AI-driven dual-task prediction model for co-stratifying efficacy and toxicity in NSCLC immunotherapy

While effective against non-small cell lung cancer (NSCLC), PD-1 inhibitors can induce immune-related adverse events (irAEs), occurring in up to 15.2% of patients and potentially fatal. Currently, effective predictive biomarkers capable of simultaneously forecasting both irAEs and immune checkpoint inhibitor (ICI) responders remain elusive. This limitation hinders the safe clinical application of these agents. This study enrolled 333 advanced NSCLC patients treated with PD-1 inhibitor monotherapy or combination therapy. CT imaging features were extracted using radiomics and deep-learning approaches. Three unimodal and two multimodal models were constructed to predict irAEs (Grade ≥3) and ICI responders in parallel. The SHAP algorithm was used to identify clinical features contributing to the prediction of both irAEs and ICI responders. The CDML-DenseNet model, integrating clinical features with deep-learning-derived radiomics features (DenseNet), demonstrated superior performance in predicting irAEs (AUC = 0.85), outperforming single-modal radiomics models. For ICI responder prediction, the CDML-DenseNet model achieved an AUC of 0.866. The Prognostic Nutritional Index (PNI) was identified as a key feature in both irAEs and ICI responder prediction models. Patients who were non-responders to ICIs but experienced irAEs had significantly lower PNI (46.8 ± 8.779, P < 0.05) compared with ICI responders without irAEs. Our multimodal CDML-DenseNet model effectively predicts both irAEs and ICI responders in NSCLC patients receiving PD-1 inhibitors. This approach provides a novel framework for balancing immunotherapy efficacy and toxicity. Furthermore, the readily available and cost-effective PNI offers clinicians a practical tool to identify potential non-responders experiencing irAEs and to refine treatment decisions.