Predicting Textbook Oncological Outcomes with Machine Learning: A Retrospective Review of Patients Who Underwent Resection for Malignant Tumours

Introduction: Textbook oncological outcome (TOO) is a composite measure used to assess the quality of care in surgical oncology. Specifically, in liver surgery, it assesses several domains including intraoperative events, postoperative recovery, liver-specific complications, early mortality, and resection margins. Historically, one-third of patients undergoing liver resection have had a non-ideal outcome that has been difficult to predict. The aim of this study is to predict the risk of achieving textbook oncological outcome (TOO) using machine learning (ML) models for patients undergoing liver resection for malignant tumours. Methods: A retrospective cohort study of patients undergoing liver resection for malignant tumours was undertaken between 1999 and 2023 from a prospectively collected database. ML models were developed using preoperative variables as candidate predictors of TOO. Feature selection was performed using stability selection with LASSO regularisation. Model performance was evaluated using discrimination and calibration metrics. Model interpretability was assessed using SHapley Additive exPlanations. Predicted probabilities were stratified into tertiles to form risk groups (low, moderate, and high). Results: From the training cohort (n = 583), the RF model demonstrated the highest discrimination (AUROC = 0.813, 95% confidence interval [CI]= 0.731–0.877) for predicting TOO in the test cohort (n = 146). Post-recalibration, the Brier score was 0.162, calibration intercept was 0.06 and calibration slope was 0.92. The top three features were the extent of the resection, the site of the tumour, and the size of the largest tumour. TOO rates decreased in low-, moderate-, and high-risk groups (91%, 80% and 38% respectively, p < 0.001). Conclusions: An ML-based model for predicting TOO following liver resection demonstrated good discrimination and calibration metrics after recalibration. Risk stratification showed a clear gradient in TOO achievement, supporting its potential use in preoperative risk assessment and patient counselling. The current exploratory nature of this study necessitates further external validation before clinical application.