%0 Journal Article
%T Explainable Multimodal Machine Learning Integrating Radiomics and Pathomics for Prediction of Pathologic Complete Response in Esophageal Squamous Cell Carcinoma Treated With Neoadjuvant Chemoimmunotherapy
%A Lucia C. Nielsen
%A Rafael Garcia
%A Juan Singh
%A Jonas Santos
%A Camila Smith
%J Asian Journal of Current Research in Clinical Cancer
%@ 3062-4444
%D 2021
%V 1
%N 2
%R 10.51847/sVlPwgIlat
%P 113-130
%X Determining, before surgery, which patients with esophageal squamous cell carcinoma (ESCC) will achieve a pathological complete response (pCR) after neoadjuvant chemoimmunotherapy (nCIT) remains a major unmet clinical need. Reliable prediction of pCR could enable risk-adapted treatment strategies and avoid unnecessary surgical intervention. The objective of this study was to design and independently validate a transparent multimodal learning system that jointly leverages radiological and histopathological imaging data to estimate pCR. We retrospectively collected data from 335 patients with ESCC treated with nCIT followed by surgical resection at three tertiary centers. One institution contributed cases that were split into model development (n=181) and internal validation (n=115) cohorts, whereas patients from the remaining centers constituted an external validation cohort (n=39). Quantitative features were extracted from contrast-enhanced CT scans and H&E-stained whole-slide images to construct radiomics-only and pathomics-only classifiers. Two strategies were implemented to integrate these modalities: a feature-level intermediate fusion approach and a prediction-level late fusion approach. Model discrimination and classification performance were evaluated using AUC, accuracy, sensitivity, specificity, and F1 score. Survival differences were explored according to both histologically confirmed and model-inferred pCR status. Model transparency was enforced through the use of interpretable feature definitions and explainable decision mechanisms. The intermediate fusion strategy demonstrated consistently superior performance compared with unimodal models and late fusion across all cohorts. In the development, internal validation, and external validation cohorts, the intermediate fusion model achieved AUC values of 0.97, 0.78, and 0.76, respectively, with corresponding accuracy values of 0.93, 0.87, and 0.77. Both true pCR and predicted pCR groups exhibited distinct overall survival trends in exploratory analyses. Importantly, the model relied on explicitly defined radiological and histomorphological attributes, and its predictions were accompanied by case-specific and population-level explanatory visualizations that clarified the underlying decision logic. A clinician-oriented graphical interface was also implemented to support real-world application. This study presents a clinically interpretable radiopathomics-based prediction framework capable of estimating pCR following neoadjuvant chemoimmunotherapy in ESCC using routinely available imaging data. The proposed approach may assist clinicians in tailoring post-treatment management, particularly when weighing active surveillance against immediate surgery.
%U https://galaxypub.co/article/explainable-multimodal-machine-learning-integrating-radiomics-and-pathomics-for-prediction-of-pathol-rcbiojdeiv6gbad