Data-driven classification of metabolic-associated steatotic liver disease subtypes predicting hepatic and extrahepatic progression

Abstract

Background:

Metabolic-associated steatotic liver disease (MASLD) is a heterogeneous condition with highly variable outcomes. We aimed to distinguish the subtypes of MASLD that were associated with varying risks of hepatic and extrahepatic outcomes.

Methods:

An innovative multi-task deep least absolute shrinkage and selection operator (LASSO) algorithm was developed for feature selection in the discovery cohort (n = 1111, 87.6% [973/1111] of biopsy-proved MASLD), followed by clustering analysis in MASLD. Validation was performed in 6172 individuals who undertook health check-ups (MASLD: 43.9% [2710/6172], mean follow-up 27.6 months) and 7406 participants from the Third National Health and Nutrition Examination Survey (NHANES III) (MASLD: 37.3%, mean follow-up 280.2 months).

Results:

Four clusters with distinct risks of hepatic and extrahepatic outcomes were identified in the discovery cohort: Cluster 1 characterized by subcutaneous adiposity, modest metabolic disorders, but lower risk of cardiovascular disease (CVD); Cluster 2 characterized by significant hyperlipidemia, substantial liver damage, and increased hepatic fibrosis risk; Cluster 3 characterized by low muscle mass, remarkable chronic systemic inflammation, and a higher risk of cardiovascular–kidney complications; and Cluster 4 characterized by severe insulin resistance, visceral adiposity, poor glucose and lipid control, and severe liver damage, conferring a high risk of cardiovascular–liver–kidney complications. Additionally, Cluster 4 exhibited the highest frequencies of PNPLA3 risk alleles. The prognostic relevance was further confirmed in external validation cohorts. Specifically, Clusters 3 and 4 had increased risks of all-cause and CVD-related mortality.

Conclusions:

We developed a novel algorithm that identified four MASLD clusters, each characterized by distinct clinical features and varying risks of hepatic and extrahepatic outcomes. This classification facilitates the precise integration of MASLD risk stratification and management within the cardiovascular–liver–kidney–metabolic framework.

Introduction

Metabolic-associated steatotic liver disease (MASLD) is a clinically heterogeneous condition with highly variable outcomes affecting more than 30% individuals globally.^[1] The situation is particularly severe in China, representing a major public health challenge.^[2] The classification of MASLD is normally based on histological progression, including simple steatosis, metabolic dysfunction-associated steatohepatitis (MASH), and fibrosis/cirrhosis. More importantly, besides hepatic progression, MASLD increases the risk of type 2 diabetes mellitus (T2DM), chronic kidney disease (CKD), and cardiovascular disease (CVD).^[3] Currently, personalized treatments and management strategies remain limited for MASLD. Thus, the prognostic stratification system for hepatic and extrahepatic outcomes governing clinical decision-making is urgently needed.

Extensive evidences indicate that heterogeneity of MASLD prognosis is influenced by multiple factors, such as age, sex, insulin sensitivity, glucose and lipid metabolism, along with genetic background.^[3] Previous studies have incorporated empirically derived serological markers to define MASLD subtypes for identifying individuals at high risk of CVD.^[4,5] However, due in part to the lack of pathological validation, current subtyping systems do not provide comprehensive or accurate risk estimates for both hepatic and extrahepatic outcomes. Moreover, the influence of genetic variants on fibrosis progression in MASLD patients remains poorly characterized within existing classification frameworks. To better understand and overcome these limitations, we hypothesized that a refined MASLD classification—derived through deep learning-based feature selection and integrating clinical, histological, and genetic data—is essential for precise risk stratification of both hepatic and extrahepatic complications, as well as for informing personalized management strategies.

Accordingly, we developed a novel multi-task deep LASSO algorithm for feature selection and subsequently performed clustering analysis in patients with biopsy-confirmed MASLD to generate a comprehensive classification system. Genetic risk variants for hepatic fibrosis were also evaluated within this classification. The prognostic utility of the model was validated in two independent cohorts: A Chinese cohort undergoing routine annual health check-ups and a U.S. cohort from the Third National Health and Nutrition Examination Survey (NHANES III).

Methods

This study complied with the Declaration of Helsinki and all other relevant ethical guidelines. The Ethics Committee of Nanjing Drum Tower Hospital approved the study’s protocol (No. 2022-444-02), which we registered in ClinicalTrials.gov (No. NCT05560997). Written informed consent was obtained from all individual participants included in the study.

Study population

In the discovery cohort, cross-sectional data were obtained from individuals who underwent liver biopsy at the time of metabolic surgery in Nanjing Drum Tower Hospital between January 2016 and June 2023. Two health examination cohorts were employed in the validation phase. The first validation cohort (health check-up cohort) included baseline and longitudinal data from individuals who undertook annual health check-ups at the medical examination center of Nanjing Drum Tower Hospital between January 2018 and December 2022. The second validation cohort consisted of participants from the NHANES III study enrolled between 1988 and 1994 [Supplementary Figure 1, http://links.lww.com/CM9/C775].

Measurements

Clinical biochemistry

Data of demographic, health history, and laboratory measurements were collected retrospectively from electronic medical records. NHANES III comprises detailed household interviews, prescription medication, examination, and laboratory information obtained from the NHANES website. The poverty income ratio (PIR), which is the ratio of family income to the poverty threshold, was extracted. The homeostatic model assessment (HOMA2) indices, fibrosis 4 (FIB-4) index, and triglyceride-glucose index (TyG) were calculated. Systemic inflammation was assessed by the neutrophil-to-lymphocyte ratio (NLR) and the systemic immune-inflammation index (SII).

Body composition

Body components were assessed using dual energy X-ray absorptiometry (DEXA/DXA) scans with an iDXA total body fan-beam densitometer in the discovery cohort. Appendicular skeletal muscle mass index (ASMI) was defined as appendicular skeletal muscle mass (ASM)/height². Fat distribution was assessed using the android fat mass to gynoid fat mass ratio (AGR).

