Fast Similar Patient Retrieval from Large Scale Healthcare Data: A Deep Learning-Based Binary Hashing Approach

摘要

Patient similarity plays an important role in precision evidence-based medicine. While great efforts have been made to derive clinically meaningful similarity measures, how to accurately and efficiently retrieve similar patients from large scale healthcare data remains less explored. Similar patient retrieval has become increasingly important and challenging as the volume of healthcare data grows rapidly. To address the challenge, we propose a coarse-to-fine approach using binary hash codes and embedding vectors derived from an artificial neural network. Experimental results demonstrated that this approach can reduce the time for retrieval by up to over 50.6% without sacrificing the retrieval accuracy. The time reduction became more evident as the data size increased. The retrieval efficiency increased as the number of bits in binary hash codes increased. Descriptive analysis revealed distinct profiles between similar patients and the overall patient cohort.