The high degree of heterogeneity in the etiology of nonalcoholic fatty liver disease (NAFLD) has made the current diagnostic and classification criteria for NAFLD no longer effective in guiding the clinical management of the disease and in reducing the disease burden associated with it. In this context, NAFLD has been renamed metabolism-associated fatty liver disease (MAFLD) to play an active role in guiding the individualized and precise treatment of fatty liver disease [1,2,3,4]. Inactivity, low levels of physical activity, nutritional imbalances, and unhealthy eating habits contribute to the prevalence of the disease [5]. Furthermore, MAFLD is not only closely associated with chronic hepatitis, cirrhosis, and hepatocellular carcinoma but also contributes to the progression of cardiovascular disease, chronic kidney disease, and extrahepatic malignancies in conjunction with other metabolism-related diseases, such as diabetes, hyperlipidemia and hyperuricemia [1,2,3,4].
Detecting MAFLD as early as possible to identify those who may have silent progressive fatty liver disease is crucial. Among the many diagnostic tools, the gold standard in diagnosing MAFLD is liver biopsy [6]. However, invasiveness is one of the drawbacks of liver biopsy. Ultrasonography (US), although inexpensive, depends on the experience of the operator and the sophistication of the technology. Other imaging tests, such as magnetic resonance spectroscopy (MRS), computed tomography (CT), and vibration-controlled transient elastography (VCTE), are too expensive for mass screening to be effective. Thus, there is a need to construct a simple, noninvasive, and efficient clinical prediction model capable of accurately screening MAFLD. Meanwhile, the screening tool should be widely adapted for the early detection of MAFLD in primary, secondary, and tertiary medical centers.
Previous literature describes several models based on demographics, laboratory factors, anthropometrics, and comorbidities for diagnosing NAFLD [7,8,9,10,11]. Among these models, the fatty liver index (FLI) has demonstrated sound diagnostic accuracy in the diagnosis of NAFLD in various populations [12, 13]. Other diagnostic models, such as the visceral adiposity index (VAI) [7], the hepatic steatosis index (HSI) [9], the ZJU index [11], and the Framingham steatosis index (FSI) [10], have also been used for NAFLD screening. Consistently, the triglyceride-glucose index (TyG), an inexpensive and reliable index for assessing insulin resistance [14], is also used to diagnose NAFLD [15]. Nevertheless, its diagnostic efficacy varies significantly between studies [11, 16, 17]. Therefore, the present study will include TyG as a variable responding to insulin resistance in the new model development and compare the new model with the above model regarding MAFLD prediction efficacy. The nomogram, a visual representation of a disease-specific prediction model based on various clinical variables, is helpful in detecting diseases early and can be easily used at all levels of medical centers [18]. Consequently, nomograms can be used to help diagnose MAFLD early. Therefore, the researchers in the present study aimed to create a novel nomogram based on demographics, laboratory factors, anthropometrics, and comorbidities to accurately detect MAFLD in the American population.
Data source
Data were included from the National Health and Nutrition Examination Survey (NHANES), which is a nationally cross-sectional and multistage study of the nonmilitary and noninstitutionalized population of the United States. Every two years, the NHANES data are released. Each participant in the survey signed an informed consent form, and ethics review board at the National Center for Health Statistics Research has approved the protocol for this survey. In addition, this study followed the same protocol as shown in Transparent Reporting of a Multivariable Predictive Model for Individual Prognosis or Diagnosis (TRIPOD) [19].
Participant selection
NHANES data (cycle 2017–2020.3) with valid vibration-controlled transient elastography (VCTE) values were used for analyses. There were 8317 subjects with valid VCTE values and ages greater than or equal to 18 years in the 2017–2020.3 NHANES database. After excluding 431 participants with no available important anthropometric data, 297 cases without key blood cell count data, 286 cases without key biochemical values, and 3 participants with no smoking data, a total of 7300 individuals were finally enrolled. An overview of the enrollment process is shown in Fig. 1.
