DSAI-04 AI, RADIOMICS, AND IMAGING-BASED MODELS FOR PREDICTING BRAIN METASTASES IN LUNG CANCER: A META-ANALYSIS STRATIFIED BY CANCER SUBTYPE

Abstract Brain metastases (BM) frequently occur in lung cancer (LC). Artificial Intelligence (AI), Machine Learning (ML), and radiomics-based models show promise for early BM risk prediction. This meta-analysis evaluates the pooled diagnostic performance of AI/ML models trained primarily on clinical data and thoracic imaging to predict future BM development in LC patients. A systematic search of PubMed, MEDLINE, and Cochrane (through March 2025) identified studies applying AI/ML to predict BM in LC patients segregated as non-small cell LC (NSCLC) and small cell LC (SCLC). Studies required extractable area under the receiver operating characteristic curve (AUC). Data on model architecture, cancer subtype, imaging modality, and demographics were collected. Random-effects meta-analysis using DerSimonian-Laird τ² was led. Thirteen retrospective studies (N=7,543) were included. NSCLC represented 88.5% and SCLC 11.5% of cases. Among NSCLC studies reporting histology, adenocarcinoma accounted for 66.3%. Gender was reported in 11 studies, with 55.1% male subjects. AI/ML methods included random forest (n=4), support vector machine (n=3), deep convolutional neural networks (n=2), and LASSO-based regression (n=4). Imaging modalities included CT, PET/CT, MRI, and digital pathology. The pooled AUC for all models was 0.860 (95% CI: 0.823–0.898; p<0.001, I²=75.2%). Subgroup AUCs were 0.866 for LUAD (95% CI: 0.823–0.909; I²=79.2%, p=0.028), 0.845 for NSCLC (95% CI: 0.799–0.891; I²=72.1%, p=0.006), 0.896 for SCLC (95% CI: 0.889–0.903; I²=0.0%, p<0.001), 0.780 for NSCLC-EGFR, and 0.901 for Mixed histology. Sensitivities ranged from 71%–96%, specificities 70%–96%, and accuracies 70%–95%, with most models above 80%. AI/ML models showed moderately strong diagnostic performance for predicting brain metastases in LC, with particularly high AUCs for SCLC and mixed histology subtypes. Heterogeneity, retrospective design, and limited validation constrain clinical applicability. Standardized, prospective studies are warranted to reduce bias.