A Configurable Multi-Modal Clinical Time-Series Platform for Predictive Modeling in Cardiology

Background/Objective: Predictive models for cardiovascular disease rely predominantly on electronic health records (EHRs), whereas wearable devices, home monitoring, and patient-reported outcomes offer new opportunities for multimodal risk prediction; however, integrating these heterogeneous sources into a reproducible, clinically usable machine learning workflow remains challenging. This study aimed to develop an end-to-end architecture for explainable multimodal time-series prediction in patients with heart failure (HF), left ventricular assist device (LVAD) recipients, and heart transplant (HT) recipients. Methods: The modular architecture harmonizes Health Level Seven Fast Healthcare Interoperability Resources (HL7 FHIR) and non-FHIR telemonitoring data into a unified event log, which leakage-aware temporal preprocessing converts into patient-day feature matrices. Configurable adverse-event tasks (HF hospitalization, cardiac decompensation, transplant rejection, and mortality) are modeled using gradient-boosted trees or temporal neural networks, with optional probability calibration. Explanations combine SHapley Additive exPlanations (SHAP) for tabular models with temporal-perturbation attribution and integrated gradients for sequence models. The pipeline is declaratively configured, tracked with MLflow, and deployable via command-line and web interfaces. It was developed alongside the prospective multicenter RETENTION trial of a telemonitoring-based care pathway and was subsequently applied to data collected within the trial. Results: Applied to the RETENTION cohort, the framework executed the complete workflow, from heterogeneous source data to per-patient-day probability estimates for each prediction task, with feature- and time-level explanations presented as clinician-facing narratives. The cohort served to demonstrate the architecture, not to establish clinical predictive performance or prospective clinical validation. Conclusions: This study provides a reproducible, end-to-end architecture that unifies data harmonization, temporal processing, machine learning orchestration, explainability, and deployment for multimodal clinical prediction. Its modular, disease-agnostic design supports adaptation to other diseases, settings, and data sources, laying a foundation for prospective validation and clinical translation.