30, 46, 36
For Chapter 30: if a model's outcome is a treatment decision, the model has learned local practice and will not transfer. Chapter 46 owns transfusion strategy and should carry this caution.
Between algorithms
Kamio T, Ikegami M, Mizuno M, Ishii S, Tajima H, Machida Y, Fukaguchi K. Transfusion 2025;65(6):1051-1060. DOI 10.1111/trf.18261
Machine learning models (LightGBM, random forest, extreme gradient boosting, support vector classification) using 47 variables, with red cell transfusion events as a surrogate marker for bleeding; SHAP analysis for feature importance
The outcome is a treatment decision, not a physiological event, so the model learns the transfusion practice of the contributing hospitals — and its top features are precisely the variables a transfusion protocol uses. Retrospective; no external validation. Abstract only.
LightGBM and random forest performed best, with ROC AUC above 0.7 for both (accuracy 70.5 percent, ROC AUC 0.703, recall 0.784; and ROC AUC 0.705). SHAP identified circulating blood volume, haemoglobin and weight as the most important predictive factors.
470 ECMO-treated intensive care patients in the Tokushukai Medical Database (2018-2022) covering 71 hospitals; 357 included for model development