Performance of risk prediction models for pancreatic cancer in patients with new-onset diabetes: a systematic review and meta-analysis
TL;DR
Existing PDAC risk prediction models for patients with NOD show moderate discrimination, however, substantial heterogeneity, limited independent external validation, incomplete calibration reporting, and inconsistent assessment of clinical utility constrain their immediate clinical translation.
Abstract
Background New-onset diabetes (NOD) can precede the clinical diagnosis of pancreatic ductal adenocarcinoma (PDAC) and may indicate occult or preclinical disease. However, identifying the small subgroup of patients with NOD who are at sufficiently high risk to justify further evaluation remains challenging. This systematic review and meta-analysis evaluated the performance of PDAC risk prediction models in patients with NOD and summarized their reported discriminative ability. Methods We systematically searched PubMed, Embase, and Web of Science from inception to March 16, 2026. Eligible studies developed or validated PDAC risk prediction models in patients with NOD. Risk of bias was assessed using the Prediction model Risk Of Bias Assessment Tool (PROBAST). Concordance indices (C-indices) were pooled using a random-effects model. For validation cohorts without reported variance estimates, standard errors were approximated using the Hanley-McNeil method. Subgroup analyses, meta-regression, and sensitivity analyses were performed to explore heterogeneity and assess the robustness of the findings. Results Fourteen studies comprising 17 prediction models were included, and 14 independent validation cohort estimates were available for quantitative synthesis. Most models had a high risk of bias in the analysis domain. The pooled C-index was 0.78 (95% confidence interval, 0.75-0.82), with a 95% prediction interval of 0.68-0.86 and substantial heterogeneity (I² = 95.0%). This estimate should be interpreted as a descriptive summary of reported discrimination across heterogeneous validation cohorts rather than as a directly generalizable estimate of clinical performance. Exploratory subgroup analyses suggested higher discrimination in prospective validation cohorts and in models with 2-year prediction windows than in models with 3-year or other/unspecified prediction windows, although several subgroups included few validation cohorts and residual heterogeneity remained substantial. Models evaluated using internal validation showed numerically higher discrimination than those evaluated using independent external validation, suggesting potential optimism. Older age, unintentional weight loss, rapid glycemic deterioration, and proton pump inhibitor use were recurrent clinical warning signals. Conclusion Existing PDAC risk prediction models for patients with NOD show moderate discrimination. However, substantial heterogeneity, limited independent external validation, incomplete calibration reporting, and inconsistent assessment of clinical utility constrain their immediate clinical translation. These models should be interpreted as tools for temporal risk stratification rather than as evidence of a causal relationship between NOD and PDAC onset. Future studies should standardize NOD definitions, improve calibration and interpretability reporting, and prioritize prospective external validation and clinical impact studies before implementation in screening pathways. Systematic Review Registration https://www.crd.york.ac.uk/PROSPERO, identifier CRD420261360099.