Machine Learning Validation

Clinical utility will be measured with decision curve analysis against treat-all, treat-none, full logistic regression, and historical hospitalisation strategies. Models included support vector machines and an ensemble voting classifier combining SVM, random forest, and histogram-based gradient boosting. The POP-score discriminated 30-day mortality better than EuroSCORE II alone in ROC analysis. Bootstrap comparisons found few statistically superior model differences after correction, although pan-cancer ensemble models tended to perform better for M0 macrophage metrics. Final performance was assessed on an independent hold-out test set, providing internal validation but not external validation. Temporal validation will evaluate model drift by training on 2012 to 2017 data and validating on 2018 to 2020 data. The observed AUC of 0.821 on the test set provided more than 90% power to detect a true AUC of at least 0.75 at a two-sided alpha of 0.05. Model performance will be assessed using discrimination and calibration. The for the five-predictor model, the events-per-variable ratio was approximately 40:1 based on training data and approximately 51:1 using all 254 events, substantial…