Predicting and elucidating the etiology of fatty liver disease: A machine learning modeling and validation study in the IMI DIRECT cohorts
Authors:
Naeimeh Atabaki-Pasdar aff001; Mattias Ohlsson aff002; Ana Viñuela aff004; Francesca Frau aff007; Hugo Pomares-Millan aff001; Mark Haid aff008; Angus G. Jones aff009; E. Louise Thomas aff010; Robert W. Koivula aff001; Azra Kurbasic aff001; Pascal M. Mutie aff001; Hugo Fitipaldi aff001; Juan Fernandez aff001; Adem Y. Dawed aff012; Giuseppe N. Giordano aff001; Ian M. Forgie aff012; Timothy J. McDonald aff009; Femke Rutters aff014; Henna Cederberg aff015; Elizaveta Chabanova aff016; Matilda Dale aff017; Federico De Masi aff018; Cecilia Engel Thomas aff017; Kristine H. Allin aff019; Tue H. Hansen aff019; Alison Heggie aff022; Mun-Gwan Hong aff017; Petra J. M. Elders aff023; Gwen Kennedy aff024; Tarja Kokkola aff025; Helle Krogh Pedersen aff019; Anubha Mahajan aff026; Donna McEvoy aff022; Francois Pattou aff027; Violeta Raverdy aff027; Ragna S. Häussler aff017; Sapna Sharma aff028; Henrik S. Thomsen aff016; Jagadish Vangipurapu aff025; Henrik Vestergaard aff019; Leen M. ‘t Hart aff014; Jerzy Adamski aff008; Petra B. Musholt aff035; Soren Brage aff036; Søren Brunak aff018; Emmanouil Dermitzakis aff004; Gary Frost aff038; Torben Hansen aff019; Markku Laakso aff025; Oluf Pedersen aff019; Martin Ridderstråle aff041; Hartmut Ruetten aff007; Andrew T. Hattersley aff009; Mark Walker aff022; Joline W. J. Beulens aff014; Andrea Mari aff043; Jochen M. Schwenk aff017; Ramneek Gupta aff018; Mark I. McCarthy aff011; Ewan R. Pearson aff012; Jimmy D. Bell aff010; Imre Pavo aff046; Paul W. Franks aff001
Authors place of work:
Genetic and Molecular Epidemiology Unit, Department of Clinical Sciences, Lund University, Malmö, Sweden
aff001; Computational Biology and Biological Physics Unit, Department of Astronomy and Theoretical Physics, Lund University, Lund, Sweden
aff002; Center for Applied Intelligent Systems Research, Halmstad University, Halmstad, Sweden
aff003; Department of Genetic Medicine and Development, University of Geneva Medical School, Geneva, Switzerland
aff004; Institute for Genetics and Genomics in Geneva, University of Geneva Medical School, Geneva, Switzerland
aff005; Swiss Institute of Bioinformatics, Geneva, Switzerland
aff006; Sanofi-Aventis Deutschland, Frankfurt am Main, Germany
aff007; Research Unit Molecular Endocrinology and Metabolism, Helmholtz Zentrum München, Neuherberg, Germany
aff008; Institute of Biomedical and Clinical Science, College of Medicine and Health, University of Exeter, Exeter, United Kingdom
aff009; Research Centre for Optimal Health, School of Life Sciences, University of Westminster, London, United Kingdom
aff010; Oxford Centre for Diabetes, Endocrinology and Metabolism, Radcliffe Department of Medicine, University of Oxford, Oxford, United Kingdom
aff011; Division of Population Health and Genomics, School of Medicine, University of Dundee, Ninewells Hospital, Dundee, United Kingdom
aff012; Blood Sciences, Royal Devon and Exeter NHS Foundation Trust, Exeter, United Kingdom
aff013; Department of Epidemiology and Biostatistics, Amsterdam Public Health Research Institute, Amsterdam UMC, Amsterdam, the Netherlands
aff014; Department of Endocrinology, Abdominal Centre, Helsinki University Hospital, Helsinki, Finland
aff015; Department of Diagnostic Radiology, Copenhagen University Hospital Herlev Gentofte, Herlev, Denmark
aff016; Affinity Proteomics, Science for Life Laboratory, School of Engineering Sciences in Chemistry, Biotechnology and Health, KTH Royal Institute of Technology, Solna, Sweden
aff017; Department of Health Technology, Technical University of Denmark, Kongens Lyngby, Denmark
aff018; Novo Nordisk Foundation Center for Basic Metabolic Research, Faculty of Health and Medical Sciences, University of Copenhagen, Copenhagen, Denmark
aff019; Center for Clinical Research and Prevention, Bispebjerg and Frederiksberg Hospital, Copenhagen, Denmark
aff020; Department of Cardiology and Endocrinology, Slagelse Hospital, Slagelse, Denmark
aff021; Institute of Cellular Medicine, Newcastle University, Newcastle upon Tyne, United Kingdom
aff022; Department of General Practice, Amsterdam Public Health Research Institute, Amsterdam UMC, Amsterdam, the Netherlands
aff023; Immunoassay Biomarker Core Laboratory, School of Medicine, University of Dundee, Ninewells Hospital, Dundee, United Kingdom
aff024; Internal Medicine, Institute of Clinical Medicine, University of Eastern Finland, Kuopio, Finland
aff025; Wellcome Centre for Human Genetics, University of Oxford, Oxford, United Kingdom
aff026; University of Lille, Inserm, UMR 1190, Translational Research in Diabetes, Department of Endocrine Surgery, CHU Lille, Lille, France
aff027; German Center for Diabetes Research, Neuherberg, Germany
aff028; Unit of Molecular Epidemiology, Institute of Epidemiology II, Helmholtz Zentrum München, Neuherberg, Germany
aff029; Steno Diabetes Center Copenhagen, Gentofte, Denmark
aff030; Department of Cell and Chemical Biology, Leiden University Medical Center, Leiden, The Netherlands
aff031; Molecular Epidemiology, Department of Biomedical Data Sciences, Leiden University Medical Center, Leiden, the Netherlands
aff032; Lehrstuhl für Experimentelle Genetik, Wissenschaftszentrum Weihenstephan für Ernährung, Landnutzung und Umwelt, Technische Universität München, Freising, Germany
aff033; Department of Biochemistry, Yong Loo Lin School of Medicine, National University of Singapore, Singapore
aff034; Diabetes Division, Research and Development, Sanofi, Frankfurt, Germany
aff035; MRC Epidemiology Unit, University of Cambridge, Cambridge, United Kingdom
aff036; Novo Nordisk Foundation Center for Protein Research, Faculty of Health and Medical Sciences, University of Copenhagen, Copenhagen, Denmark
aff037; Section for Nutrition Research, Department of Metabolism, Digestion and Reproduction, Imperial College London, London, United Kingdom
aff038; Faculty of Health Sciences, University of Southern Denmark, Odense, Denmark
aff039; Department of Medicine, University of Eastern Finland and Kuopio University Hospital, Kuopio, Finland
aff040; Clinical Pharmacology and Translational Medicine, Novo Nordisk, Søborg, Denmark
aff041; Julius Center for Health Sciences and Primary Care, University Medical Center Utrecht, Utrecht, the Netherlands
aff042; Institute of Neuroscience, National Research Council, Padua, Italy
aff043; NIHR Oxford Biomedical Research Centre, Oxford University Hospitals NHS Foundation Trust, John Radcliffe Hospital, Oxford, United Kingdom
aff044; OMNI Human Genetics, Genentech, South San Francisco, California, United States of America
aff045; Eli Lilly Regional Operations, Vienna, Austria
aff046; Department of Nutrition, Harvard School of Public Health, Boston, Massachusetts, United States of America
aff047
Published in the journal:
Predicting and elucidating the etiology of fatty liver disease: A machine learning modeling and validation study in the IMI DIRECT cohorts. PLoS Med 17(6): e32767. doi:10.1371/journal.pmed.1003149
Category:
Research Article
doi:
https://doi.org/10.1371/journal.pmed.1003149
Summary
Background
Non-alcoholic fatty liver disease (NAFLD) is highly prevalent and causes serious health complications in individuals with and without type 2 diabetes (T2D). Early diagnosis of NAFLD is important, as this can help prevent irreversible damage to the liver and, ultimately, hepatocellular carcinomas. We sought to expand etiological understanding and develop a diagnostic tool for NAFLD using machine learning.
