Extracellular vesicles as biomarkers for metabolic dysfunction-associated steatotic liver disease staging using explainable artificial intelligence

Abstract

BACKGROUND

Metabolic dysfunction-associated steatotic liver disease (MASLD) is a leading cause of chronic liver disease globally. Current diagnostic methods, such as liver biopsies, are invasive and have limitations, highlighting the need for non-invasive alternatives.

AIM

To investigate extracellular vesicles (EVs) as potential biomarkers for diagnosing and staging steatosis in patients with MASLD using machine learning (ML) and explainable artificial intelligence (XAI).

METHODS

In this single-center observational study, 798 patients with metabolic dysfunction were enrolled. Of these, 194 met the eligibility criteria, and 76 successfully completed all study procedures. Transient elastography was used for steatosis and fibrosis staging, and circulating plasma EV characteristics were analyzed through nanoparticle tracking. Twenty ML models were developed: Six to differentiate non-steatosis (S0) from steatosis (S1-S3); and fourteen to identify severe steatosis (S3). Models utilized EV features (size and concentration), clinical (advanced fibrosis and presence of type 2 diabetes mellitus), and anthropomorphic (sex, age, height, weight, body mass index) data. Their performance was assessed using receiver operating characteristic (ROC)-area under the curve (AUC), specificity, and sensitivity, while correlation and XAI analysis were also conducted.

RESULTS

The CatBoost C1a model achieved an ROC-AUC of 0.71/0.86 (train/test) on average across ten random five-fold cross-validations, using EV features alone to distinguish S0 from S1-S3. The CatBoost C2h-21 model achieved an ROC-AUC of 0.81/1.00 (train/test) on average across ten random three-fold cross-validations, using engineered features including EVs, clinical features like diabetes and advanced fibrosis, and anthropomorphic data like body mass index and weight for identifying severe steatosis (S3). Key predictors included EV mean size and concentration. Correlation, XAI, and SHapley Additive exPlanations analysis revealed non-linear feature relationships with steatosis stages.

CONCLUSION

The EV-based ML models demonstrated that the mean size and concentration of circulating plasma EVs constituted key predictors for distinguishing the absence of significant steatosis (S0) in patients with metabolic dysfunction, while the combination of EV, clinical, and anthropomorphic features improved the diagnostic accuracy for the identification of severe steatosis. The algorithmic approach using ML and XAI captured non-linear patterns between disease features and provided interpretable MASLD staging insights. However, further large multicenter studies, comparisons, and validation with histopathology and advanced imaging methods are needed.

Key Words: Metabolic dysfunction-associated steatotic liver disease; Extracellular vesicles; Non-invasive biomarkers; Machine learning; Explainable artificial intelligence; Transient elastography; Metabolic dysfunction; Hepatic steatosis

Core Tip: This study evaluated circulating plasma extracellular vesicles (EVs) as metabolic dysfunction-associated steatotic liver disease biomarkers for steatosis identification and staging using machine learning and explainable artificial intelligence. EV-based machine learning models demonstrated that mean size and concentration of EVs are key predictors that effectively distinguish the absence of significant steatosis in patients with metabolic dysfunction and the presence of severe steatosis (S3) when they are combined with clinical and anthropomorphic data. Further, large multicenter studies, comparison with advanced imaging methods, and histopathology validation are required to confirm the clinical utility of EVs.

INTRODUCTION

Metabolic dysfunction-associated steatotic liver disease (MASLD), formerly known as nonalcoholic fatty liver disease, constitutes the leading chronic hepatic disease globally, with particularly increased prevalence in Western countries[1-3]. This disease is characterized by lipid accumulation in the hepatic parenchyma, steatosis, metabolic disturbances, including dyslipidemia, obesity, type 2 diabetes mellitus (T2DM), and hypertension, and no history of significant alcohol intake[1-3]. This disease encompasses a spectrum of liver conditions including hepatic steatosis, metabolic steatotic liver, metabolic dysfunction-associated steatohepatitis with or without fibrosis, cirrhosis, and hepatocellular carcinoma. Surprisingly, the main causes of mortality in patients with MASLD are primarily cardiovascular complications, followed by extrahepatic malignancies (higher risk in MASLD vs obese) and liver-related outcomes for the advanced cases that might also require transplantation[3-6].

The rising rates of obesity, insulin resistance, and T2DM and the aging of the population are risk factors for MASLD development. They constitute a major public health concern. The increase of deaths due to cardiovascular complications and extrahepatic malignancies imply the significance of the early detection of hepatic steatosis before its advancement. Based on European Association for the Study of the Liver-European Association for the Study of Diabetes-European Association for the Study of Obesity Clinical Practice Guidelines recommendation, subjects with obesity and metabolic syndrome should be screened for steatosis as there is a high likelihood of having MASLD, and liver enzymes and/or ultrasound should be performed as routine workup[4].

