PERSONALIZED BASELINE MODELING USING MACHINE LEARNING TO DETECT ANOMALIES IN LONGITUDINAL WEARABLE SENSOR DATA

Abstract

With the promise of spotting physiological anomalies early, wearable sensors have turned out to be a better option for continuous even real-time health monitoring, supporting biology-based proactive intervention. Traditional anomaly detection methods work on population models, which neglect individual variability and have posed challenges to minor deviations. In this paper, we propose a personalized baseline modeling framework for anomaly detection in longitudinal wearable sensor data using LSTM autoencoders. The model trains on normal activity segments from an individual to learn true physiological patterns for this individual and considers deviations in terms of reconstruction error. Tested on the MHEALTH dataset, and evaluated across ten subjects, the framework achieved very high average F1 scores with a very low false-positive rate, indicating that it is able to identify abnormal changes in physiological signals without needing to label data. This unique approach is unsupervised and computationally cheap to run and adaptable to a variety of users, and can therefore be implemented in the real world for continuous health monitoring and chronic disease management.