基于医学影像的机器学习模型预测胸部肿瘤中 PD-L1 和 CD8+TILs 表达的性能：系统综述与 Meta 分析

Following PRISMA guidelines, we systematically searched PubMed, Cochrane, Embase, and Web of Science for studies published up to July 01, 2025. Studies evaluating AI-driven medical imaging for predicting thoracic tumor TIME were included. Diagnostic accuracy data were extracted, and a meta-analysis was conducted using a random-effects model to assess the predictive performance of AI models for PD-L1 expression and CD8+ TILs. Heterogeneity analysis and publication bias assessment were also performed.

A total of 68 studies were included, of which 25 were eligible for meta-analysis. The pooled area under the curve (AUC) for AI-driven medical imaging prediction of PD-L1 expression was 0.81 (95% CI: 0.78-0.85), with a sensitivity of 0.77 (95% CI: 0.74-0.80) and a specificity of 0.73 (95% CI: 0.71-0.76). For CD8+ TIL prediction, the pooled AUC was 0.86 (95% CI: 0.82-0.89), with a sensitivity of 0.81 (95% CI: 0.72-0.87) and a specificity of 0.81 (95% CI: 0.76-0.84). Subgroup analysis indicated that integrating multimodal imaging and deep learning models improved predictive performance. Nevertheless, considerable heterogeneity was observed among studies (I > 75%).

AI-driven medical imaging exhibits strong predictive capability for thoracic tumor TIME, particularly in PD-L1 expression and CD8+ TIL prediction. However, significant interstudy heterogeneity limits the generalizability of current models. Future studies should prioritize multicenter, large-scale research, standardization of imaging feature extraction, and integration of biological mechanisms to improve the clinical applicability of AI models.