Explainable multimodal fusion model integrating clinical and quantitative CT features for preoperative prediction of high-grade histologic patterns in invasive lung adenocarcinoma

Introduction Accurate preoperative prediction of high-grade patterns in invasive lung adenocarcinoma is critical for effective treatment planning. This study aimed to construct an interpretable multimodal fusion model to predict high-grade pattern status by integrating clinical information and quantitative CT features, including radiomics, intratumoral heterogeneity, and three-dimensional fractal features. Methods In this retrospective study, 743 patients with pathologically confirmed invasive lung adenocarcinoma from two centers were classified into positive and negative groups for high-grade patterns. Patients from the primary center (n=650) were divided into a training cohort (n=455) and an internal validation cohort (n=195), while the second center provided an external validation cohort (n=93). After extracting relevant clinical and imaging features, the optimal predictive model was identified from 80 combinations of various feature selection methods and machine learning algorithms. Results The resulting multimodal fusion model demonstrated robust predictive performance, achieving an area under the curve of 0.848 (sensitivity 84.7%, specificity 71.1%) in the internal validation cohort and maintaining an area under the curve of 0.815 (sensitivity 85.1%, specificity 73.9%) in the external cohort. This integrated approach outperformed individual clinical or radiomics models, as well as standard classifiers such as XGBoost, logistic regression, and multilayer perceptrons. Decision curve analysis further demonstrated favorable potential clinical utility, with the combined model generally providing a higher net clinical benefit than the treat-all and treat-none strategies across the threshold probability range examined. Discussion By employing Shapley Additive Explanations to ensure interpretability, this comprehensive predictive tool offers reliable support for preoperative risk stratification and facilitates more precise, individualized clinical decision-making.