Methods and findings
We utilized the baseline data from IMI DIRECT, a multicenter prospective cohort study of 3,029 European-ancestry adults recently diagnosed with T2D (n = 795) or at high risk of developing the disease (n = 2,234). Multi-omics (genetic, transcriptomic, proteomic, and metabolomic) and clinical (liver enzymes and other serological biomarkers, anthropometry, measures of beta-cell function, insulin sensitivity, and lifestyle) data comprised the key input variables. The models were trained on MRI-image-derived liver fat content (<5% or ≥5%) available for 1,514 participants. We applied LASSO (least absolute shrinkage and selection operator) to select features from the different layers of omics data and random forest analysis to develop the models. The prediction models included clinical and omics variables separately or in combination. A model including all omics and clinical variables yielded a cross-validated receiver operating characteristic area under the curve (ROCAUC) of 0.84 (95% CI 0.82, 0.86; p < 0.001), which compared with a ROCAUC of 0.82 (95% CI 0.81, 0.83; p < 0.001) for a model including 9 clinically accessible variables. The IMI DIRECT prediction models outperformed existing noninvasive NAFLD prediction tools. One limitation is that these analyses were performed in adults of European ancestry residing in northern Europe, and it is unknown how well these findings will translate to people of other ancestries and exposed to environmental risk factors that differ from those of the present cohort. Another key limitation of this study is that the prediction was done on a binary outcome of liver fat quantity (<5% or ≥5%) rather than a continuous one.
Conclusions
In this study, we developed several models with different combinations of clinical and omics data and identified biological features that appear to be associated with liver fat accumulation. In general, the clinical variables showed better prediction ability than the complex omics variables. However, the combination of omics and clinical variables yielded the highest accuracy. We have incorporated the developed clinical models into a web interface (see: https://www.predictliverfat.org/) and made it available to the community.
Trial registration
ClinicalTrials.gov NCT03814915.
Keywords:
Oral glucose suppression test – insulin – Fats – machine learning – Proteomic databases – Fatty liver – Metabolomics – diabetes mellitus
Introduction
Non-alcoholic fatty liver disease (NAFLD) is characterized by the accumulation of fat in hepatocytes in the absence of excessive alcohol consumption. NAFLD is a spectrum of liver diseases, with its first stage, known as simple steatosis, defined as liver fat content ≥5% of total liver weight. Simple steatosis can progress to non-alcoholic steatohepatitis (NASH), fibrosis, cirrhosis, and eventually hepatocellular carcinoma. In NAFLD, triglycerides (TG) accumulate in hepatocytes, and liver insulin sensitivity is diminished, promoting hepatic gluconeogenesis, thereby raising the risk of type 2 diabetes (T2D) or exacerbating the disease pathology in those with diabetes [1–5]. Growing evidence also links an increased risk of cardiovascular events with NAFLD [6,7].
The prevalence of NAFLD is thought to be around 20%–40% in the general population in high-income countries, with numbers growing worldwide, imposing a substantial economic and public health burden [8–11]. However, the exact prevalence of NAFLD has not been clarified, in part because liver fat is difficult to accurately assess. Liver biopsy, magnetic resonance imaging (MRI), ultrasound, and liver enzyme tests are often used for NAFLD diagnosis, but the invasive nature of biopsies, the high cost of MRI scans, the non-quantitative nature and low sensitivity of conventional ultrasounds, and the low accuracy of liver enzyme tests are significant limitations [12–14]. To address this gap, several liver fat prediction indices have been developed, but none of these has sufficiently high predictive ability to be considered a gold standard [12].
The purpose of this study was to use machine learning to identify novel molecular features associated with NAFLD and combine these with conventional clinical variables to predict NAFLD. Our models include variables that are likely to be informative of disease etiology, some of which may be of use in clinical practice.
Methods
Participants (IMI DIRECT)
The primary data utilized in this study were generated within the IMI DIRECT consortium, which includes persons with diabetes (n = 795) and without diabetes (n = 2,234). All participants provided informed written consent, and the study protocol was approved by the regional research ethics committees for each clinical study center. Details of the study design and the core characteristics are provided elsewhere [15,16].
Measures (IMI DIRECT)
A T2*-based multiecho technique was used to derive liver fat content from MRI [17,18], and the percentage values were categorized as fatty (≥5%) or non-fatty (<5%) to define the outcome variable. We elected not to attempt quantitative prediction of liver fat content, as this would require a much larger dataset to be adequately powered. A frequently-sampled 75-g oral glucose tolerance test (OGTT) or a frequently sampled mixed-meal tolerance test (MMTT) was performed, from which measures of glucose and insulin dynamics were calculated, as previously described [15,16]. Of 3,029 IMI DIRECT participants, 50% (n = 1,514) had the liver fat MRI data (503 with diabetes and 1,011 without diabetes). The distribution of the liver fat data among different centers and cohorts is shown in S1 and S2 Figs.
The list of the clinical input (predictor) variables (n = 58), including anthropometric measurements, plasma biomarkers, and lifestyle factors, are shown in S1 Table. These clinical variables were controlled for center effect by deriving residuals from a linear model including each clinical variable in each model; these residuals were then inverse normalized and used in subsequent analyses. Inverse normal transformation is a nonparametric method that replaces the data quantiles by quantiles from the standard normal distribution in order to reduce the impact of outliers and deviation from a normal distribution.
A detailed overview of participant characteristics for the key variables is shown in Table 1 for all IMI DIRECT participants with MRI data. There were no substantial differences in characteristics between these participants and those from IMI DIRECT who did not have MRI data (see S2 Table).
Genetic, transcriptomic, proteomic, and metabolomic datasets were used as input omics variables in the analyses. Buffy coat was separated from whole blood, and DNA was then extracted and genotyped using the Illumina HumanCore array (HCE24 v1.0); genotype imputation was performed using the Haplotype Reference Consortium (HRC) and 1000 Genomes (1KG) reference panels. Details of the quality control (QC) steps for the genetic data are described elsewhere [16]. Transcriptomic data were generated using RNA sequencing from fasting whole blood. Only protein-coding genes were included in the analyses, as reads per kilobase of transcript per million mapped reads (RPKM). The targeted metabolomic data of fasting plasma samples were generated using the Biocrates AbsoluteIDQ p150 Kit. Additionally, untargeted LC/MS-based metabolomics was used to cover a broader spectrum of metabolites. A combination of technologies and quantitative panels of protein assays were used to generate “targeted” proteomic data. This included Olink proximity extension assays [19], sandwich immunoassay kits using Luminex technology (Merck Millipore and R&D Systems, Sweden), microfluidic ELISA assays (ProteinSimple, US [20]), protein analysis by Myriad RBM (Germany), and hsCRP analysis (MLM Medical Labs, Germany). In addition, protein data were generated by single-binder assays using highly multiplexed suspension bead arrays [21]. This approach (denoted “exploratory” proteomics) included a combination of antibodies targeting proteins selected by the consortium given published and unpublished evidence for association with glycemia-related traits. More information about data generation and QC of the transcriptomic, proteomic, and metabolomic data is provided in S1 Text. Technical covariates for transcriptomics include guanine-cytosine mean content, insert size, analysis lane and RNA integrity number, cell composition, date, and center. Technical covariates for proteomics were center, assay, plate number, and plate layout (n = 4), and for the targeted metabolites the technical covariates were center and plate. These technical covariates were used to correct the omics data, and the residuals were then extracted from these models and inverse normalized prior to further analyses.
