Abstract:Objective: The diagnosis of multi-level central lumbar stenosis (CLS) with or without radicular symptoms is sometime challenging, especially for localizing the surgical candidacy of culprit level. The aim of the study was to develop machine learning (ML) models utilizing clinical features and CT myelograms (CTM) to predict the decompression levels for CLS patients. Methods: We retrospectively reviewed the medical records of CLS patients who underwent CTM examination between Jan 2015 and Dec 2021, and the region of interest of the lumbar dura were extracted with radiomic features by 3D Slicer. The included patient data were divided into three sub-datasets (training: validation: testing = 6: 2: 2). A total of 72 initial ML models from 6×12 combinations were developed and then we selected the best prediction model from cross-validation. Meanwhile, the area under the receiver operating characteristic curve (AUC) was regarded as the main indicator to assess all classifiers. Results: A total of 219 patients with CLS were included. The study found that recursive feature elimination (RFE) was the optimal feature selection method to identify the radiomic predictors. The 15 radiomic predictors screened by RFE included 10 texture features, 3 shape features and 2 first-order intensity features. The AUC of most ML prediction models were 0.85 and above. The top-3 ML models were logistic regression model, multilayer perceptron model and bagging model. The AUCs of the top-3 ML models were all above 0.86, and there were no significant differences in the calibration curves and clinical decision curves of the three models. Conclusion: The study found that the most radiomic predictors were texture features that were difficult to recognize with the naked eye. It also preliminarily confirmed the additive value of machine learning in CTM locating CLS culprit level.
