Semantic Annotation of Sensor Data using a Sequential Possibilistic Clustering methodology

The rapid increase of the older population has gained significant healthcare attention in recent years. Early detection of illness might result in better outcomes and reduced healthcare cost. The use of health monitoring sensors such, motion or bed or depth ones, may be able to detect early sign of illness but they are hard to interpret. To enhance caretakers understanding of health sensor data we propose to link it to signs and symptoms extracted from related electronic health record (EHR). The proposed methodology is based on detecting changes in sensor patterns using an unsupervised method (Sequential Possibilistic Gaussian Mixture Model, SPGMM), and associate health patterns from the person's EHR extracted using the clinical language annotation, modeling, and processing (CLAMP) toolkit. However, the evaluation of these health algorithms requires substantial time and effort by clinicians to examine the related electronic health record (EHR) notes. We validated our approach on three nursing home residents monitored with a collection of motion, bed, and depth sensors over several years. The obtained results show that the agreement of CLAMP-clinician and clinician-clinician evaluation are substantial, and provide strong support for our proposed methodology.