Machine Learning Reveals the Contribution of Lipoproteins to Liver Triglyceride Content and Inflammation Authors: Federica Tavaglione, Giuseppe Marafioti, Stefano Romeo, Oveis Jamialahmadi Published: 2024 ABSTRACT AbstractContextMetabolic dysfunction–associated steatotic liver disease (MASLD) is currently the most common chronic liver disease worldwide and is strongly associated with metabolic comorbidities, including dyslipidemia.ObjectiveHerein, we aim to estimate the prevalence of MASLD and metabolic dysfunction-associated steatohepatitis (MASH) in Europeans with isolated hypercholesterolemia and isolated hypertriglyceridemia in the UK Biobank and to estimate the independent contribution of lipoproteins to liver triglyceride content.MethodsWe selected 218 732 Europeans from the UK Biobank without chronic viral hepatitis and other causes of liver disease, of whom 14 937 with liver magnetic resonance imaging data available. Next, to examine the relationships between traits in predicting liver triglyceride content, we compared the predictive performance of several machine learning methods and selected the best performing algorithms based on the minimum cross-validated mean squared error (MSE).ResultsThere was an approximately 3-fold and 4-fold enrichment of MASLD and MASH in individuals with isolated hypertriglyceridemia (P = 1.23 × 10–41 and P = 1.29 × 10–10, respectively), whereas individuals with isolated hypercholesterolemia had a marginal higher rate of MASLD and no difference in MASH rate compared with the control group (P = .019 and P = .97, respectively). Among machine learning methods, the feed-forward neural network had the best cross-validation MSE on the validation set. Circulating triglycerides, after body mass index, were the second strongest independent predictor of liver proton density fat fraction with the largest absolute mean Shapley additive explanation value.ConclusionIsolated hypertriglyceridemia is the second strongest, after obesity, independent predictor of MASLD/MASH. Individuals with hypertriglyceridemia, but not with hypercholesterolemia, should be screened for liver disease.Keywords: NAFLD, NASH, MASLD, MRI-PDFF, artificial intelligence, neural network Metabolic dysfunction–associated steatotic liver disease (MASLD) (1), also defined as metabolic dysfunction–associated fatty liver disease (2), is the most common chronic liver disease worldwide, affecting up to 30% of the global population (3). MASLD encompasses a wide spectrum of liver conditions and its progressive form, metabolic dysfunction–associated steatohepatitis (MASH), is the leading indication for liver transplantation in the United States (4). A large body of evidence indicates that MASLD is strongly associated with metabolic comorbidities, mostly obesity and type 2 diabetes, but also including metabolic syndrome and dyslipidemia. Specifically, the global prevalence of MASLD and MASH among individuals with type 2 diabetes has been estimated to be 55% and 37%, respectively, whereas in those with morbid obesity they are approximately 80% to 90% and 8% to 26%, respectively (5, 6). Along this line, the prevalence of dyslipidemia among individuals with MASLD and MASH was found to be 69% and 72%, respectively (7). Within this context, the most common lipoprotein phenotype observed in individuals with MASLD is atherogenic dyslipidemia, characterized by hypertriglyceridemia, increased small dense low-density lipoprotein (LDL) particles, and decreased high-density lipoprotein cholesterol levels, resulting in an increased risk for cardiovascular disease (8). An increase in hepatic de novo lipogenesis and apolipoprotein B synthesis leading to elevation of very low–density lipoproteins secretion are fundamental mechanisms responsible for the onset of the atherogenic dyslipidemia (9). To the best of our knowledge, no studies have examined the prevalence of MASLD and MASH in European individuals with isolated hypercholesterolemia and isolated hypertriglyceridemia. Herein, we aim to (1) estimate the prevalence of MASLD and MASH in Europeans from the UK Biobank with these lipoprotein phenotypes and (2) estimate the independent contribution of lipoproteins to liver triglyceride content. Materials and Methods UK Biobank The UK Biobank is a large prospective cohort study recruiting more than 500 000 participants (age 40-69 years) between 2006 and 2010 from 22 assessment centers throughout the UK (10). Clinical, laboratory, and imaging data were collected using highly standardized procedures. Medical diagnoses were obtained through linkage of hospital admissions and death register from the National Health Service records (data-fields 41270, 40001, 40002). The UK Biobank study has been approved by the North West Multi-centre Research Ethics Committee (no. 21/NW/0157). All participants gave written informed consent to the study. Data used in this study were obtained under application number 37142. Study Cohort First, we selected unrelated participants of European ancestry based on our quality control pipeline, which has been described in detail previously (11). We added participants who self-reported as being “Irish” or “Any other White background” to the subset of White British ancestry. Next, we excluded individuals (1) with more than 10 putative third-degree relatives, (2) with a mismatch between their self-reported and genetically inferred sex, (3) having putative sex chromosome aneuploidy, (4) who had withdrawn consent, and (5) were identified by the UK Biobank as outliers based on heterozygosity and missingness. Then, we excluded individuals reporting self-reported history or medical diagnosis of chronic viral hepatitis (International Classification of Diseases 10th edition [ICD-10] B18-B19) and other causes of liver disease (ICD-10 E83.0, E83.1, E88.0, I82.0, K70, K71, K74.3, K74.4, K74.5, K75.2, K75.3, K75.4, K75.8, K75.9, K76.5, K76.8, K76.9, K83.0) from hospital admissions and death registry, and those treated with any lipid-lowering medications, including statins, fibrates, ezetimibe, and omega-3 fatty acids. A total of 218 732 individuals were included in the final analysis. We further selected a subset of 14 937 individuals with liver magnetic resonance imaging (MRI)–estimated proton density fat fraction (PDFF) and iron-corrected T1 (cT1) measurements available. Liver Magnetic Resonance Imaging Participants were scanned at the UK Biobank Imaging Centre in Cheadle (UK) using a Siemens 1.5T MAGNETOM Aera, as described in detail elsewhere (12, 13). Briefly, a shortened modified look locker inversion (ShMOLLI) was used to quantify liver T1 and multiecho-spoiled gradient echo was used to quantify liver iron and fat. Data were analyzed using LiverMultiScan Discover 4.0 software. MASLD was defined by PDFF >5.5% (12), MASH by PDFF >5.5% (12) and cT1 >800 ms (14). Definition of Diagnoses and Covariates Anthropometric measures were assessed by trained staff using standardized procedures. Height was measured using the Seca 202 height measure (Seca, Hamburg, Germany). Weight was assessed by bioimpedance using the Tanita BC-418 MA body composition analyzer (Tanita Europe, Amsterdam, Netherlands). Body mass index (BMI) was calculated by dividing the weight (kg) by the square of the height (m2). Isolated hypercholesterolemia was defined as hospital diagnosis of hypercholesterolemia (ICD-10 E78.0), total cholesterol ≥6.21 mmol/L (240 mg/dL), or LDL cholesterol ≥4.14 mmol/L (160 mg/dL) (15), with triglycerides <1.69 mmol/L (150 mg/dL). Isolated hypertriglyceridemia was defined as hospital diagnosis of hypertriglyceridemia (ICD-10 E78.1) or triglycerides ≥1.69 mmol/L (150 mg/dL) (16), with total cholesterol <6.21 mmol/L (240 mg/dL) and LDL cholesterol <4.14 mmol/L (160 mg/dL). Type 2 diabetes was defined by self-reported history or hospital diagnosis of type 2 or unspecified diabetes (ICD-10 E11, E14), treatment with insulin or oral hypoglycemic drugs, serum glucose level ≥11.1 mmol/L (200 mg/dL), and/or glycated hemoglobin (HbA1c) ≥ 48 mmol/mol (6.5%) (17). Cardiovascular disease was defined as self-reported history or hospital diagnosis of angina, myocardial infarction, stroke, or transient ischemic attack (ICD-10 I20-I25, I60-I64, I69, G45). Chronic liver disease was defined as ICD-10 I85.0, I85.9, K72.1, K72.9, K73.0, K73.1, K73.2, K73.8, K73.9, K74.0, K74.1, K74.2, K74.6, K76.0, K76.6, K76.7 from hospital admissions and death register. Data on alcohol use were collected through a touch-screen questionnaire. A detailed explanation of alcohol use pipeline has been provided elsewhere (17). Briefly, frequency of daily alcohol use (g/day) was quantified based on the average weekly or monthly alcohol intake. Alcohol grams for each type of drink were derived from the corresponding reference alcohol content (1 unit of alcohol = 8 g of alcohol). Excessive alcohol use was defined when daily alcohol intake was ≥30 g and ≥20 g for men and women, respectively (18). Statistical Analyses All statistical analyses were performed using the software R, version 4.0.4 (R Foundation for Statistical Computing, Vienna, Austria). Continuous variables were shown as mean ± SD or median (interquartile range) as appropriate. Categorical variables were shown as number(percentage). Generalized linear models were used to test the association between clinical data and lipoprotein phenotypes. All multivariable models were adjusted for age, sex, and BMI. A value of P < .05 was considered to be statistically significant. All continuous traits were rank-based inverse normal transformed before the analysis. Sensitivity analyses were performed (1) utilizing LDL cholesterol cut-off ≥130 mg/dL or ≥100 mg/dL to define isolated hypercholesterolemia, (2) excluding those with excessive daily alcohol use (≥20/30 g/day for women/men, respectively), and (3) defining MASLD according to the new nomenclature (1). Model Selection and Training Model training, validation, and selection were performed in MATLAB R2023a. To examine the complex relationships between the traits under study in predicting the liver triglyceride content, we performed a model selection approach by comparing the predictive performance of several machine learning methods, namely feed-forward neural networks, binary decision trees, support vector machine, ensemble of learners (least-square