Predicting neoadjuvant immunotherapy efficacy with machine learning models in non-small cell lung cancer: A systematic review and meta analysis

• The methodological quality and risk of bias associated with machine learning models predicting responses to neoadjuvant immunotherapy in resectable NSCLC were assessed using the PROBAST + AI tool, the IJMEDI checklist, and the RQS systems. • Quantitative analysis methods were employed to summarize the discrimination, sensitivity, specificity, and other relevant indicators of the machine learning models. Additionally, subgroup analyses were performed based on predicted outcomes, types of predictors, and machine learning methods. • By systematically identifying the methodological deficiencies observed in this study, guidance is provided for future methodological optimization and clinical application. The response of resectable non-small cell lung cancer (NSCLC) to neoadjuvant immunotherapy is heterogeneous. Machine learning can integrate multimodal data to construct predictive models, but the methodological quality, risk of bias and clinical applicability of such models have not been systematically evaluated. This study aims to systematically evaluate the methodological quality, risk of bias, and diagnostic performance of machine learning models for predicting neoadjuvant immunotherapy response in resectable NSCLC. As of August 22, 2025, 11 databases were retrieved. Two researchers independently extracted the data, and a third researcher resolved the data differences. The quality of the model, the development process and the quality of radiomics reports were evaluated respectively by probast + AI, IJMEDI checklist and RQS. Meta-analysis of the AUC, sensitivity and specificity of the model was conducted using R software, and subgroup analysis was performed according to predictors, algorithms and outcomes. Seventeen studies involving 44 models were included. Eighty-nine percent of models had relatively low quality and all had a high risk of bias – key flaws included unreasonable sample size, improper handling of missing data and defects in validation procedures – but the overall applicability was good. IJMEDI scores ranged 26.5–37 (4 high-quality, others medium); average RQS of 12 radiomics studies was 14.58 (22.22%-52.78%), with multiple deficiencies. Ten internal validation models showed that the combined internal AUC was 0.786 (95% CI: 0.740–0.826, I2 = 0%), there was no publication bias (Egger’s test), and the sensitivity was 0.763 (95% CI: (0.56–0.89), with a specificity of 0.908 (95% CI: 0.471-0.991). The predicted AUCs of MPR and PCR were 0.805 and 0.761, respectively. SVM achieved the highest AUC (0.841), and the non-radiomics model (0.869) was superior to the radiomics model (0.775). The combined external validation AUC was 0.760, among which the AUC predicted by MPR was 0.754. ML models show potential for predicting neoadjuvant immunotherapy efficacy in resectable NSCLC, with SVM and non-radiomics models superior. However, low methodological quality and high bias risk require cautious interpretation. Future work should refine methodology, address radiomics gaps, and promote clinical translation.