Machine Learning-derived Optimal Classification Tree Analysis Of An International Congenital Cardiac Surgical Database For Hospital - Specific Individual Patient Risk Prediction And Congenital Heart Surgery Program Self-assessment
George E. Sarris, MD, PhD1, Daisy Zhuo, PhD2, Jordan Levine, MEng3, Jack Dunn, PhD4, Luca Mingardi, MBAn2, Zdzislaw Tobota, MD5, Bohdan Maruszewski, MD, PhD5, Jose Fragata, MD, PhD6, Dimitris Bertsimas, PhD7.
1Athens Heart Surgery Institute - Mitera Children's Hospital, Athens, Greece, 2Alexandria Health, Cambridge, MA, USA, 3Alexandria Health, Providence, RI, USA, 4Alexandria Health, Cambridge, MA, MA, USA, 5Children's Memorial Health Institute, Warsaw, Poland, 6Hospital de Santa Marta and NOVA University, Lisbon, Portugal, 7Operations Research Center and Sloan School of Management, Massachusetts Institute of Technology, Cambridge, MA, USA.
Objective(s). We have previously shown that Artificial Intelligence (AI) Machine Learning (ML) methodology of Optimal Classification Trees (OCTs) can accurately predict risk after congenital heart surgery (CHS) and assess case-adjusted hospital-specific performance for "benchmark" procedures. We herein applied OCT methodology analysing outcomes of the full spectrum of CHS procedures aiming to provide user-friendly tools for hospital self-assessment and individual patient hospital-specific risk prediction.
Methods. Leveraging an International CHS Database, we trained OCTs segmenting procedures into similar risk cohorts. Outcomes considered were hospital mortality (HM), prolonged-postoperative-mechanical-ventilatory-support-time, and length-of-hospital-stay. Different case-mixes across hospitals were adjusted and each hospital's expected rate was calculated. Results. Analyses are presented in an interactive application where the overall hospital's risk-adjusted performance and specific cohorts of over/underperformance are highlighted. Results are further analyzed into different procedure groups, including the "ten benchmark procedures", providing detailed views of adjusted performance. For the example hospital (Figure), the OCT segments patients into 11 unique cohorts of similar HM risk. The table describes cohort performance by procedure groups. Green marks cohorts of better-than-expected performance, red of lower-than-expected. For this hospital, observed mortality is lower than expected (node1), but higher in certain procedure groups and patient subpopulations. An interactive calculator facilitates individual patient risk prediction. Conclusions. This methodology permits hospital-specific, case-adjusted hospital self-assessment of CHS performance, pinpoints areas of potential improvement, and facilitates adjusted hospital-specific individual patient risk prediction. 
Back to 2022 Abstracts