Feature selection (IMI DIRECT)
We developed a series of NAFLD prediction models composed of variables that are available within clinical settings, as well as those not currently available in most clinics (see S3 Table). We had 2 strategies for selecting the clinical variables. For models 1–3, we selected variables based on clinical accessibility and their established association with fatty liver from existing literature without applying statistical procedures for data reduction. For model 4, a pairwise Pearson correlation matrix was used for feature selection of the clinical variables by placing a pairwise correlation threshold of r > 0.8, and we then selected the variables we considered most accessible among those that were collinear. Feature selection was undertaken in the combined cohort (diabetes and non-diabetes) in order to maximize sample size and statistical power. Of 1,514 participants with liver fat data, 1,049 had all necessary clinical and multi-omics data for a complete case analysis. We used k-nearest neighbor [22] imputation with k equal to 10 as a means to reduce the loss of sample size, but found that this did not materially improve predictive power in subsequent analyses, so we decided not to include these imputed data. An overview of the pairwise correlations among the clinical variables available in these 1,049 IMI DIRECT participants is presented in Fig 1.
The high-dimensionality nature of omics data also necessitated data reduction using the feature selection tool LASSO prior to building the model. LASSO is a regression analysis method that minimizes the sum of least squares in a linear regression model and shrinks selected beta coefficients (βj) using penalties (Eq 1). Minimizing the value from Eq 1, LASSO excludes the least informative variables and selects those features of most importance for the outcome of interest (y) in a sample of n cases, each of which consists of m parameters. The penalty applied by λ can be any value from 0 to positive infinity and is determined through a cross-validation step [26].
To minimize bias (for example by overfitting), we randomly divided the dataset and used 70% (n = 735) for feature selection and 30% (n = 314) for the model generation (see below). We selected these thresholds for partitioning the dataset in order to maximize the power to select the informative features. Stratified random sampling [27] based on the outcome variable was undertaken in order to preserve the distribution of the liver fat categories in the 2 feature selection and model generation sets. We selected LASSO, as a nonlinear data reduction tool might lead to overfitting owing to the high dimensionality of omics data. LASSO was conducted with package glmnet in R [28] with a 10-fold cross-validation step for defining the λ parameter that resulted in the minimum value for the mean square error of the regression model.
Feature selection using LASSO was undertaken in each omics dataset (genetic, transcriptomic, proteomic, and metabolomic) using 70% of the available data (models 5–18). For the genetic dataset, we first performed a genome-wide association study (GWAS) prior to LASSO in order to identify single nucleotide polymorphisms (SNPs) tentatively associated with liver fat accumulation (p < 5 × 10−6). LASSO was then applied to these index variants for feature selection in 70% of the study sample. The individual SNP association analysis was conducted with RVTESTS v2.0.2 [29], which applies a linear mixed model with an empirical kinship matrix to account for familial relatedness, cryptic relatedness, and population stratification. Only common variants with minor allele frequency (MAF) greater than 5% contributed to the kinship matrix. Liver fat data were log-transformed and then adjusted for age, age2, sex, center, body mass index (BMI), and alcohol consumption. These values were then inverse normal transformed and used in the GWAS analyses. We limited our analysis to genetic MAF > 1% and imputation quality score > 0.3. S3 and S4 Figs show the resulting Manhattan plot, depicting each SNP’s association with liver fat percentage and the quantile–quantile (QQ) plot of the GWAS results for liver fat. For the genetic data, 23 SNPs were selected out of the 108 SNPs with p-values < 5 × 10−6. For the transcriptomics, 93 genes were selected out of 16,209 protein-coding genes. In the exploratory and targeted proteomics, 22 out of 377 and 48 out of 483 proteins were selected, respectively. In the targeted and untargeted metabolomic data, 25 out of 116 and 39 out of 172 metabolites were selected by LASSO, respectively.
Model training and evaluation
The remaining 30% of the data was used to develop the binary prediction models for fatty liver (yes/no) with selected features used as input variables. We utilized the random forest supervised machine learning method, which is an aggregation of decision trees built from bootstrapped datasets (a process called “bagging”). Typically, two-thirds of the data are retained in these bootstrapped datasets, and the remaining third is termed the out of bag (OOB) dataset, which is used to validate the performance of the model. To avoid overfitting and improve generalizability, 5-fold cross-validation was done for resampling the training samples and was repeated 5 times to create multiple versions of the folds. The number of trees was set to 1,000 to provide an accurate and stable prediction. Receiver operating characteristic (ROC) curves were used to evaluate model performance by measuring the area under the curve (AUC). A ROC curve uses a combination of sensitivity (true positive rate) and specificity (true negative rate) to assess prediction performance. In our analysis, the random forest model is used to derive probability estimates for the presence of fatty liver. In order to make a class prediction, it is necessary to impose a cutoff above which fatty liver is deemed probable and below which it is considered improbable. The choice of cutoff influences both sensitivity and specificity for a given prediction model. We considered the effect of different cutoffs on these performance measurements. Additionally, we calculated the F1 score, which is the harmonic mean of precision (positive predictive value) and sensitivity, derived as follows:
Balanced accuracy was also evaluated, which is the proportion of individuals correctly classified (true positives and true negatives) within each class individually. Measurements of sensitivity, specificity, F1 score, and balanced accuracy were computed and compared at different cutoffs for the diabetes, non-diabetes, and combined cohorts. The variable importance was also determined via a “permutation accuracy importance” measure using random forest analysis. In brief, for each tree, the prediction accuracy was calculated in the OOB test data. Each predictor variable was then permuted, and the accuracy was recalculated. The difference in the accuracies was averaged over all the trees and then normalized by the standard error. Thus, the measure for variable importance is the difference in prediction accuracy before and after the permutation for each variable [30]. In addition, we used the ensemble feature selection (EFS) method to determine the normalized importance value of all features [31]. With this approach, we do not rely on only random forest for the importance ranking, and we can build the cumulative importance values from different methods including Spearman’s rank correlation test, Pearson’s product moment correlation test, beta-values of logistic regression, the error-rate-based variable importance measure, and the Gini-index-based variable importance measure. Statistical analyses were undertaken using R software version 3.2.5 [32], and the random forest models were built using the caret package [33]. Fig 2 shows an overview of the different stages involved in the data processing and model training.
Comparison with other fatty liver indices
Given the accessible data within the IMI DIRECT cohorts, several existing fatty liver indices could be calculated and compared with the IMI DIRECT prediction models. These included the fatty liver index (FLI) [34], hepatic steatosis index (HSI) [35], and the NAFLD liver fat score (NAFLD-LFS) [36].
FLI
The FLI is commonly used to estimate the presence or absence of fatty liver (categorized into fatty [≥60 FLI units] or non-fatty liver [<60 FLI units]) [34]. The FLI uses data on TG, waist circumference, BMI, and serum gamma-glutamyl transpeptidase (GGTP) and is calculated as follows:
NAFLD-FLS
NAFLD-FLS was calculated using fasting serum (fs) insulin, aspartate transaminase (AST), alanine transaminase (ALT), T2D, and metabolic syndrome (MS) (defined according to the International Diabetes Federation [37]) to provide an estimate of liver fat content. A NAFLD-FLS value above −0.64 is considered to indicate the presence of NAFLD:
HSI
The HSI uses BMI, sex, T2D diagnosis (yes/no), and the ratio of ALT to AST and is calculated as follows:
HSI values above 36 are deemed to indicate the presence of NAFLD.