boosting or bootstrap aggregation), Gaussian process regression, and Gaussian kernel regression (19, 20). The dataset was split into training (80%) and test (20%) sets. All continuous predictors were normalized (mean 0, variance 1), and liver triglyceride content was rank-based inverse normal transformed. The best performed model was selected in a nested 10-fold cross-validation on the training set (80%) where in the outer fold an asynchronous successive halving algorithm (21) was used to tune the hyperparameters of all the selected machine learning algorithms in another inner 10-fold cross-validation. The best performing algorithm was then selected based on the minimum log (1 + loss), where loss is the cross-validation mean squared error (MSE), across iterations of the outer folds. The best performing machine learning method from this step was then trained on the training set with hyperparameters optimized in a 10-fold cross-validation using Bayesian optimization. Finally, the performance of the model was assessed on the remaining test set. For the final trained model, we generated individual conditional expectation (ICE) plots using 1000 randomly selected samples from the training set to visualize how individual variables influenced the model predictions. ICE plots especially highlight the nonlinearities and interactions by focusing on the impact of individual instances of each predictor while conditioning on observed values of other predictors in the model (22). Furthermore, to evaluate the contribution of each predictor in the trained model towards the predicted values, we computed feature importance using Shapley additive explanation (SHAP) values (23, 24). Briefly, SHAP values for each feature show the difference in expected value of the response with and without that predictor. The mean absolute SHAP value for 1000 randomly selected individuals (same as ICE plots) was used as a measure of each predictor importance. Results Clinical Characteristics of the Study Cohort To understand the relationship between lipoproteins and MASLD, we started by defining isolated hypercholesterolemia and isolated hypertriglyceridemia based on hospital diagnosis (hypercholesterolemia and hypertriglyceridemia, respectively) and laboratory tests (total cholesterol ≥6.21 mmol/L [240 mg/dL] or LDL cholesterol ≥4.14 mmol/L [160 mg/dL], and triglycerides ≥1.69 mmol/L [150 mg/dL], respectively). The overall cohort included a total of 218 732 individuals of European ancestry, of whom 14 937 with liver MRI data available. Individuals with isolated hypercholesterolemia were older, more represented by women, and had higher BMI than the control group. Individuals with isolated hypertriglyceridemia were older, mostly men, and had higher BMI compared with control. Consistently, individuals with both isolated hypercholesterolemia and isolated hypertriglyceridemia had higher cardiovascular disease and type 2 diabetes than the control group (Table 1). Table 1.Clinical characteristics of UK biobank participants stratified by lipoprotein phenotypes (n = 218 732) Healthy control (n = 121 021)Isolated hypercholesterolemia (n = 53 117)Isolated hypertriglyceridemia (n = 44 594) Βeta/OR95% CI P value Βeta/OR95% CI P valueAge, years53.9 ± 8.2157.8 ± 7.19.47(0.46, 0.48)<1 × 10–30755.3 ± 8.18.12(0.1, 0.13)4.6 × 10–87Men, n(%)45 886 (37.9)16 165 (30.4).67(0.65, 0.68)8.7 × 10–27327 928 (62.6)2.59(2.53, 2.65)<1 × 10–307BMI, kg/m225.9 ± 4.426.2 ± 4.19.09(0.08, 0.1)1.27 × 10–6428.7 ± 4.76.59(0.58, 0.6)<1 × –307LDL cholesterol, mmol/L3.16 ± 0.514.29 ± 0.601.42(1.4, 1.4)<1 × 10–3073.4 ± 0.45.35(0.34, 0.36)<1 × 10–307Triglycerides, mmol/L1.06 (0.50)1.22 (0.47).34(0.33, 0.35)<1 × 10–3072.24 (0.89)1.52(1.5, 1.5)<1 × 10–307 Comorbidities and lifestyle Cardiovascular disease, n(%)7459 (6.2)6188 (11.6)1.73(1.67, 1.8)1.6 × 10–1854009 (8)1.10(1.05, 1.15)2.19 × 10–05Type 2 diabetes, n (%)2880 (2.4)1513 (2.9)1.11(1.04, 1.18)2.25 × 10–32983 (6.7)1.77(1.68, 1.87)1.34 × 10–89CLD, n(%)990 (0.8)499 (0.9)1.13(1.01, 1.26)0.032736 (1.7)1.42(1.28, 1.57)1.22 × 10–11At least 1 cardiometabolic risk factor, n(%)a96 917 (85)46 171 (90)1.11(1.07, 1.16)9.30 × 10–844 572 (100)213(137.47, 331.44)3.8 × 10–126Alcohol use ≥20/30 g/day, n (%)23 309 (19.3)11 646 (22)1.25(1.22, 1.28)1.4 × 10–638497 (19.1).90(0.87, 0.93)1.19 × 10–12 Liver MRIb PDFF, %3.79 ± 3.763.88 ± 3.66.07(0.04, 0.1)5.4 × 10–56.58 ± 5.85.35(0.32, 0.39)1.17 × 10–84cT1, ms687 ± 50.1684 ± 47−0.05(−0.09, −0.01)8.88 × 10–3713 ± 59.2.25(0.21, 0.3)5.48 × 10–31MASLD, n (%)1329 (14.3)475 (15.6)1.17(1.03, 1.34).019993 (38)2.23(1.98, 2.5)1.23 × 10–41MASLDc, n (%)928 (10.8)307 (10.7).99(0.85, 1.16).940771 (29.5)2.30(2.03, 2.61)8.91 × 10–39MASH, n (%)202 (2.2)60 (2).99(0.72, 1.37).970197 (7.6)2.14(1.70, 2.70)1.29 × 10–10MASHc, n (%)156 (1.82)44 (1.53).93(0.64, 1.34).681155 (5.94)2.14(1.65, 2.77)9.06 × 10–09Continuous variables are shown as mean ± SD or median (interquartile) as appropriate. Categorical variables are shown as number(percentage). P values are from multivariable regression models comparing isolated hypercholesterolemia or isolated hypertriglyceridemia with healthy control group. P values are