Dallas Steatosis Index Identifies Patients With Nonalcoholic Fatty Liver Disease

Abstract

Background & Aims:

Tools have been developed to determine risk for nonalcoholic fatty liver disease (NAFLD) based on imaging, which does not always detect early-grade hepatic steatosis. We aimed to develop a tool to identify patients with NAFLD using ¹H MR spectroscopy (MRS).

Methods:

We collected data from the Dallas Heart Study—a multi-ethnic, population-based, probability study of adults (18–65 y) that comprised an in-home medical survey; collection of fasting blood samples; MRS images to measure cardiac mass/function, abdominal subcutaneous/visceral adiposity; and quantification of hepatic triglyceride concentration, from 2000 through 2009. NAFLD were defined as 5.5% or more liver fat and we excluded patients with more than moderate alcohol use; 737 patients were included in the final analysis. We performed binary multivariable logistic regression analysis to develop a tool to identify patients with NAFLD and evaluate interactions among variables. We performed an internal validation analysis using 10-fold cross validation.

Results:

We developed the Dallas Steatosis Index (DSI) to identify patients with NAFLD based on level of alanine aminotransferase, body mass index, age, sex, levels of triglycerides and glucose, diabetes, hypertension, and ethnicity. The DSI discriminated between patients with vs without NAFLD with a C-statistic of 0.824. The DSI outperformed 4 risk analysis tools, based on net reclassification improvement and decision curve analysis.

Conclusions:

We developed an index, called the DSI, which accurately identifies patients with NAFLD based on MRS data. The DSI requires external validation, but might be used in development NAFLD screening programs, in monitoring progression of hepatic steatosis, and in epidemiology studies.

Introduction:

Nonalcoholic fatty liver disease (NAFLD) affects one-quarter of the world population². As the hepatic manifestation of the metabolic syndrome, NAFLD patients are at risk for progression to liver cirrhosis and hepatocellular carcinoma (HCC)^3,4. The growth of the obesity epidemic is expected to compound this disease burden, and NAFLD is projected to become both the leading cause of HCC and the leading indication for liver transplantation in the United States^5,6.

Independent of these liver-related outcomes, the overall clinical course and survival of NAFLD patients is driven by how they are at an increased risk for cardiometabolic diseases such as diabetes mellitus, dyslipidemia and hypertension⁷; furthermore, those initially without manifestations of the metabolic syndrome are at an increased risk of developing diabetes mellitus and cardiovascular events^8,9. Longitudinal observational studies have demonstrated that resolution of hepatic steatosis reduces these cardiometabolic risks^9,10 and weight gain worsens the rate of progressive hepatic fibrosis¹¹, highlighting the role that early intervention could play to prevent these highly morbid, mortal and cost-burdensome outcomes^12–14.

Despite this overwhelming disease burden, neither the American nor European Associations for the Study of Liver Disease recommend screening for NAFLD in the general population^15,16. This discrepancy is based on the as-yet unclear cost-benefit analysis and risk of over-testing that a screening program could provoke. This lack of a screening recommendation may contribute to the underappreciation of NAFLD in medical records and low recognition by primary care providers^17,18.

A multi-tiered screening program where patients are first risk stratified by a clinical prediction tool and then offered imaging to clarify hepatic steatosis and fibrosis would likely improve the cost-effectiveness and therefore inform screening recommendations. Unfortunately, the previously published risk models to determine NAFLD are limited^19–22. These studies used routine liver imaging as their gold standard to diagnose hepatic steatosis, which is only reliably detected if there is ≥30% hepatic steatosis on liver biopsy^23,24. The expected consequence is that these models would miss-classify patients with earlier grades of hepatic steatosis. Other limitations of these previously published tools include the underappreciation of the ethnic/racial differences in NAFLD prevalence²⁵ or using predictors that are not reliably measured in the primary care setting.

We aimed to address these limitations by creating a novel clinical prediction tool that diagnoses patients with NAFLD derived from the population-based, multiethnic, probability sample from the Dallas Heart Study (DHS) that identified hepatic steatosis by ¹H MR spectroscopy (MRS)²⁶. We present here the Dallas Steatosis Index (DSI).

