• Ji Wu, Keyu Chen, Hao Mei, Yuankun Cai, Lei Shen, Jingyi Yang, Dongyuan Xu, Songshan Chai, and Nanxiang Xiong.
• Department of Neurosurgery, Wuhan University Zhongnan Hospital, Wuhan, People's Republic of China.
• Pain Physician. 2024 Dec 1; 27 (10): E1105E1116E1105-E1116.

BackgroundTrigeminal neuralgia (TN) is defined as spontaneous pain in the region of the trigeminal nerve that seriously affects a patient's quality of life. Percutaneous balloon compression of the trigeminal ganglion is a simple and reproducible surgical procedure that reduces the incidence of TN, but the postoperative outcome is poor in some patients, with it being ineffective or TN recurring.ObjectivesTo establish a machine learning-based clinical imaging nomogram to predict the recurrence of trigeminal neuralgia in patients treated with percutaneous balloon compression.Study DesignRetrospective study.MethodsThe clinical data of 209 patients with TN treated with percutaneous balloon compression at Zhongnan Hospital of Wuhan University from January 2017 through August 2023 were retrospectively collected and randomized into training and validation cohorts. All imaging histologic morphological features were extracted from the intraoperative x-ray balloon region using 3D slicer software. The relationship among clinical factors, least absolute shrinkage and selection operator, and 4 machine learning predictions of the patient's TN prognosis were analyzed using a one-way analysis of clinical factors. A prediction model was constructed using receiver operating characteristics curve analysis. The performance of the clinical imaging histogram of patients' TN prognoses was evaluated using a receiver operating characteristics curve and decision curve analysis. The model was finally validated using a validation cohort and a receiver operating characteristics curve.ResultsThe training group included 149 patients; 16 morphology-related imaging histological features were extracted for analysis. After one-way logistic regression analysis, least absolute shrinkage and selection operator analysis incorporated original_shape_Elongation, original_shape_MajorAxisLength, original_shape_flatness morphology-related imaging histologic features, gender, and affected side to give a total of 6 predictors. The final results were obtained for gender, affected side, and MajorAxisLength. Finally, 4 machine learning receiver operating characteristics curves for random forest tree, support vector machine, generalized linear model, and extreme gradient boosting models were obtained for the clinical and imaging features of gender, affected side, drug, original_shape_MajorAxisLength, and original_shape_flatness. The areas under the receiver operating characteristics curves were 0.990, 0.993, 0.990, and 0.986, respectively. Finally, predictive column maps of affected side, gender, original_shape_flatness, and MajorAxisLength were constructed using the support vector machine method, and the area under the receiver operating characteristics curve of the model was 0.99, which suggests that the model had good predictive ability. Decision curve analysis  and calibration curves showed high applicability of column-line diagrams in clinical practice. Our validation cohort consisting of 60 patients had an area under the receiver operating characteristics curve of 0.857.LimitationsThis study was performed in a single center. The nature of this study was retrospective rather than prospective and randomized, and it was not possible to entirely control for nuisance variables.ConclusionScreening clinical information by machine learning, combined with a clinical imaging histology nomogram, has good potential for predicting the prognosis of a patient's TN treated with percutaneous balloon compression, and is suitable for clinical application in patients with TN after percutaneous balloon compression.

15 keywords...

hide…