Role of fatty liver index in risk-stratifying comorbid disease outcomes in non-alcoholic fatty liver disease

Background & Aims Population screening for non-alcoholic fatty liver disease (NAFLD) and associated comorbidities remains an unaddressed clinical need. We aimed to assess the utility of the fatty liver index (FLI) for risk stratification of NAFLD and related comorbidities using the UK Biobank. Methods Electronic health records and liver MRI-proton density fat fraction (PDFF) were used to define NAFLD cases. FLI was calculated and individuals with high alcohol intake and other liver diseases were excluded. Using listwise deletion analysis, the area under receiver-operating characteristic curve (AUROC) of FLI for NAFLD risk was determined. Thereafter, time-dependent covariate-adjusted Cox regression models were used to estimate FLI’s risk stratification potential for comorbidities of interest. Results FLI was derived for 327,800 individuals with a median age of 58 (IQR 51.5-64.5), of whom 59.8% were females. Using Perspectum Diagnostics and AMRA protocols as references, FLI identified the risk of NAFLD with AUROCs (95% CI, n) of 0.858 (0.848-0.867, n = 7,566) and 0.851 (0.844-0.856, n = 10,777), respectively. Intermediate and high-risk FLI was associated with increased cardiometabolic and malignant disease. In the first 3 years, high-risk FLI conferred an increased risk (adjusted hazard ratio, 95% CI) of ischaemic heart disease (2.14, 1.94-2.36), hypertension (2.84, 2.70-2.98), type 2 diabetes mellitus (4.55, 4.04-5.12), dyslipidaemia (2.48, 2.32-2.64), ischaemic stroke (1.31, 1.20-1.42) and hepatic malignancy (1.69, 1.23-2.30). FLI was not associated with risk of extrahepatic malignancy but was associated with a higher risk of specific cancers (colon, upper gastrointestinal and breast). All-cause mortality was similarly stratified by FLI, independently of non-invasive fibrosis scores. Conclusions FLI identifies NAFLD and holds potential for the risk stratification of cardiometabolic and malignant disease outcomes (including some extrahepatic malignancies), as well as all-cause mortality. Its use in population screening for primary and secondary prevention of NAFLD should be considered. Impact and implications Our analysis using the UK Biobank study shows the potential of the fatty liver index as a risk stratification tool for identifying the risk of developing NAFLD, ischaemic heart disease, ischaemic stroke, type 2 diabetes mellitus, hypertension, hyperlipidaemia, hepatic malignancy, specific metabolism-related malignancies and all-cause mortality. These results suggest that the fatty liver index should be considered as a non-invasive steatosis score that may help guide primary prevention strategies for NAFLD and related outcomes.


Introduction
Non-alcoholic fatty liver disease (NAFLD) is estimated to affect 25% of the world's population and is predicted to become the most prevalent liver disease globally, contributing to healthcare cost and burden. 1,24][5][6] As such, NAFLD can be considered as part of a disease syndrome encompassing multiple disease-related comorbidities.This syndrome or "comorbidome" may largely be preventable by addressing metabolic risk factors, but NAFLD-specific primary prevention measures are lacking.In the community, NAFLD is often detected through abnormal liver function tests, such as aminotransferase levels, which are not sensitive and can be normal in advanced disease. 7,8Abdominal ultrasound, often conducted for other reasons, is another modality which NAFLD is identified, but it performs poorly at detecting mild steatosis and is operator-dependent. 9As such, identification and clinical stratification of NAFLD risk in the community remains suboptimal, with most individuals with disease being undiagnosed and patient risk factors unaddressed. 10This is in contrast to cardiovascular diseases, where primary prevention is widely practiced by addressing risk factors, such as hypertension and dyslipidaemia.
There is interest in identifying chronic liver disease, including NAFLD, at the population level. 11To achieve this, automated Keywords: NAFLD; Fatty Liver Index; Comorbidities; Risk Stratification; non-invasive tests.3][14] For the identification of NAFLD specifically, a variety of non-invasive clinical tests have also been devised to assess the risk of the presence of steatosisone such test is the fatty liver index (FLI). 15FLI uses BMI, waist circumference, gamma-glutamyltransferase (GGT) and total triglyceride levels for its calculation.The components are simple measures commonly available in primary care and not dependent on coexistent disease. 16Since comorbidities of NAFLD can develop before or after the development of liver disease, FLI has been suggested as a promising tool to identify NAFLD-prone individuals and their risk of associated comorbid conditions in primary care, enabling stratification for primary or secondary prevention.However, FLI has not been widely implemented in clinical practice, and indeed there are no national screening programmes to identify individuals with NAFLD, with significant variation in practice between different localities. 17n this study, we have utilised the UK Biobank (UKB) to explore the utility of FLI to identify individuals at risk of NAFLD, and its ability to risk stratify incident comorbidities across multiple cardiovascular, metabolic, and malignant outcomes.Further, we mimicked current NAFLD risk stratification strategies by examining the risk of mortality predicted by FLI alone and in combination with two non-invasive fibrosis scores, fibrosis-4 index (FIB4) and NAFLD fibrosis score (NFS).

