Can we be one step ahead?: A systematic review and meta-analysis on the clinical prowess of artificial intelligence (AI) in predicting the malignant potential of pancreatic cysts.

Abstract

e16008 Background: Population imaging studies show that pancreatic cystic lesions (PCLs)occur in about 2.5% of asymptomatic adults, with prevalence increasing sharply with age to nearly 10% in people over 70. Although most pancreatic cystic lesions are benign, some, such as Intraductal papillary mucinous neoplasms, carry malignant potential, making accurate diagnosis essential to guide appropriate surveillance or surgical management. Predicting malignant transformation is difficult, and the growing number of patients undergoing imaging surveillance places a significant burden on healthcare systems. Recent advances in machine learning and artificial intelligence that integrate multivariate data have demonstrated considerable potential to improve diagnostic precision beyond current clinical guidelines. Methods: The review followed PRISMA guidelines, with extensive searches of PubMed, Google Scholar, and ScienceDirect, followed by systematic qualitative and quantitative analysis of the included studies. Studies evaluating the clinical performance of Artificial Intelligence (AI) in predicting the malignant potential of pancreatic cysts using metrics such as accuracy, Area under the curve (AUC), sensitivity, and specificity were included in the final analysis. The data were analysed using the Meta, Metadata, and Metafor packages in RStudio.The study evaluated pooled AUC, sensitivity, specificity, and accuracy, along with AI model subgroups, to estimate overall clinical performance. Both the common and random effects models were considered within the linear (mixed-effects) model framework for specific statistical analyses. The Higgins’ I^2 statistic was used to assess heterogeneity across studies. Results: Eight studies met the inclusion criteria, comprising a total of 1,070 patients. The pooled AUC for the AI models was 0.93 [0.85; 1.00, 95% CI, p = 0.8832]. The overall pooled accuracy was 0.90 [0.80; 0.95, 95% CI, p < 0.0001]. Among the models, the convolutional neural network (CNN) achieved the highest accuracy at 0.94, whereas the Supervised 862 model demonstrated the lowest accuracy at 0.69. The pooled sensitivity and specificity were 0.92 [0.79; 0.98, 95% CI, p < 0.0001] and 0.93 [0.58; 0.99, 95% CI, p < 0.0001], respectively. Conclusions: The AI models evaluated in this analysis demonstrate strong predictive performance for assessing the malignant potential of pancreatic cysts, as indicated by high performance metrics. Further research is warranted to support broader clinical adoption. Pooled AI diagnostic performance. Outcome/Metric Pooled estimate 95% CI p-value Pooled AUC 0.93 0.85–1.00 0.8832 Overall accuracy (all models pooled) 0.9 0.80–0.95 <0.0001 Pooled sensitivity 0.92 0.79–0.98 <0.0001 Pooled specificity 0.93 0.58–0.99 <0.0001