instruction wording) and made improvements between rounds.

CONCLUSIONS: Testing MEMI with adults with chronic TBI revealed that this technology is highly usable and favorably rated for this population. We incorporated feedback regarding users’ preferences and plan to test the efficacy of this tool in a future clinical trial.

Presenter - Emily Morrow, PhD, MS, CCC-SLP,
Vanderbilt University Medical Center

Author - Lyndsay Nelson, PhD,
Vanderbilt University Medical Center

Author - Melissa Duff, PhD, CCC-SLP,
Vanderbilt University Medical Center

Author - Lindsay Mayberry, PhD, MS,
Vanderbilt University Medical Center

Background: African Americans (AA) have less access to smoking cessation counseling. Smoking cessation apps using conversational agents (CA; e.g. chatbots) show promise in improving access, engagement, and cessation. Artificial Intelligence (AI) using large language models enables more tailored and lifelike CA. The PASCAL project aims to develop an app using a large language model to offer tailored, just-in-time support to manage smoking cravings among AA attempting to quit smoking. The current study assesses the acceptability of the planned PASCAL app.

Methods: The study recruited AA (n=22; 54.5% Female, Mean age = 44.5 years) adults who smoke cigarettes to participate in virtual focus groups. Participants were recruited through BuildClinical, Craigslist, and current studies. Participants were asked about acceptability of AI-based cessation interventions, the proposed app design, and smoking trigger management strategies. Two analysts developed a preliminary codebook using inductive and deductive methods after reviewing two transcripts. Analysts used the codebook on the remaining transcripts.

Results: Four major themes emerged: (1) limited familiarity with AI, (2) openness to using AI to assist with smoking quit attempts, (3) desire for the app to host a community of people trying to quit, and (4) acceptability of personalized trigger management strategies. Participants voiced initial apprehension of an AI-based smoking cessation app due to concerns about risks to safety and privacy, as well as experiences with non-tailored, poorly functioning, or frustrating automated systems. However, participants expressed acceptability of PASCAL upon learning about its key features (ability to predict smoking triggers and provide just-in-time support) and potential benefits. Participants provided suggestions for app features and content including smoking triggers, design suggestions, a tracker for money saved by not smoking, and a social feed.

Conclusion: Similar to other qualitative studies of AI interventions, AA who smoke had concerns about AI, but showed acceptability upon learning more. While academic literature expresses concern over AI exacerbating biases and making recommendations that lack evidence, participants mostly focused on concerns regarding the efficiency, privacy, and empathy associated with the app AI. These data will be used to inform the development and clinical trial of the PASCAL app.

Presenter - Devyn Fernholz,
Hennepin Healthcare

Co-Author - Warren McKinney,
Hennepin Healthcare Research Institute

Co-Author - Michael Kotlyar,
University of Minnesota

Co-Author - Serguei Pakhomov,
University of Minnesota

Co-Author - Sheena Dufresne,
University of Minnesota

Co-Author - Emily Welle,
Hennepin Healthcare Research Institute

Co-Author - Sandra Japuntich, Ph.D., Ph.D.,
Hennepin Healthcare/University of Minnesota Medical School

Introduction. With the expansion of telehealth, there is a need to examine patients’ perceptions of patient centeredness, a metric of healthcare quality, in digital settings. This study explored individual-level correlates of patients’ perceptions of patient-centered communication (PCC), a core component of patient-centered care, in telehealth settings and the moderating effects of county-level preparedness for public health crises.
Methods. We conducted an online survey of 5444 US adults ≥18 years old who lived in 800 counties that were least or most vulnerable to public health crises per the Minority Health Social Vulnerability Index (MHSVI). Participants rated how often healthcare providers gave them a chance to ask questions, paid attention to their feelings, involved them in decisions, made sure they understood, explained things, spent enough time with them, and helped them deal with uncertainty. Logistic regressions quantified associations between sociodemographic, health, and technology-related factors and optimal PCC (always vs. other responses) overall and stratified by MHSVI.
Results. Patients experiencing less than optimal PCC were non-Hispanic Asians, Pacific Islanders, Native Hawaiians, and Alaska Natives, those without a primary care provider, or with limited English proficiency. For example, patients with limited English proficiency were less likely to report optimal PCC for all seven PCC items (adjusted odds ratio [aORs] 0.43, 0.66, 0.47, 0.39, 0.48, 0.56, 0.46, respectively). Digital health literacy domains such as having access to digital services that work were consistently associated with optimal PCC (aORs 2.16, 1.62, 1.52, 1.49, 1.45, 1.55, 1.64). These associations largely held across most and least vulnerable counties, but race/ethnicity was not associated with any PCC item in the most vulnerable counties. In most vulnerable counties, incremental increases in education were associated with less likelihood of reporting optimal PCC (e.g., aOR 0.86 for giving them a chance to ask questions).
Conclusion. Less than optimal PCC was associated with demographics and social determinants of health traditionally linked with disparate outcomes. Digital literacy emerged a determinant of PCC. Public health preparedness indirectly affected PCC where education and race/ethnicity related disparities differed across least and vulnerable counties. Consideration of individual and county-level factors is required to improve PCC in telehealth settings.

