Clinical Model for NASH and Advanced Fibrosis in Adult Patients With Diabetes and NAFLD: Guidelines for Referral in NAFLD

Nonalcoholic fatty liver disease (NAFLD) is the most common cause of chronic liver disease in the U.S. (1–3). Approximately 10–22% of patients with NAFLD have the progressive subtype of NAFLD termed nonalcoholic steatohepatitis (NASH), which can result in cirrhosis, hepatocellular carcinoma, and liver-related mortality (4–9). Type 2 diabetes is considered a major risk factor for advanced liver disease in patients with NAFLD (10–12).

It is estimated that ∼25.8 million Americans are afflicted with diabetes (13–15). Several studies have shown that prevalence of NAFLD in patients with diabetes is increased compared with those without diabetes (16–21). Although the exact prevalence of NAFLD in patients with diabetes is not known, previous studies suggest that it ranges between 49.6 and 74% (17,21–25). Therefore, it can be estimated that ∼13–18 million people in the U.S. have coexisting diabetes and NAFLD. This study aimed to address the clinical risk stratification that may be applied to this group of patients.

The presence of diabetes has consistently been shown to be a key predictor of NASH and advanced fibrosis in NAFLD (11,26,27). Several experts have recommended liver biopsy in select NAFLD patients with diabetes (28–32). In this era of accountable health care and increasing cost constraints, it is not feasible to recommend liver biopsy in all patients who have diabetes and NAFLD. For that reason, we aimed to determine the most reliable factors that are associated with NASH or advanced fibrosis in patients with diabetes and NAFLD in order to identify patients who should be prioritized for a liver biopsy and/or referred to a hepatologist for further evaluation.

This is a cross-sectional analysis of prospectively evaluated adult patients with biopsy-proven NAFLD who were enrolled into the NAFLD Database Study, a prospective cohort study, conducted by the National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK)-sponsored Nonalcoholic Steatohepatitis Clinical Research Network (NASH CRN) (33,34). The NASH CRN studies enrolled 1,266 patients with biopsy-proven NAFLD; patients enrolled through June 2012 were included in this analysis (33,34). Participants were enrolled by one of the eight participating medical centers in the U.S.: University of California, San Diego (La Jolla, CA); Duke University (Durham, NC); Case Western Reserve University/Cleveland Clinic Foundation (Cleveland, OH); Indiana University (Indianapolis, IN); Saint Louis University (St. Louis, MO); University of California, San Francisco (San Francisco, CA); University of Washington/Virginia Mason Medical Center (Seattle, WA); and Virginia Commonwealth University (Richmond, VA). All participants provided written informed consent prior to data collection. The subjects’ demographic characteristics, anthropomorphic measurements, alcohol consumption, medical history, medication use, clinical tests, and liver biopsy results were prospectively collected; the detailed inclusion and exclusion criteria have previously been described (34,35). Liver biopsies were obtained during the study period as clinically indicated. All protocols, consent forms, and manuals of operations were reviewed by a data safety-monitoring board established by the NIDDK for the NASH CRN and approved by the institutional review board for each site.

Participants had to meet specific criteria regarding the diagnosis of NAFLD in order to be enrolled in this study as previously published (34). Patients with alcohol consumption of >140 g/week if male or >70 g/week if female in the 2 years prior to screening were excluded. Patients with other etiologies of chronic liver disease were also excluded. For the purpose of enrollment into the observational NAFLD Database Study, the diagnosis of NAFLD was based on the histological diagnosis of NAFLD or cryptogenic cirrhosis or on imaging studies (34). For this study, only patients with liver biopsy data available within 6 months of the clinic data were included.

