Authors (including presenting author) :
Lam SJL (1), Seo M (2), Cheung KM (3), Lee SC (4), Liu HW (5), Yip ASM (6), Sung W (3), Woo PYM (7), Chow JCH (3), Ng SKK (6), Kan DMY (6), Kao S (4), Yiu HHY (3), Lam DCC (2)
Affiliation :
(1) Department of Medicine, Prince of Wales Hospital, (2) Department of Mechanical and Aerospace Engineering, HKUST, (3) Department of Clinical Oncology, (4) Surgery, (5) Department of Medicine, Queen Elizabeth Hospital, (6) Surgery, Kwong Wah Hospital, (7) Neurosurgery, Prince of Wales Hospital
Introduction :
Esophagogastroduodenoscopy (OGD) faces high demand due to dyspepsia affecting 50% of the population and the prevalence of anaemia in clinical practice. While OGD serves as the primary diagnostic tool for early gastric cancer detection, current case-finding rates remain low at 1-2% of all procedures performed. To optimize healthcare resources and waiting times, effective risk stratification is essential. Routine blood tests (RBT), including CBC and LRFT, are standard pre-OGD investigations that may indicate occult malignancy through markers of anaemia, inflammation, and cachexia. Our team developed an AI-enabled RBT-based risk stratification tool for GC (0.80 sensitivity and 0.80 specificity). Following the HOIT AI community meeting, this tool has been suggested to be explored for endoscopy triage, and this report evaluates its real-world performance in patients scheduled for OGD.
Objectives :
(1) To validate sensitivity in detecting high-risk cases for endoscopy; (2) To evaluate accuracy in identifying low-risk cases suitable for triage
Methodology :
This is a retrospective study on patients who underwent OGD between January and February 2020. Data was retrieved from CDARS, including OGD records, RBTs of CBC and LRFT performed within one month before the procedure, and confirmed gastric cancer diagnoses (ICD-10-C60).
After excluding cases with incomplete records, the final cohort comprised 241 patients, including 3 confirmed GC cases and 238 controls. The previously developed RBT-GC signature model was applied to calculate risk scores ranging from 0 to 1. Based on cutoff values calibrated in the prior HADCL big data validation study, patients were stratified into three risk categories: high risk (score≥0.64), intermediate risk (score 0.32-0.63), and low risk (score< 0.32).
Result & Outcome :
The median RBT-GC risk score for gastric cancer cases was 0.67 (0.54-0.70) versus 0.46 (0.28-0.68) for controls. With a high-risk cutoff of 0.64, RBT-GC identified 2 of 3 cases, yielding a sensitivity of 0.67. Among 182 non-high-risk patients, 181 were true negatives, led negative predictive value of 0.99.
Without RBT-GC, 241 OGDs were performed, detecting 3 gastric cancer cases (number-needed-to-scope; NNS = 80). Implementing RBT-GC and prioritizing high-risk patients could reduce NNS to 35, with 2 cases found in 70 OGDs. For the intermediate risk group (0.32 to < 0.64), 101 patients remained, leading to an NNS of 101. The low-risk group (< 0.32) had 73 patients, all controls, indicating 100% agreement. These patients could be monitored or undergo lower-priority OGDs. RBT-GC is cost-effective and ready for use, potentially shortening OGD wait times for high-risk patients (NNS = 35) and saving 30% of OGDs by observing low-risk patients.