adjusted for age, sex, and BMI and values < .05 were considered to be statistically significant.Abbreviations: BMI, body mass index; cT1, iron-corrected T1; CLD, chronic liver disease; LDL, low-density lipoprotein; MASLD, metabolic dysfunction–associated steatotic liver disease; MASH, metabolic dysfunction–associated steatohepatitis; MRI, magnetic resonance imaging; OR, odds ratio; PDFF, proton density fat fraction. a At least 1 cardiometabolic risk factor among (1) BMI ≥25 kg/m2 or waist circumference >94/80 cm for men/women, (2) fasting serum glucose ≥5.6 mmol/L or HbA1c ≥ 39 mmol/L or treatment for type 2 diabetes, (3) blood pressure ≥130/85 mmHg or antihypertensive treatment, (4) triglycerides ≥ 1.70 mmol/L or lipid-lowering treatment, (5) high-density lipoprotein cholesterol ≤1/1.3 mmol/L for men/women or lipid lowering treatment. b Liver MRI data are referred to a subset of 14 937 individuals (healthy control: n = 9278; isolated hypercholesterolemia: n = 3048; isolated hypertriglyceridemia: n = 2611). MASLD was defined as steatosis by PDFF >5.5%, MASH by PDFF >5.5% and cT1 >800 ms. c MASLD was defined as steatosis by PDFF >5.5%, no high alcohol consumption and at least 1 cardiometabolic risk factor present. MASH was defined similarly with additional inclusion criterion of cT1 >800 ms. Clinical characteristics of UK biobank participants stratified by lipoprotein phenotypes (n = 218 732) Continuous variables are shown as mean ± SD or median (interquartile) as appropriate. Categorical variables are shown as number(percentage). P values are from multivariable regression models comparing isolated hypercholesterolemia or isolated hypertriglyceridemia with healthy control group. P values are adjusted for age, sex, and BMI and values < .05 were considered to be statistically significant. Abbreviations: BMI, body mass index; cT1, iron-corrected T1; CLD, chronic liver disease; LDL, low-density lipoprotein; MASLD, metabolic dysfunction–associated steatotic liver disease; MASH, metabolic dysfunction–associated steatohepatitis; MRI, magnetic resonance imaging; OR, odds ratio; PDFF, proton density fat fraction. a At least 1 cardiometabolic risk factor among (1) BMI ≥25 kg/m2 or waist circumference >94/80 cm for men/women, (2) fasting serum glucose ≥5.6 mmol/L or HbA1c ≥ 39 mmol/L or treatment for type 2 diabetes, (3) blood pressure ≥130/85 mmHg or antihypertensive treatment, (4) triglycerides ≥ 1.70 mmol/L or lipid-lowering treatment, (5) high-density lipoprotein cholesterol ≤1/1.3 mmol/L for men/women or lipid lowering treatment. b Liver MRI data are referred to a subset of 14 937 individuals (healthy control: n = 9278; isolated hypercholesterolemia: n = 3048; isolated hypertriglyceridemia: n = 2611). MASLD was defined as steatosis by PDFF >5.5%, MASH by PDFF >5.5% and cT1 >800 ms. c MASLD was defined as steatosis by PDFF >5.5%, no high alcohol consumption and at least 1 cardiometabolic risk factor present. MASH was defined similarly with additional inclusion criterion of cT1 >800 ms. When we looked at the liver triglyceride content as assessed by liver MRI, those with isolated hypercholesterolemia had a marginal 2.4% increase in this trait compared with control, with values remaining in the normal range, namely <5.5% (mean PDFF 3.9% vs 3.8%, P < .001), whereas those with isolated hypertriglyceridemia had an almost 2-fold increase (mean PDFF 6.6% vs 3.8%, P < .001) (Table 1). Interestingly, individuals with isolated hypertriglyceridemia also had higher cT1, a measure of liver inflammation and fibrosis, than the control group (713 msec vs 687 msec, P < .001), whereas in those with isolated hypercholesterolemia cT1 was lower (684 ms vs 687 ms, P < .001). Consistently, individuals with isolated hypercholesterolemia had 9% higher rate of MASLD and no difference in MASH rate (P = .019 and P = .97, respectively) compared with control, whereas there was a nearly 3-fold and 4-fold enrichment of these conditions in individuals with isolated hypertriglyceridemia (P < .001). Utilizing LDL cholesterol cut-off ≥130 mg/dL or ≥100 mg/dL to define isolated hypercholesterolemia, excluding those with excessive daily alcohol use (≥20/30 g/day for women/men, respectively), or defining MASLD according to the new nomenclature, results were similar to the main analysis (Tables 2-4). Table 2.Clinical characteristics of UK biobank participants stratified by lipoprotein phenotypes using LDL cholesterol cut-off ≥100 mg/mL to define isolated hypercholesterolemia (n = 176 604) Healthy controlIsolated hypercholesterolemiaIsolated hypertriglyceridemia(n = 26 595)(n = 147 543)(n = 2466) Βeta/OR95% CI P value Βeta/OR95% CI P valueAge, years52.9 ± 8.3555.5 ± 8.01.31(0.30, 0.32)<1 × 10–30754.2 ± 8.630.03(−0.01, 0.07).194Men, n(%)9869 (37.1)52 182 (35.4).89(0.87, 0.92)1.25 × 10–151591 (64.5)2.76(2.53, 3.01)1.98 × 10–112BMI, kg/m225.7 ± 4.5926.0 ± 4.29.09(0.08, 0.11)2.01 × 10–4428.3 ± 5.17.46(0.42, 0.50)2.22 × 10–107LDL cholesterol, mmol/L2.29 ± 0.263.65 ± 0.671.86(1.80, 1.90)<1 × 10–3072.32 ± 0.25.11(0.06, 0.15)6.54 × 10–06Triglycerides, mmol/L0.88 (0.45)1.13 (0.49).56(0.54, 0.57)<1 × 10–3072.19 (0.88)1.66(1.60, 1.70)<1 × 10–307 