Clinical-radiographic machine learning for predicting surgical technique and approaches for lumbar spinal stenosis: Model development and multicenter validation
Jitao Yang, Dilimulati Aikeremu, Aoran Maheshati, Rui Zhang, Abudusalamu Tuoheti, Paerhati RexitiObjective
This study aims to develop a machine learning model for predicting surgical techniques and approaches for lumbar spinal stenosis (LSS) based on clinical and radiological data.
Methods
600 LSS patients from our hospital were randomly split 8:2 into training and internal validation sets, and 100 patients from another tertiary hospital during the same period were included as the external validation set. The model incorporated clinical and radiological parameters, including coronal/sagittal balance, lumbar stability, and others. Model performance was evaluated using accuracy, precision, recall, specificity, Kappa, and macro-average F1 score. Confusion matrices, ROC, PR, DCA, and calibration curves were generated to assess model performance.
Results
The model achieved accuracies of 0.93/0.91 (Kappa: 0.86/0.85) for overall prediction, 0.98 (Kappa: 0.92/0.95) for fusion versus decompression, 0.86 (Kappa: 0.71; macro-F1: 0.86) for fusion approach, and 0.97/0.95 (Kappa: 0.90/0.86; macro-F1: 0.95/0.93) for decompression approach. In the validation sets, the model achieved area under curve values ranging from 0.993 to 0.997 in the internal validation set and 0.985 to 0.991 in the external validation set. DCA showed positive net benefit across a range of threshold probabilities in both validation sets. PR curves yielded average precision values of 0.99, 0.94, and 0.97 in the internal set and 0.98, 0.91, and 0.94 in the external set, respectively. Calibration analysis for the fusion versus decompression task showed Brier scores of 0.07 in the internal and external validation sets.
Conclusions
The clinical and radiological machine learning model developed in this study demonstrates promising predictive performance for surgical decision prediction in LSS, with consistent results across internal and external validation sets. However, its clinical utility requires prospective validation, and the model should currently be regarded as an exploratory reference tool rather than a standalone clinical decision-maker.