Methods:

Ethical Statement:

The protocol and conduct of this study conformed to the 1975 Declaration of Helsinki and the Health Insurance Portability and Accountability Act. Institutional review board approval was granted through the Washington University School of Medicine in St. Louis (#201905162).

Patient Selection:

We undertook secondary analysis of the DHS: a prospective, multiethnic cohort of subjects aged 18–65 years from Dallas County, Texas. Participants were sampled based on population-based probability in 2000 to 2002 and in 2007–2009. The study conducted a comprehensive in-home medical survey, collected fasting blood and obtained MR images to measure cardiac mass/function, abdominal subcutaneous/visceral adiposity and hepatic triglyceride concentration.

Among the 3,072 participants, we excluded subjects who had not undergone MR fat quantification (n = 737). The baseline characteristics of the group without fat quantification were similar except for a slightly higher BMI (mean 31 versus 29 kg/m2) and lower age (43 vs 45 years). We also excluded those that had more than moderate amounts of alcohol intake (n = 196). Greater than moderate alcohol intake was defined by a daily ethanol intake greater than 30 grams for males and 20 grams for females averaged over one week. Unfortunately, since viral serologies were not measured as part of the DHS, this could not be used as an exclusion criterion. The analytic cohort includes 2,139 subjects.

Defining NAFLD using Liver Imaging and Fat Quantification:

The liver fat quantification was measured by MRS and used a 1.5 Tesla Gyroscan INTERA whole body system (Philips Medical System; see eSupplementary Methods). We defined NAFLD as a mean liver fat ≥5.5%. This well-established threshold corresponds to the 95th percentile of metabolically normal, lean subjects in the initial analysis of the DHS²⁶.

Candidate Predictors

We identified candidate predictor variables based on literature review of known risk factors of NAFLD that are routinely measured in the primary care setting (see Supplementary Methods for more details). Missing data in each candidate predictor was <3.5% and complete data was present for 93% of subjects. We used multiple imputations (5 times) and report the aggregate values.

Statistical Methods:

DSI Derivation and Internal Validation

Median and interquartile range for continuous variables and frequency and proportions for categorical variables were calculated. To derive the DSI clinical prediction model, binary multivariable logistic regression was performed for the outcome of NAFLD (liver fat ≥5.5%). Predictors and interactions that were statistically significant in univariable analysis were entered in a multivariable logistic regression model (see eSupplementary Methods for more model building details). Internal validation was performed by a 10-fold cross-validation strategy, which was chosen due to the increased accuracy of this methodology compared to the traditional strategy of randomly splitting the sample into training and testing populations¹. All statistical analyses utilized SPSS v25 (Armonk, NY: IBM corp).

The DSI risk thresholds for low-risk, intermediate-risk and high-risk were set at <20%, 20–50% and ≥50% predicted risk of NAFLD, respectively. These DSI risk thresholds were chosen after model derivation to give a sensitivity around 90% at the low-risk threshold to rule out NAFLD and a specificity around 90% at the high-risk threshold to diagnose NALFD.

A sensitivity analysis was performed by excluding patients that already had a diagnosis of diabetes mellitus or were already taking a statin. The goal of this sensitivity analysis was to identify the patients that would most benefit from the early identification of NAFLD since intervention (e.g. more aggressive lipid lowering goals, blood pressure control and weight loss goals) to prevent the development of diabetes mellitus or cardiovascular events would represent a more substantial change in their medical management.

DSI Compared to Previously Published Models for NAFLD

The DSI was then compared to the Framingham Steatosis Index (FSI)¹⁹, Hepatic Steatosis Index (HSI)²⁰, Fatty Liver Index derived from an Italian Cohort (I-FLI)²¹ and the updated Fatty Liver Index derived from the US population (US-FLI)²². They were compared to the DSI by their C-statistic, Net Reclassification Improvement (NRI) and decision curve analysis (see eSupplementary Methods). The standard methodology of recalibrating the other risk models before comparing them to the DSI takes into account the different prevalence of NAFLD in the respective derivation populations.

Results:

Derivation and Internal Validation of the DSI

Of the 2,139 subjects in the DHS (median age was 44 years, median BMI 28 kg/m², 54% female; see Table 1 for baseline characteristics), the prevalence of NAFLD was 31% (n = 661). The final DSI model’s multivariable predictors and associated Odds Ratios for NAFLD are shown in Table 2 with the model iteration history shown in eSupplementary Table 4.

