Artificial Intelligence Shows Promise For Identifying Undiagnosed NASH

An artificial intelligence algorithm may prove useful for detecting missed cases of non-alcoholic steatohepatitis, researchers have found.

The algorithm employs the principles of “deep learning” to recognize clinical characteristics common to patients with NASH. The initial results are encouraging, but the software is expected to become increasingly smart, improving both specificity and sensitivity, as it incorporates mo re data, said Kathryn Starzyk, MSc, the senior director of real-world evidence at OM1 Inc., the Boston-based firm that is developing the system.

The gold standard for a diagnosis of NASH is a liver biopsy, but Ms. Starzyk outlined practical limitations to this approach at the 2018 Digestive Disease Week (abstract 356).

“Liver biopsy carries risk, and it is contraindicated for some patients,” Ms. Starzyk said. Due to the high and growing prevalence of NASH—more than doubling over the past 20 years in the United States (Am J Gastroenterol 2017;112[4]:581-587)—the cost of biopsying every suspected case can also sap health care resources.

The new tool involves “a combination of sophisticated AI algorithms designed to automatically learn the characteristics of confirmed NASH cases,” Ms. Starzyk said. When applied to a learning set of patients with or without NASH, the receiver operating characteristic—a method measuring the ability of a binary classifier to discern true from false cases—proved to be “very powerful,” she added.

Validation studies in populations in which the prevalence of NASH is known to be high, such as obese people or those with diabetes, also have been encouraging. When the AI tool was applied to a database with more than 40 million U.S. individuals, the number of patients deemed to have a high likelihood of NASH was in the hundreds of thousands, Ms. Starzyk reported.

The AI algorithm could be adapted for other diseases, particularly those that are also commonly overlooked or poorly coded, she said. If therapies improve for NASH, particularly interventions that halt progression at an early stage, the need for strategies to identify asymptomatic patients will be acute.

Other groups are also pursuing a deep learning approach to detection of liver disease without biopsy. A multinational team of investigators led by Jasjit S. Suri, PhD, of Global Biomedical Technologies in Roseville, Calif., is evaluating ultrasound images with an artificial neural network.

A variation on machine learning, artificial neural networks have been compared to a biological brain that can process disparate sources of information in a nonlinear fashion to achieve a result, such as whether a CT scan is positive or negative for liver disease. Instead of being programmed to recognize characteristics, computers employing machine learning build their own knowledge. In the case of liver CT scans, machine learning labels CT scans as negative or positive after evaluating hundreds or thousands of examples to gather complex information from multiple determinants.

In a study that compared a deep learning system that employed a layered neural network and conventional machine learning in 63 patients with or without liver disease, the risk stratification was 100% for the neural network versus 82% for a conventional machine learning protocol, according to a study on which Dr. Suri was the senior author (Comput Methods Programs Biomed 2018;155:165-177).

“Deep learning systems [using neural networks] show a superior performance for liver disease detection and risk stratification compared to conventional machine learning systems,” Dr. Suri reported in his published article. Although results were drawn from a small study, he said the approach has major potential for the detection of fatty liver disease and noninvasive risk stratification.