AI helps spot HCC sooner

A novel artificial intelligence (AI) classifier can detect hepatocellular carcinoma (HCC) signatures in routine blood tests up to 1 year before clinical diagnosis, as shown in a study from Hong Kong.

Initial conception

Conceptualization of the AI classifier began following big data analysis of 20 years of medical data—ie, complete blood counts, liver and renal function tests, and clotting profiles—from more than 2 million patients. It was found that compared with chronic liver disease (CLD) patients, those with HCC had more severe liver damage, increased systemic inflammation, greater bleeding tendency, and early sign of cachexia.

“Even some blood test results [in HCC patients] remained within the normal range and were easily overlooked by clinicians,” reported lead author Kin Nam Kwok from the Department of Mechanical and Aerospace Engineering, the Hong Kong University of Science and Technology, Hong Kong SAR.

Classifier development

Kwok and colleagues then trained an AI to learn unique HCC signatures from routine blood tests, distinguish between HCC and CLD, and calculate a risk score for screening HCC using a dataset collected from 192,713 CLD patients and 31,149 HCC patients.

At a risk score cutoff of 0.4, the AI classifier had an area under the receiving operating curve of 0.89 for screening HCC. The sensitivity was 80 percent, and the specificity was 81 percent.

Performance testing

The classifier was tested in a cohort of 3,415 HCC and 10,288 CLD patients, for whom blood records within 1 year prior to the diagnosis of HCC were retrieved at intervals (1-3 months, 3-6 months, 6-9 months, and 9-12 months).

Compared with alpha-fetoprotein (AFP) at the 20 ng/mL cutoff, the routine blood-based AI classifier demonstrated a twofold increase in sensitivity for early HCC screening, Kwok noted.

The screening sensitivity was 79.4 percent with the AI classifier vs 43.7 percent with AFP at 0-30 days, 61.3 percent vs 41.9 percent at 1-3 months, 50.1 percent vs 37.9 percent at 3-6 months, 44.2 percent vs 29.8 percent at 6-9 months, and 41.3 percent vs 21.3 percent at 9-12 months before HCC diagnosis. Specificity with the AI classifier was over 75 percent at all time intervals.

The mean lead-time for detection in HCC patients was 167 days ahead of diagnosis.

“Routine blood-based AI classifier might advance the diagnosis of HCC in 40 percent of patients by 1 year,” Kwok said, highlighting its potential to create windows for timely interventions and reducing mortality in this population.

Next steps

Already, routine blood tests are about five times more commonly used than AFP during the surveillance phase, indicating high clinical adherence and the readiness of use of the AI classifier, Kwok pointed out.  

“Our team is eager to conduct prospective study on the routine blood-based AI classifier to further investigate its practicalities. We also aim to integrate the model with point-of-care testing tools to expand early detection coverage, creating a cost-effective, accurate, and timely screening solution for the public,” he added.