Prototypical Biological Signals

In this paper we present a mechanism to construct prototypes of biological activity by robustly aggregating multiple noise-corrupted and time-warped observations of patient data. This allows for a compact representation of clinical activity while improving the quality of information. Our prototype construction relies on morphology-based sequence-alignment to relate activity across a set of observations and hierarchical aggregation techniques to merge information across signals in an efficient bottom-up manner. The resulting prototype represents an averaged representation of the original observations along both the time and amplitude axes. When applied to synthetic ECG, the use of prototypes was able to reconstruct the original signal with an error of less than 6% from data corrupted with unit signal-to-noise ratio and time-warping corresponding to rate variations of 5 beats per minute. In a different experiment on a set of acoustic cardiac recordings, the use of prototypes was also able to highlight differences in patients with and without mitral regurgitation.