UK biobank study
The UKB is a prospective study that recruited 500,000 individuals aged between 40-69 years in the UK between 2006 and 2010. 18The study collected demographic, behavioural, physical, biological sampling, clinical bedside measurement and imaging data from its participants.There is also linked data to national inpatient and primary care records. 19Ethics approval was obtained from the UK North-West Multi-Centre Ethics Committee (ref: 16/ NW/0274) and informed consent was obtained from each participant for data usage in research purposes.The following data analysis stems from UKB research application ID 54764.

Study population
FLI was calculated for every individual in the UKB from measurements collected at the initial visit as published by Bedogni et al. 16 All participants with available FLI were included in the study.Individuals with FLI scores of <30, 30-59 and > − 60 were classified as having low, intermediate and high risk of steatosis, respectively.We calculated the weekly alcohol intake of participants as described previously by our group and excluded males and females who drank >21 and 14 weekly standard UK alcohol units, respectively. 20Further exclusion criteria included other causes of liver disease, including inherited, viral and alcohol, defined by ICD-9 and ICD-10 codes from linked hospital inpatient data (Tables S1 and S2) or positive serum virology results.NFS and FIB4, two non-invasive fibrosis scores, were also calculated and used to stratify patients into low, intermediate and high-risk groups as previously reported. 21,22FLD case/control definitions MRI imaging was performed on a Siemens 1.5 Tesla MAGNETOM Aera scanner (Siemens Healthineers, Erlangen, Germany).The analysis protocol for liver MRI-derived proton density fat fraction (PDFF) was developed by two separate companies, Perspectum Diagnostics and AMRA. 23,24At time of data extraction, there were 4,614 and 9,892 participants with available MRI-PDFF data from each company's protocol, respectively.We arbitrarily derived two case-control cohorts with overlap using each MRI protocol, whereby participants with > − 5% PDFF were defined as having NAFLD.To increase our sample size, ICD9/ICD10 codes and primary care data were used to identify additional NAFLD cases (Fig. S1, Table S3).Healthy individuals were defined by MRI-PDFF <5%.

Exposure, covariate and outcome variables
Covariates of interest for time-to-event modelling were selected based on previous literature for known associations with NAFLD: age, gender, Townsend deprivation index, smoking status, alcohol intake and type 2 diabetes mellitus (T2DM).We investigated ischaemic heart disease, ischaemic stroke, T2DM, hypertension, hyperlipidaemia, hepatic malignancy, extrahepatic malignancies, and all-cause mortality as events of interest.For extrahepatic malignancies, an additional sub-analysis specific to metabolism-associated cancers, namely colon (including rectal), upper gastrointestinal (GI: oesophageal and stomach) and breast cancer, was performed.Waist-hip ratio and BMI were not used as covariates due to inclusion of BMI and waist circumference in the calculation of FLI.All disease covariates and events of interest were identified through available linked clinical data registries (supplementary methods).The earliest date of recorded disease was considered the time of diagnosis.Exposures of interest are FLI classification for all outcomes and additionally non-invasive fibrosis scores when examining all-cause mortality.Participants who withdrew from the study or died were censored.The analysis was right censored using an arbitrary study end date of 31 December 2019, and the start date for each participant was defined at the UKB's initial visit date (time of FLI calculation).

