Hidden Markov Models for Sepsis Classification

Abstract Infection is not always clinically evident for early sepsis identification. Hidden Markov models (HMM) can help make inferences linking observed patient physiology to the unobserved sepsis state. 36 sepsis patient records were used to develop a HMM to model unobserved patient states, which were categorised by clinical review. A HMM was created with a two hidden state topology, an hourly transition matrix using the labelled data defined by independent (non-hierarchical) sepsis criteria, and class conditional observations defined by joint probability density profiles for cases and controls using kernel density estimates. The HMM made inferences about patient sepsis state, given the time series of observed clinical predictors. The model was updated recursively to provide a probability-based diagnosis of individual case histories. The test result was compared to the labelled patient record and diagnostic performance from the ROC curve was determined for both resubstitution (maximum performance) and repeated holdout (minimum performance) estimates. The HMM performed with 59–95% sensitivity, 61–96% specificity, 1.54–23.96 positive likelihood ratio, 0.05–0.66 negative likelihood ratio, 0.63–0.99 AUC, and 2–474 diagnostic odds ratio. This wide range of low to very high performance is conclusive, but clinically significant only towards best case performance levels, which would require a larger cohort than studied here. This HMM provides a next step in design and evaluation of bedside clinical markers for a probability-based sepsis diagnosis. Refining clinical predictor selection and clinical stage definitions with greater patient numbers would improve the model and its diagnostic performance.