Extracellular vesicles as biomarkers for metabolic dysfunction-associated steatotic liver disease staging: Methodological innovations and future perspectives

Abstract

Extracellular vesicles (EVs) have emerged as promising non-invasive biomarkers for metabolic dysfunction-associated steatotic liver disease (MASLD). Recent advances in machine learning and explainable artificial intelligence (AI) have further expanded their diagnostic potential. Recent studies explored the potential of circulating EVs as biomarkers for staging MASLD. By employing machine learning and explainable AI methodologies, the study captured non-linear relationships between EV characteristics and MASLD progression. Furthermore, combining EV features with clinical and anthropometric data significantly improves the diagnostic accuracy for identifying severe steatosis (S3). However, the study has limitations, including a high attrition rate and limited sample size, and the model performance across different MASLD subgroups remains uncertain, especially lean MASLD patients, may be associated with high uncertainty. This opinion review offers a comprehensive analysis of the methodological innovations, clinical implications, and limitations, while also suggesting avenues for future research.

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

Core Tip: Recent study have pioneered the integration of nanoparticle tracking analysis with the CatBoost algorithm and SHapley Additive exPlanations interpretability analysis, thereby uncovering non-linear relationships between extracellular vesicles (EV) characteristics and steatosis severity. The study establishes that EV mean size and concentration can function as independent predictors to distinguish S0 stage, while combination with clinical features significantly improved S3 stage identification accuracy. This provides novel insights and methodologies for non-invasive diagnosis and risk stratification of metabolic dysfunction-associated steatotic liver disease. However, limitations in sample size and unclear mechanistic understanding may hinder further clinical translation, representing urgent issues requiring resolution.

INTRODUCTION

Metabolic dysfunction-associated steatotic liver disease (MASLD) represents a continuum of liver conditions linked to metabolic syndrome, ranging from simple hepatic steatosis to metabolic dysfunction-associated steatohepatitis (MASH), with the potential to progress to liver fibrosis, cirrhosis, and hepatocellular carcinoma. MASLD has emerged as a predominant cause of chronic liver disease worldwide, with particularly high prevalence[1,2].

Recent advances in extracellular vesicles (EVs) biology and machine learning have opened new avenues for non-invasive staging of metabolic dysfunction–associated steatotic liver disease (MASLD), and the developments in artificial intelligence (AI) and machine learning have opened new avenues for disease diagnosis and stratification across multiple medical fields[3]. Accurate stratification of disease severity remains a major clinical challenge, as current diagnostic tools are either invasive or lack sufficient precision. In this context, EVs have emerged as promising non-invasive biomarkers that reflect hepatic injury and metabolic dysregulation, with potential to improve diagnostic performance and mechanistic understanding in liver diseases[4]. A recent study by Trifylli et al[5] skillfully integrates nanotechnology, machine learning, and explainable AI to propose an innovative approach for the non-invasive diagnosis of MASLD.

This review aims to systematically summarize the methodological breakthroughs, clinical application value, and potential limitations of this study, conduct an analysis of the advantages and shortcomings of EVs as biomarkers for MASLD, and clarify the key issues to be addressed in subsequent research while proposing targeted directions by integrating the current research status and technological advancements in the field. It intends to provide valuable references for the clinical translation of this technology, the improvement of the non-invasive diagnostic system for MASLD, and related academic explorations.

NOVEL DIAGNOSTIC METHODS FOR MASLD

MASLD is now increasingly evaluated through non-invasive, stepwise diagnostic pathways rather than by routine liver biopsy alone. Recent guidance from both the American Association for the Study of Liver Diseases and the joint European Association for the study of the liver, European association for the study of diabetes and European association for the study of obesity panel places strong emphasis on risk stratification with non-invasive tests, especially for identifying patients with active steatohepatitis, advanced fibrosis, cirrhosis, or increased risk of liver-related outcomes[2,6]. In this setting, the most important change has not been the emergence of a single dominant biomarker, but the development of complementary diagnostic platforms that combine imaging, serum markers, digital pathology, and data-driven algorithms to improve detection, triage, and prognostic assessment[2,6,7] (Table 1).

Table 1 Overview of current non-invasive diagnostic approaches for metabolic dysfunction-associated steatotic liver disease.