Definitions

MASLD was defined as hepatic steatosis by imaging or liver histology in addition to any of the cardio-metabolic risk factors.^[6] In the discovery cohort, hepatic steatosis was defined as the presence of at least 5% steatosis. MASH was defined as MASLD with nonalcoholic fatty liver disease activity score (NAS) ≥5. Significant liver fibrosis was defined as a fibrosis score ≥2.^[7] History of CVD was defined by self-report of a prior diagnosis of coronary atherosclerosis, myocardial infarction, heart failure, stroke, peripheral vascular disease, valvular disease, and atrial fibrillation. Among those without a history of CVD, the 30-year predicted CVD risk was calculated based on the Framingham Heart Study. High CVD risk was defined as having a history of CVDs or a 30-year risk of full CVD ≥40% based on the Framingham Heart Study.^[8] CKD was identified by self-reported physician diagnosis or estimated glomerular filtration rate (eGFR) <60 mL∙min^–1∙1.73m^–2 or urinary albumin-creatinine ratio (UACR) ≥30 mg/g (repeated on 2 occasions, at least 3 months apart).^[9]

The diagnostic criteria for T2DM were based on the updated criteria of the American Diabetes Association and included an fasting blood glucose (FBG) ≥ 7.0 mmol/L, a 2-h plasma glucose level ≥11.1 mmol/L after an oral glucose tolerance test, an hemoglobin A1c (HbA1c) level >6.5%, a prior history of T2DM, or the use of diabetes medications.^[10] Hypertension was defined as self-reported hypertension or use of antihypertensive therapy or average blood pressure ≥140/90 mmHg. Dyslipidemia was identified according to self-reported history of dyslipidemia, or taking lipid-lowering medications, or total cholesterol ≥6.2 mmol/L or triglycerides ≥2.3 mmol/L or low-density lipoprotein cholesterol ≥4.1 mmol/L or high-density lipoprotein cholesterol <1.0 mmol/L.^[11]

In the health check-up cohort, hepatic steatosis was identified by sonography or computed tomography (CT). Advanced fibrosis/cirrhosis was diagnosed via ultrasonography, radiology, or FIB-4 ≥2.67.^[6] The presence of subclinical atherosclerosis assessed by carotid ultrasonography or CT scans was defined as high CVD risk.^[12,13]

In NHANES III, the degree of hepatic steatosis was categorized as none, mild, moderate, or severe. Participants in our study were considered to have steatosis liver if mild to severe hepatic steatosis was noted.

Mortality and follow-up

We used survival data from the NHANES public-use linked mortality file up to December 31, 2019. The analysis included all-cause, cardiovascular, and cancer-related mortality. Each participant’s follow-up period spanned from the baseline examination date to their date of death or last follow-up date, whichever came first. In the health check-up cohort, follow-up time was calculated as the time from baseline until last examination.

Data processing, feature selection, and clustering analysis

Data processing

Variables with >20% missing data were excluded, leaving 52 variables for imputation using simple imputation procedures. The imputation performance was primarily evaluated using mean squared error (MSE), root mean square error (RMSE), and similarity calculations. The results showed favorable imputation performance, with an MSE of 1.758% and an RMSE of 13.26%. Additionally, Supplementary Figure 2, http://links.lww.com/CM9/C775 shows that the similarity between the original (non-imputed) dataset and the imputed dataset was 99%, suggesting that the imputed data closely matched the original data.

Feature selection

We propose a novel multi-task deep LASSO method to select features for clustering. Our method promotes sparsity in feature gradients by applying a group LASSO penalty to the gradients of the loss with respect to input features.^[14]

For train data (X, Y) with m features, the penalty is given by

$$\underset{\theta }{\min }\alpha {\displaystyle \sum }_{k=1}^K{ℒ}_{\theta }\left(X,Y^k\right)+\left(1-\alpha \right){\displaystyle \sum }_{j=1}^m{\displaystyle \sum }_{k=1}^K{\Vert \frac{\partial {ℒ}_{\theta }\left(X,Y^k\right)}{\partial X^j}\Vert }_2$$

where K is the total number of tasks and Y^t denotes the targets on the k-th task. The corresponding feature importance is provided by

$$F{I}^j={\displaystyle \sum }_{k=1}^K{\Vert \frac{\partial {ℒ}_{\theta }\left(X,Y^k\right)}{\partial X^j}\Vert }_2$$

During training, a multilayer perceptron architecture is employed as the backbone and learns a shared representation across multiple tasks. By regularizing the gradients $\frac{\partial {ℒ}_{\theta }\left(X,Y^k\right)}{\partial X^j}$ of all tasks, multi-task deep LASSO could identify and prune input features X^j that are irrelevant for predicting the multiple targets accurately. Focusing on significant fibrosis and CVD risk as our two tasks, we identify the top six variables according to their feature importance FI. We also employed five-fold cross-validation (CV) for cross-validation, in which we observed stability of top-selected features among different folds.

Clustering analysis

We assessed the variable distributions for normality and performed the cluster analysis on the normalized values using min–max normalization. K-means was performed to identify clusters of participants. The optimal number of clusters was determined using both the gap statistic method and the elbow method. To assess the robustness of the identified clusters, we conducted a comparison between K-means and hierarchical clustering. Results were verified using hierarchical clustering resulting in similar cluster distributions to K-means clustering. The cosine similarity between the feature vectors of the clusters from both methods was calculated, yielding a value of 0.90, which suggested that the clusters identified by both methods exhibit highly similar characteristics. We employed t-distributed stochastic neighbor embedding (t-SNE) for dimensional reduction and visualization. To validate clusters from the discovery cohort, k-means was applied to the health check-up cohort and NHANES III separately. Cluster stability was assessed by resampling the dataset (5000 times) and computing the Jaccard coefficient against the original clusters.

Details on inclusion and exclusion criteria, measurements and data processing are available in the Supplementary Materials, http://links.lww.com/CM9/C775.

Genotyping

Five genetic variants previously associated with hepatic cirrhosis (PNPLA3-rs738409, TM6SF2-rs58542926, HSD17B13-rs72613567, MBOAT7-rs641738, and GCKR-rs1260326) were genotyped in discovery cohort.^[15]

Statistical analysis

Descriptive statistics were summarized as median (25th and 75th percentile) for continuous variables or counts (%) for categorical variables. We generated logistic regression to estimate odds ratios (ORs), adjusted odds ratios (aORs), and 95% confidence intervals (CIs). Survival analyses were performed using two complementary approaches: (1) Cox proportional hazards models reporting crude and adjusted hazard ratios (HR/aHR). (2) Fine–Gray subdistribution hazards models for cause-specific mortality to account for competing risks, reporting subdistribution hazard ratios (sHR) with 95% CI. All regression models incorporated Tukey–Kramer multiplicity adjustments for pairwise cluster comparisons. Survival curves were generated through: Kaplan–Meier estimators for subclinical atherosclerosis, advanced fibrosis/cirrhosis, and all-cause mortality, with between-group differences assessed via stratified log-rank tests. Cumulative incidence functions for cause-specific mortality were constructed, and between-group differences were compared using Gray’s k-sample tests. We considered two-sided P values <0.05 as statistically significant. We performed all data analyses using Python version 3.6.5 (Python Software Foundation, Delaware, USA), and R version 4.3.1 (R Foundation for Statistical Computing, Vienna, Austria).

