Explainable machine learning-based prediction of early and mid-term postoperative complications in adolescent tibial fractures

Abstract

Background: Adolescent tibial fractures commonly lead to postoperative complications. Conventional coagulation markers (PT/APTT/FIB) lack combinatorial risk assessment. We developed an explainable ML model integrating coagulation and clinical features to predict adverse events. Methods: A retrospective cohort of 624 surgical patients (13-18 years) was analyzed. AutoML with Improved Harmony Search Optimization (IHSO) processed features: age, fracture classification, surgery duration, blood loss, and 24 h-postoperative labs (coagulation triad/D-dimer/CRP). Primary outcome: 90-day composite adverse events (DVT/infection/early callus formation disorder/reoperation). SHAP explained predictions. Results: > 0.05). The IHSO-optimized algorithm outperformed controls in 91.67% of CEC2022 benchmark functions. AutoML model performance significantly surpassed conventional methods: training set ROC-AUC: 0.9667, test set ROC-AUC: 0.9247 (PR-AUC: 0.8350). Decision curves demonstrated clinical net benefit across 6%-99% risk thresholds. Key feature importance ranked as: age > operative duration > fibrinogen > fracture classification > APTT > CRP > BMI > D-dimer. SHAP analysis revealed: 1) Increasing age significantly attenuates the risk contribution of surgery duration; 2) FIB >4.0 g/L + elevated CRP indicated coagulation-inflammation cascade; 3) AO-C type fractures carried highest risk. Conclusion: This AutoML model, validated through explainability techniques, confirms the core predictive value of age, operative duration, and coagulation-inflammation networks for adolescent tibial fracture risk management. Though requiring prospective validation, the three-tier warning system establishes a stepped framework for individualized intervention. Future studies should advance multicenter collaborations integrating dynamic monitoring indicators to optimize clinical applicability.

Ähnliche Arbeiten