Comorbidities and lifestyle Cardiovascular disease, n(%)1601 (6.0)12 046 (8.16)1.22(1.15, 1.29)4.81 × 10–12209 (8.48).99(0.85, 1.16).923Type 2 diabetes, n (%)879 (3.3)3514 (2.4).66(0.61, 0.72)8.12 × 10–25248 (10.1)1.96(1.67, 2.30)7.87 × 10–17CLD, n (%)283 (1.1)1206 (0.82).75(0.66, 0.86)2.12 × 10–0556 (2.27)1.41(1.04, 1.91).028Alcohol use ≥20/30 g/day, n (%)5515 (20.8)29 440 (20).97(0.94, 1).048521 (21.2)1(0.90, 1.11).992 Liver MRIa PDFF, %3.84 ± 3.933.81 ± 3.7.03(−0.01, 0.06).1955.77 ± 5.62.19(0.05, 0.34).010cT1, ms688 ± 51.5686 ± 48.9−0.05(−0.09, −0.002).041701 ± 66.1.03(−0.15, 0.21).720MASLD, n (%)316 (15.3)1488 (14.5).89(0.77, 1.04).14637 (30.1)1.45(0.89, 2.37).136MASH, n (%)43 (2.1)219 (2.1)1.01(0.70, 1.45).9717 (5.70)1.31(0.47, 3.6).606 a Liver MRI data are referred to a subset of 12 449 individuals (healthy control: n = 2071; isolated hypercholesterolemia: n = 10 255; isolated hypertriglyceridemia: n = 123). MASLD was defined as steatosis by PDFF >5.5%, MASH by PDFF >5.5% and cT1 >800 ms.Abbreviations: BMI, body mass index; cT1, iron-corrected T1; CLD, chronic liver disease; LDL, low-density lipoprotein; MASLD, metabolic dysfunction–associated steatotic liver disease; MASH, metabolic dysfunction–associated steatohepatitis; MRI, magnetic resonance imaging; OR, odds ratio; PDFF, proton density fat fraction. Clinical characteristics of UK biobank participants stratified by lipoprotein phenotypes using LDL cholesterol cut-off ≥100 mg/mL to define isolated hypercholesterolemia (n = 176 604) a Liver MRI data are referred to a subset of 12 449 individuals (healthy control: n = 2071; isolated hypercholesterolemia: n = 10 255; isolated hypertriglyceridemia: n = 123). MASLD was defined as steatosis by PDFF >5.5%, MASH by PDFF >5.5% and cT1 >800 ms. Abbreviations: BMI, body mass index; cT1, iron-corrected T1; CLD, chronic liver disease; LDL, low-density lipoprotein; MASLD, metabolic dysfunction–associated steatotic liver disease; MASH, metabolic dysfunction–associated steatohepatitis; MRI, magnetic resonance imaging; OR, odds ratio; PDFF, proton density fat fraction. Table 4.Clinical characteristics of UK Biobank participants stratified by lipoprotein phenotypes excluding those with excessive alcohol use and defining MASLD according to the new nomenclature (n = 175 280) Healthy controlIsolated hypercholesterolemiaIsolated hypertriglyceridemia(n = 97 712)(n = 41 471)(n = 36 097) Βeta/OR95% CI P value Βeta/OR95% CI P valueAge, years53.9 ± 8.2658 ± 7.2.48(0.47, 0.49)<1 × 10–30755.4 ± 8.22.14(0.12, 0.15)3.67 × 10–99Men, n (%)35 644 (36.5)11 546 (27.8).63(0.61, 0.65)4.20 × 10–26721 658 (60)2.53(2.47, 2.60)<1 × 10–307BMI, kg/m225.9 ± 4.526.1 ± 4.26.08(0.07, 0.09)3.49 × 10–4128.7 ± 4.84.60(0.59, 0.61)<1 × 10–307LDL cholesterol, mmol/L3.18 ± 0.514.32 ± 0.591.44(1.40, 1.40)<1 × 10–3073.4 ± 0.45.36(0.34, 0.37)<1 × 10–307Triglycerides, mmol/L1.07 (0.49)1.22 (0.46).35(0.33, 0.36)<1 × 10–3072.24 (0.88)1.51(1.50, 1.50)<1 × 10–307 Comorbidities and lifestyle Cardiovascular disease, n(%)6063 (6.2)4904 (11.8)1.74(1.67, 1.82)1.43 × 10–1523278 (9.1)1.09(1.04, 1.15)2.72 × 10–04Type 2 diabetes, n(%)2435 (2.5)1236 (3.0)1.12(1.04, 1.20)2.78 × 10–032509 (7.0)1.77(1.66, 1.88)2.80 × 10–75CLD, n(%)769 (0.79)379 (0.91)1.14(1, 1.29)0.047612 (1.7)1.53(1.36, 1.71)2.53 × 10–13Alcohol use ≥20/30 g/day, n (%)————————— Liver MRIa PDFF, %3.67 ± 3.653.69 ± 3.46.06(0.02, 0.096)2.67 × 10–036.36 ± 5.71.37(0.33, 0.41)3.28 × 10–75cT1, ms689 ± 49.9686 ± 46.3−.05(−0.095, −0.004).034713 ± 59.1.23(0.18, 0.28)6.59 × 10–21MASLD, n (%)1005 (13.4)331 (14.2)1.14(0.98, 1.33).099774 (36.5)2.32(2.03, 2.64)5.09 × 10–36MASLDb, n (%)928 (13.5)307 (14.1)1.12(0.95, 1.32).184771 (36.4)2.48(2.17, 2.84)2.13 × 10–40MASH, n (%)163 (2.18)45 (1.9)1(0.69, 1.45).994157 (7.4)2.19(1.69, 2.85)4.11 × 10–09MASHb, n (%)156 (2.27)44 (2.02)1.03(0.70, 1.50).897155 (7.31)2.22(1.70, 2.90)4.02 × 10–09Abbreviations: BMI, body mass index; cT1, iron-corrected T1; CLD, chronic liver disease; LDL, low-density lipoprotein; MASLD, metabolic dysfunction-associated steatotic liver disease; MASH, metabolic dysfunction-associated steatohepatitis; MRI, magnetic resonance imaging; OR, odds ratio; PDFF, proton density fat fraction. a Liver MRI data are referred to a subset of 11 939 individuals not treated with any lipid-lowering drugs and excluding those with high alcohol consumption (healthy control: n = 7486; isolated hypercholesterolemia: n = 2330; isolated hypertriglyceridemia: n = 2123). MASLD was defined as steatosis by PDFF >5.5%, MASH by PDFF >5.5% and cT1 >800 ms. b MASLD was defined as steatosis by PDFF >5.5%, no high alcohol consumption and at least 1 cardiometabolic risk factor present. MASH was defined similarly with additional inclusion criterion of cT1 >800 ms. Clinical characteristics of UK Biobank participants stratified by lipoprotein phenotypes excluding those with excessive alcohol use and defining MASLD according to the new nomenclature (n = 175 280) Abbreviations: BMI, body mass index; cT1, iron-corrected T1; CLD, chronic liver disease; LDL, low-density lipoprotein; MASLD, metabolic dysfunction-associated steatotic liver disease; MASH, metabolic dysfunction-associated