A Fiducial Scaffold for ECG Compression in Low-Powered Devices

Large volumes of physiological data can now be routinely collected using wearable devices, though a key challenge that remains is the conversion of raw data into clinically relevant and actionable information. While power constraints prevent continuous wireless streaming of large amounts of raw data for offline processing, on-board microprocessors have become sufficiently powerful for data reduction to be performed in real time on the wearable device itself, so that only aggregate, clinically interpretable measures need to be transmitted wirelessly. Here, we use the curve-length transform to extract key beat-by-beat information from the raw ECG waveform, and to identify clinically relevant timing and amplitude information. Each beat is parameterized by 12 morphological features that serve as fiducial markers, sufficient to directly reconstruct a scaffold representation of the ECG waveform. At a nominal heart rate of 70 beats/min and a sampling rate of 250 Hz, typical for wearable monitors, this represents approximately an 18-fold compression. Using difference encoding, the compression ratio improves to 21. Our algorithm computes a running exponentially-weighted average of each identified morphological feature. When any feature deviates significantly from its running average, the algorithm retains the raw waveform for five beats preceding and following the anomaly, enabling future review of the raw data. The algorithm automatically located 93.8% of the 3,615 expert-annotated QRS onsets and offsets in the PhysioNet QT-Database to within 20 ms. Similarly, it located 83.5% of all 3,194 P-wave onset and offset annotations to within 32 ms, and 89.0% of all 3,542 T-wave offset annotations within 72 ms.