External validation (UK Biobank cohort)
The UK Biobank cohort [38] was used to validate the clinical prediction models (models 1 and 2) derived using IMI DIRECT data (UK Biobank application ID: 18274). The same protocol and procedure have been used to quantify MRI-derived liver fat in IMI DIRECT and UK Biobank [18]. In addition, we validated the FLI and HSI using UK Biobank data. Field numbers for the UK Biobank variables used in the validation step can be found in the S4 Table. The data analysis procedures used for the UK Biobank validation analyses mirror those used in IMI DIRECT (as described above).
This study is reported as per the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guideline (S1 STROBE Checklist) and the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD) guideline (S1 TRIPOD Checklist).
Results
The following section describes fatty liver prediction models that are likely to suit different scenarios. We focus on a basic model (model 1), which includes variables that are widely available in both clinical and research settings. Models 2 and 3 focus on variables that could in principle be accessed within the clinical context, but that are not routinely available in the clinical setting at this time. Model 4 includes clinical variables, more detailed measures of glucose and insulin dynamics, and physical activity. Models 5 to 18 are more advanced models that include omics predictor variables alone or in combination with clinical predictor variables. See S3 Table for a full description of models.
Clinical models 1–3
We developed models 1–3 for NAFLD prediction, graded by perceived data accessibility for clinicians. These models were developed on the full dataset without applying any statistical procedures for feature selection. Model 1 includes 6 non-serological input variables: waist circumference, BMI, systolic blood pressure (SBP), diastolic blood pressure (DBP), alcohol consumption, and diabetes status. Model 2 includes 8 input variables: waist circumference, BMI, TG, ALT, AST, fasting glucose (or hemoglobin A1C [HbA1c] if fasting glucose is not available), alcohol consumption, and diabetes status. Model 3 includes 9 variables: waist circumference, BMI, TG, ALT, AST, fasting glucose, fasting insulin, alcohol consumption, and diabetes status. Clinical models 1–3 along with the FLI, HSI, and NAFLD-LFS were applied to the non-diabetes and diabetes cohort datasets separately, as well as to the combined cohort dataset; the ROCAUC results are presented in Fig 3. Model 1 yielded a ROCAUC of 0.73 (95% CI 0.72, 0.75; p < 0.001) in the combined cohort. Adding serological variables to model 2 (with either fasting glucose or HbA1c) for the combined cohort yielded a ROCAUC of 0.79 (95% CI 0.78, 0.80; p < 0.001). Model 3 (fasting insulin added) yielded a ROCAUC of 0.82 (95% CI 0.81, 0.83; p < 0.001) in the combined cohort. The FLI, HSI, and NAFLD-LFS had ROCAUCs of 0.75 (95% CI 0.73, 0.78; p < 0.001), 0.75 (95% CI 0.72, 0.77; p < 0.001), and 0.79 (95% CI 0.76, 0.81; p < 0.001), respectively, in the combined cohort. The predictive performance of clinical models 1–3, FLI, HSI, and NAFLD-LFS in the non-diabetes and diabetes cohorts is presented in S5 Table.
Performance metrics
We further investigated sensitivity, specificity, balanced accuracy, and F1 score (a score considering sensitivity and precision combined). These measurements were calculated for different cutoffs applied to the output of the random forest model (0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, and 0.9) using clinical models 1–3 in the diabetes, non-diabetes, and combined cohorts. The performance metrics for models 1 and 2 are presented in S6 and S7 Figs, and the metrics for model 3 are presented in Fig 4. We aimed to find the optimal cutoff for these models based on the cross-validated balanced accuracy. The highest balanced accuracy for models 1–3 in the non-diabetes, diabetes, and combined cohorts was observed at cutoffs of 0.4, 0.6, and 0.4, respectively (see Table 2).
Measurements of sensitivity, specificity, F1 score, and balanced accuracy were computed for the FLI, HSI, and NAFLD-LFS and compared with those of clinical models1–3. These measurements were computed at the optimal cutoff values for these indices: −0.640 for NAFLD-LFS, 60 for the FLI, and 36 for the HSI. A comprehensive overview of the prediction models’ performance metrics for all of the fatty liver indices listed above is shown in Table 2.
Validation in UK Biobank and IMI DIRECT
Liver fat data were available in 4,617 UK Biobank participants (1,011 with ≥5% liver fat and 3,606 with <5% liver fat). Of these individuals, 4,609 had all the required variables to replicate clinical model 1. To perform model 2, with either fasting glucose or HbA1c, 3,807 participants had data available for a complete case analysis. Given the limited availability of variables in the UK Biobank dataset, only models 1 and 2 of the NAFLD prediction models we developed could be externally validated. To facilitate this validation analysis, the random forest models developed in the IMI DIRECT cohorts were used to predict the liver fat category (participants with fatty liver versus non-fatty liver) for the UK Biobank participants. The performance of the FLI and HSI was also tested in the UK Biobank cohort. We validated both models 1 and 2 in the UK Biobank cohort with a similar ROCAUC as seen in the IMI DIRECT dataset. The ROCAUCs were 0.71 (95% CI 0.69, 0.73; p < 0.001), 0.79 (95% CI 0.77, 0.80; p < 0.001), and 0.78 (95% CI 0.76, 0.79; p < 0.001) for model 1, model 2 with fasting glucose, and model 2 with HbA1c, respectively. The FLI had a ROCAUC of 0.78 (95% CI 0.76, 0.80; p < 0.001), which is similar to the ROCAUC of model 2. The HSI yielded a ROCAUC of 0.76 (95% CI 0.75, 0.78; p < 0.001).
Measurements of sensitivity, specificity, F1 score, and balanced accuracy were also computed at the optimal cutoff values for these models: 0.4 for clinical models 1 and 2, 60 for the FLI, and 36 for the HSI (see Table 2).
Clinical model 4 and omics models 5–14
More advanced models using omics data were also developed. These models were generated using the omics features selected by LASSO in the combined cohort. The models include only omics or include omics plus 22 clinical variables as the input variables. Twenty-one of these clinical variables were selected based on the pairwise Pearson correlation matrix: BMI, waist circumference, SBP, DBP, alcohol consumption, ALT, AST, GGTP, HDL, TG, fasting glucose, 2-hour glucose, HbA1c, fasting insulin, 2-hour insulin, insulin secretion at the beginning of the carbohydrate challenge test (OGTT or MMTT), 2-hour oral glucose insulin sensitivity index (OGIS), mean insulin clearance during the OGTT/MTT, fasting glucagon concentration, fasting plasma total GLP-1 concentration, and mean physical activity intensity. Diabetes status (non-diabetes/diabetes) was also included as a clinical predictor in the models, given that analyses were undertaken in the combined diabetes and non-diabetes cohort. The ROCAUCs for models 4–14 are shown in Fig 5. The clinical model with the 22 selected clinical variables (model 4) yielded a ROCAUC of 0.79 (95% CI 0.76, 0.81; p < 0.001). Omics models with only the genetic (model 5), transcriptomic (model 7), proteomic (model 9), and targeted metabolomic (model 11) data as input variables resulted in ROCAUCs of 0.67 (95% CI 0.65, 0.70; p < 0.001), 0.72 (95% CI 0.69, 0.74; p < 0.001), 0.74 (95% CI 0.71, 0.76; p < 0.001), and 0.70 (95% CI 0.67, 0.72; p < 0.001), respectively. Including all the omics variables in one model (model 13) resulted in a ROCAUC of 0.82 (95% CI 0.80, 0.84; p < 0.001). Adding the clinical variables to each omics model improved the prediction ability; models with the clinical variables plus genetic (model 6), transcriptomic (model 8), exploratory proteomic (model 10), and targeted metabolomic (model 12) data resulted in ROCAUCs of 0.82 (95% CI 0.80, 0.84; p < 0.001), 0.81 (95% CI 0.79, 0.83; p < 0.001), 0.80 (95% CI 0.78, 0.83; p < 0.001), and 0.80 (95% CI 0.77, 0.82; p < 0.001), respectively. The highest performance was observed for model 14 (ROCAUC of 0.84; 95% CI 0.82, 0.86; p < 0.001). The variable importance for model 14 from the permutation accuracy importance measure, presented in Fig 6, shows that measures of insulin secretion rank amongst those having the highest variable importance of all input variables. Moreover, the importance list derived from EFS, shown in S20 Fig, is highly consistent with that derived from the random forest analysis. Rankings for the individual clinical and omics variables from the permutation accuracy importance measure and EFS are presented in S8–S19 Figs. The minor inconsistencies in results from the 2 approaches are likely to reflect the ability of the random forest analysis to detect variables that interact with others, which the linear methods are not designed to detect.
