Abstract
Goals and Background:
Using natural language processing (NLP) to create a nonalcoholic fatty liver disease (NAFLD) cohort in primary care, we assessed advanced fibrosis risk with the Fibrosis-4 Index (FIB-4) and NAFLD Fibrosis Score (NFS) and evaluated risk score agreement.
Methods:
In this retrospective study of adults with radiographic evidence of hepatic steatosis, we calculated patient-level FIB-4 and NFS scores and categorized them by fibrosis risk. Risk category and risk score agreement was analyzed using weighted kappa, Pearson correlation, and Bland-Altman analysis. A multinomial logistic regression model evaluated associations between clinical variables and discrepant FIB-4 and NFS results.
Results:
Of the 767-patient cohort, 71% had a FIB-4 or NFS score in the indeterminate or high-risk category for fibrosis. Risk categories disagreed in 43% and scores would have resulted in different clinical decisions in 30% of the sample. The weighted kappa statistic for risk category agreement was 0.41 (95% CI: 0.36 – 0.46) and the Pearson correlation coefficient for log FIB-4 and NFS was 0.66 (95% CI: 0.62, 0.70). The multinomial logistic regression analysis identified Black race (OR 2.64, 95% CI 1.84-3.78) and A1c (OR 1.37, 95% CI 1.23-1.52) with higher odds of having an NFS risk category exceeding FIB-4.
Conclusions:
In a primary care NAFLD cohort, many patients had elevated FIB-4 and NFS risk scores and these risk categories were often in disagreement. The choice between FIB-4 and NFS for fibrosis risk assessment can impact clinical decision making and may contribute to disparities of care.
Keywords: non-invasive testing, FIB-4, NFS
Graphical Abstract:
Distribution and agreement of Fibrosis-4 Index (FIB-4) and NAFLD Fibrosis scores (NFS) in a primary care NAFLD cohort. NLP=natural language processing. Dx=diagnosis. OR=odds ratio. CI=confidence interval.
Introduction
The presence of advanced fibrosis (Metavir fibrosis stages 3 and 4) in non-alcoholic fatty liver disease (NAFLD) serves as the best indicator of liver-related mortality, cardiovascular death, and overall mortality in affected patients [1–3]. Assessing fibrosis risk in the primary care setting can identify patients at greatest need for cardiovascular risk reduction and hepatology referral [4, 5]. Methods for gauging advanced fibrosis risk in patients with NAFLD have rapidly evolved from liver biopsy to non-invasive techniques that include vibration-controlled transient elastography (VCTE) and serologic risk scores derived from commonly available blood test results [6]. Due to the accessibility, performance, and ease of calculating serologic assessments, including the Fibrosis-4 index (FIB-4) and NAFLD Fibrosis score (NFS), guidelines on approaching NAFLD and abnormal liver chemistries now incorporate these scores to assist primary care providers (PCPs) in referring patients at risk for advanced fibrosis to hepatology specialists [7–9]. Guidelines suggest either FIB-4 or NFS calculation as an acceptable initial approach for fibrosis risk stratification, with second-step testing (e.g. enhanced liver fibrosis [ELF™], VCTE) for results categorized as indeterminate or high-risk [9–13]. Before PCPs implement these NAFLD care pathways, we need to better understand how non-invasive fibrosis risk assessments perform in the primary care setting.
Unfortunately, studies evaluating primary care management of NAFLD are lacking, likely due to under-diagnosis in this setting [14–18]. With under-diagnosis precluding the use of NAFLD diagnostic codes for study, one strategy for identifying patients with NAFLD includes searching for alanine aminotransferase (ALT) elevations. But, deficits in other chronic liver disease testing (Hepatitis C virus, etc.) and substance exposure documentation limit this approach [19, 20]. Another potential tactic relies on selecting patients with radiographic steatosis, but the lack of structured report data has limited the use of radiographic findings in most large datasets. Natural language processing (NLP) can obviate this barrier by converting text into structured data, allowing for computer-aided analysis of radiology reports [21].
After leveraging NLP to develop a sample of primary care patients with radiographic evidence of hepatic steatosis, we calculated FIB-4 and NFS scores for each patient and evaluated the level of agreement for these non-invasive fibrosis risk assessments. We hypothesized that many patients would have non-invasive risk scores for advanced fibrosis in the indeterminate and high-risk range and many would have discrepant FIB-4 and NFS risk assessments.
Materials and Methods
Study Design
This retrospective cohort study of electronic health record (EHR) data included primary care patients with radiographic reports of liver steatosis and no previously known, non-NAFLD, liver disease. FIB-4 and NFS scores were calculated for each patient and analyzed for agreement in advanced fibrosis risk stratification. The Institutional Review Board at the Medical University of South Carolina (MUSC) approved this study.