Table 1:

Characteristics of Study Population and Univariable Odds Ratios for Predictors

		NAFLD (N = 661) n (%)	Normal (N = 1478) n (%)	Odds Ratio (95% CI)
Body mass Index (kg/m2)	<25	64 (10)	522 (35)	ref
	25–27.49	84 (13)	297 (20)	2.3 (1.6 – 3.3)
	27.5–34.9	335 (51)	481 (33)	5.7 (4.2 – 7.6)
	35–37.49	49 (7)	74 (5)	5.4 (3.4 – 8.4)
	≥37.5	129 (19)	104 (7)	10.1 (7.0 – 14.6)
Age <50 years	Female	186 (43)	576 (56)	0.6 (0.5 – 0.7)
	Male	245 (57)	455 (44)	ref
Age >50 years	Female	135 (59)	257 (58)	1.1 (0.8 – 1.5)
	Male	95 (41)	190 (43)	ref
Race/Ethnicity	Non-Hispanic White	223 (33)	465 (32)	ref
	Black/African American	235 (36)	781 (53)	0.6 (0.5 – 0.7)
	Hispanic, Asian or Other	203	232 (15)	1.8 (1.4 – 2.3)
Statin use (yes)		68 (10)	112 (8)	1.4 (1.0 – 1.9)
Fish oil use (yes)		59 (9)	106 (7)	1.3 (0.9 – 1.8)
25-Vitamin D (ng/ml)	>30	73 (11)	202 (14)	ref
	15–30	332 (50)	728 (49)	1.3 (0.9 – 1.8)
	<15	255 (39)	547 (37)	1.3 (0.9 – 1.7)
Any Current alcohol drinking		439 (66)	1006 (68)	0.9 (0.8 – 1.1)
Smoking Status	Current	148 (22)	404 (27)	0.8 (0.7 – 0.9)
	Past	129 (20)	254 (17)	1.1 (0.9 – 1.2)
	Never	384 (58)	820 (56)	ref
Diagnosis of Hypertension		259 (39)	421 (29)	1.6 (1.3 – 2.0)
Diagnosis of Diabetes Mellitus		132 (20)	106 (7)	3.2 (2.5 – 4.3)
LDL (>160 mg/dl)		117 (18)	173 (11)	1.6 (1.3 – 2.1)
Triglycerides (>150 mg/dl)		177 (27)	102 (7)	4.9 (3.8 – 6.4)
HDL (<60 mg/dl)		358 (54)	531 (36)	2.1 (1.7 – 2.5)
ALT (IU/L)	<13.5	54 (8)	274 (25)	ref
	13.5–19.49	155 (23)	501 (34)	2.1 (1.5 – 3.0)
	19.49–39.9	333 (51)	497 (34)	4.6 (3.4 – 6.3)
	≥40	119 (18)	106 (7)	7.8 (5.3 – 11.4)

NAFLD, nonalcoholic fatty liver disease; CI, confidence interval; LDL, low-density lipoprotein; HDL high-density lipoprotein; ALT, alanine aminotransferase. The BMI categories were chosen in a data-driven fashion to have a linear-relationship to the logit for NAFLD rather than by World Health Organization convention.

Table 2.

