Ben S. Gerber

The past decade has seen machine learning play an increasingly important role in medicine. However, many of the state-of-the-art models being developed and deployed suffer from a lack of transparency, limiting integration into the clinical decision-making process and raising concerns about patient safety. Here we call attention to symbolic regression, a largely overlooked machine learning technique that aims to be both accurate and interpretable by learning readable mathematical models from data. In recent years, it has undergone rapid advancements with successful application to problems across a wide range of disciplines. In this perspective, we provide an approachable introduction to symbolic regression, highlight its strengths and limitations, and explore its potential applications in medicine.

Background: Sub-Saharan Africa remains the region most affected by the HIV epidemic, with Zambia experiencing a substantial burden of disease. Young people with HIV are disproportionately impacted. Emerging adulthood (ages 18-24 years) is characterized by increasing independence, risk-taking behaviors, identity exploration, and changing social supports, all of which may impact HIV self-management, including medication adherence and retention in care. Substance use may further undermine HIV outcomes. Innovative interventions targeting HIV self-management and substance use reduction among Zambian young people are urgently needed. Objective: This study aims to develop and pilot a multicomponent mobile health (mHealth) intervention to improve HIV self-management and reduce substance use among young people in Zambia. We will adapt the 4-session in-person Healthy Choices intervention for mobile delivery (mobile Healthy Choices [mHC]) to increase access and delivery. We will also develop motivational text messaging (MTM) to enhance intervention impact. Both components will be delivered using the Computerized Intervention Authoring System. Using the multiphase optimization strategy (MOST) framework, we will assess preliminary efficacy and identify the most effective components. Methods: The study includes 3 phases. Phase 1 will involve conducting focus groups with Zambian young people with HIV to explore barriers and facilitators of HIV self-management and substance use to inform intervention adaptation. Phase 2 will involve adaptation and beta testing of mHC and MTM for functionality and feasibility using a community advisory board. Phase 3 will consist of a pilot trial using the MOST framework to evaluate feasibility, acceptability, and preliminary efficacy of intervention components. A total of 100 young people with HIV will be randomized to one of four experimental conditions: (1) standard antiretroviral therapy (ART) counseling (SAC), (2) mHC+SAC, (3) MTM+SAC, and (4) mHC+MTM+SAC. Primary outcomes will include feasibility, acceptability, ART adherence, viral suppression, and substance use. Feasibility and acceptability will be assessed through paradata of use patterns and the System Usability Scale. Participants will complete assessments at baseline and at 3- and 6-month follow-up visits. Biomarkers of ART adherence and/or sexually transmitted infections will also be collected. Results: Recruitment for this study began in May 2024 for phase 1. Focus group discussions were completed with 48 young people with HIV in July 2024. Phase 3, the pilot feasibility trial, began in June 2025. Upon project completion, we will have developed an innovative mHealth intervention to support HIV self-management and reduce substance use among Zambian young people with HIV. Conclusions: This study addresses a critical problem—suboptimal ART adherence, substance use, and challenges with HIV self-management—among young people with HIV in Zambia. We are developing 2 technology-based, theory-driven intervention components aimed at improving HIV self-management and substance use outcomes. If successful, this study will pave the way for scale-up of mHealth interventions to improve HIV care outcomes among young people with HIV in Zambia and other low-resource settings. Trial Registration: ClinicalTrials.gov NCT06415357; https://clinicaltrials.gov/study/NCT06415357 International Registered Report Identifier (IRRID): DERR1-10.2196/99714

