Machine learning-based normal tissue complication probability model for predicting albumin-bilirubin (ALBI) grade increase in hepatocellular carcinoma patients

Abstract

Abstract Purpose: The aim of this study was to develop a normal tissue complication probability model using a machine learning approach (ML-based NTCP) to predict the risk of radiation-induced liver disease in hepatocellular carcinoma (HCC) patients. Materials and methods: The study population included 201 HCC patients treated with radiotherapy. The patients’ medical records were retrospectively reviewed to obtain the clinical and radiotherapy data. Toxicity was defined by albumin-bilirubin (ALBI) grade increase. The normal liver dose-volume histogram was reduced to mean liver dose (MLD) based on the fraction size-adjusted equivalent uniform dose (2 Gy/fraction and α/β = 2). Three types of ML-based classification models were used, a penalized logistic regression (PLR), random forest (RF), and gradient-boosted tree (GBT) model. Model performance was compared using the area under the receiver operating characteristic curve (AUROC). Internal validation was performed by 5-fold cross validation and external validation was done in 44 new patients. Results: Liver toxicity occurred in 87 patients (43.1%). The best individual model was the GBT model using baseline liver function, liver volume, and MLD as inputs and the best overall model was an ensemble of the PLR and GBT models. An AUROC of 0.82 with a standard deviation of 0.06 was achieved for the internal validation. An AUROC of 0.78 with a standard deviation of 0.03 was achieved for the external validation. The behaviors of the best GBT model were also in good agreement with the domain knowledge on NTCP. Conclusion: We propose the methodology to develop an ML-based NTCP model to estimate the risk of ALBI grade increase.