A machine learning model enhances treatment decisions for hepatocellular carcinoma, optimizing survival outcomes through personalized risk stratification.
A machine learning (ML) decision-support model developed by investigators in the Republic of Korea may be a useful tool for guiding treatment decisions in hepatocellular carcinoma (HCC) by differentiating patients by mortality risk for liver transplantation (LT) or surgical resection (SR) and consequently those who could derive survival benefit from LT, paving the way for individualized treatment, optimized clinical decision-making, and ultimately improved survival outcomes.1
“The key contribution of this study is the potential identification of distinct patient subgroups … through the integration of comprehensive clinical variables into ML algorithms,” said the study investigators, Kim et al, in the paper.1
The selected model first stratified a large retrospective cohort of patients with HCC, drawn from a nationwide cancer registry, who had undergone either LT or SR between 2008 and 2018, by 3-year mortality risk. Following risk stratification, Kaplan-Meier analysis was conducted to estimate survival outcomes.
Among patients who had undergone LT (n = 296), the model categorized 119 patients as high risk and 177 patients as low risk. The low-risk group exhibited significantly better survival compared with the high-risk group, with a hazard ratio (HR) of 0.25 (95% CI, 0.15–0.42; P <.001). Median overall survival (OS) was not reached in either group.
Among patients who had undergone SR (n = 3619), 1028 patients were classified as high risk and 2591 patients as low risk. The low-risk group also showed significantly better survival compared with the high-risk group, with an HR of 0.17 (95% CI, 0.15–.019; P <.001). The median OS for the high-risk group was 32.5 months (95% CI, 30.2–34.7); for the low-risk group, the median OS was not reached.
Based on the risk stratification findings, the model then determined the subgroup of patients who would clearly benefit from LT (LT-favorable) and which patients would have comparable outcomes with SR (LT-nonfavorable) in a survival context, upon which the individual treatment recommendations were based. Specifically, the model recommended patients the treatment yielding the lower estimated mortality risk. The model recommended SR for 74.7% of LT recipients (LT-nonfavorable); conversely, it recommended LT for 19.4% of SR recipients (LT-favorable).
Finally, in a counterfactual analysis, the model estimated and compared survival outcomes under ML-based treatment recommendations and actual recommendations. Notably, ML-based treatment was significantly associated with improved survival with an HR of 0.46 (95% CI, 0.42–0.50; P <.001), equating to a 54% reduction in mortality risk compared with actual treatment.
To ensure real-world accuracy and generalizability of the model, the model was validated against an external validation cohort (n = 614) of patients with HCC who had undergone either LT or SR between 2009 and 2020, derived from a hospital dataset. However, as a retrospective study, the investigators note that the design is subject to selection bias and residual confounding from unaccounted variables.1
“Despite evident strengths of the study, the results should be examined with caution as the analysis did not address several variables, including prior liver-directed therapies, surgical resectability, patients’ comorbidities, assessment of tumor biology, homogeneity of cohorts, and limited follow-up,” noted Varvara A. Kirchner, MD, associate professor of surgery at Stanford Medicine, and Timothy L. Pruett, MD, professor of surgery and internal medicine at the University of Minnesota, in a commentary.2
“Furthermore, the study did not take into the account the surgical resectability and anatomical distribution of the tumors for SR, and patients’ cardiovascular fitness to undergo LT … Thus, incorporation of data on the anatomical location, surgical resectability of the tumor and patients’ comorbidities would be a valuable addition,” Kirchner and Pruett added.
The current literature establishes that LT yields better long-term survival outcomes compared with SR.3 Despite this, treatment selection remains a challenge due to organ donor scarcity and patient heterogeneity.1 This study suggests that ML may potentially be leveraged to address this challenge, offering a practical tool for enhanced clinical benefit, optimized organ allocation, and even improved patient-provider communication and shared decision-making.
“With the evolution of technology and data management, future ML-based models will likely become an integral aspect of personalized medicine; however, even with perfect prediction, the decision to proceed with therapy should always respect patient autonomy and informed consent for treatment,” concluded Kirchner and Pruett.2
Stay up to date on practice-changing data in community practice.