Liver biopsy slides were stained with hematoxylin-eosin, Masson trichrome, and Perls iron stain. The NASH CRN Pathology Committee reviewed and scored the slides without any knowledge of patient’s previous diagnosis, clinical information, or laboratory values or the study for which the biopsy was being evaluated (36,37). The committee used Brunt modified classification to stage fibrosis: 0 = no fibrosis; 1a = mild, zone 3 perisinusoidal fibrosis; 1b = moderate, zone 3 perisinusoidal fibrosis; 1c = portal/periportal only fibrosis; 2 = zone 3 perisinusoidal and periportal fibrosis; 3 = bridging fibrosis; and 4 = cirrhosis (36,38,39). Advanced fibrosis was defined as stages 3–4. Patients were classified as having no NASH, possible/borderline NASH, or definite NASH as previously described (36). For the purposes of this analysis, patients with possible/borderline NASH were grouped with patients with no NASH.

The primary outcome measure was presence of definite NASH as determined by the Pathology Committee’s review of the liver biopsy (36). NASH was defined as presence of steatosis predominantly in zone 3, with varying degrees of lobular inflammation, and classic ballooning degeneration with or without presence of Mallory-Denk bodies and/or peri-sinusoidal fibrosis. The secondary outcome measure was presence of advanced fibrosis on liver biopsy defined as presence of stage 3 or 4 fibrosis on the NASH CRN histologic scoring system (36).

All data analyzed in this study were obtained within 6 months of the liver biopsy. The following variables were included: demographic features (age at enrollment [years], sex, race [white or other], ethnicity [Hispanic/Latino]), clinical data (waist circumference, waist-to-hip ratio, BMI [kg/m²], hypertension), laboratory measures (serum alanine [ALT] and aspartate [AST] aminotransferase, γ-glutamyl transpeptidase [GGT], albumin, international normalized ratio [INR], bilirubin, triglyceride, HDL, fasting serum glucose and insulin levels, and complete blood count), and presence of type 2 diabetes.

Diabetes status was based upon previous history of diabetes, use of medications to treat diabetes, and/or fasting plasma glucose ≥126 mg/dL or a 2-h glucose ≥200 mg/dL during an oral glucose tolerance test or HbA_1c ≥6.5%, consistent with the American Diabetes Association criteria for the diagnosis of diabetes (40).

Group comparisons used the two-tailed t test for continuous dependent variables and Fisher exact test for categorical variables. Multivariate logistic regression models were used to analyze the diagnostic power of clinical and biochemical characteristics to predict the presence of NASH and of advanced fibrosis. We selected the predictor variables for the clinical prediction models for NASH and for advanced fibrosis using stepwise logistic regression models derived from a large candidate set of clinical variables by application of the Akaike information criterion (AIC) to select each predictor. Predictors with higher information content were selected in the models for the probabilities of NASH. Since AIC is defined as −2log likelihood + 2p, were p = number of predictors in a model, models with more predictors are “penalized” by the 2p factor in the AIC. This means that a new predictor must improve the information in the model (AIC) more than enough to overcome the “penalty” assigned for adding a new predictor to the model. This approach results in predictors selected for models on the basis of a standard measure of information content and avoids the difficult multiplicity of comparison problems present when predictor selection uses P values.

We calculated the cross-validated areas under the receiver operating characteristic curves (AUROCs), positive predictive value (PPV), negative predictive value (NPV), the percent correctly classified, sensitivity, and specificity to measure the diagnostic test characteristics of each model. Next, we calculated low-probability cut points for “not NASH” and “not advanced fibrosis” based on the NPV of 90% (or the largest NPV defined) and high probability cut points for “NASH” and “advanced fibrosis” based on the PPV of 90% (or largest PPV defined). The “gray zone” is defined as the probability scores that fall in between the low- and high-probability cut points. Using these cut points, we calculated the number of correctly and incorrectly classified biopsies and the potential for biopsies spared, as well as the number of patient biopsies falling in between the lower and higher cut points (34).

