Predicting patient care pathway deviations and their consequences in anaesthesia using machine learning.

Deviation from the intended perioperative care pathway, particularly unplanned admission after day-case surgery, affects patient safety and the efficient organisation of care. We aimed to develop and compare interpretable machine-learning models predicting such deviations from information available before surgery. In this retrospective study, 51,112 consecutive ambulatory anaesthesia procedures from a Belgian network were analysed, with unplanned admission as the outcome. Data were partitioned into training, validation, and test sets; class imbalance was addressed with SMOTENC on the training data alone. Seven algorithms: logistic regression, decision tree, random forest, XGBoost, LightGBM, histogram-based gradient boosting, and explainable boosting machine (EBM) were compared using discrimination, precision-recall, and calibration metrics at two operating points (fixed sensitivity and fixed specificity). Unplanned admission occurred in 9.08% of patients. Discrimination varied markedly, the highest area under the receiver operating characteristic curve achieved by the explainable boosting machine (EBM, 0.826), the lowest by logistic regression (0.595). Given the rarity of the outcome, precision-recall was most informative: the EBM attained both the highest average precision (0.472) and F1-score (0.392), exceeding the black-box ensembles while remaining intrinsically interpretable. Performance was stable between validation and test sets, arguing against overfitting, although all models over-estimated risk and need recalibration. Machine learning applied to routinely collected data can identify patients at risk of deviating from their intended day-case pathway with stable performance and the potential for clinical usefulness. The explainable boosting machine is especially attractive, combining competitive accuracy with transparency. Prospective validation and recalibration of the predicted probabilities will be required before clinical deployment.