Pretreatment CT-based radiomics and machine learning models for predicting treatment response in lung cancer: A diagnostic test accuracy meta-analysis.

Abstract

e20023 Background: Lung cancer remains one of the most commonly diagnosed malignancies and the leading cause of cancer-related mortality worldwide. CT-based noninvasive predictive biomarkers, including radiomics and machine learning models, may aid in predicting treatment response and guiding therapy selection. However, heterogeneous evidence underscores the need for pooled analyses to define their clinical utility. Methods: This PRISMA-compliant meta-analysis was prospectively registered in PROSPERO. We systematically searched PubMed, Embase, Cochrane CENTRAL, and ClinicalTrials.gov from inception to January 2026 to identify studies on radiomics signatures and machine learning models for predicting treatment response (TR) and pathological response (PR), reporting poolable AUCs. Logit-transformed AUCs were used for meta-analysis, and pooled estimates were calculated using a generic inverse-variance model with REML tau estimation in R (version 2025.05.0+496; 26 Posit), with a two-sided significance threshold of p < 0.05. Results: The pooled logit AUC for radiomics signatures predicting treatment response was 1.85 [1.02–2.68]; I² = 77.0%, with combined radiomics–clinical models showing slightly improved performance (1.87 [1.02–2.71]; I² = 0%). Machine learning models demonstrated a pooled AUC of 1.42 [0.61–2.23]; I² = 94%, with the highest-performing models reaching 2.70 [0.53–4.87]. In chemotherapy-alone cohorts, the pooled AUC was 2.02 [0.78–3.26], whereas chemoimmunotherapy cohorts exhibited lower performance (0.42 [0.25–0.59]; I² = 0%, P = 0.026). Validation cohorts achieved an AUC of 1.01 [0.27–1.75]; I² = 71%, with multicenter studies reporting 0.85 [0.44–1.26] and single-center studies 1.41 [0.69–2.12]. For pathological response, radiomics signatures achieved a pooled logit AUC of 1.27 [0.92–1.63]; I² = 69.4%, with multicenter studies performing better (1.66 [1.42–1.89]; I² = 0%) and validation cohorts reaching 1.01 [0.77–1.26]; I² = 0%. Machine learning models achieved a pooled AUC of 1.40 [1.13–1.67]; I² = 42.3% in training cohorts and 0.94 [0.69–1.18]; I² = 0% in validation cohorts. Subgroup analyses did not reveal statistically significant differences. Conclusions: Radiomics signatures and high-performing machine learning models show strong predictive value for treatment response, particularly when combined with clinical features. However, their predictive performance for pathological response is moderate and inconsistent, highlighting the need for further multicenter validation before routine clinical use.