Results

Cluster analysis in the discovery cohort

A total of 1111 subjects were enrolled in the discovery cohort, including 973 with MASLD (mean age, 31.0 years; mean BMI, 38.9 kg/m²) and 138 without steatotic liver disease (SLD) (mean age, 30.4 years; mean BMI, 35.8 kg/m²; [Supplementary Table 1, http://links.lww.com/CM9/C775]).

Of the resulting set of features, the top six variables (i.e., age, BMI, HbA1c, TyG, TC/HDL, and γ-glutamyl transferase/platelet (GGT/PLT)) were selected using our proposed multi-task deep LASSO model [Supplementary Figure 3, http://links.lww.com/CM9/C775]. Then, we conducted a k-means clustering with selected variables and found k = 4 was the number to obtain the best performance [Figure 1A and B; Supplementary Figure 4, http://links.lww.com/CM9/C775]. To assess stability, we calculated the Jaccard coefficients for the four clusters: 0.861, 0.796, 0.872, and 0.854, indicating the stability of our analysis.

Figure 1.

Visualization and clinical characteristics of the four MASLD clusters identified in the discovery cohort. (A) Distributions of six selected variables. (B) Two-dimensional t-SNE representation. (C) Radar plots. Cluster 1 was labeled as low CVD risk subgroup. Cluster 2 was labeled as high fibrosis risk subgroup. Cluster 3 was labeled as high CKM risk subgroup. Cluster 4 was labeled as high CLKM risk subgroup. ALT: Alanine aminotransferase; AGR: Android fat mass to gynoid fat mass ratio; AST: Aspartate aminotransferase; ASMI: Appendicular skeletal muscle mass index; BMI: Body-mass index; CKM: Cardiovascular–kidney–metabolic, CLKM: Cardiovascular–liver–kidney–metabolic syndrome; CVD: Cardiovascular disease; CKD: Chronic kidney disease; FBG: fasting blood glucose; FM: Fat mass; GGT: γ-glutamyl transferase; HbA1c: HemoglobinA1c; HDL: High-density lipoprotein; HOMA-2IR: Homeostasis model assessment of insulin resistance; HOMA-2β: Homeostasis model assessment of β-cell function; LDL: Low density lipoprotein cholesterol; MASLD: Metabolic-associated steatotic liver disease; NLR: Neutrophil-to-lymphocyte ratio; PBG: Postprandial blood glucose; PTL: Platelet; SII: Systemic immune-inflammation index; TC: Total cholesterol; TG: Triglyceride; TyG: Triglyceride-glucose index.

These four identified MASLD clusters showed distinctive clinical and body composition characteristics [Table 1 and Figure 1C]. Cluster 1 (n = 398, 40.9%) had younger patients with lower markers of atherogenic dyslipidemia (lower TG and LDL, higher HDL) and blood glucose. They had the highest percentages of body fat and thigh fat, but the lowest levels of visceral fat (low AGR). Cluster 2 (n = 256, 26.3%) showed significant lipid profile disorders and substantial liver damage (evidenced by elevated liver enzymes), as well as mild hyperglycemia. Cluster 3 (n = 186, 19.1%) was characterized by the lowest muscle mass, obvious chronic systemic inflammation (elevated NLR and SII), moderately elevated markers of glucose and lipid metabolism and older age. Cluster 4 (n = 133, 13.7%) was distinguished by severe insulin resistance (elevated HOMA2-IR and TyG), poor glucose control, unfavorable lipid profiles, substantial liver damage, and high visceral adiposity (increased AGR). The cluster center coordinates are provided in Supplementary Table 2, http://links.lww.com/CM9/C775.

Table 1.

Characteristics of different MASLD clusters in the discovery cohort.

