Smit Deliwala, MD1, Kewan Hamid, MD1, Mahmoud Barbarawi, MD1, Yazan Zayed, MD1, Pujan Kandel, MD2, Harini Lakshman, MD1, Srikanth Malladi, MD1, Adiraj Singh, MD1, Ghassan Bachuwa, MD, MS, MHSA, FACG1, Grigoriy Gurvits, MD, FACG3, Saurabh Chawla, MD, FACG4; 1Hurley Medical Center, Flint, MI; 2Hurley Medical Center, Grand Blanc, MI; 3New York University Langone Medical Center, New York, NY; 4Emory Health Care, Atlanta, GA
Introduction: Colorectal cancer (CRC) remains a leading cause of cancer-related death in the United States. While colonoscopy based screening is the most effective of all CRC prevention strategies, it is operator dependent and can result in missed lesions, which may contribute to interval cancer. Incorporation of artificial intelligence (AI) to routine colonoscopy has been investigated in several small studies. In this meta-analysis, we attempt to collate evidence from recent randomized controlled trials (RCTs) to further define the role of AI in colonoscopy based CRC screening. Methods: A comprehensive search of MEDLINE, EMBASE, CENTRAL, and ClinicalTrials.gov from inception through May 2020 was completed. Pooled statistics using bivariate random-effects, odds ratio for binary outcomes, and standardized difference in means for continuous outcomes were used. Primary outcomes were Adenoma Detection Rate (ADR) and Polyp Detection Rate (PDR). Secondary outcomes were mean adenomas and polyps/procedure, withdrawal (WT) and cecal intubation times (CIT), and adequacy of bowel preparation. Post-hoc sensitivity analysis and subgroup analyses for adenomas and polyps was completed. Results: 6 RCTs were included. Of 4996 patients, 2487 had AI-assisted, and 2509 had routine colonoscopies. Mean age was 51.99 ± 4.43 years, 51% of males. AI had higher ADRs (p = 0.00) and PDRs (p = 0.00) [Figure 1]. Similar findings were noted for mean number of adenomas and polyps/procedure. Mean WTs favored AI when biopsy times were included. CIT and adequacy of bowel preparation were similar in both groups [Figure 2]. On subgroup analysis, AI systems had significantly better ADRs and PDRs in the transverse colon. AI had higher detection rates for adenomas < 5 mm (p = 0.00) and polyps < 10 mm (p = 0.00). However, routine colonoscopies outperformed AI in detecting pedunculated polyps (p = 0.00). PDRs had no differences based on shape. Sensitivity analysis remained unchanged for ADR, PDR, mean adenomas or polyps detected per procedure and WTs that included biopsies. Begg's funnel plots were relatively symmetrical for ADR (p – 0.25) and PDR (p - 0.20)[Figure 3]. Discussion: The use of AI has the potential to improve the sensitivity of colonoscopy based screening. Colonoscopies using AI algorithms demonstrated significantly improved detection rates for adenomas and polyps. However, research and advancements are needed to refine the AI systems to detect pedunculated polyps and incorporate optical diagnosis into the algorithms.
Figure 1 - Forest plot of primary outcomes.
Figure 2 - Forest plot of secondary outcomes.
Figure 3 - Funnel plot of the published studies.
Disclosures: Smit Deliwala indicated no relevant financial relationships. Kewan Hamid indicated no relevant financial relationships. Mahmoud Barbarawi indicated no relevant financial relationships. Yazan Zayed indicated no relevant financial relationships. Pujan Kandel indicated no relevant financial relationships. Harini Lakshman indicated no relevant financial relationships. Srikanth Malladi indicated no relevant financial relationships. Adiraj Singh indicated no relevant financial relationships. Ghassan Bachuwa indicated no relevant financial relationships. Grigoriy Gurvits indicated no relevant financial relationships. Saurabh Chawla indicated no relevant financial relationships.