Machine learning to predict survival in patients treated for hepatocellular carcinoma: a systematic review

Given the interplay of cirrhosis with malignancy, management of hepatocellular carcinoma (HCC) relies on multidisciplinary clinician judgement to guide treatment. Accurate survival prediction would facilitate decision-making and resource allocation while maintaining patient-centred care with optimized survival outcomes while avoiding futility, particularly regarding liver transplant. Machine learning may offer innovative support in clinical decision making. This systematic review is the first to our knowledge to summarize existing machine learning algorithms predicting survival in patients treated for HCC.

Studies describing a machine learning algorithm predicting survival in patients treated for HCC were included. Studies investigating only genetic or molecular factors, imaging, or disease recurrence were excluded. EMBASE, Medline and Web of Science were searched on December 4th, 2023. Included algorithms were analyzed based on input variables, sample size, algorithm type and performance as measured by the area under the receiver operating characteristic curve and concordance index.

Twelve studies including 43 algorithms were included. The algorithm type, input variables and their selection method, sample size and performance varied widely. Algorithms did not perform differently based on sample size or number of variables. Algorithms using a gradient boosting algorithm tended to perform better across short- and long-term prediction. Short-term predictions tended to perform better than long-term predictions. The area under the receiver operating characteristic curves ranged from 0.66 to 0.92.

Machine learning may offer an invaluable tool in survival prediction and individualized treatment in HCC. Existing algorithms show favourable, yet early, results. Standardized reporting, external validation and equity-focused model development are required before clinical implementation.