To develop a context-aware multi-instance learning (TransMIL) model based on whole-slide pathological images and integrate it with clinical parameters, thereby constructing a multimodal model for prognostic prediction in postoperative cervical cancer patients. Data (including pathological images and clinical information) were collected from patients diagnosed with cervical cancer via pathological examination who underwent surgery at the First Medical Centre of the Chinese People’s Liberation Army General Hospital between May 2017 and May 2023. After integration and screening, 374 patients were enrolled. To maximize the use of all available samples and obtain a robust performance estimate, we employed a rigorous patient-level nested five-fold cross-validation strategy for model development and evaluation. The CLAM algorithm was employed for attention-based, interpretable foreground segmentation. By integrating TransMIL with global context encoding techniques, spatial dependencies of tissue structures were captured to extract deep image features. And fused them with clinical features to predict the prognosis of patients. The predictive performance of each model was evaluated using the C-index. Overall survival (OS) and disease-free survival (DFS) were assessed via the Kaplan–Meier method. Model performance was further evaluated using calibration curves and decision curve analysis (DCA) curves. The predictive model integrating clinical-pathological characteristics demonstrated favourable prognostic capability. Models based on pathological images alone and combined clinical-Pathomics Models were used to predict 3-year and 5-year disease-free survival (DFS) and overall survival (OS), respectively. For OS prediction, the clinical-Pathomics Model achieved a C-index of 0.700, outperforming the pathomic model (C-index: 0.631) and the clinical model(C-index: 0.690). Similarly, for DFS prediction, the clinical-Pathomics Model (C-index: 0.626) surpassed the pathomic model (C-index: 0.549). Based on model nomogram scores, patients were effectively stratified into high-risk and low-risk cohorts, with Kaplan–Meier curves revealing statistically significant differences in survival rates between groups. Further evaluation via time-dependent ROC analysis, calibration curves, and decision curve analysis confirmed the model’s robust predictive accuracy and clinical applicability. A clinical-Pathomics Model was developed to predict 3-year and 5-year DFS and OS in postoperative cervical cancer patients. This model represents a promising approach for for patient risk stratification.
Liu et al. (Mon,) studied this question.