steatohepatitis; MRI, magnetic resonance imaging; OR, odds ratio; PDFF, proton density fat fraction. a Liver MRI data are referred to a subset of 11 939 individuals not treated with any lipid-lowering drugs and excluding those with high alcohol consumption (healthy control: n = 7486; isolated hypercholesterolemia: n = 2330; isolated hypertriglyceridemia: n = 2123). MASLD was defined as steatosis by PDFF >5.5%, MASH by PDFF >5.5% and cT1 >800 ms. b MASLD was defined as steatosis by PDFF >5.5%, no high alcohol consumption and at least 1 cardiometabolic risk factor present. MASH was defined similarly with additional inclusion criterion of cT1 >800 ms. Table 3.Clinical characteristics of UK biobank participants stratified by lipoprotein phenotypes using LDL cholesterol cut-off ≥130 mg/mL to define isolated hypercholesterolemia (n = 192 838) Healthy controlIsolated hypercholesterolemiaIsolated hypertriglyceridemia(n = 79 152)(n = 94 986)(n = 18 700) Βeta/OR95% CI P value Βeta/OR95% CI P valueAge, years53.2 ± 8.2456.7 ± 7.67.40(0.39, 0.41)<1 × 10–30754.8 ± 8.3.10(0.09, 0.12)8.74 × 10–35Men, n (%)28 547 (36.1)33 504 (35.3).91(0.89, 0.93)3.02 × 10–2011 705 (62.6)2.65(2.56, 2.74)<1 × 10–307BMI, kg/m225.6 ± 4.4226.3 ± 4.24.18(0.17, 0.19)2.25 × 10–30728.5 ± 4.92.58(0.56, 0.59)<1 × 10–307LDL cholesterol, mmol/L2.86 ± 0.374.01 ± 0.561.53(1.50, 1.50)<1 × 10–3072.97 ± 0.33.24(0.23, 0.26)3.33 × 10–173Triglycerides, mmol/L0.99 (0.48)1.2 (0.47).44(0.43, 0.45)<1 × 10–3072.22 (0.89)1.56(1.50, 1.60)<1 × 10–307 Comorbidities and lifestyle Cardiovascular disease, n (%)4622 (5.84)9025 (9.5)1.39(1.34, 1.44)6.11 × 10–641651 (8.83)1.09(1.02, 1.16)6.38 × 10–03Type 2 diabetes, n (%)2021 (2.55)2372 (2.5).83(0.78, 0.88)3.39 × 10–091415 (7.57)1.78(1.65, 1.91)1.04 × 10–50CLD, n (%)684 (0.86)805 (0.85).90(0.81, 1).054356 (1.9)1.47(1.28, 1.69)3.34 × 10–08Alcohol use ≥20/30 g/day, n (%)15 999 (20.3)18 956 (20)1.00(0.98, 1.02).9733772 (20.3).94(0.91, 0.98)7.08 × 10–03 Liver MRIa PDFF, %3.68 ± 3.773.95 ± 3.7.08(0.05, 0.11)8.22 × 10–086.14 ± 5.64.31(0.26, 0.36)2.29 × 10–31cT1, ms686 ± 50.2687 ± 48.5−.02(−0.05, 0.02).319709 ± 59.6.21(0.15, 0.27)1.15 × 10–10MASLD, n (%)832 (13.4)972 (15.9)1.12(1, 1.26).057377 (34.2)1.91(1.61, 2.27)1.66 × 10–13MASH, n (%)130 (2.1)132 (2.16).96(0.73, 1.26).74474 (6.71)1.93(1.37, 2.71)1.55 × 10–04Abbreviations: BMI, body mass index; cT1, iron-corrected T1; CLD, chronic liver disease; LDL, low-density lipoprotein; MASLD, metabolic dysfunction–associated steatotic liver disease; MASH, metabolic dysfunction–associated steatohepatitis; MRI, magnetic resonance imaging; OR, odds ratio; PDFF, proton density fat fraction. a Liver MRI data are referred to a subset of 13 429 individuals (healthy control: n = 6209; isolated hypercholesterolemia: n = 6117; isolated hypertriglyceridemia: n = 1103). MASLD was defined as steatosis by PDFF >5.5%, MASH by PDFF >5.5% and cT1 >800 ms. Clinical characteristics of UK biobank participants stratified by lipoprotein phenotypes using LDL cholesterol cut-off ≥130 mg/mL to define isolated hypercholesterolemia (n = 192 838) Abbreviations: BMI, body mass index; cT1, iron-corrected T1; CLD, chronic liver disease; LDL, low-density lipoprotein; MASLD, metabolic dysfunction–associated steatotic liver disease; MASH, metabolic dysfunction–associated steatohepatitis; MRI, magnetic resonance imaging; OR, odds ratio; PDFF, proton density fat fraction. a Liver MRI data are referred to a subset of 13 429 individuals (healthy control: n = 6209; isolated hypercholesterolemia: n = 6117; isolated hypertriglyceridemia: n = 1103). MASLD was defined as steatosis by PDFF >5.5%, MASH by PDFF >5.5% and cT1 >800 ms. Independent predictors of liver triglyceride content To understand the relative contribution of lipoproteins in determining liver triglyceride content, we first performed a multiple regression analysis by including in the model clinically well-established risk factors for MASLD/MASH (Fig. 1). As expected, BMI was the strongest independent predictor of liver triglyceride content (Beta .55, P < 1 × 10–307), followed by isolated hypertriglyceridemia (Beta .35, P = 2.7 × 10–83), male sex (Beta .33, P = 3.2 × 10–134), type 2 diabetes (Beta .30, P = 2.85 × 10–9), age (Beta .09, P = 3.1 × 10–45), and lastly isolated hypercholesterolemia (Beta .07, P = 2.9 × 10–5). Independent predictors of liver triglyceride content by MRI-PDFF in Europeans from the UK biobank (n = 14 937). Beta coefficients are from multivariable linear regression with PDFF as dependent variable. Abbreviations: BMI, body mass index; PDFF, proton density fat fraction. Features Importance of Liver Triglyceride Content by Artificial Neural Network Next, to consider the possibility of a nonlinear relationship between these traits and their complex interactions, we performed a model selection approach by comparing the predictive performance of several machine learning methods (see methods section) and selected the best performing algorithms based on the minimum cross-validated MSE (Fig. 2 and Fig. 3). Among the tested machine learning methods, feed-forward neural network had the lowest MSE in the 10-fold nested cross-validated