Category	Method/model	Principle	Key variables	Advantages	Limitations
Imaging-based	MRI-PDFF/MAST	Quantitative imaging of liver fat & fibrosis	Liver fat fraction, stiffness	High accuracy, quantitative	Expensive, limited availability
	MEFIB	MRI elastography + FIB-4	Liver stiffness + clinical data	High PPV for fibrosis	Requires MRI access
	FibroScan (FAST, Agile)	Transient elastography	CAP, liver stiffness, AST	Non-invasive, widely used	Intermediate zone exists
	Ultrasound (ATI, 2D-SWE)	Acoustic attenuation/elasticity	Echo attenuation, stiffness	Accessible, bedside use	Lower accuracy vs MRI
Serum biomarkers	FIB-4	Routine lab-based index	Age, AST, ALT, platelets	Simple, widely available	Limited specificity
	ELF test	Fibrosis marker panel	HA, PIIINP, TIMP-1	Good for advanced fibrosis	Cost, lab dependence
	NIS4/NIS2+	Multi-marker panel	miRNA, proteins, HbA1c	Detects at-risk MASH	Limited accessibility
	ADAPT/PRO-C3	ECM remodeling biomarkers	PRO-C3, platelets, diabetes	Mechanism-based	Needs validation
	CK-18/FGF21	Cell death markers	Apoptosis-related proteins	Reflects inflammation	Moderate performance
Metabolic indices	TyG index	Insulin resistance proxy	TG, glucose	Simple, scalable	Indirect marker
Machine learning	Clinical ML models	Data-driven prediction	Anthropometric + lab data	High scalability	Data dependency
Microbiome-based	Gut microbiome signature	Metagenomic profiling	Bacterial composition	Novel mechanism insight	Not standardized
Digital pathology/AI	qFIBS	ML-assisted histology	Fibrosis, inflammation	High reproducibility	Requires biopsy
EV-based (emerging)	EV size + ML	Nanoparticle + AI	EV size, concentration, clinical data	Non-invasive, innovative	Needs validation

ADAPT: Algorithm for Disease Assessment in Progressive Fibrosis Testing; ALT: Alanine aminotransferase; AST: Aspartate aminotransferase; ATI: Attenuation imaging; CAP: Controlled attenuation parameter; CK-18: Cytokeratin-18; ELF: Enhanced liver fibrosis test; EV: Extracellular vesicle; FAST: FibroScan-AST; FGF21: Fibroblast growth factor 21; FIB-4: Fibrosis-4 index; HA: Hyaluronic acid; MASH: Metabolic dysfunction-associated steatohepatitis; MAST: MRI-based assessment of steatosis and fibrosis; MEFIB: Magnetic resonance elastography and FIB-4 index; MRI-PDFF: Magnetic resonance imaging-proton density fat fraction; NIS2⁺: Non-Invasive Scoring system version 2⁺; NIS4: Non-invasive scoring system version 4; PIIINP: Procollagen III N-terminal propeptide; PRO-C3: N-terminal propeptide of type III collagen; qFIBS: Quantitative fibrosis, inflammation, ballooning, steatosis; TG: Triglyceride; TIMP-1: Tissue inhibitor of metalloproteinases-1; TyG index: Triglyceride-glucose index; 2D-SWE: 2-dimensional shear wave elastography; AI: Artificial intelligence; ML: Machine learning; HbA1c: Hemoglobin A1c.

Among imaging-based methods, magnetic resonance-based tools have become particularly important because they provide quantitative information on both steatosis and fibrosis. The FibroScan-AST (FAST) score was one of the first widely adopted composite tools designed to identify patients with steatohepatitis and clinically relevant fibrosis, using liver stiffness measurement, controlled attenuation parameter, and AST[7]. More recently, the magnetic resonance imaging (MRI)-based MAST score improved identification of patients with MASH and significant fibrosis and reduced the indeterminate zone that often limits older tests[8]. The Magnetic Resonance Elastography and FIB-4 index (MEFIB) strategy, which combines magnetic resonance elastography with fibrosis-4 index (FIB-4), also showed high positive predictive value for significant fibrosis and has been externally validated, making it useful as a staged approach for selecting patients for treatment trials or specialist referral[9,10]. A head-to-head comparison of MEFIB, MAST, and FAST further showed that these models offer different balances between performance, accessibility, and cost, which is important for a diagnostic pathway for real-world practice[11]. Beyond fibrosis staging alone, pooled analyses of MRI biomarkers suggest that iron-corrected T1 and liver fat measurements can help identify patients with MASH who are at higher risk of progression[12], and multiparametric MRI combined with blood analytes has also shown good performance for identifying MASH with significant fibrosis[13]. Taken together, these studies show that MRI-based approaches are moving from research tools toward more structured roles in clinical stratification.