We used a statistical jackknife procedure to obtain an internal but independently validated (cross-validated) and thus more realistic estimate of an AUROC. If n = number of patients, the cross-validated AUROC is obtained by fitting a total of n different models and obtaining n independently predicted probabilities (scores) of NASH or advanced fibrosis, each with a sample size of n − 1 obtained by deleting one patient at a time until the n independent scores have been obtained. These independent scores are used to predict each of the n outcomes and calculate the jackknifed AUROC (41). The jackknife procedure results in independent, and thus more valid, AUROC estimates, since its predictions are of patient results using models fit to data external to the patient being predicted. While a jackknifed AUROC is superior to an ordinary biased AUROC by virtue of predicting outcomes using models derived from these same outcomes, clinical prediction models, to be useful, must be validated in a population that is external to the population used to develop the model. We approximated this using our model for predicting advanced fibrosis in future NASH CRN patients as read by a pathologist serving each clinic rather than the central consensus reading at the histology reading center for the NASH CRN. The model for advanced fibrosis was externally validated using a cohort of NASH CRN patients not included in the primary analysis. This validation cohort consisted of 100 patients from the same studies and time period as used for the model development based upon central review of cases. Finally, we compared our clinical model for advanced fibrosis with a published model for predicting advanced fibrosis by applying this model to our data and comparing the cross-validated AUROCs, PPV, NPV, specificity, and sensitivity.

P values were considered statistically significant if P < 0.05. All analyses were conducted using SAS 9.2 (SAS Institute, Cary, NC) and Stata 13.1 (StataCorp, College Station, TX).

This study included 1,249 patients with biopsy-proven NAFLD; 435 (34.8%) of these patients had diabetes. The average age and BMI of these patients with diabetes and NAFLD was 52.4 ± 10.3 years and 35.8 ± 6.6 kg/m², respectively. Among the 346 patients with diabetes and NAFLD, the prevalence of NASH and advanced fibrosis was 69.2% and 41.0% (Table 1). Detailed baseline characteristics of the patients with diabetes are shown in Table 1.

Table 1 describes the baseline characteristics of patients with diabetes classified by the presence or absence of definite NASH. In univariate models, factors associated with presence of NASH on histology included elevated AST or ALT (P < 0.0001), serum insulin (P = 0.001), HbA_1c (P = 0.006), and HOMA of insulin resistance (HOMA-IR) (P = 0.002).

A multivariable-adjusted model (clinical model) was developed using AIC criteria. In the clinical model (Table 2), the factors associated with presence of NASH on histology included white race, BMI, waist (measured in centimeters), ALT, AST, albumin, HbA_1c, HOMA-IR, and ferritin with a cross-validated AUROC of 0.80 (95% CI 0.75–0.84, P value = 0.007). The specificity, sensitivity, NPVs, and PPVs were 90.0%, 56.8%, 47.7%, and 93.2%, respectively, and this model correctly classified 67% of patients as having NASH (Table 2).

Table 1 describes the baseline characteristics of patients with diabetes stratified by the presence of advanced fibrosis. In univariate models, the factors associated with advanced fibrosis included age (P < 0.0001), hypertension (P = 0.006), elevated AST (P = 0.02), AST-to-ALT ratio (P < 0.0001), GGT (P = 0.0006), globulin (P = 0.01), direct bilirubin (P = 0.01), serum insulin (P = 0.002), HOMA-IR (P = 0.003), LDL (P = 0.05), white blood cell count (P = 0.002), INR (P = 0.0003), and platelet count (P < 0.0001).

A multivariable-adjusted model (clinical model) was developed using AIC criteria. In the clinical model (Table 3), the factors associated with advanced fibrosis included age, Hispanic ethnicity, BMI, waist-to-hip ratio, hypertension, AST-to-ALT ratio, alkaline phosphatase, isolated abnormal alkaline phosphatase, globulin, albumin, bilirubin (total and direct), serum insulin, hematocrit, INR, and platelet count with a cross-validated AUROC of 0.80 (95% CI 0.76–0.85). The specificity, sensitivity, NPV, and PPV were 90%, 57%, 75.1%, and 80.2%, respectively, and this model correctly classified 76.6% of patients as having advanced fibrosis (Table 3).

