Abstract
Early-phase severe complications remain a major cause of morbidity and mortality during induction chemotherapy for acute leukaemia. Existing risk scores capture only limited prognostic variance and are rarely well-calibrated for clinical decision support. To develop and externally validate a machine-learning model that accurately predicts severe complications after induction, and to assess its clinical utility across key patient sub-groups. We retrospectively assembled electronic-health-record data from three tertiary haematology centres (2013-2024). After exclusion of duplicates and predefined ineligible cases, 2 870 adults with newly diagnosed AML or ALL were analysed (derivation = 2 009; external validation = 861). Forty-two candidate predictors spanning demographics, comorbidity indices, baseline laboratory values, disease biology and treatment logistics were selected via multiple imputation, Winsorised z-scaling and correlation filtering. Five supervised algorithms-including Elastic-Net, Random Forest, XGBoost, LightGBM and a multilayer perceptron-were trained using nested 5-fold cross-validation. Discrimination, calibration, decision-curve net benefit and SHAP-based interpretability were evaluated according to TRIPOD-AI and PROBAST-AI recommendations. LightGBM achieved the highest mean AUROC in derivation (0.824 ± 0.008) and maintained robust performance in external validation (AUROC = 0.801, 95% CI 0.774-0.827; AUPRC = 0.628). Calibration was excellent (slope = 0.97; intercept = - 0.03; Hosmer-Lemeshow p = 0.41). Decision-curve analysis showed superior net benefit over "treat-all," "treat-none," and a four-variable logistic benchmark across risk thresholds of 5-40%, potentially enabling targeted interventions for 14 additional high-risk patients per 100 at a 20% threshold, though clinical benefit requires prospective validation. Discrimination remained ≥ 0.80 in AML, older adults and all three centres. CRP, absolute neutrophil count, cytogenetic-risk tier, age and ferritin were the top predictors, with interpretable monotonic SHAP effects. A rigorously validated LightGBM model provides well-calibrated, interpretable prediction of early severe complications after induction therapy for acute leukaemia and provides a foundation for risk-adapted supportive care strategies, though prospective studies are needed to demonstrate clinical impact. Prospective implementation studies are warranted to confirm real-world impact.