Stacking ensemble machine learning combining tumor, peritumoral, and habitat features for predicting high-grade pattern in stage IA lung adenocarcinoma

Accurate preoperative assessment of high-grade pattern (HGP) in stage IA lung adenocarcinoma (LUAD) is crucial for clinical decision-making. This study aimed to evaluate the performance of the stacking ensemble machine learning algorithm and to develop a preoperative prediction model based on multi-scale imaging feature fusion for identifying HGP in stage IA LUAD. Of 467 patients retrospectively enrolled from two centers, those from center 1 were randomly divided into training (n = 240) and internal validation (n = 104) cohorts, and those from center 2 formed the external validation cohort (n = 123). Radiomics features from the gross tumoral volume (GTV) and 4-mm peritumoral volume (PTV4), along with clustering features from intratumoral habitats generated via unsupervised K-means clustering, were extracted from preoperative CT images. Single-scale predictive models were from these features with the stacking ensemble machine learning algorithm. A post-fusion strategy was employed to integrate the signatures from the aforementioned models, which were then combined with clinical independent risk factors to construct the multi-scale fusion model (PM Fusion ). Model performance was comprehensively evaluated using the receiver operating characteristic (ROC) curve, decision curve analysis (DCA) and calibration curve. The SHapley Additive exPlanations (SHAP) method was used for model interpretability analysis. The stacking ensemble model generally showed higher predictive performance than the single-algorithm models. PM Fusion demonstrated good predictive performance, with AUCs (95% CI) of 0.959 (0.937–0.981), 0.857 (0.783–0.931), and 0.836 (0.758–0.914) in the training, internal validation, and external validation cohorts, respectively. Its overall performance was favorable relative to the single-scale models. DCA indicated greater clinical benefit, while calibration curves verified predictive reliability. SHAP analysis further revealed that the PM Habitat -signature was the dominant contributor to the prediction of the model. By integrating information from the tumor core, peritumoral region, and intratumoral heterogeneity, and employing the stacking ensemble learning strategy, PM Fusion provides a useful tool for predicting HGP status in stage IA LUAD, with SHAP providing transparent interpretability.