On the operational definition of fatty liver

We discuss the present operational definition of fatty liver and its shortcomings.

Although liver biopsy is the reference method for the diagnosis of fatty liver, it is an imperfect gold-standard because of sampling error. A needle biopsy is in fact approximately 1/20000 th of the weight of the liver [2] and even if the most common intrahepatic fat pattern is the diffuse form, there are also heterogeneous, focal, multinodular, perilesional, perivascular, subcapsular, and lobar forms [3] .
In addition, liver biopsy is an invasive procedure and cannot be employed outside tertiary care centers and even in these centers it is nearly always performed on selected patients [4] . Importantly, liver biopsy is also the reference method for diagnosis of both intrahepatic inflammation and liver fibrosis, as we shall see below.
Proton magnetic resonance spectroscopy or quantitative fat/water selective magnetic resonance imaging (MRI) can detect an intrahepatic triglyceride content > 5% (corresponding to a proton density fat fraction > 5.6%) and could thus be used to diagnose fatty liver [1] . However, proton magnetic resonance spectroscopy is available only in few centers and mostly for research purposes. In addition, there is not a linear association between 5% fatty hepatocytes and 5% of whole liver weight fat content [5] .
Liver ultrasonography is the imaging method most commonly employed to diagnose fatty liver in the general population and in clinical practice [1] . When compared with liver biopsy, liver ultrasonography has a true positive fraction of 85% (95% confidence interval 80 to 89%), a true negative fraction of 94% (88 to 97%), a positive diagnostic likelihood ratio of 13.3 (6.4 to 27.6), and a negative diagnostic likelihood ratio of 0. 16 (0.12 to 0.22) for the detection of an intrahepatic triglyceride content > 33% [6] .
A further option to study fatty liver in the general population is the use of surrogate markers [1] . One of such markers, the fatty liver index (FLI), which we developed against liver ultrasonography in the general population of the Dionysos Nutrition & Liver Study, is based on four common measurements, i.e. body mass index, waist circumference, gamma-glutamyl-transferase and triglycerides [7] . FLI has gained increasing attention in the last decade because of its association with prevalent and incident cardio-metabolic disease. More importantly for its ability to serve as surrogate marker of fatty liver, FLI has been successfully cross-validated in external populations [8] . Although the cut-points of FLI proposed by us to rule in and rule out fatty liver in the general population of Campogalliano (Modena, Italy) are commonly employed in the literature [7] , it is much better to recalibrate FLI in (a subsample of) the population of interest, as we advised in the original paper and as it was recently done by some researchers in the Shanxi Province of China [9][10] .
Liver biopsy remains the only method able to distinguish intrahepatic inflammation from fibrosis whereas liver elastometry and multiparametric magnetic resonance provide an indirect and combined measure of the two parameters, represented by two distinct vectors that are not yet distinguishable one from the other in both US and MRI-based imaging techniques [11] [12][13] [14] .
On the dichotomization of fatty liver into alcoholic and nonalcoholic fatty liver Fatty liver is presently classified into non-alcoholic fatty liver disease (NAFLD) and alcoholic fatty liver disease (AFLD). T he European Association for the Study of the Liver (EASL) presently suggests that, after (at least) hepatitis B infection, hepatitis C infection and steatogenic drugs have been ruled out, NAFLD should be diagnosed when ethanol intake is ≤ 20 g/day in women and ≤ 30 g/day in men ( T he NAFLD vs. AFLD dichotomization has several limitations.
Besides the loss of information inherent to any dichotomization [15] , its most important criticism is that a simple cut-off of low daily alcohol intake cannot really distinguish alcohol-and non-alcohol-induced liver damage. T he toxicity from low alcohol intake is in fact determined by genetic, alimentary and lifestyle factors and some individuals may well have alcohol-induced fatty liver while drinking alcohol below the proposed cut-point [16] .
On the other hand, the gut microbiota may produce alcohol contributing to liver damage [17] .
