基于多模态融合的鼻咽癌复发风险预测OA

Objective:To integrate clinical and pathological whole slide image(WSI)with magnetic resonance imaging(MRI)data to construct a multimodal machine learning model for assessing postoperative recurrence risk in nasopharyngeal carcinoma patients.Methods:Retrospective collection of clinical data,WSI,and multi-sequence MRI from 168 nasopharyngeal carcinoma patients.MRI features and tumor region features of WSI were extracted separately via radiomics and the CTransPath+CLAM framework,respectively.The performance of unimodal and multimodal prediction models was compared using the random forest method.All models were trained and evaluated via 5-fold stratified cross-validation.The area under the receiver operating characteristic curve(AUC)served as the primary performance metric,and clinical net benefit was assessed using decision curve analysis.Results:The multimodal model integrating clinical data,WSI,and MRI demonstrated the best predictive performance,with an AUC of 0.794,representing an improvement of 0.215 compared with the clinical indicators model(AUC=0.579,P=0.109)and an increase of 0.183 compared with the AJCC anatomic staging model(AUC=0.611,P=0.015);however,the combined model of clinical indicators and staging(AUC=0.660)still showed a significant deficit compared with the multimodal model(ΔAUC=0.134,P=0.015).In head-to-head comparisons,the multimodal model also outperformed the MRI model(AUC=0.769,P＞0.05)and the WSI model(AUC=0.511,P＜0.001).Decision curve analysis(DCA)indicated that the multimodal model yielded the highest net benefit across most risk threshold ranges.Model interpretation revealed that its predictive power pri-marily stems from MRI textural features reflecting tumor heterogeneity.Conclusion:A multimodal machine learning model is successfully constructed and validated.By integrating clinical data,WSI and MRI information,it demonstrates promising clinical application potential for recurrence prediction of nasopharyngeal carcinoma.