586 posters, 8 topics, 23 sessions, 1,077 authors, 622 institutions

ISQua's 40th International Conference 2024

24th - 27th September 2024 | Istanbul, Türkiye

P167

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

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

View Poster Contact Presenter

Poster Presenter

Jia Wei KOH

Authors

Harvey Jia Wei Koh, Dragan Gasevic, David Rankin, Stephane Heritier, Mark Frydenberg, Stella Talic

Affiliations

Monash University

Part of Topic

Patient Safety & Quality Improvement

•Audit and feedback is a crucial part of improving healthcare quality.

•Comprehensive risk adjustment that consider patient risks is needed to facilitate fair benchmarking of outcome indicators.

•Current methods do not comprehensively account for the interactions between diagnosis and procedures.

•We aim to create an explainable Bayesian machine learning model that facilitates risk adjustment and explanation for better audit and feedback.

•We present VIRGO: Variational Inference Bayesian machine learning of General Outcome indicators.

Figure 2: Case study of a patient with longer than expected LOS. The top subplot compares the predicted length of stay (LOS) within a 99% probability range against the observed LOS rate. The bottom subplot elucidates the contribution of individual model variables. By adjusting for patient risk factors, procedures and diagnosis, VIRGO can provide feedback and indicate potential outliers for further investigation.

•Our model performed the best in predicting both LOS and HACs against state-of-the-art ML algorithms. (Figure 3)

Figure 4: Counterfactual analysis of a patient with surgical complications developed during open prostatectomy for prostate hyperplasia (Figure 4.1). We simulate expected outcomes if the patient did not have surgical complications and require blood transfusions . The model predicts that within 99% probability that expected LOS drops from 0 – 19 days to 0 – 12 days (Figure 4.2).

•By risk-adjusting patient risk factors, diagnosis and procedures, VIRGO can be used to provide clinically relevant feedback to healthcare professionals using only hospital administrative data.

•Furthermore, given the accurate predictions of LOS, hospital administrators can allocate hospital resources efficiently.