A K-nearest Neighbor Model to Predict Early Recurrence of Hepatocellular Carcinoma After Resection

Patients with hepatocellular carcinoma (HCC) surgically resected are at risk of recurrence; however, the risk factors of recurrence remain poorly understood. This study intended to establish a novel machine learning model based on clinical data for predicting early recurrence of HCC after resection.

A total of 220 HCC patients who underwent resection were enrolled. Classification machine learning models were developed to predict HCC recurrence. The standard deviation, recall, and precision of the model were used to assess the model's accuracy and identify efficiency of the model.

Recurrent HCC developed in 89 (40.45%) patients at a median time of 14 months from primary resection. In principal component analysis, tumor size, tumor grade differentiation, portal vein tumor thrombus, alpha-fetoprotein, protein induced by vitamin K absence or antagonist-II (PIVKA-II), aspartate aminotransferase, platelet count, white blood cell count, and HBsAg were positive prognostic factors of HCC recurrence and were included in the preoperative model. After comparing different machine learning methods, including logistic regression, decision tree, naïve Bayes, deep neural networks, and k-nearest neighbor (K-NN), we choose the K-NN model as the optimal prediction model. The accuracy, recall, precision of the K-NN model were 70.6%, 51.9%, 70.1%, respectively. The standard deviation was 0.020.

The K-NN classification algorithm model performed better than the other classification models. Estimation of the recurrence rate of early HCC can help to allocate treatment, eventually achieving safe oncological outcomes.