Integrating Multimodal Data for Precise Subtyping and Prognostication in Pulmonary Mucinous Adenocarcinoma

Background Pulmonary mucinous adenocarcinoma (PMA) represents a rare lung adenocarcinoma subtype, characterized by a lacks of comprehensive pathological classification and prognostic factors. In this study, we introduce a multimodal machine learning framework aimed at improving the accuracy of PMA subtyping and predicting the prognosis of PMA patients. Materials and methods This retrospective study enrolled 175 surgically resected primary PMA cases and demographic, histopathological, CT imaging, and genomic data of patients were collected. LASSO regularized logistic regression model were utilized for histological classification, and Cox proportional hazards model were employed for survival prediction, with internal validaition. Results Pure mucinous adenocarcinoma presented a higher prevalence of smaller tumor size, lower lobe localization, absence of lymph node metastasis, STAS, early pathological stage, CEA ≤ 5ng/mL, and EGFR ^E19 mutation ( P  < 0.001, < 0.001, < 0.001, = 0.001, < 0.001, 0.002, 0.001, respectively) compared to mix mucinous or mucin secretion adenocarcinoma. A machine learning-derived nomogram achieved discriminative accuracy (training AUC = 0.810; validation AUC = 0.785) with excellent calibration. Multivariate Cox modeling identified higher CEA levels, indistinct margin, and EGFR ^E19 mutation as independent prognostic factors in PMA ( P  = 0.043, 0.014, 0.044, respectively). Moreover, Kaplan-Meier curve revealed significantly different outcomes between low and high risk groups stratified upon the nomogram score ( P  < 0.0001). Conclusions Pulmonary pure mucinous adenocarcinoma exhibited lower malignancy compared to the mixed mucinous and mucin secretion type. The nomogram model developed and validated in this study exhibited outstanding efficacy in predicting histological subtype and survival of PMA, offering valuable .guidance for clinicians in diagnosis and treatment decision-making.