Predicting invasive risk in lung adenocarcinoma via compositional data analysis: a retrospective cohort study

Abstract

Conventional LUAD grading systems rely mainly on predominant histological patterns and may overlook intratumoral compositional heterogeneity. This study evaluated whether compositional data analysis (CoDa) of histological subtype proportions improves prediction of invasive pathological features in early-stage LUAD. Histological subtype proportions were transformed into five biologically informed ilr coordinates after zero-value replacement. CoDa-based logistic models were used to predict VPI, LVI, VI, LNM, and STAS. Model performance and robustness were assessed using ROC analysis, DCA, SHAP analysis, calibration curves, and clinicopathological-adjusted models. Among 264 patients, the global high-risk balance, z1, reflecting high-risk patterns relative to lower-risk patterns, was positively associated with all invasive features and remained independently associated with LVI and VI after clinical adjustment. The CoDa model outperformed conventional grading systems across most outcomes, with the most robust gains observed for LVI and VI after multiple-comparison correction. SHAP analysis identified z1 as the dominant contributor for LVI, VI, LNM, and STAS, and calibration curves showed acceptable agreement between predicted and observed risks. A web-based risk-assessment tool was developed ( https://mayspace.shinyapps.io/CoDaPredictionModel/ ). CoDa provides an interpretable framework for capturing histological heterogeneity in early-stage LUAD and improves prediction of invasive features beyond conventional grading systems.