Demographics, Laboratory Factors, Anthropometrics, Lifestyles and Comorbidities
Data from NHANES included variables related to MAFLD from previous studies. These variables included demographics (age, sex, and race), anthropometrics (arm circumference, waist circumference, BMI, and hip circumference), lifestyles (smoking), comorbidities (hypertension and diabetes), and biomarkers such as white blood cell (WBC), hemoglobin (HB), platelets (PLT), lymphocytes (LYM), neutrophils (NEU), fasting plasma glucose (FPG), total bilirubin (TBIL), aspartate aminotransferase (AST), alanine aminotransferase (ALT), ALT to AST ratio, alkaline phosphatase (ALP), γ-glutamyl transpeptidase (GGT), triglyceride (TG), total cholesterol (TC), high-density lipoprotein cholesterol (HDL-C), total protein (TP), globulin (GLB), albumin (ALB), estimated glomerular filtration rate (eGFR), blood urea nitrogen (BUN), creatinine (CRE), and high-sensitivity C-reactive protein (hsCRP).
This study categorized race into six categories (Non-Hispanic White, Non-Hispanic Black, Other Hispanic, Mexican American, Non-Hispanic Asian, and Others) and smoking into three groups (never, former, and current). The diagnostic criteria for diabetes were FPG ≥ 7.0 mmol/L or glycohemoglobin (HbA1c) > 6.5% or random plasma glucose ≥ 11.1 mmol/L or two-hour oral glucose tolerance test (OGTT) plasma glucose ≥ 11.1 mmol/L or under anti-diabetes treatment or self-reported diabetes [20]. A systolic blood pressure (SBP) of 140 mmHg or diastolic blood pressure (DBP) of 90 mmHg or under antihypertension treatment or self-reported hypertension was used as a diagnostic criterion for hypertension [21]. The eGFR was calculated based on the chronic kidney disease epidemiology formula (CKD-EPI) [22].
Definition of MALFD
Hepatic steatosis in this study was defined by the controlled attenuation parameter (CAP), obtained via VCTE (FibroScan®), which is a validated tool for measuring steatosis in participants with fatty liver [23]. CAP ≥ 268 dB/m was defined as significant hepatic steatosis. This cutoff value provided an AUC of 0.865 (95% CI 0.850–0.880), a sensitivity (SEN) of 0.773 (95% CI 0.690–0.838), and a specificity (SPE) of 0.812 (95% CI 0.749–0.879) [24].
A diagnosis of MAFLD was made if hepatic steatosis was present along with any of the following: overweight or obesity, diabetes, and metabolic dysfunction. At least two conditions were required for metabolic dysfunction to exist: 1) WC ≥ 102 cm in males and WC ≥ 88 cm in females, 2) hypertension, 3) hyperlipidemia (TG ≥ 1.70 mmol/L or under lipid-lowering treatment), 4) HDL-C < 1.0 mmol/L in men and < 1.3 mmol/L in women, 5) prediabetes, and 6) hsCRP > 2 mg/L [1].
Statistical analysis
R software was used for statistical analysis (version 4.1.2). For categorical data, counts and percentages were used, and for continuous data, the mean and standard deviation (SD) were used.
For model development, in a 7:3 ratio, all 7300 participants were randomly divided into two groups for training and validation (5112 and 2188 subjects, respectively) using the “caret” package. The training dataset was used to develop the model, and internal validation was performed using the validation dataset. In addition, the researchers in this study used the “glmnet” package to perform least absolute shrinkage and selection operator (LASSO) regression. This package runs a tenfold cross-validation of the included variables before selecting the optimal lambda value. Researchers chose lambda.lse from the cross-validation results because it has the best performance but the least number of variables. Then, researchers used the “rms” package to run a logistic regression analysis. By including the variables screened in the LASSO regression, a multivariable logistic regression model was constructed. For each variable, an odds ratio (OR) and 95% confidence interval (CI) were assessed. The statistical significance levels were all two-sided. In the next step, using the “rms” package, this study developed the predictive nomogram using statistically significant variables.
For model evaluation, the receiver operating characteristic curve (ROC) operation was performed using the “pROC” package and compared against existing models. Based on Delong's method, P < 0.05 was considered statistically significant when comparing the area under the receiver operating characteristic curve (AUC) values. Using the AUC, the present study could distinguish true positives from false-positives based on the quality of the risk nomogram. Because the meaning of AUC increments is not intuitive, this study calculated the NRI and IDI based on the corresponding equations [25, 26].
Furthermore, researchers used the “terms” package to calculate the calibration curve and the Brier score, which was used to assess the calibration of the newly built nomogram. The decision curve analysis (DCA) curve was conducted using the “price” package, which calculates the clinical practicability of the model based on numerous threshold probabilities.
To calculate the adequate sample size, there should be at least ten outcome events per variable (EPV) when performing prediction research. Researchers expected robust estimates as the study’s sample size and outcome events were far greater than those of the EPV method [27].