Development of a novel routine blood-based AI model for hepatocellular carcinoma screening—a territory-wide study

<h3>Background</h3> Current guidelines for hepatocellular carcinoma (HCC) screening in patients with chronic liver disease (CLD) recommend ultrasound and/or α-fetoprotein (AFP) testing every 6 months. However, both methods exhibit low sensitivity for early-stage HCC. Building on our previous study that identified a novel blood test signature for HCC, this study aims to develop a routine blood-based artificial intelligence (AI) model for HCC detection. <h3>Patients and methods</h3> Records from 2000 to 2018 were retrieved from Hospital Authority Data Collaboration Laboratory, a territory-wide database. CLD (hepatitis/cirrhosis) patients with and without HCC were identified with International Classification of Diseases codes, anti-viral drug history, virology test, and radiology reports. Decompensated cases were excluded. Blood records within 1 month before the diagnosis of HCC and CLD patients were retrieved. Data from 2000 to 2015 were split into training and testing cohorts in an 80 : 20 ratio, while data from 2016 to 2018 were used as the validation set. Machine learning models were applied and their performances were compared with AFP. <h3>Results</h3> The cohort yielded >75 000 patients including 19 670 patients with HCC. Results from the test set demonstrated promising capabilities in excluding CLD patients, achieving a sensitivity of 80% and a specificity of 81% (area under the receiver operating characteristic curve=0.894) at a cut-off value of 0.43. The AI model outperformed AFP in early-stage detection, maintaining higher sensitivity across all Barcelona Clinic Liver Cancer stages. <h3>Conclusions</h3> This novel AI model, which uses only standard blood test results, shows strong potential for improving the detection of HCC, significantly outperforming the traditional alpha-fetoprotein (AFP) test. Its straightforward design offers a practical and widely accessible screening option that could be integrated into existing healthcare pathways to aid earlier diagnosis.