S2973 AI Bridges the Diagnostic Gap in Gastrointestinal Oncology Practice: A Comprehensive Meta-Analysis Across Multiple Outcomes

Introduction: Significant variability exists in diagnostic performance between trainees and expert pathologists, workforce shortages can additionally limit access to expert interpretation. Artificial intelligence (AI) has emerged as a promising tool to enhance diagnostic consistency and accuracy. This meta-analysis aims to evaluate the role of AI in bridging the diagnostic gap in gastrointestinal oncology practice. Methods: A comprehensive literature search of PubMed, ScienceDirect, Google Scholar, PLOS ONE and ClinicalTrials.gov was conducted for studies published from January 2013 to October 2024. Studies included were required to report head-to-head comparisons between AI, expert and trainee pathologists in accuracy, sensitivity and specificity. Data extraction was performed and pooled estimates of mean differences with 95% confidence intervals (CI) were calculated for each outcome. Heterogeneity was assessed using the I² statistic. Forest plots were made for accuracy, sensitivity and specificity comparisons. Results: Out of 5249 studies retrieved, 4 met all inclusion criteria, comprising 433 patient cases. AI significantly outperformed trainees in accuracy (mean difference: +0.18; 95% CI: 0.16–0.21; P < 0.0001) and sensitivity (mean difference: +0.24; 95% CI: 0.20–0.28; P < 0.0001), with no significant difference in specificity. AI demonstrated non-inferior performance to experts across all metrics. The pooled mean difference in accuracy was +0.04 (95% CI: 0.02–0.06; P < 0.0001), favoring AI but representing a small absolute difference. Sensitivity and specificity differences were not statistically significant. Some heterogeneity was observed in most analyses likely reflecting variation in study designs. When AI assistance was provided to human readers, diagnostic result improved for trainees. Conclusion: Our meta-analysis demonstrates that AI can serve as an effective tool to bridge the diagnostic gap in Gl cancer. AI not only matched expert pathologists in accuracy and sensitivity, but also significantly outperformed trainees. These highlight the potential for AI in emergency practice where expert pathologists are unavailable or in short supply. Furthermore, AI’s robust performance is promising as a supplementary training resource to accelerate trainee development. Integrating AI in oncology practice could democratize access to expert level diagnostics, reduce interobserver variability and ultimately improve patient outcomes.