An interpretable multimodal biomechanical-radiological model for predicting fixation failure in osteoporotic hip...

This study aimed to develop and validate a multimodal prediction model integrating biomechanical and radiological variables to predict internal fixation failure in osteoporotic hip fractures, enabling individualized risk assessment and perioperative decision-making. Patients with osteoporotic hip fractures undergoing internal fixation between March 2019 and February 2024 were retrospectively enrolled and randomly divided into a training set (n = 249) and a validation set (n = 107) at a 7:3 ratio. The primary outcome was implant-related failure within 12 months post-surgery. In the training set, univariate analysis and multivariate logistic regression were performed to screen associated factors. Using independent predictors, 3 machine learning models (random forest, support vector machine, and K-nearest neighbors) were developed and compared. The model with the best discriminative ability, assessed by the area under the receiver operating characteristic curve (AUC) with internal validation (bootstrapping), calibration curves, and decision curve analysis, was selected to construct a nomogram. No significant differences in baseline characteristics were observed between the training and validation sets (P > .05). Multivariate logistic regression identified that bone mineral density, maximum fracture end displacement, peak stress distribution of the implant, fracture reduction alignment deviation, and implant insertion depth were significantly associated with fixation failure (P < .05). The nomogram demonstrated excellent performance in both the training (AUC = 0.887, 95% confidence interval: 0.835-0.939) and validation sets (AUC = 0.869, 95% confidence interval: 0.801-0.937). Calibration curves showed good agreement between predicted and observed risks (P > .05), and decision curve analysis indicated superior clinical net benefit across a wide threshold range. Stability testing confirmed no significant multicollinearity (variance inflation factor < 2, events per variable ≥ 5). The random forest model demonstrated strong discriminatory ability and calibration in predicting fixation failure in this single-center retrospective cohort. While the model shows promise for perioperative risk stratification, external validation in multicenter prospective studies is required before clinical implementation.