Characteristics	Cluster 1 (n = 398)	Cluster 2 (n = 256)	Cluster 3 (n = 186)	Cluster 4 (n = 133)
Male	115 (28.9)	124 (48.4)	44 (23.7)	65 (48.9)
Age (year)	28.00 (24.00, 32.00)	29.00 (26.00, 33.25)	43.00 (40.00, 49.00)	33.00 (29.00, 37.00)
BMI (kg/m2)	38.82 (34.72, 43.27)	39.18 (35.14, 43.75)	35.16 (32.05, 38.10)	39.01 (35.16, 43.24)
SBP (mmHg)	135.00 (125.00, 147.00)	141.00 (129.00, 155.25)	137.50 (124.25, 150.75)	145.00 (131.00, 155.00)
DBP (mmHg)	87.00 (78.25, 97.00)	92.00 (83.00, 100.00)	86.00 (79.00, 97.75)	92.00 (83.00, 102.00)
Smoker	45 (11.3)	62 (24.2)	29 (15.6)	24 (18.0)
Drink	9 (2.3)	21 (8.2)	9 (4.8)	6 (4.5)
T2DM	74 (18.6)	85 (33.2)	73 (39.2)	130 (97.7)
Hypertension	148 (37.2)	152 (59.4)	108 (58.1)	88 (66.2)
Dyslipidemia	161 (40.5)	241 (94.1)	79 (42.5)	106 (79.7)
Medication history
Anti-hypertension drugs	38 (9.5)	44 (17.2)	62 (33.3)	36 (27.1)
Lipid-lowering drugs	0	1 (0.4)	5 (2.7)	6 (4.5)
Anti-diabetic drugs	17 (4.3)	20 (7.8)	22 (11.8)	37 (27.8)
Metabolic variables
FBG (mmol/L)	5.13 (4.64, 5.82)	5.49 (4.92, 6.35)	5.67 (5.06, 6.70)	9.93 (7.93, 11.53)
INS0 (uIU/mL)	26.40 (18.65, 34.75)	32.89 (24.30, 42.00)	19.30 (14.05, 27.37)	24.20 (17.30, 35.12)
CP0 (pmol/L)	1324.00 (1073.75, 1587.25)	1604.00 (1334.00, 1878.00)	1145.00 (962.00, 1428.00)	1334.00 (1079.00, 1788.50)
PBG (mmol/L)	7.50 (6.40, 9.00)	8.30 (6.84, 10.30)	9.02 (7.50, 11.78)	16.10 (13.50, 18.70)
INS120 (uIU/mL)	142.40 (81.42, 221.00)	164.00 (106.00, 240.40)	113.70 (68.10, 163.00)	55.55 (29.94, 89.60)
CP120 (pmol/L)	4099.50 (3281.50, 5103.25)	4529.50 (3727.00, 5348.00)	3883.00 (3143.00, 4811.00)	2497.50 (1726.00, 3456.00)
HbA1c (%)	5.50 (5.30, 5.90)	5.90 (5.50, 6.30)	5.80 (5.50, 6.40)	8.70 (8.10, 9.80)
HOMA-2β	188.90 (151.60, 228.55)	190.30 (147.90, 228.25)	147.60 (104.70, 187.30)	66.75 (45.50, 104.20)
HOMA-2IR	2.96 (2.36, 3.61)	3.58 (2.96, 4.26)	2.71 (2.18, 3.32)	3.77 (2.94, 4.84)
TyG	8.56 (8.32, 8.84)	9.16 (8.89, 9.46)	8.85 (8.55, 9.21)	9.79 (9.32, 10.16)
TG (mmol/L)	1.27 (1.01, 1.59)	2.08 (1.65, 2.80)	1.56 (1.16, 2.06)	2.27 (1.75, 3.25)
TC (mmol/L)	4.47 (3.88, 4.97)	5.19 (4.70, 5.67)	4.80 (4.22, 5.33)	4.99 (4.36, 5.55)
LDL (mmol/L)	2.71 (2.24, 3.14)	3.26 (2.81, 3.73)	2.85 (2.44, 3.31)	2.89 (2.37, 3.45)
HDL (mmol/L)	1.08 (0.93, 1.26)	0.86 (0.76, 0.95)	1.10 (0.97, 1.26)	0.93 (0.83, 1.05)
ApoAI (g/L)	1.06 (0.95, 1.20)	0.97 (0.88, 1.07)	1.11 (1.01, 1.28)	1.03 (0.91, 1.14)
ApoB (g/L)	0.81 (0.69, 0.92)	1.00 (0.89, 1.13)	0.86 (0.75, 1.00)	0.95 (0.84, 1.11)
TC/HDL	4.21 (3.67, 4.67)	5.96 (5.43, 6.66)	4.21 (3.69, 4.92)	5.32 (4.70, 6.21)
ALT (U/L)	39.85 (24.95, 69.80)	58.40 (34.67, 92.67)	30.60 (20.10, 54.70)	58.10 (36.60, 106.30)
AST (U/L)	25.60 (18.20, 38.40)	32.50 (22.70, 52.62)	23.05 (16.60, 32.35)	41.70 (27.60, 60.80)
ALP (U/L)	65.35 (55.87, 77.50)	69.90 (59.27, 82.93)	64.95 (55.30, 76.57)	75.10 (64.20, 92.80)
GGT (U/L)	36.45 (26.10, 52.60)	51.45 (33.65, 76.65)	33.85 (22.63, 53.80)	63.20 (43.50, 99.30)
GGT/PLT	0.13 (0.09, 0.21)	0.20 (0.12, 0.31)	0.14 (0.09, 0.23)	0.28 (0.17, 0.42)
ALB (g/L)	42.00 (39.92, 43.60)	42.40 (40.60, 44.10)	42.05 (40.40, 43.40)	42.00 (40.00, 43.40)
GLO (g/L)	28.25 (25.80, 30.67)	28.60 (26.60, 31.63)	27.95 (25.40, 30.30)	28.60 (25.70, 30.90)
UA (umol/L)	422.50 (359.25, 482.75)	462.00 (378.00, 551.25)	381.00 (325.75, 433.75)	420.00 (356.00, 485.00)
Cr (umol/L)	51.00 (45.00, 59.00)	55.00 (47.00, 64.00)	51.00 (45.00, 60.00)	50.00 (41.00, 59.00)
eGFR (mL·min–1·1.73 m–2)	142.50 (126.12, 162.23)	138.15 (121.00, 159.45)	125.80 (110.00, 143.90)	158.70 (131.80, 182.60)
UACR (mg/g)	8.20 (5.30, 14.70)	12.30 (6.55, 28.00)	12.60 (6.47, 26.02)	24.70 (10.85, 63.05)
SII	566.74 (410.09, 746.92)	579.16 (438.67, 727.75)	584.41 (429.17, 814.88)	549.95 (406.52, 739.20)
NLR	2.00 (1.61, 2.67)	2.06 (1.69, 2.64)	2.24 (1.73, 2.89)	2.17 (1.76, 2.85)
Body composition*
Body fat mass (g)	48045.00 (41530.00, 59390.00)	49426.00 (40828.00, 59019.00)	39273.00 (34688.00, 48262.00)	46656.50 (38799.25, 56823.50)
Body fat mass (% of weight)	45.96 (42.49, 48.78)	43.25 (39.97, 46.61)	42.82 (38.80, 45.80)	43.02 (38.02, 46.90)
Android fat mass (g)	4944.00 (3957.00, 6432.00)	5240.00 (4458.00, 6996.00)	4299.00 (3521.00, 5356.00)	5632.00 (4198.50, 6741.25)
Gynoid fat mass (g)	7169.00 (5944.50, 8711.50)	6614.00 (5458.00, 8150.00)	5508.00 (4768.00, 6295.00)	5646.00 (4768.75, 7248.50)
AGR	0.71 (0.59, 0.84)	0.84 (0.70, 0.98)	0.77 (0.66, 0.95)	0.97 (0.84, 1.08)
Thigh fat mass (g)	14081.00 (11515.50, 17002.50)	12906.00 (10071.00, 15599.00)	10275.00 (8550.00, 12564.00)	10432.50 (8506.25, 13274.75)
Thigh fat mass (% of weight)	13.11 (10.90, 15.02)	11.17 (9.25, 12.88)	11.12 (8.93, 12.63)	9.62 (8.32, 11.41)
Muscle mass (g)	53193.00 (48311.00, 62396.50)	61074.00 (51901.00, 71007.00)	49839.00 (45280.00, 56166.00)	58554.50 (51018.75, 67054.25)
ASM (g)	24688.00 (21214.00, 28872.50)	26942.00 (23111.00, 32560.00)	21729.00 (19279.00, 25653.00)	25572.00 (22005.25, 30589.25)
ASMI (kg/m2)	8.93 (8.02, 10.00)	9.48 (8.59, 10.62)	8.44 (7.54, 9.29)	9.16 (8.44, 10.07)