Background: Community-based mobile health (mHealth) services are increasingly used to support chronic disease management in underserved rural populations facing workforce shortages, geographic isolation, and rapid aging. South Korea entered a super-aged society in December 2024, intensifying pressures in rural regions where multiple mHealth programs are embedded within primary care and public health systems. However, evidence on sustained use in real-world settings remains limited. Objective: This study aimed to explore user experiences and acceptance of community-based mHealth services in an underserved rural area of South Korea and identify facilitators and barriers to sustained engagement, using the Unified Theory of Acceptance and Use of Technology 2 (UTAUT2). Methods: A convergent mixed methods design was used, with qualitative and quantitative data collected in parallel, analyzed separately, and integrated at the interpretation stage. Overall, 24 participants with ≥6 months of experience using 1 of 4 publicly funded mHealth services in Pyeongchang County, Gangwon State, were purposively recruited. Semistructured interviews guided by the UTAUT2 were analyzed using directed content analysis, combining deductive and inductive coding. Structured questionnaires assessing usability and behavioral intention were analyzed using descriptive statistics. Findings were integrated through joint interpretation. Results: Participants had a mean age of 71.3 (SD 9.2) years, and 70.8% (17/24) were female; hypertension (18/24, 75%) and hyperlipidemia (15/24, 58.3%) were the most common. Perceived difficulty was low (mean 2.54, SD 2.06, on a 0‐10 scale), intention for continued use was high (23/24, 95.8%), and recommendation intention was unanimous (24/24, 100%). Willingness to pay was reported by 79.2% (19/24), most commonly KRW 1000‐5000 (US $1-3) per month. Qualitative findings identified performance expectancy, social influence, facilitating conditions, and habit as the most salient determinants of sustained use. Real-time monitoring enhanced health awareness, motivated dietary modification, and increased physical activity. Public health center nurses served as human-in-the-loop facilitators, providing continuous training, troubleshooting, and emotional support, while family and peers reinforced engagement. Habit formation emerged as a central mechanism, with 91.7% (22/24) integrating mHealth use into routines anchored to waking, exercise, and bedtime. Effort expectancy barriers among older participants were mitigated through nurse-led training, and hedonic motivation was driven by intrinsic satisfaction and peer interaction. Integrated analysis showed convergence for ease of use and behavioral intention, and partial divergence for willingness to pay. Conclusions: Community-based mHealth services were successfully integrated into daily life and supported chronic disease self-management among older adults in an underserved rural setting. Sustained engagement was driven by perceived health benefits, continuous human support, and habit formation rather than technology features alone, underscoring the importance of relationship-centered, human-in-the-loop implementation models. Strengthening intuitive design, hands-on onboarding, multidisciplinary primary care teams, and stable financing will be essential for equitable digital health adoption in rural and aging communities.

Background: Computed tomography pulmonary angiography (CTPA) is the standard imaging modality for diagnosing pulmonary embolism (PE), but diagnostic uncertainty is common due to technical limitations and vague language, leading to inconsistent interpretation and clinician frustration. Objective: This study develops a prompt-free, data-efficient method for assessing diagnostic certainty of PE in CTPA reports using small pretrained language models. Methods: This study examined 173 consecutive CTPA reports from UMass Memorial Health, each annotated by 3 radiologists for PE diagnostic certainty. We developed PECertainty, a lightweight, prompt-free model, and compared it with advanced large language model (LLM)–based methods under limited supervision settings. Baselines included prompt-free methods (support vector machine, random forest, and RoBERTa [Robustly Optimized Bidirectional Encoder Representations From Transformers Pretraining Approach]) and prompt-dependent methods (LLM fine-tuning, in-context learning, and ADAPET [A Densely-Supervised Approach to Pattern Exploiting Training]; UNC Chapel Hill) with open-source Gemma3-4B (Google DeepMind) and Llama3.2-3B (Meta), and the proprietary GPT-3.5 (OpenAI). Sensitivity analyses evaluated performance with 1 to 10 training examples per category for the top performer. Model performance was evaluated against radiologist annotations. External validation on 420 CTPA reports from the Baystate Medical Center, with validation limited to distinguishing certain from uncertain reports. Interpretability of the top-performing models (PECertainty and GPT-3.5) was evaluated using integrated gradients and prompt-based explanations reviewed by radiologists. Results: Among prompt-dependent methods, GPT-3.5 fine-tuning ( F ₁ -score 0.86; 95% CI 0.71‐1.0) and in-context learning ( F ₁ -score 0.87; 95% CI 0.71‐1.0) performed best, and the performance of in-context learning consistently outperformed 0-shot learning for Gemma3-4B ( F ₁ -score 0.63, 95% CI 0.56‐0.79 vs F ₁ -score 0.45; 95% CI 0.29‐0.56) and Llama3.2-3B ( F ₁ -score 0.54; 95% CI 0.41‐0.71 vs F ₁ -score 0.43, 95% CI 0.28‐0.62). PECertainty demonstrated numerically better or equivalent performance compared with both the top-performing prompt-dependent methods and all prompt-free baselines . Compared with fine-tuned ClinicalBERT (Bidirectional Encoder Representations From Transformers Pretrained on Clinical Text), PECertainty achieved statistically significant improvements across all metrics (paired bootstrap significance test, P <.05). RoBERTa (Robustly Optimized Bidirectional Encoder Representations From Transformers Pretraining Approach) fine-tuning lagged ( F ₁ -score 0.52; 95% CI 0.35‐0.71), and simple models such as support vector machine underperformed. In few-shot settings (10 examples/category), PECertainty ( F ₁ -score 0.80; 95% CI 0.59‐0.94) outperformed both GPT-3.5 fine-tuning ( F ₁ -score 0.74; 95% CI 0.58‐0.88) and in-context learning ( F ₁ -score 0.65; 95% CI 0.47‐0.83). External validation on the Baystate dataset showed good generalization for distinguishing certain from uncertain cases ( F ₁ -score 0.77; 95% CI 0.70‐0.83). Despite its strong performance, PECertainty was rated as less interpretable than fine-tuned GPT-3.5 by radiologists ( t test, P <.05). Conclusions: PECertainty enables accurate and data-efficient assessment of diagnostic certainty from free-text CTPA reports in low-resource settings. As an open-source, lightweight alternative to proprietary LLMs, it may support more precise communication between radiologists and referring physicians, with interpretability identified as a key direction for improvement.

