Machine learning-enabled early detection of hepatocellular carcinoma utilizing cell-free DNA mutation and fragmentation multiplicity: a prospective study

Abstract

Successful development of effective hepatocellular carcinoma (HCC) early diagnosis methods could greatly benefit disease control. Relating to the early detection of liver cancer, multifarious methods exploiting the various genetic aberrations embedded in cell-free DNA have been proposed. Multifaceted feature integration could improve model performance and interpretability. The cohort design and prospective performance validation also significantly affect the model generality. Considering the current demerits, we conducted the PRospective Early Detection In a population at high-risk for Common malignant Tumor (PREDICT) study (clinical trial number NCT04405557), which integrated mainly single nucleotide variants (SNVs) and fragmentation information in model construction on 371 retrospective participants for efficient HCC early detection. The PREDICT model reached 88.41% sensitivity and 95.65% specificity and demonstrated outstanding performance among different clinicopathological populations. Additionally, we integrated the PREDICT model into physical examination packages and prospectively recruited 720 participants from 24 medical institutions. PREDICT model reached 100% sensitivity and 86.7% specificity. Our model reaches a relative equilibrium between cost, performance as well as interpretability and offers an alternative solution for HCC risky individual regular screening and healthy population preventive screening.