Explainable Multimodal Machine Learning Predicts 90-Day Treatment Failure in Older Patients with Fragility Fractures...

Background: Fragility fractures of the pelvis (FFP) are increasingly encountered in older adults, yet early deterioration is difficult to anticipate because fracture instability interacts with frailty and systemic vulnerability. We developed and validated an admission-based multimodal framework to predict 90-day treatment failure (TF90) before definitive management. Methods: This multicentre retrospective prediction study included 1684 consecutive patients aged ≥65 years with FFP treated at five tertiary hospitals. TF90 was defined as persistent fracture-related pain or immobility, delayed conversion to operative stabilisation, secondary displacement, FFP-related unplanned readmission, revision or unplanned reoperation, or all-cause mortality within 90 days. Only predictors available within 24 h of admission and before the definitive treatment decision were eligible, including CT-defined fracture morphology, frailty, clinical characteristics and routine laboratory biomarkers; DXA and specialised bone metabolism measurements were evaluated separately in an extended model. Four prespecified models were developed in 985 patients, temporally validated in 520 patients and evaluated in a completely held-out Centre E internal-external validation cohort of 179 patients, with additional leave-one-centre-out internal-external cross-validation. Results: TF90 occurred in 307 patients (18.2%) and increased from 8.1% in FFP I to 37.1% in FFP IV. Higher risk was associated with advanced age, greater frailty, impaired prefracture mobility, bilateral posterior ring injury, greater displacement, systemic inflammation, hypoalbuminaemia and renal dysfunction. In temporal validation, AUROCs were 0.732 for the simple logistic model, 0.763 for the core logistic model, 0.759 for the core random forest and 0.755 for the extended random forest. Neither greater algorithmic complexity nor specialised skeletal measurements provided reproducible incremental value. A development-derived high-risk stratum had a TF90 incidence of 32.2% and contained 70.0% of all events. At the fixed threshold of 0.209, sensitivity was 71.3%, specificity 70.0% and negative predictive value 93.1%. Conclusions: Pretreatment integration of pelvic ring mechanics, frailty and routinely available systemic biomarkers enables clinically relevant enrichment of older patients at risk of TF90. The model is best positioned to support intensified surveillance and structured reassessment rather than determine operative treatment. Independent prospective external validation, recalibration and clinical impact evaluation are required before routine implementation.