The diagnosis of MASLD in clinical practice typically involves identifying fat infiltration in the liver in the presence of cardiometabolic risk factors after excluding other causes of steatotic liver disease. Several non-invasive imaging tests (NITs) are utilized for the identification and risk stratification of MASLD and include conventional ultrasound, transient elastography (TE), and magnetic resonance imaging (MRI). Additionally, many blood-based NITs focus primarily on fibrosis, which is prognostic for liver outcomes.

There are fewer tests available for the identification or staging of steatosis. Even though liver biopsy remains the gold standard for diagnosing metabolic dysfunction-associated steatohepatitis (histopathological finding) and fibrosis, steatosis is not routinely assessed through biopsy due to its invasive nature, cost, and associated risks[7,8]. The measurement of controlled attenuation parameter or ultrasound attenuation parameter (UAP) constitute the most frequently utilized NITs for the quantitative evaluation of steatosis stage as they are embedded in point-of-care TE devices that are widely used by clinicians[8-11]. However, the rising number of patients, as well as the lack of accessibility in such equipment, requires the development of novel blood-based diagnostic and risk stratification tools that can facilitate clinical decision making, especially in individuals presenting several cardiometabolic risk factors[1,2,8].

Several studies have highlighted the potential of extracellular vesicles (EVs) as diagnostic markers for metabolic diseases, including MASLD, due to their crucial role in disease progression and their ease of isolation from various biological fluids[12-15]. EVs constitute heterogeneous nanoparticles presenting a bilayered phospholipid membrane, which encompass a wide variety of cargoes inside them. These vesicles are sized between 40-1000 nm, depending on their biogenetic mechanism. Exosomes are the smallest in diameter subpopulation (40-150 nm), microvesicles are the medium-sized (150-1000 nm), and the apoptotic bodies the largest (> 1000 nm). They result from inward cell membrane invagination, outward membrane blebbing, and cell apoptosis, respectively[12-15].

Recent studies have demonstrated that EVs are released by a wide variety of cells in response to metabolic stress, which is attributed to several metabolic disturbances[16-18]. In the context of obesity, it has been demonstrated that the counts of total microparticles are elevated in patients with obesity compared to controls without obesity[16-18]. Similarly, another systematic review and meta-analysis of 48 studies demonstrated that the counts of total microparticles were elevated in patients with T2DM compared with patients without T2DM[19]. It has also been suggested that adipocytes constitute a major source of EVs during obesity as it was shown in mice models[20] as well as under hypertensive conditions[21].

Several studies have demonstrated the linkage between the total amount of EVs and the disease severity. A recent animal study demonstrated the correlation between EV characteristics (concentration and size) with disease severity and systemic inflammation[22]. Similarly, it was demonstrated in another human study that the median EV circulating levels were notably elevated in patients with cirrhosis and nonalcoholic steatohepatitis and less significantly in patients with nonalcoholic steatohepatitis along in comparison with healthy controls[23]. Another human study demonstrated the plasma EVs and EVs of hepatic origin were correlated with disease resolution after bariatric surgery[24].

EVs have a major key role in MASLD as they are implicated in the cell-cell communication between the parental and the recipient cells. This intercellular communication is mediated via the release of EV-contained cargoes from the former and their uptake by the latter[12-14]. Aberrations in EV quantitative and qualitative characteristics can significantly alter the functional state of the recipient cells and potentially leads to disease progression by promoting inflammation and enhancing fibrosis and angiogenesis[12-14,16].

Several cells produce a higher amount of EVs during metabolic dysfunction that eventually induces steatosis development and its progression to more severe stages. These alterations in EV characteristics have provided opportunities for patient stratification and disease monitoring[12-14,16]. Our hypothesis that alterations in circulating EV characteristics (size and concentration levels) could serve as predictors of steatosis severity in patients with metabolic dysfunction led to the design of this study. This research study aimed to investigate and evaluate EVs (mean concentration and size) as non-invasive MASLD biomarkers for identification and staging of steatosis, using UAP and integrating data science, machine learning (ML) and explainable artificial intelligence (XAI) approaches in order to elucidate complex relationships between disease features, outperforming conventional statistics and enhancing the diagnostic accuracy for MASLD[25,26].

MATERIALS AND METHODS

Study population

This observational, single-center study was conducted at the Gastroenterology and Liver Unit of the Second Academic Department of Internal Medicine, National and Kapodistrian University of Athens, at Hippocratio General Hospital. The Ethics Committee of the General Hospital of Athens “Hippocratio” (No. 24, dated 15 November 2022) granted approval of the study, and patients were enrolled from March 2023 until March 2024. We systematically collected demographic data, patient medical history, and routine blood test results from Hepatology Clinic outpatient patients with a medical history of metabolic dysfunction meeting the following criteria.

