Premium
Laboratory parameter‐based machine learning model for excluding non‐alcoholic fatty liver disease ( NAFLD ) in the general population
Author(s) -
Yip T. C.F.,
Ma A. J.,
Wong V. W.S.,
Tse Y.K.,
Chan H. L.Y.,
Yuen P.C.,
Wong G. L.H.
Publication year - 2017
Publication title -
alimentary pharmacology and therapeutics
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 3.308
H-Index - 177
eISSN - 1365-2036
pISSN - 0269-2813
DOI - 10.1111/apt.14172
Subject(s) - medicine , population , fatty liver , receiver operating characteristic , logistic regression , nonalcoholic fatty liver disease , gastroenterology , machine learning , artificial intelligence , disease , environmental health , computer science
Summary Background Non‐alcoholic fatty liver disease ( NAFLD ) affects 20%‐40% of the general population in developed countries and is an increasingly important cause of hepatocellular carcinoma. Electronic medical records facilitate large‐scale epidemiological studies, existing NAFLD scores often require clinical and anthropometric parameters that may not be captured in those databases. Aim To develop and validate a laboratory parameter‐based machine learning model to detect NAFLD for the general population. Methods We randomly divided 922 subjects from a population screening study into training and validation groups; NAFLD was diagnosed by proton‐magnetic resonance spectroscopy. On the basis of machine learning from 23 routine clinical and laboratory parameters after elastic net regulation, we evaluated the logistic regression, ridge regression, AdaBoost and decision tree models. The areas under receiver‐operating characteristic curve ( AUROC ) of models in validation group were compared. Results Six predictors including alanine aminotransferase, high‐density lipoprotein cholesterol, triglyceride, haemoglobin A 1c , white blood cell count and the presence of hypertension were selected. The NAFLD ridge score achieved AUROC of 0.87 (95% CI 0.83‐0.90) and 0.88 (0.84‐0.91) in the training and validation groups respectively. Using dual cut‐offs of 0.24 and 0.44, NAFLD ridge score achieved 92% (86%‐96%) sensitivity and 90% (86%‐93%) specificity with corresponding negative and positive predictive values of 96% (91%‐98%) and 69% (59%‐78%), and 87% of overall accuracy among 70% of classifiable subjects in the validation group; 30% of subjects remained indeterminate. Conclusions NAFLD ridge score is a simple and robust reference comparable to existing NAFLD scores to exclude NAFLD patients in epidemiological studies.