Another problem of the NAFLD vs. AFLD dichotomization is that it requires the use of an instrument accurate enough to detect small differences in ethanol intake [18] . Even the 7day weighed food record that we employed in the Dionysos Nutrition & Liver study may not be accurate enough to detect small differences in ethanol intake [19] . T o gain some understanding of the problems caused by dichotomization and/or measurement error,   At the time of the Dionysos Nutrition & Liver Study, the cut-point of ethanol intake suggested by EASL to separate NAFLD from AFLD was 20 g/day for both men and women. If we adopt the current EASL cut-point of 30 g/day, the prevalence of NAFLD in the men of the Dionysos Nutrition & Liver Study would be substantially higher. However, the (lack of) association between fatty liver and continuous ethanol intake that we have reported in the same population would obviously not change [7] . T his alerts to a clear problem with dichotomization [15] .
On the diagnosis of non-alcoholic steatohepatitis NAFLD is presently classified into uncomplicated non-alcoholic fatty liver (NAFL) and non-alcoholic steatohepatitis (NASH) [1] (Fig ure 3  Fig ure 3).  NASH is presently defined as the presence of steatosis, inflammation and ballooning at liver biopsy in a patient with NAFLD [1] . Liver biopsy cannot distinguish alcohol-from nonalcohol-induced steatohepatitis and the current separation of NASH and alcoholic steatohepatitis is based on the dichotomization of alcohol intake with the limitations described above. We refer the reader to a recent review for details and controversies on the histopathology of NASH [2] .
Here, we are interested only in the methodological consequences of the fact that liver biopsy is central to the diagnosis of steatohepatitis, independently of its etiology. T his requirement implies that NASH cannot be diagnosed in the general population. Even if accurate indirect markers of NASH were available, which are not [1] , they would have been developed in tertiary care centers where liver biopsy can be performed, most often on selected subsets of patients [4] . Accordingly, their diagnostic performance is likely to be completely different in the general population because of the different case-mix of individuals [20] . T hus, the available estimates of the prevalence of NASH in the general population should be taken with caution, especially when they are based on indirect and demonstrably unreliable markers of NASH such as altered liver enzymes in the presence On the diagnosis of liver fibrosis T he clinical relevance of NASH stems from its potential association with liver fibrosis [1] . centers [23] . Abbreviations: NAFLD = non-alcoholic fatty liver disease; NAFL = non-alcoholic fatty liver; NASH = nonalcoholic steatohepatitis.
As stated above, only liver biopsy can distinguish inflammation from fibrosis and provide an accurate grading and staging of these two histologic features. Many surrogate markers and an increasing number of elastographic techniques are nonetheless available to provide a non-invasive measure of the combined vectors of inflammation and fibrosis without being able to discriminate among them [13][24] [25] . As for NASH, these methods were calibrated against liver biopsy in tertiary care centers and most often on selected sets of patients so that their diagnostic performance is likely to be completely different in the general population because of the different case-mix of individuals [20] .
T hus, the available estimates of the prevalence of NASH-associated liver fibrosis in the general population should be taken with caution [21][22] .
Instead of the common practice of dichotomizing fibrosis, much can be gained by treating it as an ordinal outcome and by developing predictive models giving cumulative probabilities [26] . As   ordinal generalized logistic regression model (see reference [27] for details). According to this model, a patient with a liver stiffness of 12 kPa (red line) would have a 2% (95%CI 0 to 5%) probability of F0-F1, 85% (72 to 99%) probability of F2-F3, and 13% probability (1 to 25%) of F4 fibrosis. Abbreviations: Prob. = probability.

Conclusion
We have described some limitations of the current operational definition of fatty liver and NAFLD. T hey can be summarized as follows: 1. Fatty liver is presently defined as an intrahepatic triglyceride content > 5% but the most commonly most used diagnostic technique, i.e. liver ultrasonography, cannot reliably detect values of fatty liver < 10% and is much better at detecting values > 33%.
T his is likely to cause an underestimation of fatty liver; 2. Instead of dichotomizing fatty liver into NALFD and AFLD, it would be more useful if epidemiological studies could evaluate the association between fatty liver and continuous or ordinal ethanol intake with the possibility of studying its independent contribution and interaction with other factors, e.g. body mass index; 3. T he present estimates of the burden of NASH and its complications (fibrosis, cirrhosis, hepatocarcinoma) in the general populations are based on indirect assumptions and may not be reliable. T he real burden of NASH and NASH-associated fibrosis in the general population is likely to remain unknown.