Statistical analysis
The methodology used for statistical analysis is detailed in the supplementary methods.Briefly, descriptive statistics are presented as median and interquartile ranges and proportions for continuous and categorical descriptors, respectively.Performance of FLI in identifying combined incident and prevalent NAFLD was first assessed using receiver-operating characteristic (ROC) analyses, while calculated sensitivities, specificities, positive predictive and negative predictive values are presented for previously published cut-offs of FLI, against defined NAFLD cases and controls.ROC analysis was further performed for two other non-invasive steatosis tests, lipid accumulation product and hepatic steatosis index, calculated as previously published. 25,26ncidence disease rates were calculated for each FLI risk category and for misclassified individuals by FLI for NAFLD.Subsequently, Cox proportional hazard models were fitted to assess incidental disease risk stratification by FLI classification.Univariate models were first fitted to select variables to be included in subsequent multivariate models, with p <0.10 used as a loose selection threshold.Model assumptions were tested, and our models were modified to account for any violations of proportional hazards and non-linearity.Two multivariate models were fitted for each outcome.The first model included age and sex, while the second one additionally included all covariates selected from the univariate analysis.Incidental all-cause mortality was investigated in a similar manner through Cox regression modelling.In addition, modelling of FLI and one of the noninvasive fibrosis scores (NFS or FIB4) were investigated with and without adjustment for covariates.In all time-to-event analyses, a listwise deletion dataset was used.p values were adjusted by Bonferroni-correction for multiplicity of tests and usefulness of model fitting with FLI and/or fibrosis scores were tested with likelihood ratio tests where appropriate.Additional sensitivity analyses were performed for significance of components of FLI (BMI, waist circumference, triglyceride and GGT levels), and missing covariate data.Analyses was performed on R version 4.0.2 using package pROC, ggplot2, forestplot, survival and survminer.

FLI analysis cohort
From the initial UKB cohort of 502,460 participants, 327,800 had all data available for the calculation of FLI after implementing our exclusion criteria.Participant characteristics are described in Table 1.In contrast to individuals in the low-risk FLI, those with higher FLI risk tended to be older, male, smokers, and had higher BMI, waist-hip ratios, as well as greater levels of socioeconomic deprivation.Biochemistry results demonstrated that higher FLI risk was reflective of individuals with increased risk of metabolic and hepatic disease.The median hepatic MRI-PDFF values increased from low-to high-risk FLI.

Identification of NAFLD risk by FLI
The performance of FLI to identify combined incident and prevalent NAFLD in the UKB was first assessed against two NAFLD case-control definitions, which differ by the proprietary protocol used to measure MRI liver fat fraction.Using the Perspectum Diagnostics' protocol and clinical health records coding to define NAFLD in the UKB enabled an analysis of 7,656 individuals with available FLI data.Partitioning at a lower FLI cutoff (>30) resulted in a sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV) of 93.04%, 53.46%, 75.03% and 83.63%, respectively.Partitioning at the higher cut-off (>60) resulted in a sensitivity, specificity, PPV and NPV of 75.28%, 80.55%, 56.04% and 90.82% respectively (Fig. S2).Similarly, using the alternative MRI protocol derived by AMRA and clinical coding enabled an analysis of 10,777 participants with FLI data.Partitioning at the lower FLI cut-off resulted in a sensitivity, specificity, PPV and NPV of 92.49%, 53.38%, 84.15% and 72.65%, respectively.At the higher cut-off, the values were 73.39%, 80.59%, 69.34% and 83.51%, respectively.On ROC analyses, the AUROC was 0.858 (95% CI 0.848-0.867)and 0.851 (95% CI 0.844-0.856)for these two definitions of NAFLD (Fig. 1).With FLI, 24.3% and 15.6% would be classified as low risk despite having NAFLD, whereas 9.4% and 16.8% would be classified as high risk despite having a normal liver, against our two respective NAFLD definitions.This corresponds to a misclassification rate of 13.3% and 16.4%, when excluding the intermediate-risk group from the calculation.Further comparison was made with two other non-invasive steatosis scores, hepatic steatosis index and lipid accumulation product, which showed inferior ROC results compared to FLI in identifying the risk of NAFLD against these two case-control definitions (Fig. 1).

