This study proposes a machine learning approach to differentiate tumor responses to combination immunotherapies, such as pembrolizumab plus lenvatinib (PL), in hepatocellular carcinoma. However, it is difficult to determine which immune cell profiles (ICPs) will benefit from the therapy, as patients can demonstrate an objective response (responders, R) and those with tumor progression (non-responders, NR). The model was adopted and modified from a previous study, with a random forest classifier being paired with repeated stratified k-fold cross-validation, with hyperparameter optimization being achieved through Optuna. An ICP dataset including both R and NR profiles was used to train For model evaluation, receiver operating characteristic (ROC) analysis and a confusion matrix. The model demonstrated accuracy when discriminating between the R and NR groups, achieving 100% sensitivity and 66.7% specificity. It was also found that the classification of a patient as R or NR was strongly influenced by the levels of CD8 T cells, PD-1+ CD8 NK cells, and PD-L1+ monocytes. Results from the study reflect the ability of the model to recognize distinct ICPs between patients with hepatocellular carcinoma and to predict tumor response to PL therapy. Significant immune subpopulations were also identified. Overall, the findings demonstrate the potential for predicting clinical outcomes before initiating therapy.