ECG feature detection using randomly compressed samples for stable HRV analysis over low rate links

Wireless biosensors enable continuous monitoring of physiology and can provide early signs of imminent problems allowing for quick intervention and improved outcomes. Wireless communication of the sensor data for remote storage and analysis dominates the device power budget and puts severe constraints on lifetime and size of these sensors. Traditionally, to minimize the wireless communication bandwidth, data compression at the sensor node and signal reconstruction at the remote terminal is utilized. Here we consider an alternative strategy of feature detection with compressed samples without the intermediate step of signal reconstruction. Specifically, we present a compressed matched subspace detection algorithm to detect fiducial points of ECG waveform from streaming random projections of the data. We provide a theoretical analysis to compare the performance of the compressed matched detector performance to that of a matched detector operating with uncompressed data. We present extensive experimental results with ECG data collected in the field illustrating that the proposed system can provide high quality heart rate variability indices and achieve an order of magnitude better RMSE in beat-to-beat heart rate estimation than the traditional filter/downsample solutions at low data rates.