Incremental value of quantitative SPECT/CT integrated parameters in a multimodal machine-learning model for differentiating spinal tuberculosis from pyogenic spondylitis.

OBJECTIVE: To develop and internally validate a multimodal machine-learning model integrating clinical characteristics, laboratory markers, conventional imaging findings, and quantitative SPECT/CT parameters for differentiating spinal tuberculosis (STB) from pyogenic spondylitis (PS), and to evaluate the incremental diagnostic value of quantitative SPECT/CT parameters. METHODS: This retrospective study included 96 patients with STB and 69 with PS confirmed by microbiological, molecular, or histopathological examinations between September 2024 and May 2026. Demographic characteristics, laboratory markers, conventional imaging findings, and quantitative SPECT/CT parameters were collected. Two prespecified integrated quantitative SPECT/CT parameters, LBISUVmax and LBISUVmean, were used to construct a baseline model, a quantitative SPECT/CT-only model, and a fusion model. Elastic Net logistic regression was used for the primary analysis, and radial basis function support vector machine (RBF-SVM) was used for robustness analysis. Internal validation was performed using repeated stratified nested cross-validation. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), Brier score, and calibration measures. Incremental diagnostic value was assessed using 2, 000 patient-level paired bootstrap replicates. Two prespecified sensitivity analyses were also performed. RESULTS: All seven integrated quantitative SPECT/CT parameters were significantly higher in the PS group than in the STB group (all P < 0.01). In the Elastic Net analysis, the baseline, quantitative SPECT/CT-only, and fusion models achieved AUCs of 0.788, 0.871, and 0.899, respectively. Compared with the baseline model, the fusion model increased the AUC by 0.110 (95% CI, 0.054-0.172) and reduced the Brier score by 0.048 (95% CI, 0.024-0.070). The RBF-SVM fusion model achieved an AUC of 0.957; however, its incremental AUC over the corresponding baseline model was only 0.021 (95% CI, -0.003 to 0.050). Both sensitivity analyses yielded findings consistent in direction with the primary analysis, and inclusion of all seven quantitative parameters did not further improve model performance. CONCLUSION: LBISUVmax and LBISUVmean improved the discrimination between STB and PS when added to conventional clinical, laboratory, and imaging information and demonstrated consistent incremental value in the Elastic Net primary analysis and prespecified sensitivity analyses. Integrated quantitative SPECT/CT parameters may therefore provide useful complementary diagnostic information; however, their clinical applicability requires confirmation through multicenter external validation and prospective studies.