Predictive value of machine learning for radiation pneumonitis and checkpoint inhibitor pneumonitis in lung cancer patients: a systematic review and meta-analysis.

Some studies have developed machine learning (ML) models for the prediction of pneumonitis following immunotherapy and radiotherapy for patients with lung cancer (LC). However, the prediction accuracy of these models remains a topic of debate. Thus, this study aims to summarize the advantages of ML methods in the early prediction of radiation pneumonitis (RP) and checkpoint inhibitor pneumonitis (CIP) in LC patients. PubMed, Cochrane, Embase, and Web of Science were searched up to March 23, 2025. The Prediction Model Risk of Bias Assessment Tool (PROBAST) was utilized to explore the risk of bias (RoB) in the included studies. A subgroup analysis was conducted based on variables including radiomics, dosiomics, and clinical characteristics. Fifty-six studies comprising 12,803 LC patients were included. Of these, 43 studies focused on the early prediction of RP, 11 studies on CIP, and 2 studies on differentiating RP and CIP. The meta-analysis revealed that the c-index of dosiomics-based models, radiomics-based models, and models based on radiomics and clinical characteristics for predicting RP was 0.82 (95% CI: 0.76-0.87), 0.80 (95% CI: 0.71-0.89), and 0.90 (95% CI: 0.86-0.94), respectively. In the prediction of CIP, the c-index for the clinical characteristics model was 0.83 (95% CI: 0.81-0.85), while the integrated radiomics and clinical characteristics model achieved a c-index of 0.86 (95% CI: 0.80-0.92). The ML-based models exhibit strong performance for predicting RP and CIP. Models that integrate dosiomics and radiomics demonstrate superior predictive performance for RP. In addition, hybrid models combining radiomics with clinical features provide excellent predictive value for CIP.