The use of advanced machine learning to predict outcomes after atezolizumab plus bevacizumab for advanced hepatocellular carcinoma: a retrospective cohort study

BACKGROUND Combination immune checkpoint inhibitors are recommended as first-line therapy for advanced hepatocellular carcinoma. However, only a third of patients respond to treatment, and improved approaches to predict response are required. Using baseline clinical data, we aimed to use advanced machine learning models to predict overall survival and progression-free survival in patients with advanced hepatocellular carcinoma receiving atezolizumab plus bevacizumab. METHODS This retrospective cohort study was conducted at 24 centres across eight countries. Patients aged 18 years and older with a histological or radiological diagnosis of advanced hepatocellular carcinoma were included; those who had received previous systemic therapy for hepatocellular carcinoma were excluded. All patients received intravenous atezolizumab 1200 mg plus bevacizumab 15 mg/kg once every 3 weeks until disease progression. Seven supervised machine learning models, in combination with 13 feature selection techniques, were trained on 44 baseline clinical variables for the prediction of overall survival and progression-free survival. The three best-performing models, combined with their optimum feature selection techniques, were used to develop ensemble machine learning models for the prediction of overall survival and progression-free survival. The primary outcomes of the study were the predictions of overall survival, progression-free survival, and immunotherapy response using advanced machine learning. k-means clustering was used to stratify patients into two groups: those at low risk and those at high risk of either death (in the overall survival model) or disease progression (in the progression-free survival model). FINDINGS 934 patients who received immunotherapy from May 1, 2018 and were followed up until Oct 1, 2023 were screened, of whom 160 were excluded and 774 were included in the final study. Patients were divided into training (n=339), internal validation (n=146) and external validation (n=289) cohorts. Support vector machine, neural network, and naive Bayes algorithms had the best performance in the prediction of overall survival; for progression-free survival, the highest-performing algorithms were ridge regression, naive Bayes, and logistic regression. In the external validation cohort, the ensemble model for the prediction of overall survival (area under the receiver operating characteristic curve 0·75 95% CI 0·69-0·81) significantly outperformed all eight of the tested clinical benchmark variables: Barcelona Clinic Liver Cancer (BCLC) stage (0·54 0·48-0·61; p<0·0001), α-fetoprotein (AFP) concentration (0·60 0·54-0·67; p=0·0007), albumin-bilirubin (ALBI) grade (0·64 0·58-0·71; p=0·0003), neutrophil-to-lymphocyte ratio (0·56 0·49-0·62; p<0·0001), platelet-to-lymphocyte ratio (0·51 0·44-0·58; p<0·0001), combined ALBI grade and BCLC stage (0·67 0·60-0·73; p=0·0074), and two BCLC subclassifications (0·62 0·55-0·69; p=0·0007 and 0·61 0·55-0·68; p=0·0018). The ensemble model for the prediction of progression-free survival (0·64 0·59-0·70) outperformed five of the eight clinical predictors: BCLC stage (0·52 0·46-0·58; p<0·0001), neutrophil-to-lymphocyte ratio (0·53 0·47-0·59; p=0·0069), platelet-to-lymphocyte ratio (0·54 0·48-0·60; p=0·016), and two BCLC subclassifications (0·57 0·50-0·64; p=0·020 and 0·55 0·49-0·62; p=0·0091); the model did not outperform AFP concentration (0·59 0·53-0·64; p=0·14), ALBI grade (0·62 0·56-0·67; p=0·44), or combined ALBI grade and BCLC stage (0·59 0·53-0·66; p=0·12). For the overall survival model, patients stratified into the low-risk group had significantly longer median overall survival (16·4 months 95% CI 14·2-21·6) than those in the high-risk group (4·8 months 3·0-6·9; p<0·0001); similarly, patients stratified by the progression-free survival model into the low-risk group had significantly longer median progression-free survival (8·9 months 7·3-11·1) than those in the high-risk group (3·7 months 2·9-5·6; p=0·0021). INTERPRETATION Our advanced machine learning models, which use routinely collected baseline clinical variables, are robust and externally validated and outperform established clinical biomarkers for predicting clinical outcomes with atezolizumab plus bevacizumab. These data-driven models could be used to stratify patients with hepatocellular carcinoma for personalised treatment strategies. FUNDING None.