Multimodal Progressive Fusion Model Predicts Hip Fracture Risk in Elderly

Overview

The MMPro-HIP model demonstrated superior accuracy (90.94%) and AUC (0.9423) in predicting hip fracture risk among elderly patients, outperforming a global model. It effectively handles incomplete clinical and imaging data by progressively integrating multimodal features, with bone mineral density (BMD) contributing the largest predictive gain.

Background

Hip fractures in older adults carry high morbidity and mortality, with a 20–30% one-year postoperative death rate and significant disability among survivors. Accurate early risk prediction is essential to enable preventive interventions and reduce healthcare burdens. Existing clinical tools like FRAX and QFracture are limited by missing data and modular data absence, common in elderly patients with mobility and comorbidity challenges. Artificial intelligence and multimodal fusion approaches offer promising solutions to improve prediction robustness despite incomplete data.

Data Highlights

Key Findings

• The MMPro-HIP model achieved 90.94% accuracy and 0.9423 AUC on an independent test set, outperforming the global model.
• Key predictors included age, sex, bone mineral density (BMD), cholesterol, and the section modulus within BMD.
• Progressive multimodal fusion enabled robust prediction despite modular missing data common in elderly clinical records.
• Ablation studies showed demographic variables contributed 79.06% and BMD measures contributed −6.71% to predictive accuracy.
• Basic demographic data alone provided useful baseline risk stratification, though BMD added the largest performance gain.
• The model was developed on 1,287 retrospective records from a single center, highlighting the need for external validation.

Clinical Implications

The MMPro-HIP model offers a practical tool for hip fracture risk assessment in elderly patients, especially in settings with incomplete clinical or imaging data. Incorporating BMD measurements significantly enhances prediction accuracy, but even basic demographic information can guide initial risk stratification. This approach may facilitate timely preventive care and individualized management to reduce hip fracture incidence and associated morbidity.

Conclusion

The MMPro-HIP progressive fusion model demonstrates promising accuracy and robustness for predicting hip fracture risk in older adults with incomplete data. Its integration of multimodal features addresses key challenges in clinical prediction, though further external validation is warranted.

References

1. Wang et al. 2024 -- Multimodal Progressive Fusion Model for Predicting Hip Fracture Risk in the Elderly: The MMPro-HIP Approach