Abstract 442P
Background
Current tumor markers for gastric cancer (GC) such as carcinoembryonic antigen (CEA) and CA19-9 show inconsistent expression in GC, necessitating a reliable alternative. Our group developed an GC specific AI routine blood signature using >220k clinical big data, and demonstrated its high accuracy of 80% in detecting GC at diagnosis and potentially up to 1 year before clinical diagnosis. To evaluate its application as a tumor marker for treatment response assessment, an analysis was conducted to determine the correlation between post-treatment changes in AI risk score and overall survival (OS).
Methods
This study utilized territory-wide clinical big data from the Hong Kong Hospital Authority Data Collaboration Laboratory, encompassing patient records from 2000 to 2018. Patients with GC diagnosis and history of gastrectomy with or without chemotherapy (fluorouracil, capecitabine, and oxaliplatin), were included. Blood test results (complete blood count, liver/renal function test, and clotting profiles) were collected at two timepoints: 30 days before diagnosis and after completion of GC treatment. Using a light gradient boosting machine model, the AI signature processed the blood test results and assigned a GC risk score (0-1 scale). OS was calculated from treatment initiation to death and analyzed using the Kaplan-Meier method with p < 0.05 being significant.
Results
The final cohort (n=1663) was categorized into two treatment groups: gastrectomy only (n=1132) and gastrectomy followed by adjuvant chemotherapy (n=512). The median risk score in the cohort decreased from 0.77 near diagnosis to 0.49 one month after all GC treatments. The 72.4% of patients who experienced decreased risk scores had longer 5-year OS rates compared to those with increased scores (p<0.05); this trend held for patients receiving surgery alone (5-year OS 68% vs 60%) and adjuvant chemotherapy (5-year OS 61% vs 42%). No correlation was found between the single reading of risk score at diagnosis and OS (p > 0.05).
Conclusions
The AI routine blood signature exhibited the characteristics of an ideal tumor marker, accurately detecting GC at/before diagnosis and showing predictive value for post-treatment OS. Prospective validation of these findings is being conducted.
Legal entity responsible for the study
Department of Mechanical and Aerospace Engineering, HKKUST.
Funding
Has not received any funding.
Disclosure
All authors have declared no conflicts of interest.