Study Population
All patients with abdominal ultrasound (US), computed tomography (CT), or magnetic resonance imaging (MRI) results from January 1, 2012 (the introduction of the EHR) to December 31, 2018 and seen in the Internal Medicine patient-centered medical home (PCMH) at MUSC were evaluated. The practice conducts 32,000 patient visits yearly and delivers care to a diverse (39% non-white) adult (mean age 59 years) population with chronic and complex medical problems.
Using NLP, imaging reports were categorized by the presence or absence of steatosis. All abdominal imaging result notes for our patient population were extracted, those with text including “steatosis” or “steatohepatitis” were identified, and 50 characters pre- and post-location were pulled. Common affirmative (e.g. “. Hepatic Steatosis.”) and negative (e.g. “without evidence of steatosis”) phrases within the data were identified. The sequence of using a filter, find and search, and pivot table to identify and assign common phrases continued until groupings slimmed down to only one occurrence per phrase.
Primary care patients with at least one radiographic report affirming hepatic steatosis were included. Patients with no aminotransferase or platelet count (Plt) results within the 1 year preceding the first radiographic report of steatosis were excluded. Patients with markedly elevated ALT or aspartate aminotransferase (AST) values (≥ 500 IU/L) suggestive of an alternative diagnosis underwent chart review and were evaluated for more remote (still within 1 year) aminotransferase values. Those individuals with ICD—9/10 codes for other, non-NAFLD, chronic liver disease and severe liver disease outcomes (e.g. cirrhosis) at any point during the study period were also excluded (Supplement).
Outcomes
The proportion of cohort patients assigned an ICD-9/10 code for NAFLD or NASH was calculated. These codes included at least one ICD-9 assignment of 571.8 or an ICD-10 code of K75.81 or K76.0 during the study period [14]. Additionally, FIB-4 and NFS scores were calculated for each patient with all variable values coming the day of, or most proximal to (within 1 year), the first steatosis identification (Supplement) [22, 23].
The presence of diabetes was assigned based on the presence of at least one previous hemoglobin A1c ≥ 6.5%. Patient-level FIB-4 and NFS scores were categorized by advanced fibrosis risk: low risk (FIB-4 ≤ 1.3, NFS ≤ −1.455); indeterminate risk (1.3 < FIB-4 ≤ 2.67, −1.455 < NFS ≤ 0.676); and high risk (FIB-4 > 2.67, NFS > 0.676) [7, 8, 24].
The (1) proportion of patients with a FIB-4 or NFS score in the indeterminate or high-risk category; (2) the proportion of patients with discrepant advanced fibrosis risk assessments; (3) the proportion of patients with FIB-4 and NFS that would result in different clinical decisions; (4) the level of agreement between fibrosis risk categories and scores; and (5) factors associated with discrepant scores, served as the primary outcomes of interest.
The clinical decision-making outcome assumes that patients with indeterminate or high-risk scores would undergo further evaluation (i.e. VCTE or hepatology referral) and patients with low risk scores would not [4, 9, 10, 18]. Thus, patients with a high-risk FIB-4 and an indeterminate risk NFS would receive further testing regardless of which non-invasive assessment the PCP performed. Patients with a low-risk FIB-4 and an indeterminate risk NFS would not undergo further testing if only FIB-4 was applied but would get further investigation in a setting using NFS alone.
Independent Variables
The independent variables of interest included demographic, vital sign, and laboratory data, with an emphasis on metabolic syndrome (MetS) components. MetS elements were continuous variables defined by recordings of the following at the time of, or most proximal to FIB-4/NFS calculation: systolic blood pressure (mm Hg), glycosylated hemoglobin (A1c %), high-density lipoprotein (mg/dL), triglyceride level (TG, mg/dL), and body mass index (BMI, kg/m2) [10, 18]. Bilirubin and alkaline phosphatase (ALP) were treated as continuous variables, with values coming from the liver test panels accompanying the AST and ALT scores used for non-invasive test score calculations.
Gender was coded dichotomously as Male / Female. Race was a categorical variable coded as Black / non-Black, owing to a small number of non-Black, non-white patients in the sample (n=32).
Data Sources
All data came from Medical University Hospital Authority Enterprise and EPIC© (EPIC Systems Corporation, WI) Clarity databases. Clinical, laboratory, and demographic data were obtained in the ambulatory, emergency room, and inpatient settings at MUSC during the study period.
Statistical Analysis
Patient characteristics in each of the risk strata were reported as frequency counts and proportions for categorical variables, and median and interquartile range for continuous variables. The bivariate relationship between FIB-4 and NFS categories was summarized in a three by three table showing the level of agreement between the risk strata and to assess the proportion of scores that would have led to similar clinical decisions.