Data are presented as median (25th and 75th percentile) or n (%). Cluster 1 was labeled as low CVD risk subgroup. Cluster 2 was labeled as high fibrosis risk subgroup. Cluster 3 was labeled as high CKM risk subgroup. Cluster 4 was labeled as high CLKM risk subgroup. *The total sample size for body composition is 724 due to missing data; the sample sizes for clusters 1 to 4 are 307, 181, 134, and 102, respectively. ALT: Alanine aminotransferase; AST: Aaspartate aminotransferase; ALP: Alkaline phosphatase; ApoAI: Apolipoprotein AI; ApoB: Apolipoprotein B; ALB: Albumin; ASMI: Appendicular skeletal muscle mass index; AGR: Android fat mass to gynoid fat mass ratio; BMI: Body mass index; CKM: Cardiovascular–kidney–metabolic, CLKM: Cardiovascular–liver–kidney–metabolic syndrome; CVD: Cardiovascular disease; CKD: Chronic kidney disease; CP: C-Peptide; Cr: Creatinine; DBP: Diastolic blood pressure; eGFR: Estimated glomerular filtration rate; FBG: Fasting blood glucose; GGT: γ-Glutamyl transferase; GLO: Globulin; HbA1c: HemoglobinA1c; HDL: High-density lipoprotein; HOMA-2IR: Homeostasis model assessment of insulin resistance; HOMA-2β: Homeostasis model assessment of β-cell function; INS0: Fasting insulin; LDL: Low density lipoprotein cholesterol; MASLD: Metabolic-associated steatotic liver disease; NLR: Neutrophil-to-lymphocyte ratio; PBG: Postprandial blood glucose; PTL: Platelet; SBP: Systolic blood pressure; SII: Systemic immune-inflammation index; T2DM: Type 2 diabetes mellitus; TC: Total cholesterol; TG: Triglyceride; TyG: Triglyceride-glucose index; UA: Uric acid; UACR: Urinary albumin-creatinine ratio.

Association between clusters and complications in the discovery cohort

We evaluated whether the four clusters were characterized by different hepatic and extrahepatic complications. Individuals with non-SLD were set as the reference. Of the biopsy-proved MASLD, 29.1% (283/973) had MASH and 14.7% (143/973) had significant fibrosis. Approximately 26.5% (258/973) of MASLD had high CVD risk, and 20.7% (201/973) had CKD.

Patients in Cluster 1 exhibited a metabolically favorable profile, demonstrating significantly lower odds of CVD (adjusted odds ratio [aOR]: 0.385; 95% confidence interval [CI]: 0.155–0.955) with no significant associations observed for CKD or hepatic fibrosis. Cluster 2 included patients with advanced hepatic diseases, of whom 32.0% (82/256) met MASH criteria and 16.8% (43/256) exhibited significant fibrosis. This cluster showed elevated odds of significant fibrosis (aOR: 2.377; 95% CI: 1.037–5.447), but maintained comparable extrahepatic comorbidity rates to reference groups. Cluster 3 presented a distinct phenotype of systemic metabolic dysregulation, manifesting the highest burden of cardiovascular comorbidities (56.5% [105/186] prevalence; aOR: 14.651, 95% CI: 6.301–34.068) and renal manifestations (22.6% [42/186] CKD prevalence; aOR: 1.981, 95% CI: 1.008–3.890). Notably, this cluster showed no significant hepatic fibrosis associations. Cluster 4 represented the most severe hepatic phenotype, demonstrating both the highest prevalence of MASH (43.6% [58/133]) and strongest association with significant fibrosis (aOR: 3.987, 95% CI: 1.693–9.389; 28.6% [38/133] prevalence). This cluster concurrently exhibited substantial cardiovascular and renal burden, with 49.6% [66/133] CVD (aOR: 7.255) and 45.1% [60/133] CKD (aOR: 4.391) [Table 2 and Supplementary Table 3, http://links.lww.com/CM9/C775]. Inter-cluster differentials in comorbidity patterns were statistically confirmed through post hoc analyses [Supplementary Table 4, http://links.lww.com/CM9/C775].

Table 2.

Associations between different MASLD clusters and different complications for cross-sectional discovery cohort—Non-SLD patients were set as reference.

Complications	Crude model		Multi-variable adjusted model*
	OR (95% CI)	P value	OR (95% CI)	P value
Significant fibrosis (F ≥2)
Cluster 1	1.816 (0.828, 3.981)	0.137	1.590 (0.708, 3.567)	0.261
Cluster 2	3.281 (1.495, 7.196)	0.003	2.377 (1.037, 5.447)	0.041
Cluster 3	2.180 (0.940, 5.056)	0.069	1.978 (0.834, 4.695)	0.122
Cluster 4	6.500 (2.900, 14.568)	0.001	3.987 (1.693, 9.389)	0.002
High CVD risk
Cluster 1	0.597 (0.307, 1.158)	0.127	0.385 (0.155, 0.955)	0.039
Cluster 2	2.510 (1.365, 4.616)	0.003	1.121 (0.473, 2.658)	0.796
Cluster 3	10.630 (5.779, 19.551)	<0.001	14.651 (6.301, 34.068)	<0.001
Cluster 4	8.078 (4.282, 15.238)	<0.001	7.255 (2.945, 17.869)	<0.001
CKD
Cluster 1	0.891 (0.475, 1.672)	0.719	0.796 (0.408, 1.554)	0.504
Cluster 2	2.536 (1.379, 4.664)	0.003	1.790 (0.919, 3.489)	0.087
Cluster 3	2.392 (1.265, 4.521)	0.007	1.981 (1.008, 3.890)	0.047
Cluster 4	6.740 (3.570, 12.726)	<0.001	4.391 (2.187, 8.815)	<0.001

^*Adjusted for gender, smoke, drink, SBP, DBP, ALT, AST, UA, use of anti-hypertension drugs, use of lipid-lowering drugs, and use of anti-diabetic drugs. Cluster 1 was labeled as low CVD risk subgroup. Cluster 2 was labeled as high fibrosis risk subgroup. Cluster 3 was labeled as high CKM risk subgroup. Cluster 4 was labeled as high CLKM risk subgroup. ALT: Alanine aminotransferase; aOR: Adjusted odds ratio; AST: Aspartate aminotransferase; CKM: Cardiovascular–kidney–metabolic; CLKM: Cardiovascular–liver–kidney–metabolic syndrome; CVD: Cardiovascular disease; CKD: Chronic kidney disease; CI: Confidence interval; DBP: Diastolic blood pressure; MASLD: Metabolic-associated steatotic liver disease; OR: Odds ratio; SBP: Systolic blood pressure; SLD: Steatotic liver disease; UA: Uric acid.