Vibration-controlled transient elastography and related FibroScan-based models remain highly attractive because they are easier to deploy than MRI in routine care. Agile 3⁺ and Agile 4 were developed to improve the identification of advanced fibrosis and cirrhosis, respectively, by combining elastography with routine clinical variables[14]. Subsequent external validation confirmed that these scores improve performance and reduce indeterminate results compared with FIB-4 or liver stiffness alone[15]. In parallel, meta-analytic work has strengthened confidence in the FAST score across different cohorts[16]. These findings support a practical message that is increasingly reflected in guidelines: Elastography performs best not as a standalone measurement, but as part of a structured sequential strategy alongside blood-based tools. Beyond fibrosis staging, non-invasive strategies integrating routine ultrasound features with clinical variables have also been explored for predicting clinically significant portal hypertension in MASLD, although their predictive accuracy remains modest[17].

Ultrasound-based methods have also advanced beyond conventional grayscale imaging. Attenuation imaging has shown good performance for quantifying steatosis, and one biopsy-based study reported that its diagnostic performance was not materially affected by fibrosis[18,19]. Quantitative ultrasound techniques, including the ultrasound-guided attenuation parameter, have demonstrated superior diagnostic performance compared with conventional B-mode scoring methods in assessing hepatic steatosis, with improved accuracy in reflecting disease severity[20]. A prospective comparison of attenuation imaging and controlled attenuation parameter against histology and MRI-PDFF supported the diagnostic utility of attenuation imaging in patients with type 2 diabetes and MASLD[19]. For fibrosis staging, two-dimensional shear wave elastography has performed comparably to transient elastography in biopsy-based studies[21], and a multistep strategy incorporating 2-dimensional shear wave elastography has also been evaluated for detecting advanced fibrosis[22]. More broadly, a systematic review and meta-analysis concluded that elastography and MRI methods are useful for fibrosis evaluation in MASLD, while also highlighting the importance of validation in intention-to-diagnose settings[23]. Overall, newer ultrasound methods improve bedside accessibility while narrowing the gap between simple screening and quantitative disease assessment. Machine learning models integrating simple anthropometric and routine biochemical variables have been developed as effective noninvasive tools for predicting MASLD, demonstrating good diagnostic performance and potential for large-scale population screening[24].

Blood-based biomarkers remain central because they are scalable, repeatable, and easier to incorporate into primary care or endocrine settings. NIS4, a panel combining miR-34a-5p, alpha-2-macroglobulin, YKL-40, and hemoglobin A1c, was developed and globally validated for identifying MASH[25]. NIS2⁺ was subsequently derived as a streamlined optimization of the same platform and showed improved performance in detecting at-risk disease across metabolic risk groups[26]. The enhanced liver fibrosis (ELF) test continues to play a major role: A large study showed that ELF performs well for identifying advanced fibrosis in MASLD[27], while population-level work has shown that ELF, alone or combined with FIB-4, can improve liver disease screening efficiency[15]. A primary-care screening study likewise supported the use of automatically calculated FIB-4 followed by ELF as a second-line test[28]. Together, these studies suggest that the most clinically useful serum panels are those that fit into triage algorithms, not just those with strong cross-sectional area under the receiver operating characteristic curve values. In parallel, metabolic indices based on the triglyceride-glucose index and its derivatives have emerged as practical non-invasive tools for MASLD detection, risk stratification, and severity assessment, with additional value for predicting disease progression, particularly in older adults[29]. Similarly, machine learning models integrating simple anthropometric and routine biochemical variables have been developed as effective non-invasive tools for predicting MASLD, with potential utility for large-scale population screening[30].

A second important line of development is the use of mechanism-based fibrosis biomarkers. Algorithm for Disease Assessment in Progressive Fibrosis Testing (ADAPT), which incorporates PRO-C3 together with age, diabetes status, and platelet count, improved identification of advanced fibrosis in MASLD and helped demonstrate the value of extracellular matrix remodeling markers[31]. Subsequent comparison studies showed that ADAPT outperformed FIB-4 and APRI for detecting advanced fibrosis and MASH in the CENTAUR screening population[32]. Beyond ADAPT itself, a recent meta-analysis found that PRO-C3 has clinically meaningful diagnostic accuracy as a blood-based fibrosis biomarker across chronic liver diseases, including fatty liver disease[33]. These data strengthen the idea that fibrosis biomarkers reflecting active collagen turnover, rather than only indirect liver injury, may have particular value in treatment trials and longitudinal risk stratification.

