Predictive AI for Proactive Risk Detection in Future Digital Home Care Systems

Abstract

Falls and acute health deterioration in elderly care populations are often preceded by gradual, measurable changes in daily activity and sleep behaviour that unfold over days to weeks before formal system alerts are generated. Commercial wearable monitoring systems detect deterioration reactively, triggering alerts only after predefined thresholds are crossed. This thesis investigates whether machine learning applied to wrist-worn sensor data can provide earlier, more interpretable risk detection, addressing two questions: whether a model can detect if a user is currently in a behavioural deterioration state and how early (RQ1), and whether a model can predict a critical adverse event within a defined horizon (RQ2). Using pseudonymised real-world data from 20 elderly care users monitored for approximately seven months via the Vivago® wrist-worn activity device, a two-component framework was developed and evaluated under leave-one-user-out cross-validation in collaboration with Oiva Health. RQ1 is addressed through a state classification model (Task G) using change-point-derived labels; RQ2 through an event prediction model (Task C) using 14-day forward-window labels targeting red well-being flags. Task G achieves PR-AUC = 0.829, identifying deteriorating users a mean of 56.0 days before yellow well-being flags, with statistically significant separation from stable users (Mann-Whitney p = 0.029). Task C achieves PR-AUC = 0.975 (±0.020), recall = 0.885, and precision = 0.775, providing exactly 14 days of advance warning before confirmed red flags for all five deteriorating users with zero false alerts for stable users. SHAP analysis reveals complementary feature profiles; Task G is driven by 14-day mean circadian amplitude and sleep duration, while Task C by 14-day downward slopes in daytime activity. The study demonstrates that label design determines temporal detection behaviour, state-based labels unlock 56 days of lead time compared to zero days under event window labels using identical data and infrastructure, providing a transferable contribution for proactive well-being risk detection in digital home care.