BACKGROUND Chronic disease management requires sustained engagement, trust, and collaboration among patients, families, and primary care teams. In underserved rural areas with aging populations, mobile health (mHealth) services offer opportunities to support remote monitoring and self-management; however, real-world user experiences and determinants of sustained adoption remain underexplored. OBJECTIVE This study explored the lived experiences and acceptance of community-based mHealth services among residents of an underserved rural area in South Korea, using the Unified Theory of Acceptance and Use of Technology 2 (UTAUT2) framework. METHODS A convergent mixed-methods design was employed. Twenty-four participants using community-based mHealth services were recruited through public health centers. In-depth semi-structured interviews were analyzed using directed content analysis guided by UTAUT2 constructs, with inductive coding for emergent themes. Quantitative survey data on usability, service use, and willingness to pay were analyzed descriptively to complement qualitative findings. RESULTS Participants were predominantly older adults (mean age 71.3 years) with multimorbidity, most commonly hypertension, hyperlipidemia, and diabetes. Performance expectancy, social influence, facilitating conditions, and habit emerged as the strongest drivers of acceptance and sustained use. Real-time visualization of health data motivated lifestyle changes and reinforced perceived health benefits. Nurse-involved, human-in-the-loop support-including troubleshooting, interpretation of data, and group-based activities-was critical in overcoming early usability challenges and fostering trust. Habit formation was evidenced by the integration of monitoring activities into daily routines. Perceived ease of use varied by age and prior digital experience, with older participants requiring longer adaptation periods. Nineteen participants expressed willingness to pay for continued access, whereas others cited sufficient self-management confidence or device-related issues as reasons for discontinuation. CONCLUSIONS Community-based mHealth services embedded within nurse-involved and relationship-centered care models can be successfully integrated into daily life among rural older adults. Sustained adoption depends not only on technology usability but also on continuous human support, social engagement, and habit formation. Digital health strategies for underserved rural communities should prioritize intuitive design, structured onboarding, and long-term investment in primary care teams to promote equity, sustainability, and effective chronic disease self-management.

