Predicting Early Recurrence in Hepatocellular Carcinoma: Leveraging Machine Learning for Enhanced Prognosis and Personalized Treatment

Introduction Hepatocellular Carcinoma (HCC) recurrence is a significant clinical challenge, especially within the first year following treatment. Early prediction of recurrence is crucial for improved prognosis and personalized treatment strategies. Aim This study aims to establish an effective predictive model for early recurrence (i.e., within one-year post-treatment) in naive treatment patients with HCC. Materials and Methods Our study encompasses 320 patients, divided into an 80/20 training/test split, with 10-fold cross-validation employed within the training set. Various predictive models including Random Forest (RF), Multilayer Perceptron (MLP), Support Vector Machine (SVM), Logistic Regression (LR), Gradient Boosting, AdaBoost, Naive Bayes, and K-Nearest Neighbors (KNN) were compared. Predictive variables included full blood panels, liver disease etiology, portal hypertension signs, and imaging information. Results The most performant model was the RF in both the training (AUC 0.95, Accuracy 90%, Recall 0.97, Precision 0.98, F1. 0.94) and the test (AUC 0.90, Accuracy 88%, Recall 0.94, Precision 0.95, F1. 0.90) sets. The ten features that impacted the most on the model predicted risk were: tumor margin characteristics (smooth vs. irregular), arterial peritumoral enhancement, alpha-fetoprotein, wider nodule diameter (mm), platelet count, previous semestral ultrasound screening, number of nodules, portal vein thrombosis, and presence of portosystemic shunts. Conclusions Our study demonstrated the viability and robustness of the RF model in predicting early recurrence of HCC. The identified ten high-impact factors underscore the importance of comprehensive assessments involving clinical, biochemical, and radiological parameters. This model, incorporating tumor characteristics and liver function, allows for early risk stratification of recurrence, highlighting patients who might benefit from more aggressive or tailored treatments post-HCC diagnosis. These findings present a strong foundation for developing a clinical decision support system that could significantly enhance personalized patient management and ultimately improve HCC prognosis. Integrating machine learning models into clinical practice opens promising avenues for advancing the precision medicine paradigm in hepatocellular carcinoma management. Future studies should seek to validate these results in larger, diverse cohorts and explore the potential for incorporating additional dynamic and molecular markers.