MACHINE LEARNING–BASED EARLY DISEASE RISK PREDICTION USING ATHOME HEALTH PARAMETERS

Detecting early signs of health decline using self-recorded household clinical indicators can support faster medical response and lower the occurrence of preventable complications.Although wearable and remote monitoring platforms allow repeated outpatient vital acquisition, in-home physiological records remain inconsistent, person-specific, and susceptible to noise, reducing the effectiveness of generic warning systems.This work introduces a unified machine learning workflow that builds subject-level risk understanding from home-monitored vitals such as body mass, arterial pressure, and blood oxygen levels (SpO₂), and estimates short-horizon deterioration outcomes.The system forms directional trend descriptors and baselinedivergence signals, guided by clinical findings that shifts from personal normative ranges often carry stronger risk relevance than population-level thresholds.Empirical assessment confirmsthat the method reaches 92.86% predictive accuracy in two-state risk separation and 86.79% accuracy in three-level risk grouping.Cross-model benchmarking identifies that a 1D convolution-based network delivers the most reliable inference, surpassing sequence-heavy Transformers, which show weaker discrimination likely due to high model depth relative to signal complexity and personalization granularity.The CNN pipeline also maintains improved score calibration, steady convergence, and reduced inference cost, enabling human-understandable risk outputs instead of opaque alarms.The study affirms that personal baseline drift combined with recent temporal movement offers meaningful foresight for individualized risk scoring.Broader clinical applicability demands enhanced noise filtering, secure on-device inference, and forward-looking multi-site validation prior to large-scale RPM adoption.