Weakly supervised computational pathology for prognostic stratification in locally advanced nasopharyngeal carcinoma: a multicenter external validation study

Introduction Accurate prognostic stratification remains challenging in locally advanced nasopharyngeal carcinoma (LA-NPC), as conventional clinicopathological factors provide limited precision for individualized risk assessment. Although hematoxylin and eosin (H&E)-stained whole-slide images (WSIs) contain abundant morphological information, their prognostic potential in LA-NPC remains underexplored using computational pathology approaches. Material and methods Patients with stage III–IVA LA-NPC from three Chinese tertiary centers were allocated into a training cohort, an internal validation cohort, and an independent external test cohort. A CLAM-inspired attention-based multiple-instance learning (MIL) model was developed using pretreatment H&E-stained WSIs and externally validated. The model was compared with four alternative MIL architectures incorporating different feature extractors and pooling strategies. Prognostic performance for 3-year progression-free survival (PFS) was evaluated using time-dependent receiver operating characteristic analysis, the concordance index (C-index), Kaplan–Meier analysis, decision curve analysis, and Cox regression. Results A total of 253 patients contributing 265 WSIs were included: 149 patients/160 WSIs in the training cohort, 52 patients/53 WSIs in the internal validation cohort, and 52 patients/52 WSIs in the independent external test cohort. In the internal validation cohort, the CLAM-inspired Attention MIL model yielded an AUC of 0.883 (95% CI, 0.762–0.969) and a C-index of 0.802 (95% CI, 0.705–0.884) for predicting 3-year PFS. In the external test cohort, the model achieved an AUC of 0.721 (95% CI, 0.573–0.849) and a C-index of 0.710 (95% CI, 0.591–0.819). The model significantly stratified patients into high- and low-risk groups in both cohorts and showed numerically higher AUC and C-index estimates than the four comparative MIL architectures. The risk score was linked to an elevated risk of progression (HR, 1.30; 95% CI, 1.06-1.60; P = 0.0126). Conclusions A weakly supervised CLAM-inspired Attention MIL framework can extract prognostically relevant information from routine H&E-stained WSIs and provide externally validated risk stratification for patients with LA-NPC. This approach may complement conventional clinicopathological assessment and facilitate individualized prognostic evaluation in precision oncology.