An interpretable multimodal morphomolecular model refines progression-risk stratification in early-stage endometrioid endometrial carcinoma
Related Experiment Videos
Jinxin Li1, Qiong Tang2, Jinyang Chen3
1Ningbo Institute of Innovation for Combined Medicine and Engineering, The Affiliated Lihuili Hospital of Ningbo University, Ningbo, China.
Background:
Risk stratification remains incomplete for patients with FIGO stage I-II endometrioid endometrial carcinoma (EEC) despite clinicopathological assessment and molecular subtyping. Interpretable integration of diagnostic tissue morphology with molecular programmes may provide complementary information for progression-risk assessment.
Methods:
We developed a prototype-resolved interpretable survival model (PRISM), a multimodal framework integrating haematoxylin and eosin-stained whole-slide images with RNA-seq-derived pathway programmes for progression-free survival risk assessment. PRISM was developed in 287 eligible TCGA-UCEC patients with EEC, comprising a training cohort (n = 229) and a held-out internal validation cohort (n = 58). After model specification was fixed, PRISM was retrained on all 287 cases and evaluated without further tuning in an independent CPTAC-UCEC cohort of 63 patients with EEC. Prototype-pathway alignment provided tissue- and pathway-level interpretation of model predictions.
Results:
Among 350 patients with EEC and matched multimodal data, PRISM achieved C-indices of 0.770, 0.751 and 0.745 in the TCGA training, internal validation and CPTAC external validation cohorts, respectively. In CPTAC-UCEC, the 36-month time-dependent AUC was 0.727, compared with 0.685 for the clinicopathological and molecular subtype baseline model. In a complementary fixed-horizon analysis, adding PRISM to the clinicopathological model increased the 36-month AUC in CPTAC-UCEC from 0.609 to 0.719. The PRISM risk score remained associated with progression-free survival after multivariable adjustment. Prototype-resolved analyses linked higher-risk predictions to localised histomorphological patterns and pathway programmes related to proliferation, cell-cycle activity, glycolysis, hypoxia, focal adhesion and epithelial-mesenchymal transition.
Conclusion:
PRISM showed prognostic discrimination in an exploratory independent external evaluation and provided a biologically interpretable framework for integrating histopathology and transcriptomics to refine progression-risk assessment in early-stage EEC. Further validation in larger, prospective, multi-institutional cohorts is warranted to establish its generalisability and clinical utility.