Personalized Multimodal Wearable-AI for Early Prediction of Heart Failure Decompensation Using Physiological Drift and Clinical Context

Abstract

Heart failure (HF) is a progressive cardiovascular syndrome characterized by episodes of clinical stability. It interspersed with periods of worsening cardiac function and decompensation, often leading to emergency hospitalization, decreased quality of life, and increased mortality. Traditional monitoring relies heavily on periodic clinical evaluation, laboratory investigations, imaging and patient-reported symptoms that may miss subtle physiological changes occurring between clinical visits. Recent advances in wearable sensing allow for continuous, non-invasive monitoring of cardiovascular and physiological parameters, but current methods are often reliant on single physiological variables, population-level thresholds or restricted combinations of wearable measurements. Moreover, the differences in individual baseline physiology, sensor quality, comorbidities and clinical history pose a challenge to the reliable early-warning systems.
We propose a Personalized Multimodal Wearable-AI framework for early prediction of HF decompensation by integrating continuous wearable signals with patient-specific physiological baselines as well as clinical context. The proposed framework combines electrocardiography, photoplethysmography, heart rate variability, oxygen saturation, respiratory rate, activity, sleep, blood pressure, body-weight trends and, where available, bioimpedance-derived fluid-related information. We propose a new Physiological Drift Index (PDI) that measures deviations of current physiological measurements from an individual’s personalized baseline, rather than relying on fixed clinical thresholds alone. Then, temporal deep learning is used to model the evolution of physiological changes over multiple time windows. Multimodal feature fusion integrates clinical variables such as left ventricular ejection fraction, cardiovascular history, renal function, electrolytes, laboratory measurements, and medication-related information. We propose an explainable AI layer for identifying the physiological and clinical factors driving each predicted risk state.The resulting system could provide a basis for continuous, personalized and interpretable monitoring of patients at high risk of HF decompensation.