The inclusion criteria[3,4]: Willingness to participate and the ability to undergo the protocol procedures, the presence of at least 1 out of 5 cardiometabolic risk factors suggested by multisociety Delphi consensus on new nomenclature: (1) Body mass index (BMI) ≥ 25 kg/m² or increased waist circumference for males (> 80 cm) and females (> 94 cm); (2) Low high-density lipoprotein levels for males < 40 mg/dL and for females < 50 mg/dL or treatment for dyslipidemia; (3) High plasma triglycerides (> 150 mg/dL) or treatment for dyslipidemia; (4) Hypertension (≥ 130/80 mmHg) or antihypertensive drug treatment; and (5) High levels of fasting glucose, or high 2 h-postprandial glucose levels or diagnosed diabetes mellitus or high levels of glycosylated hemoglobin or anti-diabetic treatment[3]. Presence of chronic liver biochemistry disturbances, with persistent abnormalities in liver enzymes [< 3 × upper limit of normal (ULN)] as well as negative results in a large detailed panel of tests that have excluded any autoimmune diseases, viral hepatitis, cholestatic metabolic, hereditary or genetic metabolic liver diseases, and endocrine diseases that may affect liver biochemistry.

The exclusion criteria[3,4,27]: alanine aminotransferase or aspartate aminotransferase ≥ 3 × ULN, alcohol abuse (≥ 2 drinks for females and ≥ 3 drinks for males per day; 10 g of alcohol/drink), the presence of any other hepatic pathology than MASLD, past or present history of viral hepatitis or under anti-viral medication, autoimmune diseases, systemic chronic inflammatory pathologies or active inflammation, prior malignancy or active malignancy or under anti-neoplastic medications, acute/subacute cardiovascular events (e.g., myocardial infarction, stroke, etc.), endocrinopathies (e.g., Hashimoto disease), neuropathies, neurodegenerative diseases, myopathies, hereditary or genetic metabolic diseases, hemolytic and anemia-related disorders, coagulopathies, chronic kidney disease, exposure to hepatotoxic drugs, herbs, or anabolics, presence of ascites, state of pregnancy, altered circadian rhythm (sleep deprivation related to occupation), and non-compliance to fasting period prior to sample collection and/or to lack of physical exercise prior to sample collection.

It has to be underlined that all of the above conditions in the exclusion criteria could potentially increase the total amount of circulating EVs[27,28] or modify the results of the protocol procedures (e.g., ascites, transaminases ≥ 3 × ULN, etc.).

Patient enrollment

We initially enrolled 798 outpatients with metabolic dysfunction from March 2023 until March 2024. We excluded 566 patients, who met the exclusion criteria. After 38 patients voluntarily excluded themselves due to lack of interest in participating, we officially included 196 patients, who met the eligibility criteria. However, there was a large drop out of 118 patients (25 patients were lost, 27 patients did not show up, 4 patients lost interest in participating, 33 patients were newly diagnosed with a disease included in the exclusion criteria, 17 patients did not fast prior to sampling, and 12 patients presented symptomatology of acute infection/inflammation in the moment of sampling). Finally, 76 patients successfully completed all study procedures. We present the flowchart of the study participants (Figure 1), with 76 patients completing all the study protocol procedures.

Figure 1 Flowchart of study participants.

Assessment of liver steatosis and fibrosis by TE

The 76 patients were primarily assessed by upper abdominal ultrasound examination by an experienced operator for the assessment of liver parenchyma regarding the fat infiltration grade (mild, moderate, severe)[29]. Patients with inconclusive ultrasound examinations due to factors such as excessive subcutaneous or visceral fat, which may hinder sound wave penetration, or excessive bowel gas, which can limit visibility, were subsequently stratified based on the next protocol procedure. TE was further performed for all patients by an experienced operator using the “iLivTouch” FT100 device (Wuxi Hisky Medical Technologies Co., Ltd., China). This device assesses liver stiffness measurement (LSM) through TE while simultaneously measuring the UAP with a universal probe, ensuring applicability across various body sizes[9,11]. Liver steatosis and fibrosis were evaluated and quantified based on UAP and LSM, respectively. Measurements were considered valid if at least ten successful readings were obtained, with a success rate ≥ 60% and an interquartile range/LSM ratio ≤ 30% per the World Federation for Ultrasound in Medicine and Biology guidelines[30].

Steatosis severity was classified based on UAP values as follows[9]: Severe steatosis (S3): ≥ 296 dB/m; moderate steatosis (S2): ≥ 269 dB/m; and mild steatosis (S1): ≥ 244 dB/m. The median UAP values between 244 dB/m and 296 dB/m were interpreted as indicative of liver steatosis, while patients with UAP levels below 244 dB/m (S0) were considered as patients with metabolic dysfunction, without significant (absence) liver steatosis. For further analysis, we categorized patients into three groups: Severe steatosis (≥ S3, UAP ≥ 296 dB/m); no/mild-to-moderate steatosis (S0-S2, UAP < 296 dB/m); and absence of steatosis (S0, UAP < 244 dB/m)[7]. Additionally, fibrosis severity was classified based on LSM values as follows[9]: No fibrosis to mild fibrosis (F0-F1): < 7.3 kPa; moderate fibrosis (F2): < 9.7 kPa; moderate-to-severe fibrosis (F2-F3): < 12.4 kPa; advanced fibrosis and above (F3-F4): < 17.5 kPa; and cirrhosis (F4): > 17.5 kPa based on the manufacture. All patient characteristics were categorized according to the Strengthening the Reporting of Observational Studies in Epidemiology guidelines and are presented in Tables 1 and 2.

