Imaging-Based AI for Predicting Pathological Complete Response in Esophageal Cancer After Neoadjuvant Therapy: Systematic Review and Meta-Analysis
Journal of Medical Internet Research ·
Background: Imaging-based AI may help predict pathological complete response (pCR) after neoadjuvant therapy for esophageal cancer. However, an encouraging average diagnostic performance does not establish reliable performance across clinical settings or justify treatment decisions without further validation. Objective: We aimed to assess the predictive performance of imaging-based AI across neoadjuvant strategies, distinguish internal from external validation, and examine limitations affecting interpretation. Methods: PubMed, Embase, Web of Science, and the Cochrane Library were searched through August 7, 2026. Eligible English-language studies evaluated imaging-based AI against pathological assessment of pCR. PROBAST+AI (Prediction Model Risk of Bias Assessment Tool plus Artificial Intelligence) assessed study quality and risk of bias; GRADE (Grading of Recommendations Assessment, Development, and Evaluation) assessed certainty for sensitivity, specificity, and diagnostic odds ratio (DOR). Bivariate random-effects models synthesized validation groups separately. CIs described uncertainty in average performance, whereas 95% prediction intervals (PIs) described true performance in a similar new study. Results: Thirty-three studies contributed 27 internal-validation datasets with 1891 patients and 16 external-validation datasets with 777 patients. Ten publications contributed to both groups. The sum was 2668 validation observations; the unique patient count across datasets and publications could not be verified. Internal sensitivity was 0.77 (95% CI 0.72‐0.81; PI 0.57‐0.89), specificity 0.82 (95% CI 0.78‐0.86; PI 0.60‐0.93); DOR, 15.49 (95% CI 11.16‐21.50); and area under the curve (AUC), 0.86 (95% CI 0.83‐0.89). External sensitivity was 0.73 (95% CI 0.61‐0.81; PI 0.34‐0.93), specificity 0.83 (95% CI 0.73‐0.90; PI 0.36‐0.98), DOR 12.93 (95% CI 7.25‐23.07), and AUC 0.84 (95% CI 0.79‐0.89). DOR PIs were 5.40‐44.44 and 2.40‐69.62, respectively. Thirty-two studies were conducted in China, and 12 had high validation risk of bias. All GRADE-assessed outcomes had very low certainty. Selecting the first eligible imaging model reported in each article instead of the highest-AUC model showed modest decreases in average performance across 16 internal and 10 external datasets; incomplete alternative-model reporting limited interpretation. Single study algorithm estimates were reported descriptively, without ranking algorithms or interpreting their CIs as evidence of between study consistency or generalizability. Conclusions: By combining multiple neoadjuvant strategies and separating validation stages, this review extends earlier neoadjuvant chemoradiotherapy (NCRT)-focused syntheses and identifies priorities for further model evaluation. Imaging-based AI may support response assessment, but wide PIs, risk of bias, and very low evidence certainty preclude routine clinical adoption without prospective, geographically diverse validation. Future studies should evaluate prespecified models in independent populations, report clinically relevant decision thresholds, and assess calibration and patient outcomes. Until such evidence is available, these models should be considered investigational adjuncts to response assessment and should not independently guide organ preservation or omission of surgery after neoadjuvant treatment. Trial Registration: PROSPERO CRD420251207710; https://www.crd.york.ac.uk/PROSPERO/view/CRD420251207710