model selection, suggesting this model to be the best performing algorithm (25). Global (A) and local (B) features importance in predicting liver triglyceride content by MRI-PDFF based on SHAP values. The mean absolute SHAP values for 1000 randomly selected individuals are depicted to illustrate global (A) and local (B) features importance. The bar plot (A) shows the strength of the relationship between each feature and liver triglyceride content (ie, the higher mean absolute SHAP value, the stronger the association). The local summary plot (B) shows the direction of the relationship between each feature and liver triglyceride content. Positive SHAP values are indicative of a direct relationship, whereas negative SHAP values are indicative of an inverse relationship. Abbreviations: BMI, body mass index; HbA1c, glycated hemoglobin; LDL, low-density lipoprotein; PDFF, proton density fat fraction; SHAP, Shapley additive explanation. Individual conditional expectation (ICE) plots for predictors of liver triglyceride content by MRI-PDFF. ICE plots were generated using 1000 randomly selected samples from the training set. Abbreviations: BMI, body mass index; HbA1c, glycated hemoglobin; LDL, low-density lipoprotein; PDFF, proton density fat fraction. Next, we computed features importance using SHAP values as a measure of the contribution of each trait towards liver triglyceride content. The feature with the largest absolute mean SHAP value was BMI, followed by circulating triglycerides, male sex, and age. LDL cholesterol, glucose, and HbA1c were not contributing to the model (Fig. 2). Then, to visualize nonlinearity and interactions, we focused on individual observations generated by ICE plots from 1000 randomly selected individuals and adjusted for the other predictors in the model (Fig. 3). This analysis showed that BMI/liver triglyceride content had a sigmoid relationship and circulating triglycerides/liver triglyceride content had a linear relationship with a plateau. For these 2 traits, for any given increase in each of the trait there was an increase in the predicted liver triglyceride content at an individual level. For the other traits, for which the overall contribution was lower, they were either positively contributing to the predicted liver triglyceride content or had no effect on it. Discussion The major findings of this study are that (1) individuals with isolated hypertriglyceridemia, but not with isolated hypercholesterolemia, have higher risk of MASLD and MASH, and (2) circulating triglycerides are the second strongest, after BMI, independent predictor of liver triglyceride content. To the best of our knowledge, this is the first study assessing the prevalence of MASLD/MASH in individuals with common dyslipidemia, namely isolated hypercholesterolemia and isolated hypertriglyceridemia, from a large and well-characterized cohort, and the first study examining the independent contribution of these lipoprotein phenotypes to liver triglyceride content. We found that the prevalence of MASLD and MASH (as assessed by liver MRI) among individuals with isolated hypercholesterolemia was 16% and 2%, respectively, whereas among those with isolated hypertriglyceridemia was 38% and 8%, respectively. Consistently, the mean PDFF value in individuals with isolated hypercholesterolemia was within the normal range, whereas it was elevated in those with isolated hypertriglyceridemia (3.9% and 6.6%, respectively). Similarly, the mean cT1 value, an index of liver inflammation and fibrosis, was higher in isolated hypertriglyceridemia than in control individuals, whereas it was lower in those with isolated hypercholesterolemia. Could excessive alcohol intake have contributed to our results? Excess in alcohol intake is a well-known environmental risk factor contributing to steatotic liver disease (1). Interestingly, the frequency of individuals with alcohol abuse was higher in isolated hypercholesterolemia and yet there was no enrichment in MASLD/MASH. Conversely, the frequency of individuals having excessive alcohol intake was lower in isolated hypertriglyceridemia that was enriched in MASLD/MASH. These data suggest that excess in alcohol use did not contribute to our results. At a multiple regression analysis, a commonly used statistical method in clinical research, type 2 diabetes and sex were as important as circulating triglycerides in determining liver triglyceride content (Fig. 1). However, when we used a more sophisticated statistical method based on machine learning, namely a feed-forward neural network, the weight of glucose and HbA1c was largely diminished, whereas circulating triglycerides and sex remained as second and third strongest independent predictors of liver triglyceride content, respectively, after BMI (Fig. 2). Results were virtually identical when we used type 2 diabetes instead of glucose and HbA1c (data not shown). A feed-forward neural network consists of 3 fundamental components: input, output, and hidden layers, where interconnected artificial neurons operate. These neurons communicate through weighted connections, which are iteratively adjusted during the learning process to