Background: Chronic obstructive pulmonary disease (COPD) is linked to elevated hip fracture risk, but validated prediction tools integrating disease-specific pathophysiology are lacking.
Objectives: To develop a multitask deep learning model predicting hip fracture risk and acute exacerbation frequency in COPD patients, and identify key predictors of skeletal vulnerability.
Methods: This retrospective cohort study analyzed 4995 COPD patients (245 incident hip fractures) from the China Health and Retirement Longitudinal Study (CHARLS). A multitask deep survival model combined Cox proportional hazards (fracture prediction) and regression (exacerbation frequency) tasks, integrating demographic, clinical, and biomarker data. Performance was evaluated via concordance index (C-index) and mean squared error (MSE).
Results: The model achieved a C-index of 0.725 for fracture prediction and MSE of 0.522 for exacerbation frequency, outperforming conventional methods. Key predictors included acute exacerbation frequency (fracture group: 2.5 ± 4.4 vs. non-fracture: 1.1 ± 2.2 events/year; adjusted HR = 1.28 per additional event, 95 % CI: 1.19-1.38) and baseline lung function (fracture group: 262.7 ± 96.7 mL vs. non-fracture: 277.4 ± 85.7 mL). Frequent hospitalizations (≥2/year) increased fracture risk by 47 %. Systemic inflammation (elevated CRP/IL-6) and age further contributed to skeletal vulnerability.
Conclusion: This study establishes the first multitask deep learning framework for COPD-related fracture risk, demonstrating superior performance through multidimensional feature synthesis. The model enables personalized prevention by highlighting exacerbation burden, lung function decline, and inflammation as critical risk factors.

