Improving Generalizability of Hip Fracture Risk Prediction via Domain Adaptation Across Multiple Cohorts

Clinical risk prediction models often fail to be generalized across cohorts because underlying data distributions differ by clinical site, region, demographics, and measurement protocols. This limitation is particularly pronounced in hip fracture risk prediction, where the performance of models trained on one cohort (the source cohort) can degrade substantially when deployed in other cohorts (target cohorts). We used a shared set of clinical and DXA-derived features across three large cohorts - the Study of Osteoporotic Fractures (SOF), the Osteoporotic Fractures in Men Study (MrOS), and the UK Biobank (UKB), to systematically evaluate the performance of three domain adaptation methods - Maximum Mean Discrepancy (MMD), Correlation Alignment (CORAL), and Domain - Adversarial Neural Networks (DANN) and their combinations. For a source cohort with males only and a source cohort with females only, domain-adaptation methods consistently showed improved performance than the no-adaptation baseline (source-only training), and the use of combinations of multiple domain adaptation methods delivered the largest and most stable gains. The method that combines MMD, CORAL, and DANN achieved the highest discrimination with the area under curve (AUC) of 0.88 for a source cohort with males only and 0.95 for a source cohort with females only), demonstrating that integrating multiple domain adaptation methods could produce feature representations that are less sensitive to dataset differences. Unlike existing methods that rely heavily on supervised tuning or assume known outcomes of samples in target cohorts, our outcome-free approaches enable the model selection under realistic deployment conditions and improve generalization of models in hip fracture risk prediction.

💡 Research Summary

This study addresses the well‑known problem that clinical risk prediction models often lose accuracy when applied to cohorts whose data distributions differ from the training (source) cohort. Focusing on hip‑fracture risk, the authors assembled three large, well‑characterized cohorts—SOF (older women), MrOS (older men), and the UK Biobank (population‑based)—and identified a set of twelve harmonized predictors (nine clinical variables and three DXA‑derived BMD measures). After strict complete‑case selection, the source cohorts comprised 3,625 women (SOF) and 4,295 men (MrOS). The target cohorts were derived from UKB, yielding 410 women (5 fracture cases) and 210 men (3 fracture cases). To mimic a realistic deployment scenario, the UKB target samples were split into a pseudo‑training set (used only for computing unsupervised domain‑alignment losses) and a held‑out evaluation set (used for final performance reporting); the fracture outcomes of the pseudo‑training set were never used for model tuning.

The predictive architecture consists of a lightweight two‑layer multilayer perceptron (MLP) feature extractor that maps the 12‑dimensional input into a 256‑dimensional embedding, followed by a single linear logit layer with a sigmoid activation to produce fracture probabilities. Class imbalance (hip fractures are rare) is mitigated through class‑weighted binary cross‑entropy, weighted mini‑batch sampling, and a modest positive‑class augmentation.

Three unsupervised domain adaptation (UDA) techniques were implemented: (1) Maximum Mean Discrepancy (MMD) with a multi‑scale Gaussian RBF kernel whose bandwidth is set by a median‑heuristic per mini‑batch; (2) Correlation Alignment (CORAL), which aligns source and target covariance matrices; and (3) Domain‑Adversarial Neural Networks (DANN), which employs a gradient‑reversal layer and a domain discriminator to encourage domain‑invariant embeddings. The authors evaluated each method individually and in all pairwise and triple combinations, yielding seven adaptation configurations.

A key methodological contribution is the outcome‑free hyper‑parameter selection strategy. Instead of relying on target‑label information (which is unavailable in real deployments), the authors select hyper‑parameters that minimize the mean distributional discrepancy between source and target embeddings, a fully unsupervised criterion. This avoids data leakage and makes the approach directly applicable to clinical settings where outcomes are unknown at deployment time.

Performance was measured primarily by the area under the receiver‑operating‑characteristic curve (AUC). Across both gender‑specific source‑to‑target transfers, all UDA methods improved over the baseline model trained only on source data. The most notable gains were observed when the three methods were combined: the MMD + CORAL + DANN model achieved an AUC of 0.88 when the source cohort consisted of men only, and an AUC of 0.95 when the source cohort consisted of women only. These results surpass the single‑method configurations by 0.05–0.12 AUC points, demonstrating a synergistic effect where each technique addresses a different aspect of distribution shift (mean, covariance, and higher‑order feature alignment).

The study also confirms that the proposed unsupervised hyper‑parameter tuning does not compromise performance; the selected models achieve the highest reported AUCs while using only unlabeled target data for alignment. The authors discuss the practical implications: the framework can be deployed in new hospitals or population studies without needing to collect outcome data beforehand, thereby reducing the time and cost associated with model validation.

In summary, the paper makes three substantive contributions: (1) a rigorous harmonization pipeline for clinical and DXA variables across heterogeneous cohorts; (2) a systematic comparison of three popular UDA techniques and their combinations, showing that integrating multiple alignment losses yields the most robust and high‑performing models; and (3) an outcome‑free hyper‑parameter selection method that enables truly unsupervised domain adaptation in a clinical context. The findings suggest that future hip‑fracture risk tools can be made more portable across geographic regions, sexes, and measurement protocols, and that extending the approach to incorporate imaging, genomics, or longitudinal data could further enhance predictive power and clinical utility.