Agreement between FIB-4 and NFS fibrosis risk categories was analyzed using weighted kappa statistics, treating the low, indeterminate, and high-risk groups as categorical outcome variables. A Pearson correlation analysis was performed to assess linear relationship between log FIB-4 and NFS as continuous variables. Since the distribution of FIB-4 results were right-skewed, the FIB-4 results were log transformed to standardize the scale for analysis. Additionally, a Bland-Altman analysis was performed to evaluate the level of agreement between patient-level log FIB-4 and NFS scores treated as continuous variables.
A multinomial logistic regression model was developed to evaluate patient factors associated with FIB-4 and NFS disagreement. The multinomial model analyzed three outcomes: (1) FIB-4 and NFS risk category agreement (reference); (2) risk category disagreement where the FIB-4 yielded the higher risk assessment; and (3) disagreement where NFS provided the higher risk assessment (Supplement). Predictor variables were tested for the model using clinical relevance, and the predictors underwent stepwise backwards selection at an individual significance level of p < 0.1 and contribution to overall model fit. Patients without predictor variables available within the previous 1 year were considered missing. Regression models were developed for a complete case analysis and a multiple imputation model to assess the effect of missing data on the model results. The models were assessed for the effect of multicollinearity via variance inflation factor and residual analysis to assess overall goodness of fit. Statistical analyses were performed using SAS version 9.4 (Cary, NC).
Results
Of the entire cohort, a total of 8,098 unique patients possessed results for abdominal US, CT, or MRI evaluations and 1,237 (15%) had results affirming the presence of hepatic steatosis. After excluding those patients with previous diagnoses of non-NAFLD chronic liver disease and those without qualifying components for FIB-4 and NFS calculations within the preceding year, a cohort of 767 patients was created (Supplement).
Radiographic steatosis was most often identified by US (54%) or CT (40%) (Table 1). Patients had a median age of 55 years (IQR: 46-65), 64% were female, and 35% identified as Black. The median ALT and AST for those included was 30 IU/L (IQR: 20-53) and 28 IU/L (IQR: 21-45), respectively. Patients in the cohort had a median A1c of 6.1% (IQR: 5.5-7.3%) and BMI of 32 kg/m2 (IQR: 27.5-37.1 kg/m2). In this group with radiographic steatosis and no other known chronic liver disease, 235 (31%) patients received a NAFLD diagnosis code.
Table 1:
Characteristics of patient sample (N=767)
| Total | Agreement of NFS and FIB-4 | |||
|---|---|---|---|---|
| Agreement | NFS Higher | FIB-4 Higher | ||
| n=767 | n=434 | n=262 | n=71 | |
| Demographics | ||||
| Age (years)* | ||||
| Median (IQR) | 55 (46,65) | 56 (45,66) | 55 (46,63) | 57 (49,65) |
| Gender | ||||
| Female (%) | 487 (63.5) | 270 (62.2) | 182 (69.5) | 35 (49.3) |
| Male (%) | 280 (36.5) | 164 (37.8) | 80 (30.5) | 36 (50.7) |
| Race | ||||
| Black (%) | 265 (34.6) | 114 (26.3) | 129 (49.2) | 22 (31.0) |
| Non-Black (%) | 502 (65.4) | 320 (73.7) | 133 (50.8) | 49 (69.0) |
| Clinical variables | ||||
| Median ALT (IQR) | 30 (20,53) | 33 (22,57) | 24 (17,34) | 66 (40,109) |
| Median AST (IQR) | 28 (21,45) | 30.5 (22,47) | 22 (18,29) | 70 (46,118) |
| Median Platelets (IQR) | 238 (195,292) | 247 (191,303) | 237.5 (203,281) | 212 (151,259) |
| Median Albumin (IQR) | 3.7 (3.4,4.0) | 3.8 (3.4,4.1) | 3.6 (3.3,3.8) | 3.9 (3.5,4.1) |
| Median ALP (IQR) | 83 (66,105) | 83.5 (68,107) | 82 (64,101) | 81 (64,114) |
| Median Bilirubin (IQR) | 0.6 (0.4,0.9) | 0.6 (0.4,0.9) | 0.5 (0.4,0.8) | 0.8 (0.5,1.2) |
| MetS Components | ||||
| Median SBP, mmHg (IQR) | 132 (120,144) | 130 (118,143) | 135 (126,146) | 132 (121,147) |
| Median A1c % (IQR) | 6.1 (5.5,7.3) | 5.9 (5.4,6.6) | 6.8 (6.0,8.6) | 5.5 (5.1,6.0) |
| Median HDL (IQR) | 44 (37,55) | 46 (38,57) | 42 (35,50) | 49 (38.5,62.5) |
| Median Triglycerides (IQR) | 134 (93,203) | 128.8 (88,191.5) | 136 (98.5,208.5) | 146.5 (85,218) |
| Median BMI (IQR) | 32 (27.5, 37.1) | 30.1 (26.7,35.3) | 36.3 (32.3,42.2) | 26.5 (23.4,30.6) |
| Imaging modality | ||||
| Ultrasound | 409 (53.3) | |||
| CT | 310 (40.4) | |||
| MRI | 48 (6.3) | |||
| Median days between image and risk score (IQR) | 5 (0,37) | |||
| NAFLD Diagnoses † | 235 (30.6) | |||
Age is at the time of the radiographic image;
NAFLD and NASH diagnosis codes (571.8, K75.81, or K76.0).