To further explore the influence of genetic variants on fibrosis, we conducted an analysis examining the association between SNP genotypes and phenotypes in a subset of individuals (n = 442, Supplementary Table 5, http://links.lww.com/CM9/C775). Cluster 4 exhibited the highest frequencies of risk alleles in PNPLA3 (CG and GG: 73.9%), TM6SF2 (CT and TT: 23.2%), and MBOAT7 (CT and TT: 53.6%), followed by Cluster 2. PNPLA3 rs738409 C >G variant carriers showed a 3.2-fold increase in significant fibrosis among those with the PNPLA3 CG genotype (aOR: 4.214; 95% CI: 1.798–9.878) and a 2.7-fold increase among those with the PNPLA3 GG genotype (aOR: 3.737; 95% CI: 1.411–9.899) [Table 3].

Table 3.

The association between genetic variants and significant fibrosis.

SNPs	Crude model		Multi-variable adjusted model*
	OR (95% CI)	P value	aOR (95% CI)	P value
PNPLA3 rs738409
CC	Reference		Reference
CG	4.655 (2.000, 10.834)	<0.001	4.214 (1.798, 9.878)	0.001
GG	4.133 (1.582, 10.800)	0.004	3.737 (1.411, 9.899)	0.008
TM6SF2 rs58542926
CC	Reference		Reference
CT and TT	4.824 (2.422, 9.609)	<0.001	4.718 (2.337, 9.525)	<0.001
GCKR rs1260326
TT	Reference		Reference
TC	1.146 (0.580, 2.265)	0.695	1.046 (0.523, 2.094)	0.899
CC	0.766 (0.325, 1.805)	0.542	0.764 (0.320, 1.823)	0.544
MBOAT7 rs641738
CC	Reference		Reference
CT and TT	1.907 (1.068, 3.405)	0.029	1.841 (1.022, 3.314)	0.042
HSD17B13 rs72613567
TT	Reference		Reference
TTA	0.581 (0.312, 1.081)	0.087	0.573 (0.306, 1.074)	0.083
TATA	0.445 (0.130, 1.526)	0.198	0.469 (0.135, 1.629)	0.233

^*Adjusted for gender, age, BMI, and drink status. aOR: Adjusted odds ratio; BMI: Body mass index; CC: Homozygous C allele; CG: Heterozygous C/G allele; CI: Confidence interval; CT: Heterozygous C/T allele; GG: Homozygous G allele; TT: Homozygous T allele; TC: Heterozygous T/C allele; TTA: Heterozygous allele with T/TA insertion; TATA: Homozygous allele with TA duplication; OR: Odds ratio.

External validation in the health check-up cohort and NHANES III

To test whether our classification for MASLD could be applied to a broader population, we replicated the cluster analysis in a health check-up cohort, including 6172 adults (MASLD: 43.9% [2710/6172], mean follow-up was 27.6 months) and NHANES III, including 7406 participants (MASLD: 37.3% [2765/7406], mean follow-up was 280.2 months). Overall, we found consistent clusters in both cohorts (k = 4, Supplementary Tables 6 and 7, http://links.lww.com/CM9/C775).

In the health check-up cohort, subclinical atherosclerosis was most frequent in Cluster 3 (65.3% [363/556]), and then in Cluster 4 (44.4% [231/520]). We observed that advanced fibrosis/cirrhosis was more frequent in Cluster 4 (3.1% [16/520]) and Cluster 2 (1.3% [12/932]) [Supplementary Table 8, http://links.lww.com/CM9/C775]. Then, Cox proportional hazards regressions were applied to identify differences in subclinical atherosclerosis and advanced fibrosis/cirrhosis among clusters. Patients in Cluster 4 had significantly increased advanced fibrosis/cirrhosis risk (aHR: 5.004; 95% CI: 2.220–11.278) and an elevated risk of subclinical atherosclerosis (aHR: 1.534; 95% CI: 1.296–1.815). Cluster 3 patients exhibited the highest risk of subclinical atherosclerosis (aHR: 2.853; 95% CI: 2.489–3.272), while Cluster 2 patients showed a modestly elevated risk of advanced fibrosis/cirrhosis (aHR: 2.446; 95% CI: 1.094–5.469). In contrast, patients in Clusters 1 had better clinical outcomes, with an approximately 54% lower risk of subclinical atherosclerosis (aHR: 0.463; 95% CI: 0.366–0.586) and non-significant increased risks of advanced fibrosis/cirrhosis [Table 4]. These results were in line with those we obtained in the discovery cohort. The Kaplan–Meier analysis revealed that the cumulative incidence of subclinical atherosclerosis in Cluster 3 was the highest of all, with a median time to onset of 23.8 months (95% CI: 22.6–25.0 months), followed by that in Cluster 4, with a median time to onset of 35.0 months (95% CI: 33.7–36.4 months). The cumulative incidence of advanced fibrosis/cirrhosis was significantly higher in Cluster 2 and Cluster 4 (log-rank: P <0.05) than in any other cluster [Figure 2A, B].

Table 4.

Incidence and risk of different complications/mortality in different MASLD clusters for external validations—non-SLD patients were set as reference.