Table 1 Patient characteristics.

Feature	Count	Missing (%)	Mean	SD
Age	74	2.63	55.34	13.20
Height	75	1.32	1.68	0.09
Weight	75	1.32	80.81	14.30
Mean size	74	2.63	138.16	19.76
50-150 nm	74	2.63	8.27 × 1010	4.37 × 1010
> 150 nm	74	2.63	5.84 × 1010	4.85 × 1010
Sum (50-1000 nm)	74	2.63	1.41 × 1011	7.04 × 1010

Table 2 Summary of categorical features.

Feature	Count	Missing (%)	Percentage
Sex	75	1.32	58.67 (males)
Diabetes	76	0	19.74
S0	21	0	27.63
S1	20	0	26.32
S2	14	0	18.42
S3	21	0	27.63
F0-F1	40	0	52.63
F2	23	0	30.26
F2-F3	5	0	6.58
F3-F4	2	0	2.63
F4	6	0	7.89

Preanalytical blood sample processing for EV isolation and characterization

To minimize variability in EV analysis, we followed the same preanalytical protocol for fasting blood sample collection and processing, including the time of collection (up to 5 min), post-collection tube inversion (ten times) for proper anticoagulation, blood processing interval, which was up to 1 h, as well as the handling. Blood samples were collected in citrate tubes (Vacuette sodium citrate 3.2%, volume 3.5 mL, Greiner Bio-One), which were selected for their suitability in EV analysis[31].

For the reduction of variability due to physiological factors that modify vesiculation, all patients were fasted, they did not present any symptom of acute infection, and they did not have any physical activity or sleep deprivation prior to sampling [27].

Firstly, cellular debris and larger particles were removed via a 20-min centrifugation at 3000 g at 4 °C, with the supernatant being immediately preserved and stored at -80 °C for the preservation of the sample’s integrity. Further analysis included a room temperature thawing and dilution in particle-free PBS (0.02 μm filtered, Cytiva Whatman, United Kingdom). For a gentle, simplified, and less time-consuming EV isolation, sequential filtrations (microfiltration) were used, which have been also applicable for EV isolation, starting with passage of the supernatant through a filter pore of 0.8-μm for the retaining of large particles, cell fragments, and platelets and eventually reducing the filter pore diameter[27,32].

EV characterization was mediated via nanoparticle tracking analysis (NTA), which was conducted using the NanoSight NS300 instrument (Malvern Instruments, Amesbury, United Kingdom)[32,33]. The conditions during the measurements were constant, including the flow (syringe speed at 100 μL/second), temperature (25 °C), sCMOS camera, green laser, no evidence of vibration, which can alter the sizing of particles, and autofocus adjustment to ensure clarity by avoiding indistinct particles. Five 30-sec videos were recorded per measurement and analyzed using NanoSight NTA 3.4 build 3.4.4 software (Malvern, 2020) in script control mode, comprising 1500 frames per sample. Each sample was measured five times, and the size distribution data were averaged. Additionally, we evaluated the possible interday variability of the method by performing NTA for each sample on different days, and it did not show any significant difference between the values. Lastly, we obtained data regarding the mean EV size for each patient and the number of vesicles 50-150 nm, > 150 nm, and 50-1000 nm by which we acquired data about their total concentration levels, size, and distribution in the sample[33].

Data science

We used a personal computer for our experiments, which runs Linux OS Version: 74-20.04.1-Ubuntu, OS Release: 5.15.0-67-generic. Hardware: Architecture: AMD^® Ryzen 93950x 16-core processor × 32, RAM (GB): 125.71, CUDA Version: 12.6, cuDNN Version: 90501, GPU Name: NVIDIA GeForce RTX 3060. Software: Library Versions: Python: 3.10.16, pandas: 2.1.4, numpy: 1.26.4, sklearn: 1.4.1. post1, catboost: 1.2.3, shap: 0.46.0, matplotlib: 3.8.2, plotly: 5.18.0, scipy: 1.11.4.

Based on the known alterations in EV quantity, size, and quality in MASLD pathogenesis, we applied ML algorithms and feature engineering finite element (FE) techniques to analyze non-linear feature relationships[34]. FE enhances ML performance by generating new features, while feature selection (FS) identifies the most important ones[34]. Using the CatBoost algorithm[35-37], which by default does not easily over-fit[35] (by definition the probability of over-fitting is high when the performance of an ML model is better on the training dataset than on the validation/test sets). Using FE, XAI methods, and FS techniques[34,37,38], we developed a data science pipeline to predict non-severe or severe steatosis in MASLD patients for two cases: Case 1 (C1): S0 vs S1, S2, S3; and Case 2 (C2): S0, S1, S2 vs S3. UAP cutoff values were employed to label steatosis stages, and circulating plasma EV levels were quantified for C1 and C2.