Comorbidities according to ICD-9/10 Elixhauser coding algorithm. ALT=alanine transaminase; AST=aspartate transaminase; ALP=alkaline phosphatase; BMI=body mass index (kg/m2); SBP=systolic blood pressure; CT=computed tomography; HDL=high density lipoprotein; IQR=inter-quartile range; MetS=metabolic syndrome; MRI=magnetic resonance imaging; NAFLD=nonalcoholic fatty liver disease.
The median number of days between imaging and risk score calculation was 5 (IQR: 0-37). Of the included patients, 55%, 32%, and 13% had low, indeterminate, and high-risk FIB-4 scores, respectively (Table 2). NFS scores were low-risk in 33%, indeterminate-risk in 48%, and high-risk in 19% of patients. For 29% of the sample, risk scores agreed and were in the low risk category, whereas 8% of the sample had concordant high-risk fibrosis risk scores. Advanced fibrosis risk assessment categories agreed in 57% of patients. FIB-4 and NFS scores would have resulted in the same clinical decision in 70% of patients, and different clinical decisions in 30% of patients.
Table 2:
Patient FIB-4 and NFS scores, categorized by advanced fibrosis risk.
| NAFLD Fibrosis Score | |||||
|---|---|---|---|---|---|
|
| |||||
| Low | Indeterminate | High | |||
|
| |||||
| Fibrosis-4 Index | NFS ≤ −1.455 | −1.455 < NFS ≤ 0.676 | NFS > 0.676 | Total (%) | |
| Low | FIB-4 ≤ 1.3 | 222 | 177 | 20 | 419 (55%) |
| Indeterminate | 1.3 < FIB-4 ≤ 2.67 | 30 | 152 | 65 | 247 (32%) |
| High | FIB-4 > 2.67 | 4 | 37 | 60 | 101 (13%) |
|
| |||||
| Total (%) | 256 (33%) | 366 (48%) | 145 (19%) | 767 | |
The weighted kappa statistic to assess FIB-4 and NFS agreement in fibrosis risk group categorization was 0.41 (95% CI: 0.36 – 0.46). After log transforming the patient-level FIB-4 scores, a correlation analysis of NFS and log FIB-4 values provided a Pearson correlation coefficient of 0.66 (Figure 1a, 95% CI: 0.62, 0.70). The Bland-Altman analysis plotted the standardized mean NFS and log FIB-4 values by the difference between the standardized NFS and log FIB-4 scores, with numerous assessments falling outside the 95% confidence interval (Figure 1b).
Figure 1:
Plots of patient-level log FIB-4 scores by NFS scores for the Pearson correlation (a); and the Bland-Altman analysis assessing agreement between patient-level standardized log FIB-4 and NFS scores (b).
The multinomial logistic regression analysis included a complete case analysis (n=586) model and a model with multiple imputation to address patients with missing predictor variable data (n=767, Table 3). Gender and triglyceride variables were found to be not significant and removed from the model. Black race (OR 2.64, 95% CI 1.84-3.78) and A1c (OR 1.37, 95% CI 1.23-1.52) were associated with higher odds of having a higher NFS risk category than FIB-4, compared to those patients with risk scores in agreement and controlling for the other predictor variables. Increasing HDL (OR 0.98, 95% CI 0.97-0.99) and bilirubin (OR 0.59, 95% CI 0.40-0.88) were associated with lower odds of having a higher NFS risk than FIB-4 risk category, compared to those patients with concordant risk. Bilirubin (OR 1.68, 95% CI 1.07-2.62) was associated with higher odds of having FIB-4 risk greater than NFS risk, compared to those patients with scores in agreement (in the complete case analysis).
Table 3.