Co-Author - Lydia Tesfaye,
National Institute on Minority Health and Health Disparities

Co-Author - Zahra Ansari,
Dell Medical School at University of Texas Austin

Background: Accurate monitoring of eating behaviors is crucial for managing conditions like obesity, diabetes, and eating disorders. Wearable devices like smartwatches offer a promising solution for continuous, objective monitoring, which increase user adherence compared to self-report. However, implementing real-time eating detection on these devices is challenging due to their limited processing power, memory, and battery life. To address these issues, we've developed a new approach for efficient, accurate real-time eating detection on wearable devices. Our solution uses a dataset of typical eating cycle patterns collected on wrist-worn devices to create models that overcome device limitations while providing reliable, continuous monitoring.
Method: We proposed an advanced machine learning approach called Neural Architecture Search (NAS) to find the best model design for detecting eating behaviors. This process explored various deep learning architectures that recognize eating gestures in inertial measurement unit (IMU) data from the wrist. The search process was guided by a strategy that balanced the tradeoff between accuracy and the device's limitations, such as memory size and processing speed. The data we used came from motion sensors (accelerometer and gyroscope) on the wrist collected from 12 participants for 246 minutes. We analyzed this data in 5-second intervals, with each interval overlapping the previous one by half.
Result: Our approach identifies the best model architecture (2-D convolutional neural network (CNN) with long short-term memory (LSTM) layers), achieving high accuracy in eating detection (84.81% F1-score) while meeting strict on-device constraints. The best model requires less than 225.38 KB of memory, performs real-time inference with latency under 36.74 ms on typical wrist-worn hardware, and maintains battery life around 25.7 hours under continuous monitoring.
Conclusion: Clinically, our proposed method allows healthcare providers to detect patients' eating patterns using wrist-worn devices, enabling accurate assessment of dietary adherence and timely treatment adjustments. Continuous detection allows identification of eating triggers over time, enabling personalized behavior modification strategies and just-in-time Interventions. For conditions like diabetes or obesity, this facilitates precise insulin dosing or immediate feedback on caloric intake, potentially reducing complications and improving health outcomes.

Author - Boyang Wei,
Northwestern University

Co-Author - Glenn Fernandes,
Northwestern University

Co-Author - Farzad Shahabi,
Northwestern University

Co-Author - Christopher Romano,
Northwestern University

Co-Author - Annie W. Lin, PhD, RD,
University of Minnesota

Co-Author - Nabil Alshurafa, PhD,
Northwestern University

Introduction
Chronic pain in youth affects 1 in 4 children and significantly impairs functioning.[1] Parental emotional and behavioral responses to their child’s pain significantly impact child functioning and it is imperative to incorporate parents into behavioral pain management interventions.[2,3] Graded Exposure Therapy (GET) is an evidence-based intervention based on the interpersonal fear avoidance model targeting pain-related fear and avoidance in youth and their caregivers. The current study aims to examine parent responses to the digital version of GET Living (iGET Living).[2]

Methods
Participants (N = 5) were caregivers of youth diagnosed with chronic pain who enrolled in a pilot investigation of iGET Living. Caregivers engaged in short (5-10 minute), self-paced, daily modules for ~10 weeks. Modules focused on education on chronic pain, parent distress in the context of chronic pain, and applied exercises on chronic pain management and functional recovery. Measures assessing parent pain-related fear and avoidance were completed with the Parent Fear of Pain Questionnaire (PFOPQ) and parent catastrophizing was reported with the Parent Catastrophizing Scale (PCS-P) at baseline and discharge.[4, 5]

Results
Three caregivers demonstrated greater than a 30% decrease in PFOPQ from pre- to post-intervention (P1 decreased 40%, , P8 decreased 52%, P11 decreased 49%). Two caregivers were nearing the 30% clinical change threshold (P4 decreased 29%, P7 decreased 24%). On the PCS-P, two caregivers demonstrated a greater than 30% change: P8 decreased 52%, P11 changed 59%. Omitted here due to space, but qualitative responses from treatment modules will be included in the poster.

Conclusions
Caregivers reported a reduction in pain-related fear and catastrophizing after completing the iGET Living intervention. The education and exercises lowered parent fears and worries about their child’s pain. iGET Living appears to be a promising approach to help parents cope with their child’s pain. Since parent and child mental health are interconnected, using a transdiagnostic and biopsychosocial approach is crucial.

Presenter - Dylan Mayanja, BA,
Charles R. Drew University College of Medicine

Co-Author - Lara A. Minassians, MS,
Stanford University School of Medicine, Department of Anesthesiology, Perioperative and Pain Medicine

Co-Author - Katrina Guardino, BS,
Duke University School of Medicine

Co-Author - Emma F. Gaydos, BA,
Stanford University School of Medicine, Department of Anesthesiology, Perioperative and Pain Medicine

Co-Author - Laura E. Simons,
Stanford University School of Medicine, Department of Anesthesiology, Perioperative and Pain Medicine

Co-Author - Lauren E. Harrison, PhD,
Stanford University School of Medicine, Department of Anesthesiology, Perioperative and Pain Medicine