The linear associations between features and steatosis groups in C1 and C2 cases were explored via the performance of correlation analysis using the Point-Biserial test[39,40]. Correlations that exceeded a threshold of 0.5 were considered meaningful. For our analysis we utilized the following EV and clinical features as well as the anthropometric data: Mean. Average vesicle size (nm); 50-150 nm. The count of vesicles sized between 50 to 150 nm; > 150 nm. The count of vesicles sized above 150 nm; Sum. Total vesicle count sized between 50 and 1000 nm; Anthropometric data. Sex, age, height, weight, and BMI; and Clinical features. Adv_Fibrosis and diabetes. We trained/tuned 20 CatBoost models, leveraging XAI and FS to identify key features. Model robustness was evaluated with non-stratified 5CV for C1, 3CV for C2, and test sets. Given the small dataset size, we repeated cross-validation with ten different random seeds to mitigate splitting randomness effects in cross-validation folds[34]. For C1 and C2, we created six and eight sub-cases, respectively (Supplementary material).

For C2 we used the cutoff UAP value of ≥ 296 to label the target feature, which was severe steatosis (S3) and not severe steatosis (S0, S1, S2). For the C1, we used the cutoff UAP value of ≥ 244. We dropped data instances with null values, and no normalization/scaling was applied. Given the minimal proportion of missing data (maximum 2.63%), we opted to exclude these records rather than apply imputation techniques. This decision was made deliberately to maintain data integrity, minimize bias, and preserve the validity of subsequent analyses. For both cases, we randomly split the dataset (74 or 72 instances depending on the subcase) into 80% training and 20% testing sets, ensuring similar target distributions in both sets for fairness. A fixed random seed was used for reproducibility, and patients in the training set were excluded from the testing set to prevent data leakage[41]. This random split approach aligns with TRIPOD-AI guidelines[42], ensuring the ML models are applicable to the center where the data originated. FS[38], utilized in both datasets, ensured no information leakage[41] due to its inductive learning methodology[43]. We employed two approaches to examine the predictive and associative roles of EV motivated by: (1) Limited bibliographic data on EV assessment for MASLD using XAI and ML; and (2) The small dataset size. For FE subcases, we divided EV features by 10⁹, raised numerical features to the 11^th power, calculated their square roots, and computed pairwise feature products.

Applying the FE on the sub-cases C2d, C2f, and C2h, lead us to 5, 11, and 21, respectively, most important features. Analyzing these features using SHapley Additive exPlanations (SHAP)[44] and the feature importance of CatBoost provided us with insights regarding the features in C2 cases. All ML models were tuned using 3CV and 5CV on the training datasets for C1 and C2, respectively, and validated on the test set for each case. Hyperparameters for each CatBoost model were manually tuned using systematic grid search guided by domain expertise. The specific hyperparameters tested and final values selected are comprehensively detailed in Supplementary Tables 1 and 2. Each model could represent a method for automated classification of individuals into classes for each case. The iterators, learning rate, depth, and weights were tuned in CatBoost, while the other hyperparameters were set to default values. We developed a deterministic algorithm to create two predictive ML models for cases C1 and C2 and to identify the most important features and their associations. In Figure 2, we demonstrate the flowchart of the algorithm, while the steps of the algorithm are described in the Supplementary material. Additionally, we demonstrate a graphical abstract that summarizes the methodology and principal findings (Figure 3).

Figure 2 Flowchart of the algorithm. EV: Extracellular vesicle; UAP: Ultrasound Attenuation Parameter; SHAP: SHapley Additive exPlanations; C1: Case 1; C2: Case 2.

Figure 3 Graphical abstract. MASLD: Metabolic dysfunction-associated steatotic liver disease; EVs: Extracellular vesicles; ROC-AUC: Receiver operating characteristic-area under the curve; C1: Case 1; C2: Case 2; UAP: Ultrasound Attenuation Parameter; ML: Machine learning. Created in BioRender (Supplementary material).

RESULTS

In C1 the training set included 59 individuals, and the test set had 15. Six CatBoost models were trained. Performance metrics, including sensitivity, specificity, ROC-AUC, and F1 scores, are summarized in Supplementary Table 1. The CB-C1a model performed best (Table 3) with a mean (± SD) ROC-AUC of 0.70 (± 0.15) and an F1 score of 0.73 (± 0.14), showing balanced sensitivity (0.66 ± 0.17) and specificity (0.74 ± 0.14). Correlation analysis using Point-Biserial (Supplementary Figure 1) revealed no correlations exceeding the 0.5 threshold, suggesting weak and possibly complex relationships between features and steatosis severity. Violin plots (Figure 4) of EV feature distributions showed no distinguishable patterns, indicating these features fail to differentiate steatosis stages when analyzed linearly. However, SHAP scatter plots (Figure 4) revealed non-linear relationships, with SHAP values highlighting the impact of EV features on predictions.