Table 4 provides the probability of presence of NASH and advanced fibrosis at various cut points. It also shows the cut points that could be used in clinical practice to determine when to consider a biopsy for the diagnosis of NASH; a model parameter of >0.75 would result in a PPV of 90% for the presence of NASH. Similarly, for advanced fibrosis, a cut point >0.85 would result in a PPV of 89.5% for advanced fibrosis.

Internal cross-validation was done and is shown in Table 3 using jackknife procedures (as explained in 2research design and methods). Using an independent validation cohort of 100 patients recruited from the NASH CRN sites as part of the same studies, we showed that the results remained consistently robust with AUROC for NASH and advanced fibrosis in the validation cohort of 0.83 (95% CI 0.75–0.92) and 0.84 (0.76–0.92), respectively (as shown in Table 5).

Finally, we compared the diagnostic accuracy of the current model (developed specifically for patients with diabetes) with the NAFLD Fibrosis Score (as shown in Supplementary Table 1). The models developed for the diabetic population were significantly more accurate than the previously published NAFLD Fibrosis Score applied to this population for the diagnosis of advanced fibrosis, with a cross-validated AUROC of 0.80 vs. 0.76 (P < 0.05).

With a large, well-characterized cohort of patients with biopsy-proven NAFLD and diabetes, we demonstrate that routinely available clinical and biochemical factors can be used to accurately determine the likelihood of NASH (AUROC 0.80, P = 0.007) and advanced fibrosis (AUROC 0.80, P < 0.001) in patients with diabetes and NAFLD. These data can guide clinicians regarding when to refer patients with diabetes who have NAFLD for a liver biopsy. The application of these prediction models accurately classified 67% of our study set with NASH and 77% with advanced fibrosis. The models are clinically stringent and weighted to having high PPVs with the trade-off of lower NPVs. Thus, clinical judgment and further testing, including liver biopsies, may still be needed in patients determined not to be at high risk for NASH or advanced fibrosis using these models but would correctly classify three-quarters of patients with advanced fibrosis.

Prior studies have used similar clinical and laboratory measures to identify patients with NAFLD to predict the presence or absence of advanced fibrosis in NAFLD patients. One example is the Fatty Liver Index, which uses triglyceride level and waist circumference to predict NAFLD (42,43). Other studies of NAFLD patients have demonstrated that the presence of metabolic syndrome and hypertriglyceridemia, higher AST-to-ALT ratio, and lower platelet count are associated with more advanced liver disease (34,44). Clinical prediction rules have also been created to identify NAFLD patients with and without advanced fibrosis. One example is the well-validated NAFLD fibrosis score, which consists of age, BMI, impaired fasting glucose or diabetes, AST-to-ALT ratio, platelet count, and albumin (45,46). Our model performed better than the NAFLD fibrosis score. Unlike prior studies, the current model proposed in this study focuses on patients with diabetes, a population known to have higher risk of NASH, advanced fibrosis, and mortality (10,26,47–50). The models developed in this study can thus help to identify patients with diabetes at high risk for the presence of NASH or advanced fibrosis and help guide clinicians when to refer patients with diabetes for a liver biopsy and appropriate management. Future studies combining the clinical prediction rules with other noninvasive imaging methods (51) need to be performed to further improve the diagnostic accuracy.

