Development and validation of interpretable Bayes machine learning model for risk adjustment of general outcome indicators: An example from Urology

Abstract

Abstract Background: Outcome quality indicators (QIs) are often used to measure quality of care, particularly when disease-specific QIs are underdeveloped, an example being urological interventions. Without risk-adjustment, outcome QIs cannot be used for feedback and benchmarking. Administrative data captures multiplediseases and risk factors at a population level and may be a valuable resource in developing risk-adjusted QIs for quality improvement. The aim of the current study was to employ novel variational Bayes to risk adjust outcome QIs, explain how predictors affect outcome QIs, and detect outliers by using large administrative data sets in urological disease. Methods: The study included all urological admissions in Victoria, Australia from 2009 – 2019. The model used demographic variables, procedure, and diagnosis codes to predict common outcome QIs: length of stay (LOS) and hospital acquired complication (HACs) rate. Bayesian zero-inflated binomial regression was used to predict and explain the QIs. For comparison, we tested it against two models, LASSO, and random forest on a test dataset and an external dataset. The model’s ability to detect meaningful outliers was also reported. Findings: There were 108,453 urological patients, totalling 239,067 admissions. When tested both the test and external dataset, The Bayesian model was on par with random forest and better at predicting LOS and HACs when compared to LASSO. We found that age and emergency admissions, were more likely to be attributed to longer LOS and HACs. We also found that blood and platelet transfusions were likely to result in longer LOS and HACs and demonstrated how the model can be utilised for outlier detection. Interpretation: Our model provides a useful tool that explain parameters and uncertainty estimates of patient factors that drive higher LOS and HACs, while maintaining accurate predictions of outcomes when compared to other contemporary models, facilitating risk-adjustment.