Can AI Guide the Decision to Transplant or Resect for Hepatocellular Carcinoma?
Menée à partir de données coréennes portant sur 3 915 patients atteints d'un carcinome hépatocellulaire traité entre 2009 et 2020, cette étude évalue la performance de modèles d'apprentissage automatique pour estimer la survie des patients en fonction du traitement reçu (greffe hépatique ou résection chirurgicale) et orienter ainsi le choix thérapeutique
Liver transplantation (LT) offers superior outcomes compared with surgical resection (SR) in the subgroup of patients with hepatocellular carcinoma (HCC), particularly those with underlying synthetic dysfunction, cirrhosis, and portal hypertension. The drawbacks of LT are the limited availability of donors, making waiting time for curative treatment unpredictable, and long-term morbidity that is associated with lifelong immunosuppression. The study by Kim et al aimed to develop and validate a machine learning (ML) decision model to optimize the selection of patients with HCC for LT vs SR. Data for derivation cohort of 3915 patients who underwent LT or SR for HCC between 2008 and 2018 were obtained from the Korea Central Cancer Registry. The external validation cohort data were retrospectively collected from Seoul St Mary’s Hospital for 614 patients who underwent treatment for HCC between 2009 and 2020. Six ML algorithms were applied to estimate 3-year overall survival (OS) by using demographics, clinical, and tumor characteristics that were collected prior to treatment initiation. Patients were stratified into high- and low-risk groups for each treatment, which was followed by identification of LT-favorable and LT-nonfavorable groups. To compare OS based on ML-guided prediction vs clinical practice treatment decision, counterfactual analysis was used. Key factors to model outcomes using ML algorithms were tumor number, platelet count, and protein induced by vitamin K absence-II (PIVKA-II), creatinine, alpha-fetoprotein (AFP), and total bilirubin levels for the LT cohort, whereas maximum tumor size, international normalized ratio, and AFP, albumin, PIVKA-II, and sodium levels were key factors for the SR cohort. In the derivation cohort, the model recommended SR for 221 of 296 (74.7%) patients who originally underwent LT, and LT for 701 of 3619 (19.4%) patients who originally underwent SR. Application of ML-guided treatment demonstrated improved OS (hazard ratio, 0.46 [95% CI, 0.42-0.50]; P < .001). The external validation cohort confirmed the performance of the model, with the area under the receiver operating characteristic curve (AUROC) of 0.75 (95% CI, 0.70-0.80) for the LT and 0.80 (95% CI, 0.75-0.85) for the SR cohorts. The authors concluded that ML-based strategies may enhance individualized decision to treat and optimize transplant resource utilization for patients with HCC.
JAMA Network Open , éditorial en libre accès, 2025