Figure 4 Explainable artificial intelligence SHapley Additive exPlanations for case 1. Plots of SHapley Additive exPlanations scatter plots and extracellular vesicle distribution for the train set of the machine learning CB-C1a model. SHAP: SHapley Additive exPlanations.

Table 3 Performance on the CatBoost-C1a (a) and CatBoost-C2h-21 (b).

Model		spec_cv_m	sens_cv_m	roc_cv_m	f1_cv_m
CB-C1a	Mean	0.76	0.66	0.71	0.72
	SD	0.05	0.06	0.03	0.05
	Min	0.65	0.53	0.67	0.64
	25%	0.73	0.63	0.68	0.68
	50%	0.76	0.66	0.71	0.73
	75%	0.79	0.70	0.73	0.75
	Max	0.83	0.74	0.77	0.78
CB-C2h-21	Mean	0.80	0.81	0.81	0.71
	SD	0.06	0.06	0.04	0.05
	Min	0.68	0.74	0.77	0.64
	25%	0.78	0.76	0.78	0.67
	50%	0.81	0.80	0.79	0.69
	75%	0.85	0.85	0.82	0.73
	Max	0.89	0.92	0.89	0.81

Using ten times iterative 5CV and 3CV for Case 1 and Case 2, respectively. Here x_cv_m for x in {spec, sens, roc, f1} corresponds to {specificity, sensitivity, receiver operating characteristic-area under the curve, F1}. roc: Receiver operating characteristic; spec: Specificity; sens: Sensitivity; CB-C2h-21: CatBoost C2h-21; CB-C1a: The CatBoost C1a; C1: Case 1; C2: Case 2; Min: Minimum; Max: Maximum.

Adding anthropometric and clinical features (e.g., BMI, advanced fibrosis, and diabetes) or applying FE did not improve the performance of the model. In the S1-S3 group advanced fibrosis was present in 6/53 cases, diabetes in 13/53, and both in 5/53 (Table 4), supporting the sufficiency of EV features for distinguishing S0 from higher steatosis stages. Feature importance and SHAP analyses (Figure 4) identified mean, 50-150 nm, > 150 nm, and Sum as the top predictors, with mean ranked highest. SHAP plots underscored the need for advanced AI techniques to capture these complex relationships. Moreover, Figure 5 presents a SHAP beeswarm plot, showcasing the contributions of key EV features to the predictions made by the CB-C1a ML model for differentiating steatosis stages (ROC-AUC train: 0.70, test: 0.86). The features, including Sum, 50-150 nm, mean, and > 150 nm, are ranked based on their importance, with Sum being identified as the most impactful feature.

Figure 5 Each dot represents an individual data point, with the x-axis denoting the SHapley Additive exPlanations value (impact on model output) and the color gradient indicating feature values (blue for low values, red for high values). The Sum feature showed the strongest influence, with higher values correlating with positive SHapley Additive exPlanations values, indicating its critical role in predicting steatosis stages. The beeswarm distribution highlights the non-linear and complex relationships between features and predictions, emphasizing the importance of explainable artificial intelligence in interpreting the contributions of extracellular vesicle characteristics in diagnosing metabolic dysfunction-associated steatotic liver disease. SHAP: SHapley Additive exPlanations.

Table 4 Advanced fibrosis and diabetes.

Group	Total patients	Advanced fibrosis	Diabetes	Advanced fibrosis and diabetes
S0	21	0/21 (0.0)	0/21 (0.0)	0/21 (0.0)
S1-S3	53	6/53 (11.3)	13/53 (24.5)	5/53 (9.4)
S3	19	3/19 (15.8)	9/19 (47.4)	3/19 (15.8)
S0-S2	55	3/55 (5.5)	4/55 (7.3)	2/55 (3.6)
S1-S2	34	3/34 (8.8)	4/34 (11.8)	2/34 (5.9)

Data are presented as n (%).

In C2 the training set included 58 individuals, and the test set consisted of 14. Fourteen CatBoost models were tuned, with the CB-C2h-21 model achieving the best performance (Table 3) with a ROC-AUC of 0.89 (± 0.03) on the training set (3CV) and 1.00 on the test set. Sensitivity and specificity values of 0.92 (± 0.12) and 0.87 (± 0.10) demonstrated its reliability in distinguishing mild-to-moderate steatosis (S0-S2) from severe steatosis (S3). Figure 6 displays a SHAP, illustrating the contributions of interaction and engineered features to the predictions made by the CB-C2h-21 ML model for distinguishing severe steatosis (S3) from non-severe stages (S0-S2).

