Abstract 968P
Background
Current guidelines for hepatocellular carcinoma (HCC) screening in patients with chronic liver disease (CLD) recommend ultrasound with or without alpha-fetoprotein (AFP) every 6 months. The sensitivity of AFP and USG are low for early stage HCC. Following our previous study which identified a novel routine blood test signature for HCC, this study aims to develop a high-dimensional data classifier which accounts for the intricate interactions between these blood parameters for the detection of HCC. Its performance is then compared with AFP.
Methods
Records from 2000 to 2018 were retrieved from a population-based clinical database of >200k patients, the Hong Kong Hospital Authority Data Collaboration Laboratory. CLD (hepatitis/ cirrhosis) patients with and without HCC were identified with ICD codes, antiviral drug history, virology test and radiology reports. Decompensated cases were excluded. Blood records within one month before the diagnosis of HCC and CLD patients were retrieved. The records retrieved included all components of CBC, LFT, RFT and clotting profile. Data from 2000-2015 were split into training and testing cohorts in an 8:2 ratio while data from 2016-2018 were used as the validation set. Machine learning models were applied and their performances were compared with AFP.
Results
The cohort yielded 223,862 patients including 31,149 with HCC (45.9% hepatitis-associated) and 192,713 without. Results from the test set (n = 12528; 3279 HCC) showed promise in safely excluding non-HCC patients with a sensitivity of 90% and negative predictive value of 97%. The model’s high sensitivity maintained steady from late to early stage HCC while that of AFP halved. Our model increases HCC detection rate via USG from 19% to 50% compared to current protocol, indicating potential use in imaging prioritization and cost-saving in HCC surveillance.
Conclusions
An AI classifier for HCC detection is developed from routine blood test patient data. Its performance tested using an ultra-large cohort is consistently high without degradation across all stages with marked superiority to AFP for early stage HCC detection. Currently, multicenter prospective validation is being performed in Hong Kong.
Clinical trial identification
Editorial acknowledgement
Legal entity responsible for the study
The Hong Kong University of Science and Technology.
Funding
Has not received any funding.
Disclosure
All authors have declared no conflicts of interest.
Resources from the same session
832P - Characterisation of EXS73565: A potent and selective MALT1 inhibitor with low drug-drug interaction risk and potential in lymphoma
Presenter: Major Gooyit
Session: Poster session 18
833P - New targets for adult T cell leukemia/lymphoma (ATLL): A map for ATLL immunotherapy
Presenter: Zahra Rezaei Borojerdi
Session: Poster session 18
834P - Phase II clinical study of VR-CAP regimen for first-line treatment of marginal zone lymphoma
Presenter: Junfeng Chu
Session: Poster session 18
835P - A safe and effective immunochemotherapy with oral sobuzoxane and etoposide for untreated diffuse large B cell lymphoma patients aged 80 and over
Presenter: Kaname Miyashita
Session: Poster session 18
838P - Matching-adjusted indirect comparison (MAIC) of axicabtagene ciloleucel (axi-cel) and epcoritamab (epcor) in relapsed/refractory (R/R) large B cell lymphoma (LBCL) after at least two prior systemic therapies (3L+)
Presenter: Olalekan Oluwole
Session: Poster session 18
840P - Genomic landscape, immune characteristics and prognostic mutation signature of extranodal NK/T cell lymphoma, nasal type in China
Presenter: Yue Chai
Session: Poster session 18
841P - Ki67-revised risk index to risk-stratify patients with extra-nodal natural killer/T cell lymphoma
Presenter: Shuo Li
Session: Poster session 18