Background: Heart failure (HF) is a major public health problem associated with frequent hospitalizations, high mortality, and substantial health care costs. Self-care is fundamental to improving health outcomes; yet, self-care is commonly poor among patients with HF. SMS text messaging interventions may provide a simple, scalable, and accessible strategy to support HF self-care, particularly among older adults who may face barriers to using more complex digital health technologies. However, the efficacy of text messaging as a standalone intervention for patients with HF remains underexplored. Objective: This protocol paper describes the rationale and design of a pilot randomized controlled trial examining the feasibility, acceptability, and preliminary efficacy of an individually Tailored Text Messaging Intervention to Improve Self-Care in Adults with HF (Text4HF). Methods: This study is a single-site, stage I, parallel-group randomized controlled trial. Participants (n=30) are community-dwelling adults aged 50 years or older with stage C HF and suboptimal self-care, defined as a score of 3 or less on at least 2 items of the Self-Care of Heart Failure Index (SCHFI v7.2). Participants are randomized (1:1) to either a 12-week tailored text messaging intervention (Text4HF) plus usual care or usual care alone. Text messages are triggered based on patient responses to validated instruments assessing evidence-based, modifiable behavioral factors of HF self-care. Feasibility (recruitment and retention) and acceptability of the intervention are assessed as key process outcomes. The main exploratory patient-reported outcome is HF self-care (SCHFI v7.2). Other patient-reported outcomes include medication adherence, adherence to a heart-healthy diet, HF knowledge, health-related quality of life, self-efficacy, and health beliefs. Results: This study was funded in June 2022, and participant recruitment began in September 2024. A total of 26 participants have been enrolled and randomized to the intervention (n=13) and control (n=13) groups. Participants have a mean age of 60 (SD 6.6) years, 46% (12/26) are female, and 73% (19/26) identify as non-Hispanic Black. Half of the participants are individuals with reduced ejection fraction. Study completion is anticipated in June 2026. Conclusions: This protocol describes an important step toward evaluating a scalable, low-cost text messaging intervention designed to improve self-care in patients with HF. Study findings will provide critical data on feasibility and acceptability to guide a future fully powered efficacy trial of Text4HF.

BACKGROUND Heart failure (HF) is a serious chronic condition and is one of the most common hospital discharge diagnosis among adults in the United States. Self-care adherence is poor among middle aged and older adults with HF. Effective interventions targeting modifiable behavioral risk factors are critically needed given the increasing prevalence and worsened outcomes for patients with HF. Text4HF is a pilot study guided by the situation specific theory of HF self-care. The objective is to evaluate the feasibility and preliminary efficacy of an individually tailored text messaging intervention promoting self-care among middle aged and older adults with HF. OBJECTIVE This protocol aims to describe a randomized controlled trial designed to evaluate the feasibility and preliminary efficacy of individually tailored text messages on heart failure self-care. METHODS A single site, prospective, two-arm randomized controlled trial will be conducted. Participants (n= 30) will be recruited and randomized (1:1) to the intervention (Text4HF) or control group (usual care). The intervention group will receive 5 text messages per week for 12 weeks. Text message content reflects behavior consistent with adequate HF self-care such as medication adherence, heart healthy diet, HF symptoms self-monitoring, and self-efficacy in addition to usual care. The primary outcome is HF self-care, assessed with the Self-care Heart Failure Index (SCHFI v7.2). Feasibility (recruitment and retention) will be assessed over 12 weeks. Secondary outcomes include medication adherence, heart healthy diet adherence, HF knowledge, health related quality of life, self-efficacy, and health beliefs. RESULTS The trial was registered in June 2024. Recruitment is currently underway and began in September 2024 to the present, with 23 participants enrolled. Data collection will continue until early 2026. Results are expected to be published by the end of 2026. CONCLUSIONS This study represents a crucial step in testing a potentially scalable and low-cost intervention that can augment usual care to improve self-care among adults (≥50 years) with HF. CLINICALTRIAL NCT06473532

Background: Comorbidities worsen cancer survival, but patterns of preexisting and new-onset comorbidities among cancer survivors are unknown. Methods: We investigated self-reported and clinically diagnosed comorbidity among cancer survivors in the All-of-Us program’s national database. Eight highly prevalent comorbidities were identified using self-reported data from the personal health history survey among cancer survivors (n = 20 534) and noncancer adults (n = 113 628) and validated among cancer survivors (n = 26 978) using data from electronic health records (EHRs). Among 5-year survivors (n = 9174) documented in EHR, we further estimated the incidence of new-onset comorbidities. Results: The most prevalent comorbidities identified in personal health history data were hypertension (40.5%), osteoarthritis (28.4%), depression (28.0%), and obesity (23.2%). EHR data identified preexisting comorbidities: hypertension (43.3%), osteoarthritis (29.4%), depression (19.4%), and obesity (19.1%). During 5-year survival, more than 50% of cancer survivors developed at least one new comorbidity, and more than 25% developed two or more. The onset of new comorbidities showed a sharp increase in the first-year postdiagnosis. Incidence rates varied by age, race, and ethnicity. Conclusion: Future research is needed to develop effective strategies to prevent new-onset comorbidities during and after cancer treatment.

