Application of Explainable Machine Learning for Individualized Risk Stratification and Surgical Optimization in Geriatric Intertrochanteric Fractures

Mechanical failure remains a critical complication in the fixation of intertrochanteric fractures. While the Tip-Apex Distance (TAD) is a well-established prognostic metric, the independent biomechanical impact of the nail entry point remains controversial due to confounding surgical variables. This study aimed to evaluate the stability offered by a medialized entry point using interpretable machine learning. A retrospective cohort of 628 patients treated with cephalomedullary nailing was analyzed. A Random Forest model was developed using 12 clinical and radiographic variables to predict implant failure. SHapley Additive exPlanations (SHAP) were used to disentangle confounding factors and visualize non-linear interactions between the entry-point strategy and TAD. The medialized entry point group exhibited a significantly lower failure rate compared to the standard group (25.3% vs. 44.2%; p < 0.001). The predictive model demonstrated robust performance (AUC: 0.877) on an independent testing set. Crucially, SHAP interaction analysis revealed a "buffering" effect: while risk profiles were comparable at optimal TAD levels, the medialized trajectory significantly attenuated the probability of failure when the TAD exceeded 25 mm, potentially mitigating the increased risk associated with suboptimal screw placement. The findings identify the medialized entry point as a significant independent predictor of success, increasing the biomechanical tolerance of the fixation construct. This surgical modification offers a valuable safety margin, particularly in cases where anatomical reduction or ideal screw positioning is challenging.