Univariate analysis of incident outcomes
We next assessed whether FLI and potential covariates may predict the selected comorbid outcomes of NAFLD.The incident rates for low-to high-risk FLI (per 100 person-years) increased from 0.31 to 0.93 for ischaemic heart disease, 0.11 to 0.20 for ischaemic stroke, 1.07 to 3.25 for hypertension, 0.56 to 1.65 for dyslipidaemia, 0.25 to 1.03 for T2DM, 0.01 to 0.02 for hepatic malignancy and 1.22 to 1.56 for extrahepatic malignancy (Figs 2, 3, Table S4).Having determined the misclassification rates of FLI, we also calculated the subgroup absolute incidence rates of individuals with NAFLD classified as low-risk FLI and healthy persons with high-risk FLI (Table S5).
Using a listwise deletion approach, over 190,000 individuals were available for univariate and downstream multivariable timeto-event analyses across disease outcomes of interest, with a median follow-up of >10 years (Fig. S3).Univariate analysis showed FLI was associated with all disease outcomes.All covariates selected on univariate analysis also passed the threshold for selection (p <0.10) in multivariate models, with some exceptions (Table S6).These were Townsend index (p = 0.69) and alcohol intake (p = 0.55) for hepatic malignancy and T2DM (p = 0.16) for extrahepatic malignancy.Similarly, for all-cause mortality, all covariates passed the selection threshold for multivariable analysis.For  ) had increased risk of all-cause mortality compared to low-risk counterparts (Fig. 3 and Fig. S4, Table S7).

Multivariate analysis of FLI for incident disease outcomes
To assess the ability of FLI to stratify the incidence of selected NAFLD comorbidities, Cox proportional hazards models were fitted with covariates selected from univariate analysis.Of the disease outcomes investigated, FLI met the proportional hazards assumption required for Cox regression on examination of scaled Schoenfeld residuals for ischaemic stroke and hepatic malignancy.Given this, time-dependent coefficient modelling was employed for FLI for all other disease outcomes investigated.For the two outcomes where time-dependent models were not applied, individuals with intermediate-risk FLI had an increased incident risk of ischaemic stroke (HR [95% CI] 1.21 [1.11-1.33],p = 2.55x10 -5 ), but similar risk of hepatic malignancy (HR [95% CI] 1.01 [0.70-1.45],p = 0.965) when compared to those with lowrisk FLI.However, high-risk FLI risk-stratified both diseases (HR [95% CI] 1.31 [1.20-1.42],p = 1.26x10 -9 and 1.69 [1.23-2.32],p = 0.001, respectively) (Fig. 4).
For time-dependent modelling of FLI, we arbitrarily split the follow-up period into 3-yearly intervals to allow for more refined estimation of risk by FLI during follow-up.Our analysis shows that the incident risk of ischaemic heart disease, hypertension, dyslipidaemia, and T2DM was informed by the non-invasive steatosis score, for both the intermediate-and high-risk category at all intervals of the follow-up period, but not for extrahepatic malignancy (Fig. 4, Table S8).To illustrate, an intermediate-risk FLI was associated with a 1.52-fold [95% CI 1.37-1.70]increased risk of ischaemic heart disease in the first 3 years of follow-up, which reduced to 1.31-fold [95% CI 1.17-1.47]at >9 years of follow-up, when compared to individuals with low-risk FLI.Similarly, highrisk FLI was associated with a 2.14-fold [95% CI 1.94-2.36]and 1.66-fold [95% CI 1.49-1.85]increase in risk when compared to low-risk FLI for these two periods of follow-up, respectively.There was a general trend in the reduction of risk with both intermediate-and high-risk FLI with longer periods of follow-up, except for T2DM.The reduction in estimated risk was most prominent when 3 JHEP Reports 2023 vol. 5 j 100896   In risk-stratifying T2DM, our analysis showed a bimodal pattern of fold-change in the incident risk given by FLI, with higher hazard ratios during the first 3 years and after 6 years of follow-up.When compared to individuals with low-risk FLI (