Outcomes	Crude model		Multi-variable adjusted model
	HR (95% CI)	P value	aHR (95% CI)	P value
Health check-up cohort
Subclinical atherosclerosis
Cluster 1	0.523 (0.421, 0.650)	<0.001	0.463 (0.366, 0.586)*	<0.001
Cluster 2	1.156 (1.003, 1.332)	0.046	0.991 (0.853, 1.152 *	0.910
Cluster 3	4.035 (3.565, 4.567)	<0.001	2.853 (2.489, 3.272)*	<0.001
Cluster 4	1.907 (1.648, 2.206)	<0.001	1.534 (1.296, 1.815)*	<0.001
Advanced fibrosis/cirrhosis
Cluster 1	1.354 (0.453, 4.052)	0.588	0.481 (0.102, 2.268)*	0.355
Cluster 2	2.884 (1.364, 6.097)	0.006	2.446 (1.094, 5.469)*	0.029
Cluster 3	1.989 (0.729, 5.431)	0.179	1.629 (0.570, 4.658)*	0.362
Cluster 4	6.591 (3.296, 13.181)	<0.001	5.004 (2.220, 11.278)*	<0.001
NHANES III
All-cause mortality
Cluster 1	0.209 (0.161, 0.273)	<0.001	0.320 (0.241, 0.424)†	<0.001
Cluster 2	0.705 (0.609, 0.815)	<0.001	0.704 (0.600, 0.826)†	<0.001
Cluster 3	3.529 (3.225, 3.862)	<0.001	2.182 (1.969, 2.418)†	<0.001
Cluster 4	3.578 (3.064, 4.178)	<0.001	2.240 (1.843, 2.722)†	<0.001
Cardiovascular mortality
Cluster 1	0.133 (0.066–0.268)‡	<0.001	0.228 (0.111–0.472)†,‡	<0.001
Cluster 2	0.729 (0.542–0.982)‡	0.037	0.809 (0.596–1.098)†,‡	0.173
Cluster 3	2.496 (2.077–2.999)‡	<0.001	1.393 (1.129–1.719)†,‡	0.002
Cluster 4	3.851 (2.937–5.050)‡	<0.001	1.854 (1.273–2.699)†,‡	0.001
Cancer-related mortality
Cluster 1	0.319 (0.196–0.519)‡	<0.001	0.409 (0.251–0.667)†,‡	<0.001
Cluster 2	1.037 (0.792–1.357)‡	0.794	1.101 (0.828–1.464)†,‡	0.507
Cluster 3	2.435 (2.000–2.964)‡	<0.001	2.113 (1.696–2.633)†,‡	<0.001
Cluster 4	1.533 (1.016–2.313)‡	0.042	1.575 (1.006–2.465)†,‡	0.047

^*Adjusted for gender, SBP, DBP, ALT, AST, and UA. ^†Adjusted for gender, ethnicity, smoker, drinker, SBP, DBP, ALT, AST, UA, PIR, use of anti-hypertension drugs, use of lipid-lowering drugs, and use of anti-diabetic drugs. ^‡Subdistribution hazard ratios based on a Fine–Gray competing risk model. Cluster 1 was labeled as low CVD risk subgroup. Cluster 2 was labeled as high fibrosis risk subgroup. Cluster 3 was labeled as high CKM risk subgroup. Cluster 4 was labeled as high CLKM risk subgroup. ALT: Alanine aminotransferase; AST: Aspartate aminotransferase; aHR: Adjusted hazard ratio; CKM: Cardiovascular–kidney–metabolic, CLKM: Cardiovascular–liver–kidney–metabolic syndrome; CVD: Cardiovascular disease; CKD: Chronic kidney disease; CI: Confidence interval; DBP: Diastolic blood pressure; HR: Hazard ratio; MASLD: Metabolic-associated steatotic liver disease; NHANES III: the Third National Health and Nutrition Examination Survey; PIR: Poverty income ratio; SBP: Systolic blood pressure; SLD: Steatotic liver disease; sHR: Subdistribution hazard ratio; UA: Uric acid.

Figure 2.

Progression of disease over time by MASLD cluster in the health check-up cohort and NHANES III. (A) Cumulative incidence of subclinical atherosclerosis. (B) Cumulative incidence of advanced fibrosis/cirrhosis. (C) Cumulative incidence of all-cause deaths. (D) Cumulative incidence function estimates for cardiovascular deaths. (E) Cumulative incidence function estimates for cancer-related deaths. Cluster 1 was labeled as low CVD risk subgroup. Cluster 2 was labeled as high fibrosis risk subgroup. Cluster 3 was labeled as high CKM risk subgroup. Cluster 4 was labeled as high CLKM risk subgroup. CKM: Cardiovascular–kidney–metabolic, CLKM: Cardiovascular–liver–kidney–metabolic syndrome; CVD: Cardiovascular disease; CKD: Chronic kidney disease; MASLD: Metabolic-associated steatotic liver disease; NHANES III: the Third National Health and Nutrition Examination Survey.

In NHANES III, Clusters 3 and 4 were associated with higher incidences of all-cause, cardiovascular, and cancer-related deaths. Specifically, Cluster 4 showed increased risks of all-cause death (aHR: 2.24; 95% CI: 1.854–2.722), cardiovascular death (sHR: 1.854; 95% CI: 1.273–2.699), and cancer-related death (sHR: 1.575; 95% CI, 1.006–2.465), as did Cluster 3 (aHR: all-cause death 2.182, sHRs: cardiovascular death 1.393, cancer-related death 2.113) [Supplementary Table 9, http://links.lww.com/CM9/C775; Table 4]. Additionally, survival analyses using Kaplan–Meier curves and cumulative incidence functions further supported our findings [Figure 2C, D, E]. Post hocP-values from the log-rank test or Gray’s k-sample test were used to assess survival differences across clusters in both the health check-up cohort and NHANES III. The results remained consistent with our previous findings [Supplementary Table 10, http://links.lww.com/CM9/C775].

Discussion

In our present study, based on a novel developed algorithm-selected variables, we identified four MASLD clusters showing distinct hepatic (e.g. hepatic fibrosis) and extrahepatic (e.g. metabolic disorders, CVD, or CKD) complications. Clinical characteristics, body composition, genetic information, together with all-cause, and cause-specific mortalities were further integrated and described in detail for those four clusters from three independent cohorts. Our findings provide insights into the heterogeneity of MASLD and may be helpful for the individualized management of MASLD.

CVDs are the leading cause of mortality among individuals with MASLD.^[3] Meanwhile, MASLD is significantly associated with an increased risk of CKD.^[16] Both CKD and CVD share several common risk factors, and each predisposes to and exacerbates the other. Recently, the American Heart Association introduced the cardiovascular–kidney–metabolic (CKM) health framework to standardize comprehensive screening and management to improve patient outcomes.^[17] In our study, Cluster 3 exhibited an increased risk of both CVDs and CKD, without a corresponding increase in liver-related complications, leading us to refer to it as the “high CKM risk subgroup”. Cluster 3 was characterized by glucose and lipid metabolism disorders, consistent with prior studies that categorized MASLD into subtypes with varying CVD risk profiles, primarily based on metabolic characteristics.^[4,5] Notably, Cluster 3 exhibited pronounced chronic systemic inflammation, lower muscle mass, and advanced age. Long-term exposure to chronic inflammation is a risk factor for CKM.^[18] Circulating pro-inflammatory and pro-oxidative stress factors exacerbate pathophysiological processes involved in atherosclerosis, myocardial injury, renal hypoxia, and fibrosis. Moreover, a recent study has highlighted that elevated levels of inflammatory markers correlate with reduced muscle mass and diminished muscle strength,^[19] which, compounded by age-related muscle decline, further increased cardiovascular risk.^[20] Skeletal muscle plays a critical role in regulating lipid and glucose uptake and utilization, which agrees with the metabolic disorders observed in this cluster. The major challenge in managing MASLD individuals in Cluster 3 is the reduction of CKM risk. Therefore, based on our above findings, besides effective metabolic control, the strategy on anti-inflammatory and muscle-building should be considered, especially among the elderly.

