MMPro-HIP Model Improves Hip Fracture Risk Prediction Accuracy Compared to Global Machine Learning Model

Background and objectivesHip fractures, often termed the “last fracture in life,” are associated with a 20–30% one-year postoperative mortality and about 50% of survivors endure permanent disability and loss of independence. Early, accurate risk prediction enables timely preventive measures and individualized care, reducing incidence. However, older adults frequently present with mobility limitations, multiple chronic comorbidities, and heterogeneous hospital resources, leading to missing clinical or imaging data. Consequently, developing hip-fracture risk models robust to incomplete data is essential to improve generalizability, feasibility, and clinical decision-making in resource-constrained settings.Materials and methodsWe retrospectively analyzed 1,287 elderly clinical records from Beijing Jishuitan Hospital, Capital Medical University, including 643 patients with hip fractures and 644 controls without fractures. Baseline variables showed notable between-group differences in sex, age, fracture history, hemoglobin levels, and bone mineral density (BMD). A global machine learning model was developed using complete-case data, with performance evaluated by accuracy and the area under the receiver operating characteristic curve (AUC). To address the frequent problem of modular missing data, we further proposed a progressive fusion model (MMPro-HIP). This model integrates features from multiple modalities stepwise, allowing for robust prediction despite incomplete information.ResultsThe global model reached an accuracy of 84.67% and an AUC of 0.8064. Key predictors included age, sex, BMD, and cholesterol, with the section modulus within BMD emerging as an important but previously underutilized factor. The MMPro-HIP model achieved superior performance on the independent test set, with an accuracy of 90.94% and an AUC of 0.9423. By capturing cross-modal interactions, this approach outperformed the global model. Ablation experiments confirmed the contribution of demographic variables (79.06%) and BMD measures (−6.71%) to predictive accuracy.ConclusionMMPro-HIP showed favorable predictive performance for hip fracture risk assessment in older adults with incomplete clinical data in this single-center retrospective cohort. BMD contributed the largest performance gain, while basic demographic information alone still provided useful baseline stratification. These findings suggest that progressive multimodal fusion with residual correction may be a practical strategy for structured clinical prediction under modular missingness, although external validation is still needed.