FLI risk stratification of metabolism-related cancer
As the value of FLI risk stratification was minimal for overall extrahepatic malignancy, we surmised that non-metabolismrelated cancer risk is poorly captured by the components of FLI, which correlate strongly with metabolic disease states, and may therefore dilute estimates of metabolism-related cancers.Additional analysis was therefore performed for selected cancers, which have a known association to metabolic syndrome: colorectal, upper GI (oesophageal and stomach) and breast cancer (Fig. 5).High-risk FLI conferred an increased risk of developing colorectal, upper GI and breast cancer (HR [95% CI] 1.2 [1.08-1.34],1.52 [1.25-1.86]and 1.17 [1.09-1.26],respectively), when compared to low-risk FLI in the covariate-adjusted model.An intermediate-risk FLI was associated with a 1.09-fold [1.01-1.17]higher risk of breast cancer but not with an increased risk of the luminal GI cancers tested.FLI and non-invasive fibrosis scores stratify all-cause mortality Finally, all-cause mortality was examined in a similar manner.In the covariate-adjusted model, high-risk FLI was associated with a 1.2-fold [95% CI 1.14-1.27]increased risk of death from any cause, whereas individuals with intermediate-risk scores had a similar risk of death as those with low-risk FLI (Fig. 4).As current advanced liver disease risk assessment involves confirming the presence of fibrosis, which is independently associated with mortality, we examined whether FLI provides any independent information over non-invasive fibrosis scores (NFS or FIB4), a surrogate to the presence of fibrotic liver disease.To do this, we first fitted separate stratified Cox models by FLI risk category, which showed that high-risk NFS and FIB4 predicted an increased risk of all-cause mortality in individuals with intermediate-or high-risk, but not low-risk, FLI (Fig. 6 and Fig. S7).Incorporation of FLI with each non-invasive fibrosis score into a unified covariate-adjusted model showed that only the high-risk category of each non-invasive fibrosis score was significantly associated with a higher risk of death from all-causes.In each of these models, individuals with a high-risk FLI had a 1.2-fold [95% CI: 1. 13     S9).

Sensitivity analysis for components of FLI and missing data
Given the strong risk differences in FLI risk groups for the disease outcomes investigated, sensitivity analyses were performed to investigate whether components of FLI vary in their risk predictive capacity for these different outcomes.Using multivariate analysis to determine the role of the individual FLI components showed that ischaemic heart disease, T2DM, hypertension, dyslipidaemia and extrahepatic malignancy were informed by all components of FLI.The only outcome associated with three FLI components was all-cause mortality (waist circumference, BMI and GGT levels).Hepatic malignancy and ischaemic stroke were only significantly associated with two components of FLI (Table S10).Of note, waist circumference and BMI exhibited opposite effects towards the risk of extrahepatic malignancy.A further sensitivity analysis was performed for the main outcome analysis to investigate effects of missing data when using listwise deletion.To crudely investigate this, we ran the same time-toevent regression analysis for disease outcomes of interest with a listwise deletion cohort excluding alcohol intake as a variable, which had the largest proportion of missingness (of up to 20%).Comparing it with the main analysis cohort, little difference in risk estimates was shown for FLI, whereby estimates of risk were consistently more conservative in the final models used (Table S11).

