Evaluating parameter optimization in locality-sensitive hashing for high-dimensional physiological waveforms

Abstract

We develop and evaluate a theoretical architecture to inform parameter choice for locality-sensitive hashing methods used towards identifying similarity in physiological waveform time-series data. The goal is to achieve increased probability of successful patient outcomes in emergency rooms by tackling the problem of efficient information retrieval within massive, high-dimensional medical datasets. To solve this problem, we explore the relationship between a number of data inputs and elements of locality-sensitive hashing schemes in order to drive optimal choice of parameters throughout the pipeline from raw data to locality-sensitive hashing output. We achieve significant increases in retrieval times while generally maintaining the prediction accuracy achieved by naive retrieval methodologies.