The Dallas Steatosis Index Risk Model for Nonalcoholic Fatty Liver Disease

	Model Coefficient	Odds Ratio (95% Confidence Interval)	P Value
Postmenopausal female (0 if male or age <50 y; 1 if female and age ≥50 y)	0.316	1.4 (1.1–1.8)	.034
Diabetes mellitus diagnosed	2.430	11.4 (4.0–32)	<.001
If not diabetic, glucose	0.019	1.02 (1.01–1.03)	<.001
Hypertension diagnosed	0.288	1.3 (1.1–1.7)	.026
Alanine aminotransferase
<13.5 IU/L	ref	ref	ref
13.5–19.49 IU/L	0.408	1.5 (1.1–2.2)	.034
19.5–39.9 IU/L	1.107	3.0 (2.1–4.3)	<.001
≥40 IU/L	1.515	4.5 (2.9–7.1)	<.001
Body mass index
If not black/African American
<25 kg/m2	Ref	Ref	Ref
25–27.49 kg/m2	0.692	2.0 (1.3–3.1)	.001
27.5–34.9 kg/m2	1.429	4.2 (2.9–6.0)	<.001
35–37.49 kg/m2	1.933	6.9 (3.4–13.9)	<.001
≥37.5 kg/m2	2.643	14.0 (7.9–25)	<.001
If black/African American
<25 kg/m2	ref	ref	ref
25–27.49 kg/m2	−0.163	0.8 (0.5–1.6)	.605
27.5–34.9 kg/m2	0.882	2.4 (1.7–3.5)	<.001
35–37.49 kg/m2	0.759	2.1 (1.1–4.0)	.020
≥37.5 kg/m2	1.806	6.1 (3.8–9.7)	<.001
Hispanic or Asian	0.495	1.6 (1.2–2.2)	.001
Natural log of triglycerides	0.999	2.7 (2.2–3.4)	<.001
Constant	−9.388

The DSI when expressed as a LOGIT equation is:

$${\text{DSI}}^{\text{LOGIT}}=-9.4+0.316(\text{if age}\ge \text{50 and female})+2.4(\text{if known diabetes})+0.02*(\text{equals }0\text{ if diabetic};\text{if not diabetic equals the glucose concentration in mg}/\text{dl})+0.3(\text{if known hypertension})+0.5(\text{if Hispanic/Asian/Other race/ethnicity})+\text{Ln }(\text{triglycerides in mg}/\text{dl})+0.4(\text{if ALT }13.5–19.49\text{ IU/L})+1.1(\text{if ALT }19.5-40\text{ IU/L})+1.5(\text{if ALT}>40\text{ IU/L})+0.7(\text{if not black and BMI }25–27.49\text{ kg}/\text{m}2\text{ and})+1.4(\text{if not black and BMI }27.5–34.9\text{ kg}/\text{m}2)+1.9(\text{if not black and }35–37.49\text{ kg}/\text{m}2)+2.6(\text{if not black and}>37.5\text{ kg}/\text{m}2)-0.2(\text{if black and BMI }25–27.49\text{ kg}/\text{m}2\text{ and})+0.8(\text{if black and BMI }27.5–34.9\text{ kg}/\text{m}2)+0.8(\text{if black and }35–37.49\text{ kg}/\text{m}2)+1.8(\text{if black and}>37.5\text{ kg}/\text{m}2)$$

The DSI model had good discrimination for subjects with NAFLD with a C-statistic of 0.824 in the overall population. When stratified by race/ethnic group, the C-statistic was 0.805 for non-Hispanic whites, 0.765 for black/African American and 0.845 for Asian/Hispanic/Other. The agreement between the observed and predicted probability of NAFLD was excellent with a calibration slope of 0.986, calibration intercept of −0.048 and a non-significant Hosmer and Lemeshow test (p = 0.27). The model explained 37% of the variability for NAFLD by the Nagelkerke R² statistic. After performing 10-fold cross validation, the average optimism of the C-statistic was 0.010 making the optimism corrected C-statistic 0.814.

At a threshold of <20% risk, 45% of subjects would be classified as low-risk and NAFLD could be excluded with 86% sensitivity and 90% negative predictive value. At a threshold of ≥50% risk, 23% of subjects would be classified as high-risk and NAFLD could be diagnosed with 90% specificity and 69% positive predictive value. The operating characteristics to diagnose or exclude NAFLD at other risk thresholds is shown in Table 3.

Table 3.

