Analysis of locality-sensitive hashing for fast critical event prediction on physiological time series

We apply the sublinear time, scalable locality-sensitive hashing (LSH) and majority discrimination to the problem of predicting critical events based on physiological waveform time series. Compared to using the linear exhaustive k-nearest neighbor search, our proposed method vastly speeds up prediction time up to 25 times while sacrificing only 1% of accuracy when demonstrated on an arterial blood pressure dataset extracted from the MIMIC2 database. We compare two widely used variants of LSH, the bit sampling based (L1LSH) and the random projection based (E2LSH) methods to measure their direct impact on retrieval and prediction accuracy. We experimentally show that the more sophisticated E2LSH performs worse than L1LSH in terms of accuracy, correlation, and the ability to detect false negatives. We attribute this to E2LSH's simultaneous integration of all dimensions when hashing the data, which actually makes it more impotent against common noise sources such as data misalignment. We also demonstrate that the deterioration of accuracy due to approximation at the retrieval step of LSH has a diminishing impact on the prediction accuracy as the speed up gain accelerates.