Deep-Learning Tool Helps Identify Patients with Cirrhosis

Researchers from the Medical University of South Carolina (MUSC) have developed a deep-learning (DL) approach that can identify patients with cirrhosis using clinical text in EHRs, according to a study published in the Journal of Clinical Gastroenterology.

Cirrhosis, or severe scarring of the liver, is a serious and potentially life-threatening condition that results from various liver conditions, such as disease or injury, according to the Mayo Clinic. Cirrhosis is the final stage of chronic liver disease. Centers for Disease Control and Prevention data indicate that the condition is the ninth leading cause of death in the US.

Predicting which patients will progress to cirrhosis can be a challenge but is crucial to facilitate care management and improve outcomes, the study authors explained. Currently, predicting cirrhosis progression relies on International Classification of Diseases (ICD) codes, with limited success.

“If we could identify a patient predicted to have cirrhosis a year or two earlier, that would give clinicians time to treat and prevent scar tissue from accumulating in the liver but also to prevent complications of cirrhosis,” said Don Rockey, MD, director of the Digestive Disease Research Core Center at MUSC, in a press release discussing the study.

The researchers turned to artificial intelligence (AI) to help, leveraging machine-learning models to evaluate whether such approaches could accurately identify patients with cirrhosis. The research team tested a Naive Bayes and a Random Forest approach as baseline models and a DL-based convolutional neural network (CNN) to see which would perform best.

Models were trained using discharge summaries from known cirrhosis patients gathered from a patient registry, alongside a group of random controls without cirrhosis or its complications, based on relevant ICD codes and manually reviewed by clinicians. The training set comprised 446 cirrhosis patients and 689 controls, and the test set contained 139 cirrhosis patients and 152 controls.

Overall, the CNN achieved the highest performance of any model, with an area under the receiver operating characteristic curve of 0.993 and a precision of 0.965. The Naive Bayes model reached an area under the receiver operating characteristic curve of 0.879 and a precision of 0.787, compared with the Random Forest model, which achieved 0.981 and 0.958, respectively.

The CNN’s performance can be explained, in part, by how the model mimics neurons in the brain to analyze information.

“Neural networks give more power than a typical statistical AI model because each of these artificial neurons behaves like a model on its own, allowing you to come up with a more sophisticated way of distinguishing and predicting things,” said Jihad Obeid, MD, a professor in biomedical informatics at MUSC and co-first author on the study, in the press release.

The study’s findings indicate that similar AI-based approaches can help provide clinical decision support for providers assessing disease burden and progression.

“I think it's exciting that [the CNN] was successful at identifying cirrhosis using just the text in the discharge summaries, as is the idea of taking it to the next level to see if we can apply it for earlier identification,” Obeid concluded.

Rockey echoed this, noting that the findings showcase how AI can advance chronic disease management.

“I think AI is going to be used to predict how severe the disease will become, in combination with genetics and imaging data,” he said. “I don't know how long that's going to take to figure it out, but I can't imagine that it's not going to happen.”