Publications

Abstract : The increasing number of wearable sensors for monitoring of various vital parameters such as blood pressure (BP), blood glucose, heart rate (HR), etc., has opened up an unprecedented opportunity for personalized real-time monitoring and prediction of critical health conditions of patients. This, however, also poses the dual challenges of identifying clinically relevant information from vast volumes of sensor time series data and of storing and communicating it to health-care providers especially in the context of rural areas of developing regions where communication bandwidth may be limited. One approach to address these challenges is data summarization, but the danger of losing clinically useful information makes it less appealing to medical practitioners. To overcome this, we develop a data summarization technique called RASPRO (Rapid Active Summarization for effective PROgnosis), which transforms raw sensor time series data into a series of low bandwidth, medically interpretable symbols, called “motifs”, which measure criticality and preserve clinical effectiveness benefits for patients. We evaluate the predictive power and bandwidth requirements of RASPRO on more than 16,000 minutes of patient monitoring data from a widely used open source challenge dataset. We find that RASPRO motifs have much higher clinical efficacy and efficiency (20 − 90% improvement in F1 score over bandwidths ranging from 0.2–0.75 bits/unit-time) in predicting an acute hypotensive episode (AHE) compared to Symbolic Aggregate approXimation (SAX) which is a state-of-the-art data reduction and symbolic representation method. Furthermore, the RASPRO motifs typically perform as well or much better than the original raw data time series, but with up to 15-fold reduction in transmission/storage bandwidth thereby suggesting that less is better.