Novel machine learning model effectively predicts NASH
Click Here to Manage Email Alerts
BOSTON — A machine learning approach demonstrated a relatively high sensitivity rate for predicting the probability of non-alcoholic steatohepatitis in at-risk patients, according to data presented at The Liver Meeting 2019.
“I have been involved in NASH research for the last 15 years and one of the striking findings is the low awareness of the disease; both from patients, but also from a primary care physician’s perspective,” Jörn M. Schattenberg, MD, of the University Medical Center in Mainz, Germany, said during a press conference.
Schattenberg noted that the researchers asked whether novel machine learning tools could help identify patients at risk for NASH.
“That’s relevant because patients with NASH will progress and eventually ... develop end-stage liver disease with occurring complications,” he said.
Schattenberg and colleagues conducted an exploratory analysis, feature extraction, model training and parameter tuning on the NAFLD Adult Database from the National Institute of Diabetes, Digestive Diseases (NIDDK). The database consisted of 422 patients with histologic NASH and 282 patients confirmed to not have NASH.
The researchers then tested the best-performing model from NIDDK on the Optum HER database to understand model performance.
Data from 1,016 patients with NASH confirmed by liver biopsy within the Optum HER database were then used to evaluate model performance.
The model, known as NASHmap, includes 14 variables deemed the most important features for predicting NASH.
The model then ranked the features, as follows, in order of importance: HbA1c, AST, ALT, total protein, AST/ALT, BMI, triglycerides, height, platelets, WBC, hematocrit, albumin, hypertension and gender.
A simplified version of the model was also developed to include five features: HbA1c, AST, ALT, total protein and triglycerides.
The 14-feature model correctly identified 81% of patients (AUC = 0.82) with NASH in the NIDDK database. The simplified five-feature model correctly identified 76% of patients (AUC = 0.8) with NASH in that database.
In the Optum HER database, the 14-feature model identified 72% of patients with NASH (AUC = 0.76), while the five-feature model identified 66% of patients with NASH (AUC = 0.74).
“While a lot of talk and a lot of focus is on the degree of fibrosis as being the most important predictor of outcomes in these patients, we are not doing a good job at this time to identify at-risk populations and actually propose screening to them,” Schattenberg said. “The way I see the algorithm ... in the end working, is you do apply this to a large dataset that is available and ... you use every day clinical parameters that the doctor has to identify at-risk population ... to then order a specific test and refine the at-risk population.”
Schattenberg stressed that it’s not considered a diagnostic test, but rather should be used complementary to other tests.
“This will allow us to use resources wisely instead of testing every diabetic patient, for example, that is out there for advanced fibrosis,” he said. “So, while HbA1c is one of the parameters that’s important in the algorithm, the algorithm refines so finely that it will give us a subgroup and allow us a more sophisticated risk stratification.” – by Ryan McDonald
Reference: Schattenberg J, et al. Abstract 190. Presented at: The Liver Meeting; Nov. 7-12, 2019; Boston.
Disclosure: Schattenberg reports serving as a consultant for AbbVie, Boehringer Ingelheim, Bristol-Myers Squibb, Genfit, Gilead, Intercept Pharmaceuticals, MSD and Novartis.
Perspective
Back to Top
Arthur McCullough, MD
While many doctors feel uncomfortable with artificial intelligence, it is coming. It’s going to be used in both the academic and clinical setting and people have to get comfortable with it.
This is an important study performed by excellent investigators because it attempts to distinguish NASH from non-NASH.
NASH has the potential to progress to cirrhosis, which is why it’s critical to diagnose which patients have just steatosis vs. NASH.
It is also important to diagnose the presence and degree of fibrosis in NAFLD, but there are already many models and radiologic techniques that do that. However, there are no models, with the exception of perhaps our study in 2016, that have investigated methods to differentiate NASH from non-NASH.
This study addresses that crucial issue. The researchers took a training set using supervised machine learning, which means they used a machine to differentiate patients who have biopsy-proven NAFLD or steatosis vs. biopsy-proven NASH.
This study used well characterized patients with NAFLD from the NIH funded clinical research network as the training set with data from the Optum health record data base as the validation set.
One concern is the accuracy of the histologic diagnosis in the Optum validation set. In the training set, the biopsies were performed and reviewed by expert pathologists who were trained on how to read NASH. In the Optum health record database, there were over 1,000 biopsies but that database included patients from 150,000 providers, and I’m sure that different pathologists were reading the biopsy slides, possibly in different ways.
A potential point of interest that wasn’t addressed in this study was to perform non-supervised or lightly supervised machine learning. Instead of telling the computer to compare two preselected groups – steatosis or NASH – have the machine evaluate one group, all NAFLD. Additionally, let the machine look for differences and commonalities to determine potential additional factors that an investigator might not have previously considered. For example, there may differences in the EKG or mean cellular volume on a CBC that separate NASH from non-NASH – items that we don’t usually think of in terms of separating the two groups.
We’re all learning about the potential benefits and limitations of AI, including myself. Expert computer and statistical analysts are needed to perform AI studies and results need to be replicated by multiple sites and data sets. We also should compare results obtained using AI with results obtained using non-artificial intelligence to see if there is any benefit of AI in the item being studied. For example, the data in the current study did not provide any additional accuracy using area under the curve compared with our data using non-AI technology. So, we always should ask the question, what is the added value to a test; even when performed with AI?
Read more about
Click Here to Manage Email Alerts