Estimated Odds Ratios (OR) from Multinomial Logistic Regression.
| Sample | Complete Case Analysis (n=586) | Multiple Imputation Analysis (n=767) | ||||||
|---|---|---|---|---|---|---|---|---|
| OR | 95% CI | OR | 95% CI | |||||
| Predictors * | Outcomes | |||||||
| Agreement (reference) | ||||||||
| Black | FIB-4 Higher | 0.808 | 0.387 | 1.689 | 1.193 | 0.668 | 2.131 | |
| NFS Higher | 2.311 | 1.543 | 3.462 | 2.635 | 1.836 | 3.782 | ||
| HDL | FIB-4 Higher | 1.016 | 1.000 | 1.032 | 1.009 | 0.994 | 1.025 | |
| NFS Higher | 0.983 | 0.970 | 0.995 | 0.981 | 0.969 | 0.994 | ||
| SBP | FIB-4 Higher | 1.013 | 0.998 | 1.029 | 1.009 | 0.996 | 1.022 | |
| NFS Higher | 1.012 | 1.001 | 1.022 | 1.011 | 1.002 | 1.020 | ||
| Bilirubin | FIB-4 Higher | 1.678 | 1.073 | 2.622 | 1.190 | 0.911 | 1.555 | |
| NFS Higher | 0.550 | 0.352 | 0.858 | 0.594 | 0.400 | 0.880 | ||
| ALP | FIB-4 Higher | 1.002 | 0.998 | 1.005 | 1.002 | 0.999 | 1.005 | |
| NFS Higher | 0.990 | 0.986 | 0.995 | 0.991 | 0.987 | 0.995 | ||
| A1c % | FIB-4 Higher | 0.811 | 0.627 | 1.050 | 0.836 | 0.663 | 1.054 | |
| NFS Higher | 1.400 | 1.247 | 1.571 | 1.370 | 1.233 | 1.522 | ||
Gender and triglyceride variables were not significant, did not contribute to overall model fit, and were removed;
Outcomes were FIB-4 and NFS scores in risk category agreement; ALP=alkaline phosphatase; CI=confidence interval; SBP=systolic blood pressure; HDL=high-density lipoprotein.
The relationship between Black race and higher odds of an NFS risk score exceeding the FIB-4 was further explored by stratifying the risk score categories by race (Table 4). Non-Black patient risk scores agreed 64% of the time, compared to 43% in Black patients. Black patients had NFS risk values that exceeded the FIB-4 scores in 49% of the sample, compared to 26% of Non-Black patients.
Table 4:
Patient FIB-4 and NFS scores by race, categorized by advanced fibrosis risk.
| Black patients | ||||
|---|---|---|---|---|
| NAFLD Fibrosis score | ||||
| Fibrosis-4 Index | NFS ≤ −1.455 | −1.455 < NFS ≤ 0.676 | NFS > 0.676 | Total (%) |
| FIB-4 ≤ 1.3 | 47 | 79 | 20 | 141 (53%) |
| 1.3 < FIB-4 ≤ 2.67 | 8 | 41 | 35 | 84 (32%) |
| FIB-4 > 2.67 | 1 | 13 | 26 | 40 (15%) |
| Total (%) | 56 (21%) | 133 (50%) | 76 (29%) | 265 |
| Non-Black patients | ||||
| NAFLD Fibrosis score | ||||
| Fibrosis-4 Index | NFS ≤ −1.455 | −1.455 < NFS ≤ 0.676 | NFS > 0.676 | Total (%) |
| FIB-4 ≤ 1.3 | 175 | 98 | 5 | 278 (55%) |
| 1.3 < FIB-4 ≤ 2.67 | 22 | 111 | 30 | 163 (33%) |
| FIB-4 > 2.67 | 3 | 24 | 34 | 61 (12%) |
| Total (%) | 200 (40%) | 233 (46%) | 69 (14%) | 502 |
| Agreement | NFS Higher | FIB-4 Higher | Total | |
| Black | 114 (43%) | 129 (49%) | 22 (8%) | 265 |
| Non-Black | 320 (64%) | 133 (26%) | 49 (10%) | 502 |
NFS=NAFLD Fibrosis score. FIB-4=Fibrosis-4 index.
Discussion
This study offers a striking example of NAFLD under-diagnosis in clinical practice by utilizing a sample of patients with radiographic liver steatosis and no other known chronic liver disease [14–17]. Improving NAFLD diagnosis in primary care may not always require additional testing and cost if clinicians can make use of previously performed abdominal radiographic examinations [25]. Harnessing NLP technology to take advantage of EHRs and primary care patient registries may make previously latent NAFLD diagnostic information actionable [21]. Alarmingly, 71% of this patient cohort had either a FIB-4 or NFS score in the indeterminate or high-risk category for advanced fibrosis, reinforcing the clinical significance of this missing diagnostic data.
