Clinical pathway clustering using surrogate likelihoods and replayability validation

Abstract

Modelling clinical pathways from Electronic Health Records (EHRs) can optimize resources and improve patient care, but current methods for generating pathway models using clustering have limitations including scalability and fidelity of the clusters. We propose a novel pathway modelling approach using Maximum Likelihood (ML) data clustering on Markov chain representations of clinical pathways. Our method is calibrated to produce clusters with low inter-cluster variability across the pathways. We use machine learning with Stochastic Radial Basis Functions (SRBF) kernels for surrogate optimization to handle non-convexity and propose an incremental optimization method to improve scalability. We also define a methodology based on novel replayability scores to help analysts compare the fidelity of alternative clustering results. Results show that our ML method produces clusters that have higher fidelity in terms of replayability scores than k-means based clustering and in capturing queueing contention, which is important for bottleneck identification in healthcare.