Dallas Steatosis Index Risk Thresholds to Detect Nonalcoholic Fatty Liver Disease

DSI Risk Threshold		Subjects Above Risk Threshold (N = 2139), n (%)	NAFLD (N = 661)	Normal Liver (N = 1478)	Sensitivity (%)	Specificity (%)	PPV (%)	NPV (%)	Net Benefit (Screen At Risk Threshold)	Net Benefit (Screen All)
Logit	Probability
≥ −2.2	>10%	1565 (73)	625	940	95	36	40	94	0.243	0.232
≥ −1.4	>20%	1181 (55)	569	612	86	59	48	90	0.194	0.136
≥ −0.8	>30%	913 (43)	503	410	76	72	55	87	0.153	0.013
≥−0.4	>40%	671 (31)	421	250	63	83	63	84	0.119	−0.152
≥ 0	>50%	491 (23)	339	152	51	90	69	81	0.087	−0.382
≥ 0.4	>60%	355 (17)	266	89	40	94	75	78	0.062	−0.727
≥ 0.8	>70%	209 (10)	163	46	25	97	78	74	0.026	−1.303
≥ 1.4	>80%	101 (5)	89	12	14	99.2	88	72	0.019	−2.455
≥ 2.2	>90%	27 (1)	26	1	4	99.9	96	70	0.008	−5.910

DSI, Dallas Steatosis Index; NAFLD, nonalcoholic fatty liver disease; PPV, positive predictive value; NPV, negative predictive value. Net Benefit is the difference between the proportion of true positives and the proportion of false positives*(risk probability/1-risk probability).

Comparison of the DSI to Previously Published Models

The C-statistic for the DSI (0.824; 95% CI 0.805–0.843) was greater than the recalibrated FSI (0.782; 95% CI 0.761–0.802), HSI (0.746; 95% CI 0.724–0.768) and I-FLI (0.781; 95% CI 0.761–0.801); however, the 95% confidence intervals overlap when compared to the US-FLI (0.810; 95% CI 0.791–0.829). The receiver operating curves are shown in Figure 1.

Figure 1: Receiver Operating Curves to Diagnose Nonalcoholic Fatty Liver Disease.

The area under the receiver operating curve (C-statistic) for the Dallas Steatosis Index (0.824; 95% CI 0.805–0.843) is greater compared to the Framingham Steatosis Index (0.782; 95% CI 0.761–0.802), Hepatic Steatosis Index (0.746; 95% CI 0.724–0.768) and the Italian Fatty Liver Index (0.781; 95% CI 0.761–0.801). However, the C-statistic for the US Fatty Liver Index (0.810; 95% CI 0.791–0.829) is not statistically different.

Compared to the other risk scores, the DSI more correctly classify patients into low, intermediate and high risk with an NRI ranging from 0.11–0.26 (see Figure 2 and eSupplementary Table 5A). The Decision Curve Analysis shows higher net benefit to screen compared to the other previously published risk models (see Figure 3).

Figure 2: The Dallas Steatosis Index Has Greater Net Reclassification Improvement.

DSI, Dallas Steatosis Index; RI, reclassification improvement. This bar graph is expressed with the comparison risk model as reference; therefore, positive numbers demonstrate superiority of the DSI (all p values <0.001).

Figure 3: Decision Curve Analysis for Nonalcoholic Fatty Liver Disease Screening.

The net benefit of a risk model is used to determine the value of using the model to inform a screening program (see eSupplementary Methods for more details).

Sensitivity Analysis: Patients without Diabetes or Statin Use

After excluding patients with a clinical diagnosis of diabetes or statin use (n = 373), the sensitivity analysis included 1766 patients (median age 43 years, median BMI 28 kg/m², 54% women) with a NAFLD prevalence of 28% (n = 487).

At the threshold of <20% risk, 50% of patients would be at low-risk and NAFLD could be excluded with 85% sensitivity and 91% negative predictive value. At the threshold of ≥50%, 18% of patients would be at high-risk and NAFLD could be diagnosed with 93% specificity and 70% positive predictive value.

The DSI continued to have the highest discriminatory power by the C-statistic (0.828; 95% CI 0.806–0.849) but this was only greater when compared to the recalibrated FSI (0.779; 95% CI 0.755–0.802) and the HSI (0.741; 95% CI 0.715–0.766) while the 95% confidence intervals overlap for the I-FLI (0.784; 0.761–0.807) and the US-FLI (0.817; 0.796–0.839). Compared to the other risk scores, the DSI more correctly classify patients into low, intermediate or high risk with an NRI ranging from 0.09–0.26 (see eSupplementary Table 5B, all comparisons’ p values <0.05).