Efforts to implement reliable non-invasive fibrosis risk assessment of primary care patients with NAFLD must also evolve. FIB-4 and NFS address this need in NAFLD registries, but their performance in primary care is less well known [4, 5, 13, 26]. A recent study applying FIB-4 to primary care patients with NAFLD demonstrated the utility of this approach, reducing unnecessary referrals by 80%, and improving the detection of advanced fibrosis and cirrhosis in a primary care population [4]. Another study showed that FIB-4 scores, in combination with VCTE, could reduce unnecessary further investigation in patients with low-risk assessments by 87% [5]. In our sample, 71% of patients had non-invasive test scores in the indeterminate or high-risk categories for fibrosis, results that would prompt further evaluation and consideration of specialty referral, but 43% of FIB-4 and NFS advanced fibrosis risk assessments disagreed. Perhaps more importantly, 30% of patients had risk scores that would have resulted in different clinical responses depending on which tool (FIB-4 or NFS) the PCP applied. Problems resulting from discordant signals span the spectrum from delayed and missed diagnoses to inappropriate, costly testing and specialty referrals. With the rising and already massive burden of NAFLD, this clinical deviation could affect care for a significant number of patients.
The relationship between Black race and NFS risk exceeding FIB-4 poses another critical issue in primary care NAFLD management. Non-invasive test selection in this population holds the potential to either narrow or exacerbate racial disparities of chronic liver disease care [27–29]. FIB-4 application could lead to under-diagnosis of NAFLD fibrosis, while NFS use could result in unnecessary testing and excessive cost. Understanding why this relationship exists will play a critical role in developing a comprehensive primary care NAFLD fibrosis risk assessment strategy.
While FIB-4 and NFS can both provide information on fibrosis risk, their different origins likely contribute to discrepant performance. FIB-4 was developed to predict fibrosis in patients with chronic Hepatitis C infection, whereas NFS derived from a sample of patients with biopsy-proven NAFLD and incorporates measures of metabolic syndrome (BMI and insulin resistance) into its equation [22, 23]. With these contributions from metabolic disease and Black patients suffering greater burdens of obesity and diabetes historically, higher NFS scores in this group seem likely, but whether these values predict fibrosis is not known [30, 31]. Given the endogeneity of BMI in the NFS outcome, we could not control for this variable in the regression models. Issues of endogeneity also limited our ability to assess the relationship between abnormal ALT values and fibrosis risk agreement. Though ALT results can remain normal in the presence of advanced fibrosis, the performance of FIB-4 and NFS may change for different ALT values[32]. Understanding how FIB-4 and NFS scores change in accordance with patient characteristics may help PCPs select the right test for the right patient, demonstrate a need for improved access to secondary fibrosis risk assessment (ELF™ testing or VCTE) in primary care, or lead to the development of more comprehensive tools in this setting.
We recognize the limitations of this study. The selection of the patient sample relied upon radiographic reports of steatosis rather than evaluation of the images directly, which could invite reporting bias. Further, we were unable to control for the reason the patients underwent imaging, likely resulting in selection bias. We attempted to address these issues by applying NLP to the entirety of the radiographic report, not simply the summary statement, thus capturing both targeted (e.g. in response to abnormal liver chemistries) and incidental steatosis findings. Reassuringly, demographic and clinical characteristics in Table 1 were consistent with what we expected (i.e. high burdens of obesity, metabolic syndrome, and no marked transaminase abnormalities) for a NAFLD cohort in primary care. The inclusion strategy assumed an under-reporting of NAFLD ICD-9/10 codes (which the results confirmed), but it deserves mention that ICD-9/10 codes for the exclusion criteria of alcohol use and other chronic liver diseases are likely under-reported as well [33]. Thus, characterization of some included patients as NAFLD may be inaccurate. Conversely, our cohort selection did not account for dual liver disease etiologies, and it is likely that some of those categorized as viral hepatitis may also have some component of NAFLD. Only data present in the EHR was available for analysis, missing lab results obtained outside of the health system. Lastly, this study comes from a single institution which could threaten its generalizability. However, its primary care focus with a diverse clinical population may combat this concern.
NAFLD is under-diagnosed in primary care and NLP may help to combat this problem in patients with previous radiographic studies. Improving NAFLD diagnosis could lead to enhanced primary care management, development of real-life cohorts for retrospective research, and identification of patients for future prospective studies. Non-invasive fibrosis risk assessment with FIB-4 and NFS can help to discern those patients requiring no additional workup from those that may benefit from further investigation. Agreement between FIB-4 and NFS in primary care patients with NAFLD is not perfect, and the non-invasive test selected can impact the clinical decisions made as a result.
Supplementary Material
Acknowledgments
We thank Dr. Don Rockey for his committed mentorship in Dr. Schreiner’s ongoing K23 Award and his continued contributions to the generation and communication of new research.
