USING MACHING LEARNING TO POPULATE A MARKOV MODEL BY MINING BIG DATA DIRECTLY FROM A HOSPITAL EHR

Purpose: Obtaining transition probabilities for Markov modeling in economic evaluation is generally a cumbersome process of systematic review and meta-analysis, obtained from clinical trials, or from observational research. Real-world big data accessible through many hospital electronic health record (EHR) systems offer an opportunity to collect frequently updated information to populate economic models. Outcomes researchers could collect transition probabilities from large cohorts efficiently to populate models with generalizable data. Using a supervised machine learning approach, the objectives of this study were: (a) to mine a hospital EHR for transition probabilities of high-risk patients for developing hospital-acquired pressure ulcers (HAPUs); and (b) to compare efficiency and accuracy of predictive methods between Markov modeling and Bayesian inference with EHR data.

Method: This study used a de-identified panel of patient hospitalizations since 2010 in a U.S. tertiary academic medical center EHR to study Braden scores of patient risk for developing HAPUs. The study focused on patients hospitalized for ³5 days and at least two Braden scores. Braden scores were converted from an ordered scale into five categories (i.e. minimal risk; at risk; moderate risk; high risk; very high risk). A 10-stage Markov model was constructed via supervised machine learning using R software designating the five Braden categories as transition states, as well as end-states for discharge or HAPU incidence. Findings of the Markov approach were age-adjusted and compared to prior probabilities of HAPU risk derived from na•ve and full Bayesian inference. Measures of computational accuracy and efficiency were derived to compare analytical approaches.

Result: The EHR provided a panel of over 34,787 patients. The Markov model yielded transition probabilities for each of 7 transitions. Patient risk for developing a HAPU is highly predictable after approximately 4-6 iterations. The very high-risk cohort had a clinically meaningful increase in risk for HAPU development of 2.35% compared to a minimal risk transition probability of 0.05% (p<0.001). Neither of the Bayesian classifiers provided accurate comparisons.

Conclusion: Real-world big data from an EHR enables outcomes researchers to mine transition probabilities using supervised machine learning. These results can be obtained to efficiently populate Markov models for cost-effectiveness and decision analysis.