Cross-stage validation of a multimodal machine learning model to predict pathological complete response to neoadjuvant chemotherapy or chemoimmunotherapy in resectable stage III non–small cell lung cancer.

Abstract

8022 Background: Pathological complete response (pCR) after neoadjuvant chemoimmunotherapy is associated with improved outcomes in resectable non–small cell lung cancer (NSCLC), yet reliable tools to predict treatment response before surgery are lacking. Machine learning models have shown promise in advanced disease, but their ability to generalize across disease stages remains uncertain. We evaluated the performance of a machine learning model developed in stage IV NSCLC when applied to a cohort of patients with surgically resectable stage III disease treated with neoadjuvant chemotherapy or chemoimmunotherapy. Methods: The DEEP-Lung-IV study (NCT04994795) developed and validated machine learning models for personalized risk prediction based on multimodal data in patients with stage IV NSCLC treated with first-line pembrolizumab and/or chemotherapy. The models incorporated multimodal clinical routine data such as clinical, biological and CT-scan images. For the present study, the trained pre-treatment model was applied without retraining to patients enrolled in the NADIM trials (NCT03081689, NCT03081689) who received neoadjuvant chemotherapy alone (CTx), or in combination with nivolumab (Nivo+CTx), followed by surgery. Model performance was assessed using the area under the ROC curve (AUC) by considering only the last chest CT-scan image before surgery, or by combining it with other multimodal data. Results: A total of 103 patients with available clinical, biological, imaging data and who experienced surgery was considered for validation analysis (N = 88 patients with Nivo+CTx, N = 15 patients with CTx). When applied to the stage III NADIM cohort, the stage IV–derived model demonstrated strong predictive performance for pCR, with an AUC of 0.68 (95% CI, 0.57–0.78) in all patients using the CT-scan image only, rising to 0.76 (95% CI, 0.67–0.84) when additionaly leveraging on clinical and biological data. Performance was consistent across clinically relevant subgroups, especially in the patients treated with Nivo+CTx, with AUC estimates from 0.65 (95% CI, 0.54–0.76) using CT-scan image alone to 0.72 (95% CI, 0.61–0.82) when combined with the other multimodal data. The model was capable of handling missing predictor values, and preserved predictive value despite differences in disease stage and treatment intent. Conclusions: A machine learning model trained in advanced-stage NSCLC accurately predicted pathological response to neoadjuvant chemo- or chemoimmuno-therapy in a resectable stage III cohort, supporting the biological continuity of treatment response across disease stages. This cross-stage generalizability highlights the potential of machine learning–based tools to guide treatment personalization in earlier-stage NSCLC and warrants prospective validation.