Internal validation of an 11-yr prediction model for new vertebral fractures using the vertebral bone quality score: a prospective cohort study

TL;DR

This study validated a model that predicts new vertebral fractures over 11 years using bone quality scores from MRI scans.

Contribution

The study internally validated and corrected a prediction model for vertebral fractures using regularization to avoid overfitting.

Findings

The original model had an AUC of 0.84 for predicting new vertebral fractures.

Regularization improved the model by eliminating overfitting without reducing its predictive power.

29 out of 157 participants developed new vertebral fractures during the 11-year follow-up.

Abstract

The vertebral bone quality score is used to assess bone quality using magnetic resonance imaging. This study aimed to perform internal validation of a previously developed prediction model for the occurrence of new vertebral fractures (NVFs) over 11 yrs using the vertebral bone quality score. A prospective cohort of 157 participants from the Minami-Aizu study, with lumbar magnetic resonance imaging follow-ups over 11 yrs, was analyzed from an initial 200-participant cohort, applying the exclusion criteria. The primary outcome was the presence or absence of NVFs over 11 yrs. The predictors included age, sex, existing vertebral fractures at baseline, and VBQ scores. The prediction model, constructed using multiple logistic regression analysis and the Akaike information criterion, was evaluated for its discrimination power using the receiver operating characteristic curve and area under the curve (AUC). Internal validation was performed using the bootstrapping method with model reconstruction through regularization to correct for overfitting. In the multiple logistic regression analysis, predictors with a p-value of <.10 in multiple logistic regression analysis and <.05 in other analyses were considered statistically significant. New vertebral fractures occurred in 29 of the 157 participants. The AUC of the original model was 0.84 (95% confidence interval [CI]: 0.77-0.92). Bootstrapping revealed overfitting, which led to model reconstruction with regularization. The AUC of the regularized model was 0.84 (95% CI, 0.77-0.91), with no significant overfitting. The regularized model showed discrimination power equivalent to that of the original model without overfitting. A prediction model corrected for overfitting may be effective for long-term VF prediction. Future studies should investigate the external validity and clinical impact of the model.

Graphical Abstract

Figures