Abstract 437P
Background
A four-week delay in cancer treatment delay was shown to increase mortality. Gastric cancer survival can be improved if the cancer is detected early. A big-clinical-data AI model was developed using routine blood test data in > 275,000 dyspepsia patients within 30 days of diagnosis in the Hong Kong Hospital Authority Data Collaboration Laboratory. The model demonstrated that 80% of GC patients and 80% of dyspepsia patients could be detected accurately. This study assesses the early detection capabilities of this model using a new cohort.
Methods
This study included patients with prescribed medications for dyspepsia for over 2 months. GC patients were identified with ICD codes, and cases with other cancers were excluded. Blood records (complete blood count, liver/renal function test, and clotting profiles) were retrieved from three time intervals: 1-3, 3-6, and 6-12 months before diagnosis. A risk score with a range from 0 to 1 was calculated for each period and scores > 0.35 were categorized as GC positive. The lead time between correctly predicted positive and diagnosis date were calculated for each patient. The sensitivity and specificity at each time interval were calculated to determine the early detection capability.
Results
Overall, 1,456 GC and 25,531 dyspepsia patients had records from the specified periods. GC patients exhibited an increasing trend in the risk scores as the diagnosis date approached. The model found 365 out of 576 (64%), 203 out of 406 (50%), 205 out of 474 (43%) GC patients early, in the corresponding time intervals. The median lead time for GC detection was 7.1 months ahead of diagnosis. The median specificity over the three periods was 70%. The risk score for negative controls remained negative with high precision (SD 0.03).
Conclusions
The data showed that the AI-enabled routine blood test model could detect GC by 7.1 months earlier than GC diagnosis. Up to 40% of gastric cancer patients could have their GC diagnosed 1 year earlier. This creates a meaningful time window for curative intervention and therefore, indicates its potential in reducing cancer mortality.
Legal entity responsible for the study
Department of Mechanical and Aerospace engineering, HKUST.
Funding
Has not received any funding.
Disclosure
All authors have declared no conflicts of interest.