Automated Health Event Detection in Smart Homes
MetadataShow full item record
Unsupervised anomaly detection techniques can extract from data a wealth of information about unusual events. While this information can at times offer valuable insights on critical health concerns, often anomaly detection methods yield an abundance of results that are not of interest, obscuring the anomalies of true clinical relevance. In this work we improve upon traditional anomaly detection methods by introducing indirectly-supervised anomaly detection that reduces false positives and increases detection of important anomalies for a target application. We also introduce a novel synthetic data generation technique to generate synthetic smart home data with realistic synthetic health events. To validate our technique, we employ clinician-guided indirect supervision to automatically detect fall, nocturia, and depression health events from smart home sensor data. Our indirect supervision approach employs Bayesian optimization to automatically select the appropriate parameters, time scale, feature set, and anomaly detector, for detecting anomalous events of the indicated types. To decrease the number of evaluations required to learn the appropriate parameters, we also introduce a technique to warm start Bayesian optimization that requires fewer evaluations but achieves similar results compared to the standard technique. We also analyze another type of health event, weakness, in our Bayesian warm start experiments. Finally, we present a case study about the design of a smart home-powered mobile intervention and evaluate how our indirect supervision and warm start methods could be adapted to provide anomaly based prompting for in home healthcare. Our results show that indirect supervision reduces false positives and is better able to detect anomalies of clinical relevance in comparison with both a standard unsupervised anomaly detector and a fully-supervised approach. Our results also indicate that warm-started Bayesian optimization can successfully reduce the number of evaluations needed to automatically learn parameters for unsupervised anomaly detection.