Riassunto analitico
Background and Aims: Most scores for HCC prediction can assess at most 3- or 5-year HCC risk, as the observation period of the derivative cohort is usually short. We aimed to develop a 5- and 10-year HCC risk score from a prospective cohort of patients with compensated advanced chronic liver disease (cACLD) of any aetiology followed up for 12 years.
Methods: 545 patients with cACLD, HCC-free, prospectively enrolled from 2011 to 2022, using a convenience sampling, underwent at enrolment upper G.I. endoscopy, liver ultrasound/elastography, HVPG measurement, lab tests. Cox proportional models were used to assess the association between esophageal varices, adjusted for all the selected covariates, and HCC incidence. Random Survival Forest (RSF), a machine learning (ML) prediction model, was used as a sensitivity analysis to test prediction power of the same covariates, considering all the possible interactions and non-linear relationships with HCC incidence as the outcome.
Results: Median follow-up time was 5.9 years. We observed 78 incident HCCs (14.3%). In the fully adjusted Cox proportional models, after the adjustment for covariates, patients with large esophageal varices had 4 times the risk of developing HCC (HR:4.02; 95% C.I.: 2.42-6.68) than patients with small/without varices. The covariates, viral aetiology (HR 2.61;95% C.I.: 1.57-4.35), LSM (for each kPa)(HR: 1.01;95 % C.I. 1.01-1.03), male sex (HR:1.94 C.I. 95%: 1.10-3.41), were also meaningfully associated with HCC risk. As a sensitivity analysis we performed the RSF selection algorithm to rank all the variables of the Cox models, according to their prediction power (using minimal depth metric) for the incidence of HCC. Large esophageal varices had the best prediction power for HCC, followed by LSM, albumin, BMI and viral aetiology. Interestingly, RSF prediction power was in line with the magnitude of association with Cox model, but ML further identified BMI and albumin as related and excluded sex. The score built with the RSF-selected variables (Esophageal Varices [EV] score) had excellent discrimination and calibration in assessing both 5- (AUROC 0.823) and 10-year (AUROC 0.792) HCC risk irrespective of aetiology, with a significantly better overall performance at both time points than aMAP score, built on the same data.
Conclusions: The machine learning approach, used to build this score, allowed us to identify large varices as the most important predictor for HCC risk, underlining the critical pathogenetic role of longstanding and severe portal hypertension in HCC development. This score also obtained better prediction for 5- and 10-year HCC development than aMAP score (i.e. the best score available so far for HCC prediction independently from aetiology) tested in the same dataset. The proposed score is highly reliable. Being based on routine clinical data of the patients with cACLD it can be easily applied worldwide.
|