Additional proteomic and metabolomic analyses (models 15–18)
Data from targeted proteomic and untargeted metabolomic data were further utilized to develop the omics models separately or in combination with the clinical data. However, as some participants lacked these omics data, their models were developed using a smaller data subset and were, hence, not included in the advanced (model 14) analyses. The complete case analysis was primarily defined on the availability of the 22 selected clinical variables (n = 1,049). Within this complete case set, 511 had a complete set of untargeted metabolomic data, and 686 had a complete set of targeted proteomic data. The models with targeted proteomic data only and with proteomic and clinical variables combined resulted in ROCAUCs of 0.81 (95% CI 0.78, 0.84; p < 0.001) and 0.84 (95% CI 0.81, 0.87; p < 0.001), respectively. The untargeted metabolomic model alone had a ROCAUC of 0.66 (95% CI 0.63, 0.69; p < 0.001), which increased to 0.78 (95% CI 0.75, 0.80; p < 0.001) when the 22 clinical variables were added.
Discussion
Using data from the IMI DIRECT consortium, we developed 18 diagnostic models for early-stage NAFLD. These models were developed to reflect different scenarios within which they might be used: These included both clinical and research settings, with the more complex (and less accessible) models having the greatest predictive ability. The models were successfully validated in the UK Biobank where data permitted such analysis (clinical models 1 and 2). Overall, the basic clinical variables proved to be stronger predictors of fatty liver than more complex omics data, although adding omics data yielded the most powerful model, with very good cross-validated predictive ability (ROCAUC = 0.84).
NAFLD is etiologically complex, rendering its prevention and treatment difficult, and diagnosis can require invasive and/or relatively expensive procedures. Thus, noninvasive and cost-effective prediction models with good sensitivity and specificity are much needed. This is especially important because if NAFLD is detected early, treatment through lifestyle interventions can be highly effective [39]. However, simple steatosis is usually asymptomatic, and many patients only come to the attention of hepatologists when serious complications arise [40].
To date, several prediction models have been developed to facilitate the diagnosis of steatosis (thoroughly reviewed elsewhere [13]). The FLI is one of the most well-established and commonly used fatty liver indices, initially developed using ultrasound-derived hepatic steatosis data [34]. The FLI yielded similar predictive performance in the diabetes and non-diabetes cohorts of IMI DIRECT (both ROCAUCs approximately 0.75).
Though commonly used for liver fat prediction, the FLI has a similar discriminative ability as waist circumference alone [41]. Better discrimination can be obtained by incorporating additional serological and hemostatic measures, which is the case with NAFLD-LFS [14], the SteatoTest [42], and the HSI [35], for example. Notwithstanding the added complexity and cost of these scores, the FLI, HSI, and NAFLD-LFS yielded similar predictive ability in a series of liver-biopsy-diagnosed NAFLD cases (n = 324) [36].
Omics technologies have been used in a small number of studies to identify molecular biomarkers of NAFLD [43–45]. These include tests utilizing genetic data such as FibroGENE for staging liver fibrosis [46], and tests using metabolomic data derived from liver tissue to differentiate simple hepatitis from NASH [47], as well as a multi-component NAFLD classifier using genomic, proteomic, and phenomic data [45]. Machine learning models based on lipidomic, glycomic, and free fatty acid data were also developed for the diagnosis of NASH and liver fibrosis [48,49]. In a recent retrospective case series of patients with obesity, EFS was applied for feature selection, using a set of sociodemographic and serum variables to predict the presence or absence of NASH [50].
Using data from IMI DIRECT, we explored the predictive ability of genetic, transcriptomic, proteomic, and metabolomic data from blood in the diagnosis of NAFLD. The top 20 features of each omics model are presented in S9–S14 Figs. The details of the LASSO selected features are summarized in S7 Table. Reassuringly, several of the features that ranked highest have been previously described for their association with liver fat content or closely related traits; these include PNPLA3 gene variants [44,51], fetal liver tyrosine kinase-3 (FLT3) transcripts [52], IGFBP1 [53–55] and lipoprotein lipase (Lpl) [56] proteins, and the metabolite glutamate [57]. In the analysis of the targeted metabolites, phosphatidylcholines (including PC.aa.C32, PC.aa.C38, PC.aa.C40, and PC.aa.C42), glycerophospholipids, and valine were amongst the highest-ranked metabolites that are known for their correlation with NAFLD and metabolic disorders [58,59]. For exploratory proteomics, the most important variables were proteins secreted into the blood, expressed by the liver as well as those leaking from the blood cells [60]. The prediction model that only included targeted proteomic data (model 15) performed well (ROCAUC = 0.81), rendering it an interesting candidate biomarker for future clinical tests. Among the top 20 most important proteins were many secreted into the blood or leaked by the liver, as well as the pancreas, fat, or muscle tissue [61].
Our intention by including all features in the same model (model 14) was to maximize predictive power by leveraging interactions between features. Moreover, we explored the value of boosting ensemble algorithms for each data source. The purpose of this was to enhance predictions. We trained a stochastic gradient boosting algorithm for each data source separately and then applied a weighted averaging on the probabilities of observations. The optimal weighting was observed at 0.5 for the clinical data and 0.125 for each omics data layer (i.e., genetic, transcriptomic, exploratory proteomic, and targeted metabolomic). The ensemble prediction model of omics and clinical datasets resulted in a ROCAUC of 0.83 (95% CI 0.78, 0.87; p < 0.001), which is not materially different from the ROCAUC derived for the advanced model 14 (described in the Results), which includes all the omics and clinical features in a single model (ROCAUC of 0.84; 95% CI 0.82, 0.86; p < 0.001). The models developed here may be useful for screening for NAFLD, and this should be evaluated in future clinical studies.
In order to stratify people into groups of those unlikely and likely to have NAFLD, the latter of whom might subsequently undergo more invasive and/or costly clinical assessments, it would be important for the prediction model to have high sensitivity. However, the predictive utility of a given model can be further improved by selecting model cutoffs that optimize sensitivity or specificity, as the 2 metrics rarely perform optimally at the same cutoff. This issue was apparent for models 1–3 in the current analyses, where we selected cutoffs that maximized balanced accuracy (considering both sensitivity and specificity); these features are especially important in screening algorithms, where the cost of false negatives can be high. Models 1–3 resulted in higher sensitivity in the diabetes cohort than the non-diabetes cohort, whereas the specificity was higher in the non-diabetes and combined cohorts than in the diabetes cohort.
The linear LASSO method was used to minimize overfitting that can occur with high-dimensionality data, while random forest analysis was used to identify nonlinear associations where data structure allowed. We also considered several other machine learning approaches including generalized linear model, stochastic gradient boosting, support vector machines, and k-nearest neighbor, and the random forest analysis yielded similar or better results compared with any of these other approaches (see S6 Table).