Grant Support:
National Institute of Diabetes and Digestive and Kidney Diseases (NIH/NIDDK K23DK118200 PI: Schreiner). This project is also supported in part by the SSCI Research Scholar Award. The content is solely the responsibility of the authors and does not necessarily represent the official views of the Southern Society for Clinical Investigation (SSCI). This project was also supported by the South Carolina Clinical & Translational Research Institute with an academic home at the Medical University of South Carolina CTSA National Center for Advancing Translational Sciences (NCATS) of the National Institutes of Health (NIH) under UL1 TR001450. The content is solely the responsibility of the authors and does not necessarily represent the official views of the NCATS or NIH.
Abbreviations:
- ALT
alanine aminotransferase
- AST
aspartate aminotransferase
- BMI
body mass index
- EHR
electronic health record
- FIB-4
Fibrosis-4 Index
- HDL
high density lipoprotein
- ICD
International Classification of Diseases
- IQR
inter-quartile range
- MetS
metabolic syndrome
- NAFLD
nonalcoholic fatty liver disease
- NFS
NAFLD fibrosis score
- PCMH
patient-centered medical home
- Plt
platelet
Footnotes
Disclosures: All authors report no conflicts of interest with this work.
References
- 1.Dulai PS, Singh S, Patel J, et al. Increased risk of mortality by fibrosis stage in nonalcoholic fatty liver disease: Systematic review and meta-analysis. Hepatology. 2017;65:1557–1565. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Ekstedt M, Franzen LE, Mathiesen UL, et al. Long-term follow-up of patients with NAFLD and elevated liver enzymes. Hepatology. 2006;44:865–873. [DOI] [PubMed] [Google Scholar]
- 3.Taylor RS, Taylor RJ, Bayliss S, et al. Association Between Fibrosis Stage and Outcomes of Patients With Nonalcoholic Fatty Liver Disease: A Systematic Review and Meta-Analysis. Gastroenterology. 2020;158:1611–1625 e1612. [DOI] [PubMed] [Google Scholar]
- 4.Srivastava A, Gailer R, Tanwar S, et al. Prospective evaluation of a primary care referral pathway for patients with non-alcoholic fatty liver disease. J Hepatol. 2019;71:371–378. [DOI] [PubMed] [Google Scholar]
- 5.Davyduke T, Tandon P, Al-Karaghouli M, et al. Impact of Implementing a “FIB-4 First” Strategy on a Pathway for Patients With NAFLD Referred From Primary Care. Hepatol Commun. 2019;3:1322–1333. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Siddiqui MS, Yamada G, Vuppalanchi R, et al. Diagnostic Accuracy of Noninvasive Fibrosis Models to Detect Change in Fibrosis Stage. Clin Gastroenterol Hepatol. 2019;17:1877–1885 e1875. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.McPherson S, Stewart SF, Henderson E, et al. Simple non-invasive fibrosis scoring systems can reliably exclude advanced fibrosis in patients with non-alcoholic fatty liver disease. Gut. 2010;59:1265–1269. [DOI] [PubMed] [Google Scholar]
- 8.Shah AG, Lydecker A, Murray K, et al. Comparison of noninvasive markers of fibrosis in patients with nonalcoholic fatty liver disease. Clin Gastroenterol Hepatol. 2009;7:1104–1112. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Tsochatzis EA and Newsome PN. Non-alcoholic fatty liver disease and the interface between primary and secondary care. Lancet Gastroenterol Hepatol. 2018;3:509–517. [DOI] [PubMed] [Google Scholar]
- 10.European Association for the Study of the L, European Association for the Study of D and European Association for the Study of O. EASL-EASD-EASO Clinical Practice Guidelines for the management of non-alcoholic fatty liver disease. J Hepatol. 2016;64:1388–1402. [DOI] [PubMed] [Google Scholar]
- 11.Newsome PN, Cramb R, Davison SM, et al. Guidelines on the management of abnormal liver blood tests. Gut. 2018;67:6–19. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Budd J and Cusi K. Nonalcoholic Fatty Liver Disease: What Does the Primary Care Physician Need to Know? Am J Med. 2020;133:536–543. [DOI] [PubMed] [Google Scholar]
- 13.Dokmak A, Lizaola-Mayo B and Trivedi HD. The Impact of Nonalcoholic Fatty Liver Disease in Primary Care: A Population Health Perspective. Am J Med. 2021;134:23–29. [DOI] [PubMed] [Google Scholar]