minimize errors. This dynamic adaptation empowers the model to adeptly accommodate diverse datasets, fostering improved generalization (26) and accounting for nonlinear relationship and complex interactions between traits. The main strengths of this study are that (1) the analysis was carried out in a well-phenotyped population-based cohort, (2) liver triglyceride content was assessed by MRI-PDFF, an accurate and reproducible biomarker for hepatic steatosis accepted as a noninvasive reference standard to liver biopsy to diagnose MASLD (27-29), and (3) the use of machine learning methods to establish independent predictors of liver triglyceride content. It is important to bear in mind that the concept of independent predictor does not imply causality. In fact, circulating triglyceride levels in fasting conditions are a proxy of the amount of secreted very low–density lipoprotein particles that are directly correlated to the amount of hepatocyte triglyceride content (30). Therefore, hypertriglyceridemia is a consequence rather than a cause of increased liver triglyceride content. An additional limitation of the study is the inclusion of only Europeans and, therefore, these results may not be translated in other ethnic groups. In conclusion, in individuals with isolated hypertriglyceridemia there is a nearly 3-fold and 4-fold enrichment of MASLD and MASH, respectively. Furthermore, we demonstrated that isolated hypertriglyceridemia is the second strongest, after obesity, independent predictor of these conditions. Our results suggest that individuals with hypertriglyceridemia, but not with hypercholesterolemia, are at high-risk for MASLD/MASH and should be screened for this chronic liver disease. Acknowledgments We thank the staff and the participants of the UK Biobank study. This research has been conducted using the UK Biobank resource (application #37142). Abbreviations body mass index iron-corrected T1 International Classification of Diseases 10th edition glycated hemoglobin individual conditional expectation low-density lipoprotein metabolic dysfunction–associated steatohepatitis metabolic dysfunction–associated steatotic liver disease magnetic resonance imaging mean squared error proton density fat fraction Shapley additive explanation Contributor Information Federica Tavaglione, Operative Unit of Clinical Medicine and Hepatology, Fondazione Policlinico Universitario Campus Bio-Medico, 00128 Rome, Italy; Research Unit of Clinical Medicine and Hepatology, Department of Medicine and Surgery, Università Campus Bio-Medico di Roma, 00128 Rome, Italy. Giuseppe Marafioti, Clinical Nutrition Unit, Department of Medical and Surgical Sciences, University Magna Graecia, 88100 Catanzaro, Italy. Stefano Romeo, Clinical Nutrition Unit, Department of Medical and Surgical Sciences, University Magna Graecia, 88100 Catanzaro, Italy; Department of Molecular and Clinical Medicine, Institute of Medicine, Sahlgrenska Academy, Wallenberg Laboratory, University of Gothenburg, 41345 Gothenburg, Sweden; Department of Cardiology, Sahlgrenska University Hospital, 41345 Gothenburg, Sweden; Department of Medicine, Karolinska Institutet, Huddinge, 14152 Stockholm, Sweden. Oveis Jamialahmadi, Department of Molecular and Clinical Medicine, Institute of Medicine, Sahlgrenska Academy, Wallenberg Laboratory, University of Gothenburg, 41345 Gothenburg, Sweden. Funding This research was funded by the Cancerfonden (22 2270 Pj), the Swedish Research Council (Vetenskapsradet (VR), 2023-02079), the Swedish state under the Agreement between the Swedish government and the county councils (the ALF agreement, ALFGBG-965360), the Swedish Heart Lung Foundation (20220334), Knut and Alice Wallenberg Stiftelse (KAW 2017.0203), the Novonordisk Distinguished Investigator Grant—Endocrinology and Metabolism (NNF23OC0082114), the Novonordisk Project grant (NNF20OC0063883) (SR). Author Contributions F.T., S.R., and O.J. contributed to study concept and design. F.T., S.R., and O.J. contributed to drafting the manuscript. O.J. contributed to perform the statistical analysis. All authors contributed to analysis and interpretation of data, critically revised the manuscript for important intellectual content and approved the final version for submission. All authors had fully access to all the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis. Disclosures In the last 3 years: S.R. has served as a consultant or been lecturer for AstraZeneca, Sanofi, AMGEN, Novartis, Ribocure, Regeneron, Foresite Labs, Ultragenyx and received research grant from AstraZeneca; O.J. has served as consultant for Ribocure; F.T. and G.M. have no conflict of interest to declare. Data Availability UK Biobank data are publicly available through a procedure described at http://www.ukbiobank.ac.uk/using-the-resource/. All other data supporting the findings of this study are available within the article. References Associated Data Data Citations Data Availability Statement UK Biobank data are publicly available through a procedure described at http://www.ukbiobank.ac.uk/using-the-resource/. All other data supporting the findings of this study are available within the article.