Non-Alcoholic Fatty Liver Disease (NAFLD) is a progressive liver disease that affects up to 30% of worldwide population, of which up to 25% progress to Non-Alcoholic SteatoHepatitis (NASH), a severe form of the disease that involves inflammation and predisposes the patient to liver cirrhosis. Despite its epidemic proportions, there is no reliable diagnostics that generalizes to global patient population for distinguishing NASH from NAFLD. We performed a comprehensive multicohort analysis of publicly available transcriptome data of liver biopsies from Healthy Controls (HC), NAFLD and NASH patients. Altogether we analyzed 812 samples from 12 different datasets across 7 countries, encompassing real world patient heterogeneity. We used 7 datasets for discovery and 5 datasets were held-out for independent validation. Altogether we identified 130 genes significantly differentially expressed in NASH versus a mixed group of NAFLD and HC. We show that our signature is not driven by one particular group (NAFLD or HC) and reflects true biological signal. Using a forward search we were able to downselect to a parsimonious set of 19 mRNA signature with mean AUROC of 0.98 in discovery and 0.79 in independent validation. Methods for consistent diagnosis of NASH relative to NAFLD are urgently needed. We showed that gene expression data combined with advanced statistical methodology holds the potential to serve basis for development of such diagnostic tests for the unmet clinical need.