The following is the summary of “An AI Approach for Identifying Patients With Cirrhosis” published in the January 2023 issue of Clinical gastroenterology by Obeid, et al.

The purpose of this research was to assess the efficacy of applying deep learning to EHR clinical text to identify cirrhotic patients. In order to conduct reliable epidemiological, health services, and outcomes research, it is crucial that cirrhosis be correctly identified in EHR. International Classification of Diseases (ICD) codes are currently used in these kinds of initiatives, but they are only partially successful.

Researchers used discharge summaries from patients in a patient registry who were diagnosed with cirrhosis to train a number of machine learning models, while ICD-code-based controls served as a positive control group. In addition, patients whose discharge summaries were manually evaluated were employed as the gold standard test set for model validation. Baseline models like Naive Bayes and Random Forest, as well as a deep learning model based on word embeddings and a convolutional neural network (CNN), were all put to the test. As a result, the gold standard test set consisted of 139 cirrhosis patients and 152 controls, while the training set contained 446 cirrhosis patients and 689 controls. 

The CNN machine learning model outperformed the Naive Bayes and Random Forest models in terms of area under the receiver operating characteristic curve (0.993), precision (0.965), and recall (0.9978). (precision 0.787 and 0.958, and recall 0.878 and 0.827). Recall for cirrhosis ICD codes was 0.978, and accuracy was 0.883. Investigators concluded that a convolutional neural network (CNN) model trained on discharge summaries accurately recognized cirrhosis patients. The potential improvement in disease burden assessment across studies using this method for cirrhosis phenotyping in the EHR is substantial.

Source: journals.lww.com/jcge/Abstract/2023/01000/An_AI_Approach_for_Identifying_Patients_With.10.aspx