Prediction Models for Pathological Response to Neoadjuvant Chemoradiotherapy in Locally Advanced Rectal Cancer: A Systematic Review and Meta-Analysis

Background: Accurate pretreatment prediction of pCR is critical for stratifying patients with LARC and selecting candidates for organ-preserving approaches. However, pCR is a postoperative pathological endpoint, whereas watch-and-wait strategies are based on cCR. Despite a rapidly expanding literature encompassing gene expression signatures, MRI radiomics, deep learning, and multi-modal fusion models, no prior systematic review has comprehensively evaluated predictive performance across all modalities using contemporary meta-analytic techniques for diagnostic test accuracy. Methods: Systematic review and random- effects meta-analysis of studies developing or validating predictive models for pCR in LARC treated with neoadjuvant chemoradiotherapy (nCRT) were conducted. Summary AUC was estimated using logit-transformed random-effects models. Heterogeneity was quantified by I2 and τ2. Subgroup analyses were performed according to model modality and validation strategy. Publication bias was assessed with Egger’s linear regression test. Sensitivity analyses were performed using leave-one-out cross-validation. Results: Forty-three studies with 67,302 patients were included. The pooled AUC across all models was 0.874 (95% CI: 0.855–0.892), with substantial heterogeneity (I2 = 94.3%, τ2 = 0.247). Deep learning models showed the highest pooled AUC (0.919, 95% CI: 0.877–0.948), with MRI radiomics (0.885, 95% CI: 0.852–0.911), multi-modal (0.875, 95% CI: 0.813–0.919), gene expression (0.855, 95% CI: 0.797–0.898), and clinical factors (0.820, 95% CI: 0.788–0.849) showing progressively lower pooled estimates. External validation cohorts had a lower pooled AUC (0.864) than internal validation (0.884). Significant publication bias was detected (Egger’s test p < 0.001). Leave-one-out sensitivity analysis confirmed robustness of the overall estimate. Conclusions: Deep learning and MRI radiomics models demonstrate promising diagnostic accuracy for predicting pCR after nCRT in LARC. However, the predominance of retrospective, single-center designs with limited external validation represents a critical translational gap, and the lack of studies in the total neoadjuvant therapy era limits generalizability. Prospective, multicenter validation studies incorporating standardized endpoints and class imbalance-appropriate metrics are urgently needed.