Discussion:

The Dallas Steatosis Index (DSI) accurately identifies patients with NAFLD as diagnosed by ¹H MR spectroscopy using predictors that are routinely available to the primary care physician. The DSI has superior discrimination as determined by net reclassification improvement and a superior net benefit to screen when compared to previously published risk tools for NAFLD. The superiority of the DSI is likely a function of using a multi-ethnic population in its derivation, the use of a more accurate diagnostic gold standard for NAFLD and the use of biologically plausible interaction effects.

A timely clinical application of the DSI is to guide a multi-tiered NAFLD screening program where patients at intermediate- or high-risk are offered liver imaging to evaluate both hepatic steatosis and fibrosis (e.g. transient elastography with controlled attenuation parameter measurements). Such a strategy would achieve relative cost-savings since it would avoid both the direct cost of screening those at low-risk but also the indirect cost of further working up those patients that are more likely to be false-positives. A recent study suggests such a screening program without risk stratification may already be cost-effective in Europe and Asia²⁹; therefore, adapting a multi-tiered approach is likely to further improve the cost-effectiveness if implemented in the United States.

Another application of the DSI is to use the high-risk category as a diagnostic surrogate for NAFLD in epidemiologic studies, analogous to how the previously published risk tools have most often been used^30–34. The most appropriate comparator for this purpose is the US-FLI as that model has the greatest discriminatory power. In addition to not requiring the measurement of waist circumference and insulin concentration, the major benefit of the DSI compared to the US-FLI is its higher specificity and positive predictive value of the high-risk threshold.

A final clinical application of the DSI is to better appreciate the positive and negative predictive value of hepatic steatosis when it is found incidentally on imaging. For example, if a patient has hepatic steatosis on liver ultrasound but only a 10% risk of NAFLD by the DSI, a Bayesian analysis would yield a post-test positive predictive value of only 61% (based on the pooled sensitivity/specificity of 85%/94%³⁵). Therefore, further efforts to exclude alcohol abuse or to confirm the diagnosis of steatosis (e.g. MR with proton density fat fraction method for fat quantification or even going to a liver biopsy) would be appropriate.

The primary limitation of this study is based on the argument that NAFLD alone is not a clinically important outcome and our efforts should be limited to identifying the more severe subgroup, nonalcoholic steatohepatitis (NASH). We feel that this limitation is not particularly relevant since the anticipated use of the DSI would be to follow a positive screen with a fibrosis assessment, which is a better predictor for the natural history of NAFLD than the grade of steatohepatitis^36,37. Additionally, even patients with simple steatosis in the absence of NASH can have progression of biopsy-proven fibrosis³⁸. Finally, there is clearly prognostic value for NAFLD as diagnosed by imaging due to the increased cardiometabolic outcomes^8,39; however, we do acknowledge that those studies used liver ultrasound and not MR spectroscopy. Therefore, validation of the DSI for cardiometabolic outcomes is appropriate and such studies are currently underway.

A primary strength of the DSI is that it was derived in a population-based, multi-ethnic, probability-sample, which should limit the sick-patient bias or bias by indication inherent to models derived from clinic populations. This sampling strategy also suggests that it will be applicable to other populations; however, this will need formal external validation. Another strength is that the diagnosis of NAFLD is as close to the gold-standard (liver biopsy) as we can ethically achieve in a screening population. This means that we are not systematically misclassifying patients with earlier grades of hepatic steatosis. Despite these strengths, the DSI could likely still be improved if we added variables that were not measured in the DHS but that would be routinely available to a primary care physician (e.g. viral serologies).

In conclusion, the DSI accurately identifies patients with NAFLD as diagnosed by ¹H MR spectroscopy using routinely obtained clinical variables that are available in a primary care encounter. Future directions of this research include external validation, outcome validation and the use of the DSI to promote a cost-effective NAFLD screening program.

Need to Know.

Background:

Tools have been developed to determine risk for nonalcoholic fatty liver disease (NAFLD) based on imaging, which does not always detect early-grade hepatic steatosis. We aimed to develop a tool to identify patients with NAFLD using ¹H MR spectroscopy (MRS).

Findings:

We developed an index, called the DSI, which accurately identifies patients with NAFLD based on MRS data.

Implications for patient care:

After external validation, the DSI might be used to identify patients with NAFLD and as an endpoint in clinical trials.