Development and external validation of an explainable machine-learning model for the prediction of unfavourable clinical-radiological outcome at three months in thoracolumbar osteoporotic vertebral fractures

Scritto il 15/11/2026
da Raquel Gutiérrez-González

Arch Osteoporos. 2026 Nov 15;21(1):165. doi: 10.1007/s11657-026-01809-3.

ABSTRACT

No externally validated tools stratify three-month prognosis after thoracolumbar osteoporotic vertebral fracture. An explainable logistic-regression model with nine predictors achieved an external AUC-ROC of 0.744 (sensitivity 90.7%). The counterfactual analysis of vertebral augmentation, the only modifiable predictor, is hypothesis-generating and should not be interpreted as a treatment recommendation.

PURPOSE: Thoracolumbar osteoporotic vertebral fractures (OVF) carry substantial morbidity and mortality, yet no externally validated tool stratifies three-month prognosis. We developed and externally validated an explainable machine-learning model for the three-month unfavourable clinical-radiological outcome.

METHODS: Retrospective observational study reported according to TRIPOD. Unfavourable outcome was defined as persistent pain, progression of vertebral height loss and/or worsening of independence in daily activities. A derivation cohort of 482 patients aged ≥50 years with acute thoracolumbar OVF treated at a tertiary centre (2017-2021) was used to train and internally validate seven algorithms with hyperparameter optimisation and stratified fivefold cross-validation. An independent two-centre cohort (n = 74) was used for external validation without retraining. Interpretability was based on SHAP values.

RESULTS: The unfavourable outcome occurred in 53.9% of the derivation cohort and 58.1% of the external cohort. A nine-variable logistic-regression model with Platt calibration was selected (cross-validated AUC-ROC 0.675, 95% CI 0.622-0.728; external AUC-ROC 0.744, 95% CI 0.628-0.844). At the pre-specified internal Youden threshold (0.556), external sensitivity was 60.5% and specificity 71.0%; at a post-hoc external threshold (0.432), 90.7% and 54.8% (Brier 0.211, expected calibration error 0.096). Vertebral augmentation was the strongest and only modifiable predictor (mean |SHAP | = 0.278); counterfactual analysis showed an absolute risk reduction of 17.1-26.6 percentage points (median 25.3).

CONCLUSION: The model predicts the three-month outcome of thoracolumbar OVF with an external AUC-ROC of 0.744 and preserved calibration. The counterfactual analysis of vertebral augmentation is hypothesis-generating and not a treatment recommendation; larger external validation, long-term outcomes and harm assessment are required before clinical use.

PMID:43034954 | DOI:10.1007/s11657-026-01809-3