Scholarlink

<p>Osteoporotic fractures impose a heavy clinical and economic burden, yet conventional risk tools such as FRAX rely on narrow feature sets and cohort-specific calibration, while richer cross-hospital data remain locked behind privacy and regulatory walls. We propose ADP-FedRE, a four-layer federated framework that aligns heterogeneous electronic records, DXA imaging, biochemistry panels, and longitudinal visit sequences across institutions without transferring raw data. A hierarchical preprocessing pipeline combines privacy-preserving entity resolution with attention-weighted multi-modal fusion; an adaptive differential privacy scheduler routes record-level Gaussian noise according to each client’s gradient heterogeneity under an explicit lower-bound floor that stabilises cold-start rounds, with per-client Rényi accounting underlying the global guarantee. A Paillier-based secure aggregation protocol with threshold decryption shields individual updates from the orchestrator, and the trust assumptions of that protocol are treated as first-class design objects. Evaluation rests on a simulated eight-site federation built from OsteoLaus, NHANES, and SOF, complemented by a leave-one-cohort-out audit and an external MrOS cohort that participated in neither training nor federation simulation. Under the simulated federation, ADP-FedRE reached an AUC of 0.894 (95% CI 0.885–0.903) on the external test fold—within roughly one point of a centralised oracle and 2.8 points above FedProx—while keeping the empirical membership-inference advantage below 0.04 at the chosen privacy ceiling. Calibration matched the oracle in the clinically actionable range and outperformed FRAX on Brier score (0.118 versus 0.156). Robustness panels showed graceful degradation under 50% client dropout and Byzantine poisoning, with AUC losses below 1.8 points at adversary ratios up to one in four; an extended attack panel covering shadow-model and loss-based membership inference, the Carlini likelihood-ratio attack, Geiping gradient inversion, Ganju property inference, and a colluding-client probe confirms the same envelope. Subgroup audits by sex, age band, and self-reported ethnicity revealed performance and privacy gaps within acceptable bounds. Within these bounds, the framework offers a deployment-oriented template for cross-hospital fracture screening that reconciles regulatory constraints with clinical utility; the layered design transfers naturally to other chronic-disease risk problems sharing tabular-plus-imaging substrates, though real multi-centre deployment remains an open task.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A federated privacy-preserving framework for osteoporotic fracture risk assessment over multi-source heterogeneous medical data

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月31日

查看全文
  • Fan Cheng

摘要

Osteoporotic fractures impose a heavy clinical and economic burden, yet conventional risk tools such as FRAX rely on narrow feature sets and cohort-specific calibration, while richer cross-hospital data remain locked behind privacy and regulatory walls. We propose ADP-FedRE, a four-layer federated framework that aligns heterogeneous electronic records, DXA imaging, biochemistry panels, and longitudinal visit sequences across institutions without transferring raw data. A hierarchical preprocessing pipeline combines privacy-preserving entity resolution with attention-weighted multi-modal fusion; an adaptive differential privacy scheduler routes record-level Gaussian noise according to each client’s gradient heterogeneity under an explicit lower-bound floor that stabilises cold-start rounds, with per-client Rényi accounting underlying the global guarantee. A Paillier-based secure aggregation protocol with threshold decryption shields individual updates from the orchestrator, and the trust assumptions of that protocol are treated as first-class design objects. Evaluation rests on a simulated eight-site federation built from OsteoLaus, NHANES, and SOF, complemented by a leave-one-cohort-out audit and an external MrOS cohort that participated in neither training nor federation simulation. Under the simulated federation, ADP-FedRE reached an AUC of 0.894 (95% CI 0.885–0.903) on the external test fold—within roughly one point of a centralised oracle and 2.8 points above FedProx—while keeping the empirical membership-inference advantage below 0.04 at the chosen privacy ceiling. Calibration matched the oracle in the clinically actionable range and outperformed FRAX on Brier score (0.118 versus 0.156). Robustness panels showed graceful degradation under 50% client dropout and Byzantine poisoning, with AUC losses below 1.8 points at adversary ratios up to one in four; an extended attack panel covering shadow-model and loss-based membership inference, the Carlini likelihood-ratio attack, Geiping gradient inversion, Ganju property inference, and a colluding-client probe confirms the same envelope. Subgroup audits by sex, age band, and self-reported ethnicity revealed performance and privacy gaps within acceptable bounds. Within these bounds, the framework offers a deployment-oriented template for cross-hospital fracture screening that reconciles regulatory constraints with clinical utility; the layered design transfers naturally to other chronic-disease risk problems sharing tabular-plus-imaging substrates, though real multi-centre deployment remains an open task.