Federated Deep Survival Learning for Hip Fracture Risk Prediction from Pelvic X-rays in the Study of Osteoporotic Fractures

Niklas Christoph Koser, Jesse Lindstädt, Marten Johannes Finck, Claus-C. Glüer, Soren Pirk

Keywords: Federated Learning, Hip Fracture Risk, Osteoporosis, Deep Learning

Registration Requirement: Yes

Abstract: Hip fractures are a major cause of morbidity and mortality among older adults. Although deep learning models show promise for predicting hip fracture risk, the development of robust and generalizable models is constrained by cross-institutional data-sharing restrictions. Federated learning (FL) addresses this challenge by enabling collaborative training without exchanging raw patient data. We present the first FL framework for hip fracture risk prediction from pelvic X-rays, training a DeepSurv-based survival model across four clinical sites of the Study of Osteoporotic Fractures (SOF). We compare FedAvg, FedBN, FedProx, FedNova, and a site-specific personalization adapter in pairwise and triple-site settings. All FL strategies outperform isolated site-specific training, and the best methods match or exceed centralized pooled training, demonstrating the potential of privacy-preserving multi-center risk prediction. Code will be released upon acceptance.

Visa & Travel: No

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 73