S84 Machine Learning-Based Prediction Models for Pancreatic Cancer Risk in New-Onset Diabetes: A Systematic Review and Meta-Analysis of Clinical/Biochemical and Genetic Indicators

Introduction: Pancreatic cancer progresses silently, often presenting after disease has advanced beyond curative stages. A subset of patients with newly diagnosed diabetes may exhibit early signs of pancreatic malignancy. Machine learning algorithms have emerged as tools for analyzing complex clinical, biochemical, and genetic markers to identify high-risk individuals. Despite growing popularity, the predictive accuracy and consistency of these models remain unclear. Methods: We conducted a systematic review and meta-analysis following PRISMA guidelines. Data from 14 studies across PubMed, Scopus, and Google Scholar (inception to 2025) that developed machine learning models to predict pancreatic cancer in new-onset diabetes patients were combined. The primary outcome was area under the receiver operating characteristic curve (AUC). Random-effects meta-analysis pooled diagnostic accuracy measures, with heterogeneity assessed using I² statistics. Leave-one-out sensitivity analysis evaluated robustness of findings. Results: Fourteen studies with over 2.2 million participants from 6 countries were included. Studies employed various machine learning approaches including random forest, logistic regression, and neural networks. Data sources included population biobanks and healthcare databases. The pooled AUC was 0.77 (95% CI: 0.72-0.82), indicating good discriminative performance. Individual AUCs ranged from 0.71 to 0.93, with substantial heterogeneity (I² = 98.95%, P < 0.01). Leave-one-out sensitivity analysis confirmed robustness with consistent estimates of 0.77 (95% CI: 0.72-0.82). Studies utilized diverse predictors including biochemical parameters, demographics, genetic variants, and clinical indicators. Models incorporating clinical and genetic data achieved higher performance. Conclusion: Machine learning models demonstrate good diagnostic accuracy for pancreatic cancer risk prediction in new-onset diabetes (pooled AUC 0.77). Despite heterogeneity, sensitivity analyses confirmed robust findings. These models show promise for clinical implementation as risk stratification tools, enabling earlier detection in high-risk populations. Future research should focus on external validation and clinical integration.