Interpretable PET/CT habitat radiomics integrates novel metabolic metrics for postoperative risk stratification in stage I lung adenocarcinoma

This study aimed to develop and validate an interpretable machine learning (ML) model that integrates clinical data, normalized distance from the metabolic hotspot to the tumor centroid (NHOC) / normalized distance from the metabolic hotspot to the tumor perimeter (NHOP) metrics, and PET/CT habitat radiomics to predict recurrence in lung adenocarcinoma (LUAD) and perform postoperative risk stratification. In this retrospective study, we analyzed patients with resectable stage Ⅰ LUAD from two hospitals and a public dataset, comprising a training cohort ( n = 253) and an internal validation cohort ( n = 106), including a subset with radiogenomics data ( n = 26). NHOC and NHOP were extracted using LIFEx software. Nine machine learning algorithms were evaluated to identify the optimal radiomics features. The best-performing algorithm was used to construct predictive models incorporating clinical variable, NHOC/NHOP, and habitat-based radiomics features. Shapley Additive Explanations (SHAP) were employed to interpret model contributions, while paired RNA-sequencing explored potential biological pathways. Smoking history and NHOCpeak were identified as independent predictors constituting the clinical model. The combined model demonstrated moderate discrimination performance for recurrence prediction. achieved an AUC of 0.876 (95% CI: 0.824–0.929) in the training cohort (Sensitivity: 0.890, Specificity: 0.720, PPV: 0.565, NPV: 0.942) and an AUC of 0.794 (95% CI: 0.671–0.918) in the internal testing cohort (Sensitivity: 0.833, Specificity: 0.750, PPV: 0.568, NPV: 0.919). NHOCpeak was the highest-ranked individual predictor in the SHAP analysis, whereas RadScore summarized the aggregate contribution of the radiomic features. The hybrid model successfully stratified patients into distinct risk groups for recurrence-free survival (RFS). Notably, the high-risk group showed enrichment in pathways related to motor proteins and neuroactive ligand-receptor interactions. This study presents exploratory findings of an interpretable machine learning model with moderate discrimination for the postoperative prediction of recurrence in stage I LUAD. NHOC emerged as a promising, novel independent predictor, providing significant added prognostic value beyond classical PET parameters.