A limitation of the analytical approach used here is that the methods required a complete case analysis, which diminishes sample size considerably; although imputing missing data here helped preserve sample size, it did not improve the prediction ability of the models, and we hence elected to use the complete case analysis. Heavy alcohol consumption is a key determinant of fatty liver, but is unlikely to be a major etiological factor in IMI DIRECT owing to the demographics of this cohort. Nevertheless, a further limitation of this analysis is that alcohol intake was self-reported and may lack validity. To address this limitation, we removed all self-reported heavy alcohol consumers from the UK Biobank cohort and undertook sensitivity analyses, but this did not materially affect the results.
Here we considered lifestyle variables, but not medications. The use of medicines affecting liver fat is likely to be less in the non-diabetes than in the diabetes cohort, yet the models fit better in the latter, suggesting that glucose-lowering medication use in the IMI DIRECT cohorts did not have a major detrimental impact on prediction model performance.
A further consideration for future work is the impact lifestyle and medications are likely to have on the prediction of NAFLD. Furthermore, this study was undertaken in people of European ancestry, and the extent to which the results will generalize to other ethnic groups is unknown. Moreover, the prediction is for a binary liver fat outcome (<5% or ≥5%), and neither fully quantifies liver fat volume nor elucidates the degree of liver damage (cirrhosis). These key limitations of the current work will be the focus of future research.
Our finding that a model focused on proteomic data yielded high predictive utility may warrant further investigation. Our analysis also suggests that insulin sensitivity and beta-cell dysfunction may be involved in liver fat accumulation, which are at present not considered as features of conventional NAFLD risk models.
In summary, we have developed prediction models for NAFLD that may have utility for clinical diagnosis and research investigations alike. A web interface for the diagnosis of NAFLD was developed using the findings described above (https://www.predictliverfat.org), which renders clinical models 1–3 developed here accessible for the wider community of clinicians and researchers.
Supporting information
S1 Fig [tiff]
Violin plot showing the distribution of liver fat percentage for the diabetes and non-diabetes cohorts of IMI DIRECT.
S2 Fig [tiff]
Distribution of liver fat percentage among the different centers contributing to the IMI DIRECT cohorts.
S3 Fig [tiff]
Manhattan plot showing SNPs associated with liver fat level (approximately 18 million imputed SNPs) in the IMI DIRECT cohorts.
S4 Fig [tiff]
Quantile–quantile (QQ) plot showing results of genome-wide association study (GWAS) for liver fat content in the IMI DIRECT consortium (1,514 individuals).
S5 Fig [tiff]
Details of the feature selection step for models 4–14 and models 15–18 using the IMI DIRECT data.
S6 Fig [tiff]
Measurements of sensitivity, specificity, F1 score (a score considering sensitivity and precision combined), and balanced accuracy at different cutoffs for model 1 in the diabetes, non-diabetes, and combined cohorts of IMI DIRECT.
S7 Fig [tiff]
Measurements of sensitivity, specificity, F1 score (a score considering sensitivity and precision combined), and balanced accuracy at different cutoffs for model 2 in the diabetes, non-diabetes, and combined cohorts of IMI DIRECT.
S8 Fig [tiff]
Variable importance for the clinical model via a permutation accuracy importance measure.
S9 Fig [tiff]
Variable importance for the genetic model via a permutation accuracy importance measure.
S10 Fig [tiff]
Variable importance for the transcriptomic model via a permutation accuracy importance measure.
S11 Fig [tiff]
Variable importance for the exploratory proteomic model via a “permutation accuracy importance” measure.
S12 Fig [tiff]
Variable importance for the targeted metabolomic model via a permutation accuracy importance measure.
S13 Fig [tiff]
Variable importance for the targeted proteomic model via a permutation accuracy importance measure.
S14 Fig [tiff]
Variable importance for the untargeted metabolomic model via a permutation accuracy importance measure.
S15 Fig [s_cor]
Variable importance for the clinical model derived from ensemble feature selection (EFS).
S16 Fig [s_cor]
Variable importance for the genetic model derived from EFS.
S17 Fig [s_cor]
Variable importance for the transcriptomic model derived from EFS.
S18 Fig [s_cor]
Variable importance for the exploratory proteomic model derived from EFS.
S19 Fig [s_cor]
Variable importance for the targeted metabolomic model derived from EFS.
S20 Fig [s_cor]
Variable importance for the clinical plus multi-omics model (clinical = 22, genetic = 23, transcriptomic = 93, exploratory proteomic = 22, and targeted metabolomic = 25) derived from EFS.
S1 STROBE Checklist [docx]
The STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) checklist.
S1 Table [xlsx]
The list of the clinical input variables with the abbreviation used in the analyses and their meaning.
S2 Table [xlsx]
Characteristics of the study in the non-diabetes, diabetes, and combined cohorts separated for participants from IMI DIRECT who had MRI data versus those who did not have MRI data.
S3 Table [xlsx]
Variables used to construct each of the NAFLD prediction models developed in IMI DIRECT.
S4 Table [xlsx]
UK Biobank field number with the description used in the analyses.
S5 Table [xlsx]
Receiver operating characteristic area under the curve (ROCAUC) with 95% confidence interval for clinical models 1–3, fatty liver index (FLI), hepatic steatosis index (HSI), and NAFLD liver fat score (NAFLD-LFS) in the non-diabetes and diabetes cohorts of the IMI DIRECT separately.
S6 Table [xlsx]
Receiver operating characteristic area under the curve (ROCAUC) with 95% confidence interval of each separate dataset obtained from random forest (RF), generalized linear model (GLM), stochastic gradient boosting (GBM), support vector machine (SVM), and -nearest neighbor (KNN) analyses in the cross-validated test data of the IMI DIRECT combined cohort.
S7 Table [xlsx]
The details of the LASSO (least absolute shrinkage and selection operator)–selected features of the omics layers in separate sheets (genetic, transcriptomic, exploratory proteomic, targeted proteomic, targeted Metabolomic, and untargeted metabolomic).
S1 Text [docx]
QC of the transcriptomic, proteomic, and metabolomic variables in the IMI DIRECT datasets.
S1 TRIPOD Checklist [docx]
The TRIPOD (Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis) checklist.
Zdroje
1. Tilg H, Moschen AR. Insulin resistance, inflammation, and non-alcoholic fatty liver disease. Trends Endocrinol Metab. 2008;19(10):371–9. doi: 10.1016/j.tem.2008.08.005 18929493
2. Sattar N, Gill JM. Type 2 diabetes as a disease of ectopic fat? BMC Med. 2014;12:123. doi: 10.1186/s12916-014-0123-4 25159817
3. Sattar N, Forrest E, Preiss D. Non-alcoholic fatty liver disease. BMJ. 2014;349:g4596. doi: 10.1136/bmj.g4596 25239614
4. Lucas C, Lucas G, Lucas N, Krzowska-Firych J, Tomasiewicz K. A systematic review of the present and future of non-alcoholic fatty liver disease. Clin Exp Hepatol. 2018;4(3):165–74. doi: 10.5114/ceh.2018.78120 30324141
5. Fazel Y, Koenig AB, Sayiner M, Goodman ZD, Younossi ZM. Epidemiology and natural history of non-alcoholic fatty liver disease. Metabolism. 2016;65(8):1017–25. doi: 10.1016/j.metabol.2016.01.012 26997539
6. Targher G, Byrne CD, Lonardo A, Zoppini G, Barbui C. Non-alcoholic fatty liver disease and risk of incident cardiovascular disease: a meta-analysis. J Hepatol. 2016;65(3):589–600. doi: 10.1016/j.jhep.2016.05.013 27212244
7. Mahfood Haddad T, Hamdeh S, Kanmanthareddy A, Alla VM. Nonalcoholic fatty liver disease and the risk of clinical cardiovascular events: a systematic review and meta-analysis. Diabetes Metab Syndr. 2017;11(Suppl 1):S209–16.
