Multi-modal Stacking Machine Learning Model for Predicting Lymph Node Metastasis in Cervical Cancer.
Authors
Affiliations (4)
Affiliations (4)
- College of Clinical Medicine, Ningxia Medical University, 692 Shengli Road, Yinchuan, China (H.Y., Y.C.). Electronic address: [email protected].
- Department of Radiology, General Hospital of Ningxia Medical University, 804 Shengli Road, Yinchuan 750004, China (W.Y.); State Key Laboratory of Pathogenesis, Prevention and Treatment of High Incidence Diseases in Central Asia, Xinjiang Medical University, China (W.Y.). Electronic address: [email protected].
- Department of Pathology, General Hospital of Ningxia Medical University, 804 Shengli Road, Yinchuan, China (H.T.). Electronic address: [email protected].
- College of Clinical Medicine, Ningxia Medical University, 692 Shengli Road, Yinchuan, China (H.Y., Y.C.). Electronic address: [email protected].
Abstract
To develop and validate a multi-modal stacking machine learning model integrating intratumoral and peritumoral habitat radiomics from multiparametric Magnetic resonance imaging (MRI) with clinicopathological variables for noninvasive preoperative prediction of lymph node metastasis (LNM) in early- stage cervical cancer (ECC). This retrospective study with prospective validation analyzed 623 ECC patients divided into training (n = 311), internal validation (n = 187), and external validation (n = 125) cohorts. Preoperative T2-weighted imaging (T2WI), diffusion-weighted imaging (DWI), and contrast-enhanced T1-weighted imaging (CE-T1WI) were used to cluster tumor voxels into homogeneous habitats via K-means. Habitat radiomic features were extracted from intratumoral and a 3-mm peritumoral expansion. A stacking ensemble framework was built with five base classifiers (logistic regression, extreme gradient boosting (XGBoost), Elastic Net logistic regression, Depthwise Separable Convolutional Attention (DSCA), and support vector machine (SVM) as first-level learners and an XGBoost-based Meta-model as the second-level learner. Performance was assessed by area under the curve (AUC), calibration curves, decision curve analysis, net reclassification improvement (NRI), and integrated discrimination improvement (IDI). Among base learners, the SVM-based ITH_integrated_score model performed best, with AUCs of 0.882, 0.865, and 0.845 in the training, internal validation, and external validation cohorts. The Meta-learner achieved AUCs of 0.915, 0.895, and 0.875, and DeLong test confirmed it significantly outperformed all base learners (all P < 0.05). The Meta- model demonstrated excellent calibration and the highest clinical net benefit. NRI and IDI analyses confirmed significant incremental value over each base learner. This multimodal stacking ensemble model improves preoperative LNM prediction compared with unimodal approaches and may help guide individualized surgical and surveillance strategies.