LASSO regression-based machine learning model for differentiating spinal tuberculosis, pyogenic spondylitis, and endplate osteochondritis: development and clinical application.

INTRODUCTION

Accurate differentiation of spinal tuberculosis, pyogenic spondylitis, and endplate osteochondritis remains a clinical challenge due to overlapping clinical and radiological manifestations. This study aimed to construct and validate an interpretable machine learning model to assist the differential diagnosis of the three commonly confused spinal diseases.

METHODS

A total of 481 patients were retrospectively recruited from Liuzhou People's Hospital between June 2020 and December 2024, including 247 cases of spinal tuberculosis, 92 cases of pyogenic spondylitis, and 142 cases of endplate osteochondritis. All enrolled participants were randomly divided into a training cohort and an internal validation cohort at a 7:3 ratio, with 338 and 143 patients respectively. Six machine learning algorithms were comprehensively evaluated, and the optimal LASSO regression model was constructed based on 12 screened core features. SHAP analysis was employed to interpret global feature importance and individual predictive contributions. Additionally, an open-access web-based prediction calculator was developed to support clinical decision-making.

RESULTS

In the validation cohort, the LASSO model achieved AUC values of 0.838 for spinal tuberculosis, 0.683 for pyogenic spondylitis, and 0.897 for endplate osteochondritis. SHAP analysis quantified the predictive contribution of each feature and enhanced model interpretability. A user-friendly online prediction tool was successfully constructed for clinical auxiliary use.

CONCLUSION

This study established an interpretable machine learning model and a supporting web calculator. The model exhibits good diagnostic efficiency for spinal tuberculosis and endplate osteochondritis and provides a practical auxiliary tool for clinical differential diagnosis of the three spinal conditions.