Purpose: Health-related quality of life (HRQL) assessment provides insights into the lived experiences of diverse populations. This study evaluated the convergent validity and responsiveness of the EQ-5D-3L in racial/ethnic minority populations with type 2 diabetes and elevated hemoglobin A1c (HbA1c). Methods: Secondary data from a clinical trial of a diabetes adherence support intervention for African-American and Latinx patients with type 2 diabetes (NCT02990299) were analyzed. Clinical (HbA1c, systolic blood pressure [SBP], body mass index [BMI]), psychological (brief 4-item diabetes distress scale [DDS4] for diabetes-related distress, 9-item patient health questionnaire [PHQ-9] for depressive symptoms), and HRQL (EQ-5D-3L) data were collected at baseline and every 6 months for 2 years. Convergent validity was assessed by examining the strength of associations between clinical/psychological measures and EQ-5D scores. A responder analysis using generalized estimating equation models was used to evaluate the EQ-5D's sensitivity to changes in clinical/psychological factors over time. Results: Among 221 individuals analyzed, HbA1c and SBP levels did not differ across those reporting varying levels of problems in any EQ-5D dimension. Individuals with higher BMI were more likely to report problems with mobility, usual activities, pain/discomfort, and anxiety/depression. DDS4 scores were moderately correlated with EQ-5D anxiety/depression dimension and index score, while PHQ-9 scores were strongly correlated with both. EQ-5D index score was insensitive to improvements in clinical measures but sensitive to improvements in psychological measures over time. Conclusion: The EQ-5D-3L captured variations in BMI and psychological measures but not in HbA1c and SBP. This study provides HRQL estimates that can be compared with other populations or studies.

Background: The COVID-19 pandemic led to increased demand for remote management of type 2 diabetes using secure messaging, or patient-provider text-based communication. Prior research on secure messaging has described the content of messages sent for type 2 diabetes management and demonstrated its impact on clinical outcomes. However, there is a gap in knowledge about how secure messaging performs as a communication medium for specific tasks in clinical care (eg, prescription management and discussing medical questions). Additional research is needed to understand physicians’ experiences using secure messaging to communicate with patients about clinical tasks that support diabetes management. Objective: This study aims to investigate physicians’ experience using secure messaging to communicate with patients about specific clinical tasks for type 2 diabetes management. Methods: We interviewed a sample of endocrinologists and internists from 2 different medical facilities who have used secure messaging to communicate with adult patients about type 2 diabetes management. Semistructured interviews were used to solicit physicians’ experience using secure messaging for 6 specific tasks that support diabetes management: refilling prescriptions, answering nonurgent medical questions, scheduling appointments, discussing test results, making referral requests, and discussing visit follow-up. Interviews were conducted until we achieved saturation of themes for these tasks. Interview data were collected between 2021 and 2023. Qualitative data were analyzed using the framework method for thematic analysis. Results: We interviewed 6 internists and 4 endocrinologists (n=10). Physicians reported spending between 2 and 5 hours per day messaging with patients. They observed that secure messaging increased the frequency and timeliness of communication, which improved care coordination and facilitated care delivery between visits. This served as a time-efficient way to iterate specific components of treatment plans, including discussing test results, visit follow-up, scheduling, and prescription refill. Physicians were frustrated with the unstructured nature of secure messages. Patients wrote messages that were often disorganized, confusing, or did not have enough information for the provider to take action. This often made answering nonurgent medical questions difficult. In many cases, poorly structured secure messages resulted in lengthy back-and-forth communications between patients and physicians, which sometimes required a phone call or an office visit to resolve. Conclusions: Physicians reported that secure messaging supports a longitudinal model of care, where patients can iterate their treatment plan between visits. For tasks with well-defined information boundaries, such as scheduling and prescription refill, physicians reported that secure messaging improved the time efficiency of care delivery. Providers experienced challenges using secure messaging for more complex tasks and often reported not receiving sufficient clinical information. We identified a demand for workflow technologies to process incoming secure messages to improve clarity and ensure that messages have sufficient information to inform decisions on the best course of action.