Discussion
In one of the largest studies undertaken to date, we have demonstrated that a higher FLI from a single measurement performed in the UKB identified individuals at risk of existent and incident NAFLD and its comorbidities.A previous metaanalysis of 27,221 individuals showed that the sensitivity, specificity, PPV, and NPV for FLI were 81%, 65%, 53%, and 84%, respectively, for the lower cut-off, while the corresponding values at the higher cut-off were 44%, 90%, 67%, and 76%, respectively. 27Our data compare favourably with these estimates, but with slightly higher sensitivity values and slightly lower specificity at both cut-off thresholds.A previous study comparing FLI with 1 H-magnetic resonance spectroscopy showed an AUROC of 0.79 compared with our value of 0.85, suggesting FLI has good diagnostic value. 28However, in contrast to its use for point of care diagnosis in the previous studies, our results suggest that FLI may be capable of identifying NAFLD and associated comorbidities over a lifetime period up to middle and early old-age.This is reflected by the age group recruited in the UKB and reflects risk stratification of a one-off score between the ages of 40-69.
A misclassification rate of FLI for NAFLD is also observed from our analysis.This is expected given only a small proportion of risk is captured by the index for the complex trait disorder.As such, it is important to distinguish that our time-to-event analysis results do not reflect a direct assessment of NAFLD towards the risk of developing the disease, but rather the risk from metabolic factors associated with NAFLD development, which it shares with many other metabolic conditions.FLI thus provides a validated score to link NAFLD with incident disease outcome assessment, which has not been considered in the development of many established risk stratification tools in use, such as QRISK3 for cardiovascular disease. 29Misclassification may   hamper the prediction capacity of disease outcomes, but from our results, there were generally comparable comorbid event rates in NAFLD cases with a low-risk FLI and those with healthy livers and a high-risk FLI.
Bearing this in mind, our results show that the risk of ischaemic heart disease, stroke, diabetes, hypertension, hyperlipidaemia, hepatic malignancy, and all-cause mortality was stratified by FLI longitudinally over a median follow-up >10 years, after adjustment of known covariates associated with NAFLD development.1][32][33][34] However, we have also shown that FLI predicts of the risk of hepatic and certain extrahepatic malignancies as well, which is a novel finding.A previous study has examined FLI stratification of cardiac events and stroke within the UKB, in which the authors used FLI deciles for modelling. 35This may hamper ease of clinical interpretation given the index was originally designed as a tripartite classification system (low, intermediate and high risk for NAFLD).The presented results show a similar direction of the estimated effect, with our use of time-dependent models potentially providing more precise risk estimates in an updated analysis with a longer follow-up period.Separately, FLI has also been shown to risk stratify incident hypertension and dyslipidaemia in comparatively smaller Asian cohorts. 31,36Our results therefore provide further information on FLI's utility for risk stratification, specifically in a UK cohort composed mainly of Caucasians.
We show that a high-risk FLI (>60) was predictive of a higher risk of hepatic malignancy when compared to an index score of <30.This conforms well with the current understanding that NAFLD-related metabolic risk factors increase the risk of disease progression.However, there was no overall risk prediction of incident extrahepatic malignancy, which is contrary to the association of NAFLD with the development of and mortality from extrahepatic malignancy. 37,38The discrepancy is likely related to differences in the outcome investigated, whereby cancer-related mortality is more strongly predicted by metabolic comorbidities reflected by FLI but potentially less so with incidence of all types of cancer.Supporting this, in a recent study investigating incident cancer and cancer-related mortality, using raised alanine aminotransferase levels as a surrogate marker for NAFLD in three independent Scottish cohorts, it was shown that FLI predicted risk of mortality from cancer. 39However, the differences in modelling and covariates with the present study could also influence these results.1][42] We also examined specific incidences of metabolism-related malignancies, namely colorectal, breast and upper GI cancers (oesophageal and gastric), and show that high-risk FLI was able to identify increased risk relative to low-risk scoring individuals, independent of common risk factors for NAFLD.
Lastly, we demonstrate that high-risk FLI was also associated with increased all-cause mortality independent of age, sex, diabetes, alcohol intake and socioeconomic deprivation, consistent with NAFLD and metabolic risk factors being predictive factors.To mimic the current two-step strategy of clinical NAFLD assessment which first stratifies risk of disease and then subsequently risk of advance disease, we explored whether FLI combined with non-invasive fibrosis scores can be used to assess mortality risk.Results show that non-invasive fibrosis scores and FLI independently stratified risk of all-cause mortality.Furthermore, in individuals with intermediate-and high-risk FLI, both fibrosis scores predicted all-cause mortality in their high-risk category in our adjusted model.This suggests FLI adds information to risk assessment of all-cause mortality and can potentially be used in tandem with fibrosis scores for this purpose.
Our study has certain limitations.The use of a listwise deletion cohort may have created bias, but our sensitivity analysis suggests our estimates were relatively precise when discounting for the largest missing variable (alcohol intake).Given the observational nature of the UKB, FLI's use in the general population and in individuals with NAFLD as a means of follow-up should be directly assessed in future intervention studies.Lastly, cost-effectiveness analysis will be required to build a case for implementation, which is beyond the scope of our study.With different components of FLI showing different directions of risk in certain disease outcomes assessed, further refinement of the index for specific outcomes may be attractive, but this must be balanced against the fact that diversification of risk stratification tools may increase complexity and thereby difficulty in clinical implementation.
Overall our results suggest that the FLI is an attractive potential option to identify individuals at risk of NAFLD and related comorbidities within the community, particularly during routine follow-up for other metabolic diseases in primary care.Whether the degree of misclassification we have identified is clinically acceptable warrants further debate, but it is probably better than the current standard of care which varies with geography and mostly amounts to doing nothing. 17Clearly, as we have observed, the risk effect estimates observed from a single FLI may change over time, and thus repeated FLI measurements may provide better indication of comorbid disease and mortality risk, as shown by a previous study. 43n summary, the FLI identified prevalent and incident NAFLD and stratified the risk of incident cardiovascular and metabolic diseases, hepatic malignancy, and some extrahepatic cancers.FLI alone or in combination with NFS or FIB-4 independently enabled the risk assessment of all-cause mortality.FLI shows potential as a tool to stratify individuals for further assessment and to guide the development of prevention strategies for NAFLD and its related comorbidities in the population.
bootstrapping data 2000 times at intervals of 5% for specificities and the distribution smoothened with the loess method.