- 14.Alexander M, Loomis AK, Fairburn-Beech J, et al. Real-world data reveal a diagnostic gap in non-alcoholic fatty liver disease. BMC Med. 2018;16:130. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Cleveland ER, Ning H, Vos MB, et al. Low Awareness of Nonalcoholic Fatty Liver Disease in a Population-Based Cohort Sample: the CARDIA Study. J Gen Intern Med. 2019;34:2772–2778. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Loomba R, Wong R, Fraysse J, et al. Nonalcoholic fatty liver disease progression rates to cirrhosis and progression of cirrhosis to decompensation and mortality: a real world analysis of Medicare data. Aliment Pharmacol Ther. 2020. [DOI] [PubMed] [Google Scholar]
- 17.Blais P, Husain N, Kramer JR, et al. Nonalcoholic fatty liver disease is underrecognized in the primary care setting. Am J Gastroenterol. 2015;110:10–14. [DOI] [PubMed] [Google Scholar]
- 18.Chalasani N, Younossi Z, Lavine JE, et al. The diagnosis and management of nonalcoholic fatty liver disease: Practice guidance from the American Association for the Study of Liver Diseases. Hepatology. 2018;67:328–357. [DOI] [PubMed] [Google Scholar]
- 19.Husain N, Blais P, Kramer J, et al. Nonalcoholic fatty liver disease (NAFLD) in the Veterans Administration population: development and validation of an algorithm for NAFLD using automated data. Aliment Pharmacol Ther. 2014;40:949–954. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Schreiner AD, Bian J, Zhang J, et al. The Association of Abnormal Liver Tests with Hepatitis C Testing in Primary Care. Am J Med. 2020;133:214–221 e211. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Pons E, Braun LM, Hunink MG, et al. Natural Language Processing in Radiology: A Systematic Review. Radiology. 2016;279:329–343. [DOI] [PubMed] [Google Scholar]
- 22.Angulo P, Hui JM, Marchesini G, et al. The NAFLD fibrosis score: a noninvasive system that identifies liver fibrosis in patients with NAFLD. Hepatology. 2007;45:846–854. [DOI] [PubMed] [Google Scholar]
- 23.Vallet-Pichard A, Mallet V, Nalpas B, et al. FIB-4: an inexpensive and accurate marker of fibrosis in HCV infection. comparison with liver biopsy and fibrotest. Hepatology. 2007;46:32–36. [DOI] [PubMed] [Google Scholar]
- 24.Xiao G, Zhu S, Xiao X, et al. Comparison of laboratory tests, ultrasound, or magnetic resonance elastography to detect fibrosis in patients with nonalcoholic fatty liver disease: A meta-analysis. Hepatology. 2017;66:1486–1501. [DOI] [PubMed] [Google Scholar]
- 25.Wright AP, Desai AP, Bajpai S, et al. Gaps in recognition and evaluation of incidentally identified hepatic steatosis. Dig Dis Sci. 2015;60:333–338. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Tapper EB, Hunink MG, Afdhal NH, et al. Cost-Effectiveness Analysis: Risk Stratification of Nonalcoholic Fatty Liver Disease (NAFLD) by the Primary Care Physician Using the NAFLD Fibrosis Score. PLoS One. 2016;11:e0147237. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Purnell TS, Calhoun EA, Golden SH, et al. Achieving Health Equity: Closing The Gaps In Health Care Disparities, Interventions, And Research. Health Aff (Millwood). 2016;35:1410–1415. [DOI] [PubMed] [Google Scholar]
- 28.Singal AG, Li X, Tiro J, et al. Racial, social, and clinical determinants of hepatocellular carcinoma surveillance. Am J Med. 2015;128:90 e91–97. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Kanwal F, Kramer JR, El-Serag HB, et al. Race and Gender Differences in the Use of Direct Acting Antiviral Agents for Hepatitis C Virus. Clin Infect Dis. 2016;63:291–299. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30.Ogden CL, Carroll MD, Kit BK, et al. Prevalence of childhood and adult obesity in the United States, 2011–2012. JAMA. 2014;311:806–814. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31.Zhu Y, Sidell MA, Arterburn D, et al. Racial/Ethnic Disparities in the Prevalence of Diabetes and Prediabetes by BMI: Patient Outcomes Research To Advance Learning (PORTAL) Multisite Cohort of Adults in the U.S. Diabetes Care. 2019;42:2211–2219. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32.Gawrieh S, Wilson LA, Cummings OW, et al. Histologic Findings of Advanced Fibrosis and Cirrhosis in Patients With Nonalcoholic Fatty Liver Disease Who Have Normal Aminotransferase Levels. Am J Gastroenterol. 2019;114:1626–1635. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 33.Quan H, Li B, Saunders LD, et al. Assessing validity of ICD-9-CM and ICD-10 administrative data in recording clinical conditions in a unique dually coded database. Health Serv Res. 2008;43:1424–1441. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.