Figure 6 Each dot represents an individual prediction, with the X-axis denoting the SHapley Additive exPlanations value (impact on model output), while the color gradient indicates the feature value (blue for low values, red for high values). Features such as > 150_pow_5, Height_x_Weight, and Sum_x_BMI also significantly contribute to the predictive capability of the model, showcasing complex interactions between anthropometric, clinical, and extracellular vesicle features. In Supplementary Table 2 we demonstrate the information regarding all the machine learning models of C2 and in Table 3 performances on the CB-C1a and CB-C2h-21, using ten times iterative 5CV and 3CV for C1 and C2, respectively. SHAP: SHapley Additive exPlanations; BMI: Body mass index.

The features are ranked by their importance, with Mean_x_Weight emerging as the most influential predictor. The beeswarm plot highlighted the non-linear and intricate relationships captured by advanced AI techniques, reinforcing the relevance of combining EV features with engineered clinical and anthropometric data to enhance diagnostic accuracy for severe MASLD stages. This visualization underscores the interpretability provided by XAI methods in complex predictive modeling.

Furthermore, violin plots (Figure 7) showed no clear separation between groups, highlighting the limitations of traditional visualization methods. In contrast, SHAP scatter plots (Figure 7) revealed non-linear relationships, with features like 50-150 nm, FE weights, and BMI interactions contributing significantly to predictions. Correlation analysis identified associations for features like 50-150 nm, > 150 nm, and BMI-related interactions, but none exceeded the 0.5 threshold, emphasizing the multiparametric nature of MASLD. Unlike C1, the inclusion of anthropometric and clinical features improved model performance in C2, raising the ROC-AUC from 0.81 to 0.89. The S3 group had higher rates of advanced fibrosis (3/19), diabetes (9/19), and their co-occurrence (3/19) (Table 4), underscoring their relevance in predicting severe steatosis. SHAP and feature importance analyses (Figure 6) highlighted key predictors such as mean × Weight and Sum × Age, reflecting complex non-linear relationships between features and outcomes.

Figure 7 Explainable artificial intelligence SHapley Additive exPlanations C2. A: Plots of SHapley Additive exPlanations scatter plots; B: Extracellular vesicle distribution for the training set of the machine learning CB-C2h-21 model. SHAP: SHapley Additive exPlanations; BMI: Body mass index.

Across both cases EV features consistently demonstrated diagnostic value. In C1 predictors like mean and 50-150 nm were most significant (Figure 5), while in C2 engineered features such as mean × Weight and Sum_sqrt played key roles. SHAP scatter plots (Figures 4 and 7) revealed the importance of non-linear ML models in capturing intricate feature interactions. XAI analysis provided interpretable insights, identifying mean and 50-150 nm as top predictors in C1 and mean × Weight and Weight⁷ in C2 (Figures 5 and 6). These findings reinforce the physiological relevance of EV as biomarkers in MASLD diagnosis. The CB-C1a model effectively distinguished S0 from S1-S3 using only EV features, while the CB-C2h-21 model demonstrated that combining EV with anthropometric and clinical features enhanced predictions for severe steatosis in C2. These results highlight the utility of EV-related features and XAI in advancing clinical AI applications in hepatology.

DISCUSSION

To the best of our knowledge, this study constituted one of the first projects that evaluated EV features using UAP and integrating advanced techniques, such as ML and XAI, to shed light on the complex relationships between disease features that could not be elucidated via performing conventional statistics based on linear associations.

Focusing on the findings of our data analysis, the lack of mild correlations (≥ 0.5) in both C1 and C2 highlights the complexity of MASLD where individual features provide limited predictive power. Moderate correlations observed across the Point-Biserial analyses, emphasize the need for integrative approaches that leverage multiple features. These findings align with clinical observations of MASLD progression driven by metabolic, genetic, and environmental interactions rather than isolated markers. The correlation analyses provided a foundation for advanced predictive modeling using XAI.

Moreover, this study underscored the potential of EV as a non-invasive biomarker for MASLD. Using XAI, we demonstrated the diagnostic value of EVs, particularly their size distributions and concentrations. Traditional statistical methods and visualization techniques, such as violin plots, failed to capture meaningful patterns in the steatosis stages. In contrast, SHAP scatter plots revealed complex, non-linear relationships critical for differentiating steatosis stages. This emphasizes the need for advanced XAI approaches to address the multifactorial nature of MASLD. EV features, such as size distributions (mean, 50-150 nm, > 150 nm) and concentrations (Sum), were identified as key predictors. In C1 EV alone achieved strong diagnostic performance, with the CB-C1a model yielding a ROC-AUC of 0.70 in 5 CV and 0.86 on the test set. Conversely, in C2 adding anthropometric and clinical features like advanced fibrosis and diabetes improved the ROC-AUC of the CB-C2h-21 model from 0.81 to 0.89. This highlights the standalone diagnostic value of EVs for early steatosis and its complementary role in advanced stages. Additionally, the results demonstrate the promise of AI-enhanced EV analysis for non-invasive diagnostics, achieving robust sensitivity, specificity, and ROC-AUC values. Rigorous iterative 3-fold and 5-fold cross-validation and standardized EV analysis protocols ensure reliability.

