Comorbidity Scores and Machine Learning Methods Can Improve Risk Assessment in Radical Cystectomy for Bladder Cancer.

Background: Pre-operative risk assessment in radical cystectomy (RC) is an ongoing challenge especially in elderly patients.

Objective: To evaluate the ability of comorbidity indices and their combination with clinical parameters in machine learning models to predict mortality and morbidity after RC.

Methods: In 392 patients who underwent open RC, complication and mortality rates were reported. The predictive values of the age-adjusted Charlson Comorbidity index (aCCI), the Elixhauser Index (EI), the Physical Status Classification System (ASA) and Gagne's combined comorbidity Index (GCI) were evaluated using regression analyses. Various machine learning models (Gaussian naïve bayes, logistic regression, neural net, decision tree, random forest) were additionally investigated.

Results: The aCCI, ASA and GCI showed significant results for the prediction of complications (χ² = 8.8, p < 0.01, χ² = 15.7, p < 0.01 and χ² = 4.6, p = 0.03) and mortality (χ² = 21.1, p < 0.01, χ² = 25.8, p < 0.01 and χ² = 2.4, p = 0.04) after RC while the EI showed no significant prediction. However, areas under receiver characteristic curves (AUROCs) revealed good performance only for the prediction of mortality by the aCCI and ASA (0.81 and 0.78, CGI 0.63) while the prediction of complications was poor (aCCI 0.6, ASA 0.63, CGI 0.58). The combination of ASA, age, body mass index and sex in machine learning models showed a better prediction. Gaussian naïve bayes (0.79) and logistic regression (0.76) showed the best performance using a hold-out test set.

Conclusions: The ASA and aCCI show good prediction of mortality after RC but fail predicting complications accurately. Here, the combination of comorbidity indices and clinical parameters in machine learning models seems promising.