Screening for advanced fibrosis was recommended in each MASLD patient.^[7] However, the existing cluster classifications failed to distinguish patients with potential high risks of adverse liver-related outcomes.^[4,5] Cluster 2 exhibited a higher risk of fibrosis, without a corresponding increase in adverse cardiovascular or kidney outcomes. Accordingly, Cluster 2 was labeled as the “high fibrosis risk subgroup”. It was characterized by severe lipid metabolism disorder and liver damage. Hepatocyte lipotoxicity is a primary event in the development of liver injury and contributes to inflammation, leading to hepatic fibrosis.^[21,22] We speculate this subgroup most likely represents a transient phenotype and longer follow-ups are needed to confirm the long-term outcomes. Patients in this cluster would benefit from aggressive lipid-lowering therapy and improvement in insulin sensitivity to further mitigate their risk of adverse outcomes over the long term.

Cluster 4 is characterized by severe insulin resistance, liver injury, and visceral adiposity. A large amount of visceral abdominal fat is considered a risk factor for liver diseases and insulin resistance, which further aggravates the progression of liver fibrosis.^[23] Long-term exposure of hepatocytes to hyperglycemia or/and hyperlipidemia induced oxidative stress, leading to inflammatory cytokines release and collagen deposition, promoting fibrosis.^[3,21] Previous study indicated that PNPLA3-rs738409 genotype in addition to diabetes dramatically increased the risk stratification of cirrhosis in MASLD subjects.^[24] Consistently, owing to highest frequencies of PNPLA3 risk alleles and up to 98% with diabetes in this cluster, cluster 4 has highest proportion of MASH or significant fibrosis. Although genetic risk is often perceived as immutable, strict glycemic control and improved insulin sensitivity might be important to prevent fibrosis in Cluster 4. Moreover, severe insulin resistance not only accelerates the progression of liver fibrosis, but also promotes the incidence of cardiovascular and renal diseases.^[25] Therefore, in addition to having a higher risk of liver fibrosis as Cluster 2, Cluster 4 demonstrates a significantly increased risk of cardiovascular and renal complications, as well as a higher incidence of all-cause and cardiovascular mortalities. Considering profound impact of metabolic risk factors on liver fibrosis, CKD, and CVD, and the critical role of the liver in maintaining metabolic homeostasis, we proposed patients in Cluster 4 should be included in the management of “cardiovascular–liver–kidney–metabolic (CLKM) syndrome” to effectively implement comprehensive prevention and control strategies. Thus, we named Cluster 4 as “high CLKM risk subgroup”.

Of note, the presence of steatosis is not always independently associated with a clinically meaningful increase in the risk of CVDs.^[26] There is an urgent need for a prognostic classification capable of identifying low-risk patients, thereby focusing healthcare resources toward those patients at risk for worse outcomes. Patients in Cluster 1 had more thigh adipose tissue distribution, but significantly lower abdominal visceral fat. Their phenotypes, including lower plasma markers of atherogenic dyslipidemia, higher insulin sensitivity and greater beta cell function, resemble metabolically healthy obese.^[27] As expected, this cluster conferred lower risks of cardiovascular complications, as well as all-cause and cause-specific mortality. Accordingly, this cluster was labeled as “low CVD risk subgroup”. A growing body of evidence suggests that subcutaneous fat is protective of cardiometabolic diseases and mortality.^[28] This is consistent with our findings, suggesting that fat distribution may be more relevant to CVD risks than BMI. Therefore, this cluster assists in identifying MASLD patients characterized as “relatively metabolically healthy”. Another important issue, we have to point that the individuals in Cluster 1 were relatively young, which may also contribute to the low CVD risk. Recognizing their distinct metabolic profiles and body fat distribution enables the development of tailored, comprehensive treatment plans aimed at maintaining their metabolic health over prolonged periods of time.

There were some limitations to our study. First, whether the four identified clusters represent different stages in the progression of MASLD, and whether patients in one subgroup can transfer into another one over time need to be assessed in future prospective studies with continuous follow-up. Second, cohort heterogeneity, particularly in MASLD prevalence and follow-up duration, may affect the comparability of the findings. While similar clusters were observed in external cohorts with different MASLD prevalence rates (43.9% and 37.3%), these differences should be considered when interpreting the results. Additionally, when attempting to validate the association of the four clusters with CKD risk, we were unable to replicate the findings due to the lack of relevant data in the health check-up cohort and NHANES datasets. Future research incorporating these datasets or other relevant cohorts may address this gap. Lastly, instead of hard endpoint outcomes, we utilize surrogate endpoints that have been validated by guidelines and authorities. More large-scale studies are needed given the potential biases, and longer-term follow-up is required to confirm the observed associations.

In this study, we identified four MASLD clusters with distinct clinical characteristics associated with different risks of hepatic and extrahepatic outcomes. This refined classification not only aids in identifying and stratifying MASLD patients based on their clinical phenotypes but also represents a pivotal advancement toward personalized treatment strategies. Furthermore, this study highlights the interactions among metabolic risk factors, liver fibrosis, CKD, and CVD, which have a profound impact on the complications and mortality of MASLD.

Funding

This work was supported by grants from the Key Research and Development Program of Jiangsu Province of China (Nos. BE2022666 and BE2023774), the Nanjing Health Science and Technology Development Project (No. ZKX22027), the National Natural Science Foundation of China (Nos. 82030026, 82270883, 82370841, and 82401007), the Natural Science Foundation of Jiangsu Province of China (No. BK20240232), and Clinical Trials from the Affiliated Drum Tower Hospital, Medical School of Nanjing University (No. 2021-LCYJ-DBZ-10).

Acknowledgements

The authors thank all volunteers for their participation in this study and medical personnel from the Department of Endocrinology, Department of Health Management Center, Department of Pathology, and National Institute of Healthcare Data Science of Nanjing University for their valuable assistance.

Conflicts of interest

None.

Data availability statement

Some or all datasets generated and/or analyzed during the current study are not publicly available, but are available from the corresponding author on reasonable request.

Supplementary Material

cm9-139-1062-s001.docx ^{(503.8KB, docx)}