Dynamic Clustering-driven Pruning Framework for Efficient Machine Learning-based Anterior Cruciate Ligament Injury Risk Biomarker Classification

Abstract

Background: Anterior cruciate ligament (ACL) injuries are common in athletes and can lead to long-term physical impairment. Machine learning models offer promising tools for predicting injury risk using high-dimensional biomarker data. However, traditional ensemble methods suffer from redundancy and high computational cost, limiting their use in real-time or resource-constrained clinical settings. Methods: We propose a novel Dynamic Clustering-Driven Pruning Framework (DCDPF) that enhances computational efficiency while maintaining high prediction accuracy. The framework incorporates: 1. A Graph Attention Network for selecting relevant biomarkers by capturing nonlinear interactions. 2. Spectral clustering to group gradient boosted decision trees based on decision boundary divergence, identifying and reducing redundancy. 3. An Artificial Bee Colony (ABC) optimiser that dynamically prunes weak learners while maximising validation performance. 4. A Transformer-based calibration module to refine probabilistic outputs and ensure robustness against dataset shifts in longitudinal studies. Results: DCDPF achieved an AUC of 0.923 on ACL injury risk classification tasks, outperforming baseline models while using only ~39% of the ensemble size. The system reduced training time by over 60% and memory usage by nearly 70%. In clinical thresholds (90% sensitivity), it improved positive predictive value and maintained stability across datasets and seasons. Conclusion: DCDPF offers a scalable, efficient and accurate solution for ACL injury risk prediction. Its modular design allows seamless integration with existing biomarker pipelines, supporting real-time clinical decision-making in sports medicine.

Ähnliche Arbeiten