8. Bellentani S. The epidemiology of non-alcoholic fatty liver disease. Liver Int. 2017;37(Suppl 1):81–4.
9. Younossi ZM, Koenig AB, Abdelatif D, Fazel Y, Henry L, Wymer M. Global epidemiology of nonalcoholic fatty liver disease—meta-analytic assessment of prevalence, incidence, and outcomes. Hepatology. 2016;64(1):73–84. doi: 10.1002/hep.28431
10. Younossi ZM. Non-alcoholic fatty liver disease—a global public health perspective. J Hepatol. 2019;70(3):531–44. doi: 10.1016/j.jhep.2018.10.033
11. Younossi ZM, Tampi R, Priyadarshini M, Nader F, Younossi IM, Racila A. Burden of illness and economic model for patients with nonalcoholic steatohepatitis in the United States. Hepatology. 2019;69(2):564–72. doi: 10.1002/hep.30254 30180285
12. Castera L, Friedrich-Rust M, Loomba R. Noninvasive assessment of liver disease in patients with nonalcoholic fatty liver disease. Gastroenterology. 2019;156(5):1264–81.e4. doi: 10.1053/j.gastro.2018.12.036 30660725
13. Castera L. Diagnosis of non-alcoholic fatty liver disease/non-alcoholic steatohepatitis: non-invasive tests are enough. Liver Int. 2018;38(Suppl 1):67–70.
14. Kotronen A, Peltonen M, Hakkarainen A, Sevastianova K, Bergholm R, Johansson LM, et al. Prediction of non-alcoholic fatty liver disease and liver fat using metabolic and genetic factors. Gastroenterology. 2009;137(3):865–72. doi: 10.1053/j.gastro.2009.06.005 19524579
15. Koivula RW, Heggie A, Barnett A, Cederberg H, Hansen TH, Koopman AD, et al. Discovery of biomarkers for glycaemic deterioration before and after the onset of type 2 diabetes: rationale and design of the epidemiological studies within the IMI DIRECT Consortium. Diabetologia. 2014;57(6):1132–42. doi: 10.1007/s00125-014-3216-x 24695864
16. Koivula RW, Forgie IM, Kurbasic A, Vinuela A, Heggie A, Giordano GN, et al. Discovery of biomarkers for glycaemic deterioration before and after the onset of type 2 diabetes: descriptive characteristics of the epidemiological studies within the IMI DIRECT Consortium. Diabetologia. 2019;62(9):1601–15. doi: 10.1007/s00125-019-4906-1 31203377
17. Thomas EL, Fitzpatrick JA, Malik SJ, Taylor-Robinson SD, Bell JD. Whole body fat: content and distribution. Prog Nucl Magn Reson Spectrosc. 2013;73:56–80. doi: 10.1016/j.pnmrs.2013.04.001 23962884
18. Wilman HR, Kelly M, Garratt S, Matthews PM, Milanesi M, Herlihy A, et al. Characterisation of liver fat in the UK Biobank cohort. PLoS ONE. 2017;12(2):e0172921. doi: 10.1371/journal.pone.0172921 28241076
19. Assarsson E, Lundberg M, Holmquist G, Bjorkesten J, Thorsen SB, Ekman D, et al. Homogenous 96-plex PEA immunoassay exhibiting high sensitivity, specificity, and excellent scalability. PLoS ONE. 2014;9(4):e95192. doi: 10.1371/journal.pone.0095192 24755770
20. Aldo P, Marusov G, Svancara D, David J, Mor G. Simple Plex(TM): a novel multi-analyte, automated microfluidic immunoassay platform for the detection of human and mouse cytokines and chemokines. Am J Reprod Immunol. 2016;75(6):678–93. doi: 10.1111/aji.12512
21. Drobin K, Nilsson P, Schwenk JM. Highly multiplexed antibody suspension bead arrays for plasma protein profiling. Methods Mol Biol. 2013;1023:137–45. doi: 10.1007/978-1-4614-7209-4_8 23765623
22. Altman NS. An introduction to kernel and nearest-neighbor nonparametric regression. Am Stat. 1992;46(3):175–85.
23. Matsuda M, DeFronzo RA. Insulin sensitivity indices obtained from oral glucose tolerance testing: comparison with the euglycemic insulin clamp. Diabetes Care. 1999;22(9):1462–70. doi: 10.2337/diacare.22.9.1462 10480510
24. Mari A, Pacini G, Murphy E, Ludvik B, Nolan JJ. A model-based method for assessing insulin sensitivity from the oral glucose tolerance test. Diabetes Care. 2001;24(3):539–48. doi: 10.2337/diacare.24.3.539 11289482
25. Stumvoll M, Mitrakou A, Pimenta W, Jenssen T, Yki-Järvinen H, Van Haeften T, et al. Use of the oral glucose tolerance test to assess insulin release and insulin sensitivity. Diabetes Care. 2000;23(3):295–301. doi: 10.2337/diacare.23.3.295 10868854
26. Tibshirani R. Regression shrinkage and selection via the lasso. J R Stat Soc Series B Stat Methodol. 1996;58:267–88.
27. Setia MS. Methodology series module 5: sampling strategies. Indian J Dermatol. 2016;61(5):505–9. doi: 10.4103/0019-5154.190118 27688438
28. Friedman J, Hastie T, Tibshirani R. Regularization paths for generalized linear models via coordinate descent. J Stat Softw. 2010;33(1):1–22. 20808728
29. Zhan X, Hu Y, Li B, Abecasis GR, Liu DJ. RVTESTS: an efficient and comprehensive tool for rare variant association analysis using sequence data. Bioinformatics. 2016;32(9):1423–6. doi: 10.1093/bioinformatics/btw079 27153000
30. Strobl C, Boulesteix AL, Zeileis A, Hothorn T. Bias in random forest variable importance measures: illustrations, sources and a solution. BMC Bioinformatics. 2007;8:25. doi: 10.1186/1471-2105-8-25 17254353
31. Neumann U, Genze N, Heider D. EFS: an ensemble feature selection tool implemented as R-package and web-application. BioData Min. 2017;10:21. doi: 10.1186/s13040-017-0142-8 28674556
32. R Core Team. R: a language and environment for statistical computing. Version 3.2.5. Vienna: R Foundation for Statistical Computing; 2013.
33. Kuhn M. caret: classification and regression training. Version 6.0–71. Comprehensive R Archive Network; 2016.
34. Bedogni G, Bellentani S, Miglioli L, Masutti F, Passalacqua M, Castiglione A, et al. The Fatty Liver Index: a simple and accurate predictor of hepatic steatosis in the general population. BMC Gastroenterol. 2006;6:33. doi: 10.1186/1471-230X-6-33 17081293
35. Lee JH, Kim D, Kim HJ, Lee CH, Yang JI, Kim W, et al. Hepatic steatosis index: a simple screening tool reflecting nonalcoholic fatty liver disease. Dig Liver Dis. 2010;42(7):503–8. doi: 10.1016/j.dld.2009.08.002 19766548
36. Fedchuk L, Nascimbeni F, Pais R, Charlotte F, Housset C, Ratziu V, et al. Performance and limitations of steatosis biomarkers in patients with nonalcoholic fatty liver disease. Aliment Pharmacol Ther. 2014;40(10):1209–22. doi: 10.1111/apt.12963 25267215
37. Alberti KG, Zimmet P, Shaw J, Group IDF Epidemiology Task Force Consensus Group. The metabolic syndrome—a new worldwide definition. Lancet. 2005;366(9491):1059–62. doi: 10.1016/S0140-6736(05)67402-8
38. Sudlow C, Gallacher J, Allen N, Beral V, Burton P, Danesh J, et al. UK biobank: an open access resource for identifying the causes of a wide range of complex diseases of middle and old age. PLoS Med. 2015;12(3):e1001779. doi: 10.1371/journal.pmed.1001779 25826379
39. Lean ME, Leslie WS, Barnes AC, Brosnahan N, Thom G, McCombie L, et al. Primary care-led weight management for remission of type 2 diabetes (DiRECT): an open-label, cluster-randomised trial. Lancet. 2018;391(10120):541–51. doi: 10.1016/S0140-6736(17)33102-1 29221645
40. Araujo AR, Rosso N, Bedogni G, Tiribelli C, Bellentani S. Global epidemiology of non-alcoholic fatty liver disease/non-alcoholic steatohepatitis: what we need in the future. Liver Int. 2018;38(Suppl 1):47–51.
