Machine learning outperforms noninvasive tests in NAFLD
Click Here to Manage Email Alerts
Machine learning may be a reasonable solution for screening for fatty liver in the general population in areas where more niche noninvasive tests are unavailable, according to a presenter at the Digital International Liver Congress.
“Risk stratification is key to reducing unnecessary testing in liver disease screening. While transient elastography [TE] could be considered the noninvasive gold standard for liver fibrosis diagnosis before biopsy, evidence regarding which features or scoring systems ... help to discriminate at-risk patients in the general population is built upon heterogenous and small studies,” Miquel Serra-Burriel, PhD, a researcher from the Center for Research in Health and Economics, Barcelona, Spain, said in his presentation. “Therefore, we aimed to evaluate which risk factors better predict elevated TE values by means of using supervised learning models.”
Serra-Burriel presented data from six prospective cohorts from the general population who underwent screening for liver fibrosis with TE. In total, 32 variables, including demographics, biometrics and analytical parameters, were used in training a diagonal discriminant analysis model. TE values above 9.1 kPa were considered positive cases for significant liver fibrosis ( F2).
The analysis included 6,199 patients (mean age 54.7 years, 52% women), with 5,856 negative and 343 positive cases.
The final machine learning (ML) model, with 32 variables, presented an area under the curve of 0.88 (95% CI, 0.87-0.9). Analyzing individual variables, Serra-Burriel said that the AUC for abdominal perimeter was 0.71 (0.68-0.74), aspartate aminotransferase 0.73 (0.7-0.76) and albumin 0.68 (0.64-0.71). The ML model had higher accuracy than the FIB-4 score or the NAFLD fibrosis score (NFS) in this population (AUC 0.88 [0.86-0.9] for ML, 0.68 [0.65-0.71] for FIB-4, 0.71 [0.68-0.74] for NFS).
“We see that we are outperforming these algorithms by a magnitude of between 20% to 80%, so it is quite significant,” he said.
If a different ML model was constructed using a cutoff value of TE of 14 kPa, suggestive of cirrhosis, the accuracy of the model was 0.95 (0.94-0.97), compared with 0.79 (0.75-0.83) and 0.78 (0.74-0.82) for FIB-4 and NFS, respectively.
“All in all, the ability to predict pretest individual risk of liver fibrosis may prove useful as a means of reducing unnecessary testing. Our results suggest that pretest risk stratification is feasible and very accurate with routine check-up data,” Serra-Burriel said. “Abdominal perimeter alone might have the same predictive value than common liver function tests in general populations for the assessment of liver fibrosis. That is especially useful in settings where these tests are not performed.”