Multicenter machine learning model using clinical and radiomic features for prediction of postoperative residual status in meningioma.
Authors
Affiliations (10)
Affiliations (10)
- 1Department of Neurological Surgery, University of Wisconsin School of Medicine & Public Health, Madison, Wisconsin.
- 3Department of Neurosurgery, University of Health Sciences, Istanbul Training and Research Hospital, Istanbul, Turkey.
- 2Icahn School of Medicine at Mount Sinai, Biomedical Engineering and Imaging Institute, New York, New York.
- 4Department of Radiology, Gazimagusa State Hospital, Gazimagusa, Cyprus.
- 5Department of Radiology, University of Health Sciences, Istanbul Training and Research Hospital, Istanbul.
- 6Department of Neurosurgery, Gazi University Faculty of Medicine, Ankara.
- 7Department of Neurosurgery, University of Health Sciences, Bakırkoy Prof. Dr. Mazhar Osman Training and Research Hospital for Psychiatry, Neurology, and Neurosurgery, Istanbul.
- 8Department of Neurosurgery, Istanbul Medeniyet University, Prof. Dr. Suleyman Yalcin City Hospital, Istanbul.
- 9Department of Neurosurgery, Istanbul Health and Technology University, Istanbul; and.
- 10Acıbadem University School of Medicine, Istanbul, Turkey.
Abstract
Gross-total resection is the primary surgical objective in meningioma management; however, predicting resectability preoperatively remains challenging, particularly for meningiomas located in the skull base that exhibit complex anatomical relationships. There is a lack of validated reproducible models that integrate known anatomical factors for surgery; similarly, the value of radiomics as a stand-alone predictor is still unclear. The aim of this study was to develop a machine learning model that combines clinical and radiomic features to estimate early postoperative residual meningioma. This retrospective multicenter study included 369 patients who underwent meningioma resection from 2020 to 2024 and had available preoperative contrast-enhanced T1-weighted MRI. Patient data from 3 centers (n = 307) were used to develop the model using leave-one-center-out cross-validation, and an independent cohort (n = 62) was used for external validation. Radiomic features were derived from manually segmented meningiomas, filtered for reproducibility, and integrated with 6 predefined clinical variables. The clinical-only, radiomics-only, and combined models were trained using 4 machine learning classifiers. Model performance was assessed through cross-validation, independent external validation, and receiver operating characteristic analysis. Formal incremental benefit analyses were performed on the common overlap external subset, with predictions available for all compared models. Residual meningioma occurred in 23.5% of patients in the development cohort and 14.5% in the external cohort. Lesion location and venous sinus involvement were significantly associated with residual status. Following the feature selection, 2 stable radiomic texture features were retained. In external validation, the combined radiomics-clinical k-nearest neighbors model achieved the highest area under the curve (0.821), with sensitivity of 0.889, specificity of 0.706, and accuracy of 0.733. Clinical variables provided most of the predictive value, whereas radiomic features provided only limited incremental value when added to the clinical model. Decision curve analysis revealed net benefit for the combined model within a narrow range of low thresholds. In this multicenter study, which included external validation, clinical variables provided most of the predictive value for postoperative residual meningioma. Radiomic features alone showed limited discrimination and only modest added value beyond clinical predictors. These combined models could serve as decision-support tools for preoperative risk assessment in meningioma surgery but are not yet suitable for routine stand-alone clinical use.