Many hospitals provide patients with online access to their own radiology reports to foster more informed patients and greater shared decision making. However, the complexity of these documents provides significant challenges for patient comprehension. This study explores the feasibility of using machine learning to translate the results of follow-up imaging in patients with hepatocellular carcinoma (HCC) into layperson’s language. HCC surveillance CT examinations performed from January 1, 2010 – October 30, 2017 for patients in IR clinic were mined using the Montage Search and Analytics platform. In the first task, the summary from each radiology report was extracted and the burden of disease was categorized as either stable or not stable. Four models were constructed to predict the stability of disease from the report text utilizing natural language processing (NLP) and demographic information. Model performance was measured with accuracy, sensitivity and specificity as well as positive and negative predictive power. In the second task, a deep neural network model was trained to automatically translate these reports into layperson’s terms. Translation performance was measured by comparing each generated sentence to a reference sentence (BLEU score). Stable disease was seen in a majority of patients (1,865/2,512, 74%). In the first task, the overall accuracy for the XGBoost and support vector machine models was 86% compared to the accuracy for naive bayes (79%) and the random forest (84%) models. Table 1 presents a more detailed list of model performance metrics. In the second task, the deep neural network model achieved a BLEU score of 0.22. This preliminary study demonstrates machine learning models are able to predict the stability of disease from radiology reports with high accuracy. Deep neural network models demonstrate fair performance in the translation of radiology report summaries, likely limited by the size of the dataset. In the future, similar machine translation models may provide patients with access to their results in a format that is easier to understand and allows greater participation in decision making during IR clinic visits.Table 1Model Performance MetricsAccuracy (%)Sensitivity (%)Specificity (%)PPV (%)NPV (%)NB7985628759RF8493598676XGB8693608775SVM8694618781NB, naive Bayes; RF, random forest; SVM, support vector machines; XGB, XGBoost. Open table in a new tab