The NASH CRN cohort is a multiethnic and multicenter study including eight sites across the U.S. This ethnic and geographic variation is a strength that may allow the results to be applied to other NAFLD patients in the U.S. Additionally, the NASH CRN cohort includes prospective cohort data. The histology was also subject to blinded analysis by a committee of expert pathologists who used the accepted and validated NASH CRN histology scoring system. A series of validation procedures were used to confirm the reproducibility of findings. We performed jackknife internal cross-validation of AUROC, as it is superior to an ordinary biased AUROC. In order for the models to be generalizable, they must be validated in a population that is external to the population used to develop the model. We approximated this using our model for predicting advanced fibrosis in future NASH CRN patients as read by a pathologist serving each clinic rather than the central consensus reading at the histology reading center for the NASH CRN. The model for advanced fibrosis was externally validated using a cohort of NASH CRN patients not included in the primary analysis. This validation cohort consisted of 100 patients from the same studies and time period as used for the model development based upon central review of cases. The results remained statistically significant and robust and consistent with the model development cohort. One limitation of the study is the recruitment of patients from tertiary care centers. The associations between diabetes-related phenotypes—diabetes or diabetes-related treatments, such as, but not exclusively, insulin, metformin, sulfonylureas, or statins and fibrates and the duration of diabetes—and their possible effects on the development of NASH and advanced fibrosis were not examined. This study population may not represent the spectrum of patients in the general population seen in primary care. Further studies would be needed to externally validate these results.

This study may help to guide further research on potential relationships between NASH and diabetes. It may help identify high-risk patients and target interventions in order to prevent progression of NASH to cirrhosis and hepatocellular carcinoma, especially in populations with diabetes.

Currently, NAFLD Practice Guidelines recommend the use of the NAFLD Fibrosis Score for the screening for advanced fibrosis. In this report, we demonstrate that the proposed model is better than NAFLD Fibrosis Score in assessing advanced fibrosis in patients with diabetes. If validated in an independent cohort, the current model will replace the NAFLD fibrosis score in the NAFLD Practice Guidelines in future. These data could also be used to screen patients with diabetes who should be screened for NASH prior to enrollment in a clinical trial. This is emerging to be an important unmet need, and these findings provide a clinically useful tool that can be applied directly in clinical practice using routinely available data.

Using a large, diverse cohort of patients with biopsy-proven NAFLD and diabetes, we developed a clinical prediction guide to identify patients with diabetes at risk for having NASH using readily available clinical data such as BMI, presence of hypertension, and routine laboratory values. This guide could potentially impact an estimated 10 million people residing in the U.S. who have coexisting diabetes and NASH by allowing for early identification of high-risk patients. These models may help inform the decision as to who should be considered for liver biopsy and/or referred to a hepatologist for further evaluation of NAFLD. Further studies using additional biomarkers are needed to improve the clinical models and to better understand the pathogenesis of NASH and its relationship with diabetes.

Funding. This work was supported in part by the American Gastroenterological Association (AGA) Foundation–Sucampo–Association of Specialty Professors (ASP) Designated Research Award in Geriatric Gastroenterology and by a T. Franklin Williams Scholarship Award. Funding was also provided by The Atlantic Philanthropies, the John A. Hartford Foundation, the ASP, the AGA (to R.L.), and the NIDDK, National Institutes of Health (grant K23-DK-090303). This research was funded in part with the support of the University of California, San Diego, Digestive Diseases Research Development Center, Public Health Service grant DK-080506. NASH CRN is supported by the NIDDK (grants U01-DK-061718, U01-DK-061728, U01-DK-061731, U01-DK-061732, U01-DK-061734, U01-DK-061737, U01-DK-061738, U01-DK-061730, and U01-DK-061713), by the Eunice Kennedy Shriver National Institute of Child Health and Human Development, and by the National Institutes of Health.

Duality of Interest. No potential conflicts of interest relevant to this article were reported.

Author Contributions. J.B. and R.L. drafted the manuscript. M.D., L.W., J.T., and R.L. analyzed and interpreted data. M.D., B.A.N.-T., D.K., E.M.B., L.W., E.D., J.L., and J.T. critically revised the manuscript for important intellectual content. M.D., L.W., J.T., and R.L. performed statistical analysis. R.L. conceived the study concept and design, acquired data, provided technical or material support, and supervised the study. R.L. is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.