Our findings support that circulating EV characteristics constitute promising non-invasive biomarkers for steatosis prediction regarding its presence as well as its severity. EV-based ML models demonstrated that circulating plasma EV features, such as mean size and concentration, can effectively distinguish the absence of significant steatosis (S0) from the presence of steatosis (S1-S3 stages). The diagnostic accuracy of EV features for severe steatosis was improved when they were combined with clinical and anthropomorphic data.

Although the CB-C1a model achieved an ROC-AUC of 0.71, this accuracy aligns well with other established non-invasive tests currently in clinical use for hepatic steatosis evaluation, such as the controlled attenuation parameter (ROC-AUC approximately 0.70-0.75)[11], nonalcoholic fatty liver disease-LFS (ROC-AUC approximately 0.72-0.80)[45], and the widely recommended fibrosis marker FIB-4 (ROC-AUC approximately 0.70-0.80)[46,47]. Additionally, our model demonstrated strong predictive accuracy (ROC-AUC = 0.86) on the independent test set, indicating good generalizability. Given the moderate variance (SD ± 0.10) and the narrow 95% confidence interval of 0.51-0.78 around the ROC-AUC from cross-validation, we see robust potential for our EV-based diagnostic approach. Future research with larger cohorts and external validation will further confirm and potentially enhance these promising results.

Even though our study identified EV size and concentration as key diagnostic features, detailed biological interpretation remains limited. EV alterations observed in patients with MASLD could reflect metabolic stress, inflammatory signaling, or cellular injury mechanisms. Future studies employing molecular profiling (proteomics, lipidomics) to characterize EV cargo could further elucidate the biological significance of these findings, enhancing both diagnostic precision and mechanistic understanding.

Our current study specifically focused on evaluating the standalone diagnostic value of circulating EV characteristics. Although clinically relevant, biochemical markers such as alanine aminotransferase and aspartate aminotransferase were deliberately excluded to test whether EV features alone, alongside simple clinical and anthropometric variables, have predictive power. Nonetheless, future studies should consider integrating these commonly available biochemical markers, as their inclusion could further enhance the diagnostic accuracy and clinical relevance of the developed models.

This study has some acknowledgeable limitations, such as the variability presented in blood sample processing and the isolation method. More particularly, the freeze-thaw cycle of some of the samples may alter the integrity of the samples. The isolation procedure and the possible contamination with lipoproteins are potential limitations[48-50]. Additionally, the primary aim of this study was to evaluate and investigate EV characteristics (concentration and size) in patients with MASLD after excluding a wide variety of conditions that could alter the total amount. A logical next step for future research is the molecular profiling of these EVs to identify the exact cell origin and cargo, which are closely implicated in the disease progression and will permit the inclusion of more patients[51]. These incorporated data will possibly increase the diagnostic accuracy of the EV-based models, paving the way for more precise, non-invasive disease monitoring and stratification. Moreover, further research is needed to validate EVs as biomarkers for steatosis staging independently of the machine/method (MRI, FibroScan, etc.) by which we labeled steatosis before their clinical application as NITs. The integration and/or comparison of EV characteristics with advanced imaging techniques, such as MRI-proton density fat fraction, and their validation with histopathology findings are essential to enhance diagnostic precision. This approach paves the way for the development of routine, non-invasive clinical diagnostic EV-based tools.

A key limitation of our study was the absence of a healthy control group. Although our primary goal was to identify EV biomarkers differentiating steatosis severity specifically among patients with metabolic dysfunction, including healthy individuals would help clarify the specificity of these EV alterations in MASLD. Future studies should thus incorporate healthy cohorts, which will enhance generalizability, improve diagnostic specificity, and reduce potential biases in EV biomarker validation.

Additionally, another limitation of our study was the absence of external validation in an independent cohort, which is essential to ensure the generalizability of our predictive models. Future multicenter studies should externally validate our findings across diverse populations and clinical settings, confirming the reliability, robustness, and clinical utility of these EV-based biomarkers for MASLD staging.

CONCLUSION

Mean size and concentration of circulating plasma EVs constitute key predictors for distinguishing the absence of significant steatosis (S0), while combining EV features with clinical (presence of advanced fibrosis and T2DM) and anthropomorphic data (sex, age, height, weight, BMI) can identify severe steatosis stage among the patients with metabolic dysfunction. The integration of advanced techniques such as ML and XAI effectively outperformed conventional linear statistical analysis by capturing non-linear patterns. This algorithmic approach using ML and XAI for steatosis provided interpretable insights for MASLD staging. However, larger multicenter studies, comparison with other imaging or blood-based NITs, histopathology validation, as well as molecular profiling of these vesicles are needed for further clinical utilization of this promising biomarker, which may increase the diagnostic accuracy of ML models.

ACKNOWLEDGEMENTS

We gratefully acknowledge the contributions of all study participants whose involvement was instrumental in making this research possible.