Abstract
BACKGROUND: Digital health tools integrating electronic patient-reported outcome and experience measures (ePROMs/ePREMs) enable longitudinal monitoring of health-related quality of life (HRQoL), psychological well-being, and treatment satisfaction in pre-exposure prophylaxis (PrEP) users. However, determinants of sustained engagement with digital follow-up platforms remain insufficiently characterized. OBJECTIVE: To describe the feasibility of the Naveta-Phemium digital platform for longitudinal monitoring of safety, HRQoL, and treatment satisfaction among PrEP users, and to develop and internally validate a machine learning framework to characterize and predict engagement with digital follow-up. METHODS: A prospective observational study was conducted using the Naveta digital follow-up platform. HIV-negative adults at high risk of HIV infection received tenofovir disoproxil fumarate (245 mg) plus emtricitabine (200 mg). Clinical safety, HRQoL, and satisfaction were assessed using laboratory parameters and validated ePROMs/ePREMs (Hospital Anxiety and Depression Scale, Patient-Reported Outcome Measurement Information System Profile-29, Treatment Satisfaction Questionnaire for Medication, and Person-Centered Coordinated Care Experience Questionnaire). Engagement was defined at the questionnaire level and analyzed using the ALGOPROMIA-Classification framework with repeated stratified cross-validation. Model explainability was assessed using permutation-based Shapley Additive Explanations. RESULTS: A total of 81 participants contributed repeated questionnaire-level observations (mean PrEP duration 689 d). PrEP was well tolerated, with no moderate or severe adverse events; mild transient symptoms were mainly gastrointestinal (31/45, 68.9%) and neurological (26/45, 57.8%). Renal function remained stable (creatinine: 0.86 [SD 0.13] mg/dL; estimated glomerular filtration rate: P=.498). Psychological well-being and HRQoL remained stable (Hospital Anxiety and Depression Scale<7; Patient-Reported Outcome Measurement Information System Profile-29 near population norms). Treatment satisfaction was consistently high (Treatment Satisfaction Questionnaire for Medication≈85-87), and satisfaction with the NAVETA telemedicine model remained stable (8/10). Engagement showed clear sociodemographic and behavioral gradients. Ensemble-based machine learning models achieved good discrimination in predicting engagement (area under the curve≈0.82) across ≈12,300 questionnaire-level observations. Random forest was retained for robustness and consistency. Shapley Additive Explanations analysis highlighted lifestyle-related variables as the most influential predictors, with heterogeneous individual-level effects. CONCLUSIONS: Naveta enabled feasible telemedicine-based PrEP follow-up with preserved HRQoL, high satisfaction, and stable safety. Combining longitudinal ePROMs/ePREMs with explainable machine learning allowed detailed characterization of digital engagement, supporting digitally supported PrEP care optimization and informing future comparative and cost-effectiveness studies.