Prediction of pathological complete response to neoadjuvant chemo-immunotherapy in locally advanced lung adenocarcinoma: a multimodal machine learning approach

Predicting pathological complete response (pCR) following neoadjuvant chemo-immunotherapy in locally advanced lung adenocarcinoma (LUAD) remains clinically challenging. We aimed to develop an explainable machine learning framework to predict pCR and stratify patient prognosis. This retrospective study included 627 driver-negative LUAD patients (2020–2025). Fifteen baseline clinicopathological, metabolic, and inflammatory features were integrated to evaluate four algorithms. The optimal model was interpreted using SHapley Additive exPlanations (SHAP) and validated for progression-free survival (PFS). The Extreme Gradient Boosting (XGBoost) model exhibited superior discriminative performance (AUC = 0.872), significantly outperforming traditional logistic regression ( P < 0.001). SHAP analysis identified PET-CT SUVmax, CEA, and PD-L1 expression as the primary predictors, unveiling their non-linear thresholds and synergistic interactions. Decision curve analysis confirmed its clinical net benefit. Furthermore, patients stratified into the high-probability group by the AI model demonstrated significantly prolonged PFS, even within highly malignant sub-cohorts ( P < 0.001). This SHAP-enhanced XGBoost framework shows significant potential as a proof-of-concept model for predicting pCR and stratifying long-term survival in LUAD. However, given the artificially enriched cohort, prospective multicenter external validation is required before it can be considered a definitive clinical tool for optimizing individualized neoadjuvant therapeutic strategies.