Time to Event Analysis
Cox proportional hazards model was used for time to event analyses.Study start date was defined at recruitment to UK Biobank when FLI score was derived.Listwise deletion dataset was used in Cox regression modelling and individuals who have experienced the event of interest prior to study start date were excluded from the respective time to event analysis.Univariate regression was first performed for each time to event outcome for exposures of interest.A p-value of <0.1 was used as threshold for inclusion for downstream multivariate analysis.To assess linearity of Cox regression model, penalised splines were applied to continuous covariates to assess for significant non-linear component in univariate cox regression.Exposure covariates with significant non-linear component (p < 0.05) were subsequently regressed by applying a natural cubic spline for better model fitting.T2DM, as an exposure of interest, was fitted as a time dependent covariate for all models, except in the model where T2DM is the event of interest.Subsequently, two models were fitted to assess risk stratification by FLI: the first model is termed the "Basic Model", which included covariates age and sex, and the second model included additional covariates brought forward from univariate regression.Finally, proportional hazards assumption was tested in these models using Schoenfeld residuals.If breaches of assumptions were found, defined again by significance level of p<0.05, time dependent coefficients were created using a step function, which divided follow-up from FLI into 0-3, 4-6, 7-9 and >9 years.This division in follow-up period as a step function from UK Biobank recruitment was chosen from the variation in median >10 years observed in our UK Biobank participants in all our time to event analyses.Models whereby FLI was treated as categorical variable was additionally assessed with likelihood ratio test between models with and without FLI as a covariate.This allowed testing whether FLI provided additional predictive risk stratification for event of interest as a single covariate.Both dummy variable p-values within the model and model comparison tests p-values are reported.P-values <0.05 were considered significant after Bonferonni correction for multiple testing (this is 26 models/tests for FLI, p < 0.0019)

Additional and Sensitivity Analyses
An additional analysis was performed to look at incidence rates of misclassified individuals, with high-risk FLI but normal livers on MRI and low-risk FLI with evidence/diagnosis of NAFLD.Two sensitivity analyses were performed.Firstly, examining time to event analysis using components used to calculate FLI (BMI, waist circumference, serum GGT and total triglyceride levels) in multivariate Cox regression.This is to allow estimate of significance and risk estimate direction for each of these components toward the incident disease outcome.Secondly, listwise deletion cohort with/without accounting for missing data from alcohol intake for each time to event analyses was examined.Alcohol has the highest overall proportion of missing data of 31% and allows best chance to see examine whether exclusion of missing broadly altered the model risk estimates.

Incident All-cause Mortality with FLI and Fibrosis Scores
For incident all-cause mortality, a similar approach was adopted for Cox regression analysis investigating risk stratification by FLI and non-invasive fibrosis scores (NAFLD Fibrosis Score and FIB4).Univariate modelling and assumptions testing was performed similarly as detailed above for FLI on incident disease outcomes.
Separate multivariate modelling for fibrosis scores was further performed after stratifying individuals by FLI risk to examine NFS and FIB4's performance in these subgroups.Further multivariate modelling included FLI, and a single fibrosis score (NFS or FIB4) was used or with all exposure covariates brought forward by univariate analysis.This allowed examining differences in risk effect estimates given by fibrosis scores and the effect exposure covariates have.P-values <0.05 were considered significant after Bonferonni correction for multiple testing.and their associated p-value.Bolded and italicised values indicate that a time-dependent coefficient was fitted for the variable and the values indicate the first three years of follow-up from time of FLI calculation.Splines were not fitted in this instance to allow for ease of interpretation of the significance of the FLI component in predicting the comorbid outcome and its general direction of effect.Caution needs to be had with interpreting this due to likely correlation between these components and the magnitude of its risk effect, which is likely to be imprecise.

2.0
Please confirm for randomised controlled trials all versions of the clinical protocol are included in the submission.These will be published online as supplementary information.