41. Motamed N, Sohrabi M, Ajdarkosh H, Hemmasi G, Maadi M, Sayeedian FS, et al. Fatty liver index vs waist circumference for predicting non-alcoholic fatty liver disease. World J Gastroenterol. 2016;22(10):3023–30. doi: 10.3748/wjg.v22.i10.3023 26973398
42. Poynard T, Ratziu V, Naveau S, Thabut D, Charlotte F, Messous D, et al. The diagnostic value of biomarkers (SteatoTest) for the prediction of liver steatosis. Comp Hepatol. 2005;4:10. doi: 10.1186/1476-5926-4-10 16375767
43. Baranova A, Liotta L, Petricoin E, Younossi ZM. The role of genomics and proteomics: technologies in studying non-alcoholic fatty liver disease. Clin Liver Dis. 2007;11(1):209–20. doi: 10.1016/j.cld.2007.02.003 17544980
44. Eslam M, Valenti L, Romeo S. Genetics and epigenetics of NAFLD and NASH: clinical impact. J Hepatol. 2018;68(2):268–79. doi: 10.1016/j.jhep.2017.09.003 29122391
45. Wood GC, Chu X, Argyropoulos G, Benotti P, Rolston D, Mirshahi T, et al. A multi-component classifier for nonalcoholic fatty liver disease (NAFLD) based on genomic, proteomic, and phenomic data domains. Sci Rep. 2017;7:43238. doi: 10.1038/srep43238 28266614
46. Eslam M, Hashem AM, Romero-Gomez M, Berg T, Dore GJ, Mangia A, et al. FibroGENE: a gene-based model for staging liver fibrosis. J Hepatol. 2016;64(2):390–8. doi: 10.1016/j.jhep.2015.11.008 26592354
47. Alonso C, Fernandez-Ramos D, Varela-Rey M, Martinez-Arranz I, Navasa N, Van Liempd SM, et al. Metabolomic identification of subtypes of nonalcoholic steatohepatitis. Gastroenterology. 2017;152(6):1449–61.e7. doi: 10.1053/j.gastro.2017.01.015 28132890
48. Perakakis N, Polyzos SA, Yazdani A, Sala-Vila A, Kountouras J, Anastasilakis AD, et al. Non-invasive diagnosis of non-alcoholic steatohepatitis and fibrosis with the use of omics and supervised learning: a proof of concept study. Metabolism. 2019;101:154005. doi: 10.1016/j.metabol.2019.154005 31711876
49. Katsiki N, Gastaldelli A, Mikhailidis DP. Predictive models with the use of omics and supervised machine learning to diagnose non-alcoholic fatty liver disease: A “non-invasive alternative” to liver biopsy? Metabolism. 2019;101:154010. doi: 10.1016/j.metabol.2019.154010
50. Canbay A, Kalsch J, Neumann U, Rau M, Hohenester S, Baba HA, et al. Non-invasive assessment of NAFLD as systemic disease—a machine learning perspective. PLoS ONE. 2019;14(3):e0214436. doi: 10.1371/journal.pone.0214436
51. Danford CJ, Yao ZM, Jiang ZG. Non-alcoholic fatty liver disease: a narrative review of genetics. J Biomed Res. 2018;32(5):389–400. doi: 10.7555/JBR.32.20180045 30355853
52. Al-Fayoumi S, Hashiguchi T, Shirakata Y, Mascarenhas J, Singer JW. Pilot study of the antifibrotic effects of the multikinase inhibitor pacritinib in a mouse model of liver fibrosis. J Exp Pharmacol. 2018;10:9–17. doi: 10.2147/JEP.S150729 29785143
53. Hagstrom H, Stal P, Hultcrantz R, Brismar K, Ansurudeen I. IGFBP-1 and IGF-I as markers for advanced fibrosis in NAFLD—a pilot study. Scand J Gastroenterol. 2017;52(12):1427–34. doi: 10.1080/00365521.2017.1379556
54. Petaja EM, Zhou Y, Havana M, Hakkarainen A, Lundbom N, Ihalainen J, et al. Phosphorylated IGFBP-1 as a non-invasive predictor of liver fat in NAFLD. Sci Rep. 2016;6:24740. doi: 10.1038/srep24740 27091074
55. Adamek A, Kasprzak A. Insulin-like growth factor (IGF) system in liver diseases. Int J Mol Sci. 2018;19(5):1308.
56. Chen Y, Huang H, Xu C, Yu C, Li Y. Long non-coding RNA profiling in a non-alcoholic fatty liver disease rodent model: new insight into pathogenesis. Int J Mol Sci. 2017;18(1):21.
57. Gaggini M, Carli F, Rosso C, Buzzigoli E, Marietti M, Della Latta V, et al. Altered amino acid concentrations in NAFLD: impact of obesity and insulin resistance. Hepatology. 2018;67(1):145–58. doi: 10.1002/hep.29465 28802074
58. Imhasly S, Naegeli H, Baumann S, von Bergen M, Luch A, Jungnickel H, et al. Metabolomic biomarkers correlating with hepatic lipidosis in dairy cows. BMC Vet Res. 2014;10:122. doi: 10.1186/1746-6148-10-122 24888604
59. Koch M, Freitag-Wolf S, Schlesinger S, Borggrefe J, Hov JR, Jensen MK, et al. Serum metabolomic profiling highlights pathways associated with liver fat content in a general population sample. Eur J Clin Nutr. 2017;71(8):995–1001. doi: 10.1038/ejcn.2017.43 28378853
60. Uhlen M, Karlsson MJ, Hober A, Svensson AS, Scheffel J, Kotol D, et al. The human secretome. Sci Signal. 2019;12(609):eaaz0274. doi: 10.1126/scisignal.aaz0274
61. Uhlen M, Fagerberg L, Hallstrom BM, Lindskog C, Oksvold P, Mardinoglu A, et al. Proteomics. Tissue-based map of the human proteome. Science. 2015;347(6220):1260419. 25613900
Článek vyšel v časopise
PLOS Medicine
2020 Číslo 6
- Prof. Libor Grubhoffer: Nová varianta mpox je příležitostí pokročit v poznání původce varioly
- Co obsahují nové pokyny WHO pro dozor nad antibiotickou rezistencí?
- AI místo lékaře? Výzkum zkoumal ochotu Čechů konzultovat své zdraví s umělou inteligencí
- Prof. Jan Škrha: Metformin je bezpečný, ale je třeba jej bezpečně užívat a léčbu kontrolovat
- Ibuprofen jako alternativa antibiotik při léčbě infekcí močových cest
Nejčtenější v tomto čísle
- Distinct subtypes of polycystic ovary syndrome with novel genetic associations: An unsupervised, phenotypic clustering analysis
- Fatty acids in the de novo lipogenesis pathway and incidence of type 2 diabetes: A pooled analysis of prospective cohort studies
- The potential impact of COVID-19 in refugee camps in Bangladesh and beyond: A modeling study
- Chronic pain diagnosis in refugee torture survivors: A prospective, blinded diagnostic accuracy study