Fig. 1 .
Fig. 1.Performance of non-invasive steatosis scores estimating combined prevalent and incident NAFLD risk.Receiver-operating characteristic curves of FLI (A, B), hepatic steatosis index (C, D) and lipid accumulation product (E, F) against two case-control definitions of NAFLD.First definition (left column) and second definition (right column) based on MRI results derived from Perspectum Diagnostics and AMRA, respectively.AUC and best binary cut-off shown, with sensitivity and specificity values.95% CI were derived from bootstrapping (n = 1,000) and shown in shaded areas.FLI, fatty liver index; NAFLD, non-alcoholic fatty liver disease.5 JHEP Reports 2023 vol. 5 j 100896

Fig. 4 .
Fig.4.FLI risk stratification of incident comorbidities.Forest plot shows the HR estimates and 95% CIs comparing intermediate-and high-risk FLI with low-risk individuals in multivariable Cox regression analysis, adjusted for covariates selected from univariate regression.The p value in the diagram shows log-likelihood ratio test between models with and without FLI as a variable in the regression, assessing the value of FLI in risk prediction.FLI, fatty liver index; HR, hazard ratio.

Fig. 5 .
Fig. 5. FLI risk stratification of selected incident metabolism-related cancers.Shown are Kaplan-Meier curves for (A) colon, (B) breast and (C) upper GI cancer (oesophageal and gastric) for low-, intermediate-and high-risk FLI, as defined by scores of <30, 30-59 and > − 60. 95% CIs are plotted for general interpretation (shaded area) with + signs indicating data censoring.(D) The forest plot shows the HRs derived from Cox proportional hazards regression adjusted for selected covariates for each of these cancers.FLI, fatty liver index; HR, hazard ratio.

Fig. S2 -
Fig. S2 -Diagnostic Performance at Low and High Cut-offs of FLI The 1 st definition denotes case control definition derived from MRI proton density fat fraction from Perspectum Diagnostics and 2 nd definition originates from AMRA; both cohorts haring clinically/primary care coded cases.

Fig. S5 -
Fig. S5 -Incident All-cause Mortality By Fibrosis Scores Stratified By FLI Risk

Table 1 .
Characteristics of patient cohort.Table showing characteristics of UKB participants with data available to calculate FLI and also by the index risk categories.Summary statistics are shown in median and interquartile ranges for continuous values and percentages for categorical values, with the number of available for analysis.

Table 2 .
Risk stratification of all-cause mortality when combining FLI and non-invasive Table shows time-dependent Cox regression results of FLI and one non-invasive fibrosis scores (FIB4 or NAFLD fibrosis score), adjusted and not adjusted for covariates, which include age, sex, Townsend index, type 2 diabetes mellitus, smoking status, and alcohol intake.

Unadjusted Incidence Rates of Comorbidites in Individuals Misclassified by FLI with and without NAFLD.
Table showing incidence rate of comorbid outcomes in misclassified individuals whereby they have low-risk FLI but found to have hepatic steatosis and high-risk FLI, but normal liver based on MRI-PDFF and ICD9/10 classification.The definition of NAFLD is based on MRI-PDFF values derived by Perspectum Diagnostics (our first definition) and AMRA (our second definition), whereas the diagnostic code is obtained from a common source: diagnostic codes from hospital inpatient records.

S6 -Univariate Cox Regression: FLI and Covariates for Disease Outcomes Table
showing univariate cox regression modelling outputs (hazards ratios and 95% CI) with associated with p-values disease outcomes with all exposures of interest.

Multivariable Covariate Adjusted Regression Outputs with Components of FLI Table
Table output of multivariate cox regression analysis for incident ischaemic heart disease, stroke, hypertension, dyslipidaemia, hepatic malignancy, and extrahepatic malignancy.showing multivariate Cox regression model outputs for components of FLI (BMI, waist circumference, GGT and total triglyceride levels)

Sensitivity Analysis of Listwise Deletion Estimates Accounting for Missing Data in Alcohol Intake Table
below shows hazard ratios for intermediate and high-risk FLI with reference to low-risk FLI derived from multivariable of Cox proportional hazards models fitted without alcohol intake as covariate, for all comorbid outcomes of interest.Alcohol intake has the highest rate of missing and therefore allow assessment of risk estimates effect towards missing data.Models for each time to event outcome used a listwise deletion dataset accounting (not shaded) and not accounting (shaded) for missing data in alcohol intake to examine differences in risk effect estimation.(1.59 -1.84) 3.01E-48 2.51 (2.35 -2.67) 1.36E-167 4-6 years 1.53 (1.42 -1.64) 1.77E-31 2.11 (1.98 -2.25) 7.93E-115 7-9 years 1.56 (1.45 -1.68) 4.44E-33 2.17 (2.03 -2.32) 1.71E-117