Other circulating markers remain relevant, especially where a review section needs to acknowledge the broader biomarker landscape. A classic study showed that a two-step approach using cytokeratin-18 (CK-18) followed by fibroblast growth factor 21 improved non-invasive diagnosis of MASH[34]. More recent work has also shown that CK-18 fragment levels, especially when combined with FIB-4, can help identify steatohepatitis among patients with biopsy-confirmed MASLD[35]. In Asian cohorts, WFA-positive Mac-2-binding protein has shown diagnostic value for fibrosis severity in MASH and MASLD[36]. Although none of these markers alone has replaced broader composite panels, they remain important because they capture distinct biological processes and may still be useful in multimarker or disease-enrichment settings.

Newer diagnostic directions also include digital pathology, microbiome-based models, and AI. Quantitative fibrosis, inflammation, ballooning, steatosis demonstrated that machine learning-assisted digital pathology can quantify fibrosis, inflammation, ballooning, and steatosis on biopsy with improved reproducibility[37]. At the non-invasive end of the spectrum, a gut microbiome-based metagenomic signature has shown that fecal microbial profiles can help detect advanced fibrosis in biopsy-proven MASLD[38]. Gut microbiome dysbiosis, characterized by increased Bacteroides and decreased Bifidobacterium, together with elevated transforming growth factor-β levels, has also been proposed as a potential non-invasive diagnostic approach for MASLD, further emphasizing the role of the gut-liver axis[39]. Ultrasound AI is also advancing quickly; a recent study found that deep learning could grade hepatic steatosis on ultrasound with performance comparable to, and in some contexts better aligned than, routine reader assessment[40]. These approaches are not yet replacements for established care pathways, but they illustrate where the field is heading: Toward multimodal, data-integrated diagnostics that combine biologic specificity with scalable deployment.

In summary, the most important development in MASLD diagnostics is the movement from isolated tests to integrated, layered pathways. MRI-based scores, including MAST and MEFIB offer high diagnostic precision in specialist settings, whereas FibroScan-based tools, including FAST and Agile are better suited to broad clinical deployment]. Quantitative ultrasound continues to improve the bedside assessment of steatosis and fibrosis. Serum tests, including NIS4, NIS2⁺, ELF, ADAPT, and PRO-C3-based approaches are making blood-based triage increasingly practical. Finally, AI, digital pathology, and microbiome-informed methods are expanding the field beyond conventional imaging and biochemistry. In routine practice, liver biopsy will remain important for selected cases, but the center of gravity has clearly shifted toward non-invasive, sequential, and risk-based diagnosis.

POTENTIAL BIOLOGICAL MECHANISMS UNDERLYING EVS ALTERATIONS IN MASLD

Understanding the biological mechanisms by which MASLD progression modulates EV characteristics is crucial for establishing EVs as clinically relevant biomarkers. In MASLD pathogenesis, hepatocyte lipid accumulation triggers endoplasmic reticulum stress and mitochondrial dysfunction, stimulating the release of pro-inflammatory EVs[41], which are key drivers of altered EVs biogenesis. Recent studies have demonstrated that lipotoxicity-induced cellular stress in MASLD activates multiple EVs biogenesis pathways, including the endosomal sorting complexes required for transport (ESCRT)[42] machinery and ceramide-dependent pathways, leading to increased multivesicular body formation and subsequent EVs release[41].

The observed increase in EVs concentration with steatosis severity may reflect enhanced hepatocyte-derived EVs secretion as a cellular response to lipid overload and oxidative stress. Hepatocytes under metabolic stress release EVs enriched in damage-associated molecular patterns, inflammatory cytokines, and lipotoxic lipid species, which can activate Kupffer cells and hepatic stellate cells via signaling pathways mediated by miRNA[43], perpetuating inflammation and fibrogenesis[44]. Furthermore, the shift in EVs size distribution observed across MASLD stages may be attributed to differential activation of EVs biogenesis pathways: Smaller EVs (exosomes, 30-150 nm) are predominantly generated through the ESCRT-dependent pathway, while larger EVs (microvesicles, 100-1000 nm) are produced via direct plasma membrane budding, a process enhanced during cellular activation and apoptosis[42].

Recent evidence suggests that lean MASLD patients exhibit distinct metabolic and inflammatory profiles compared to obese MASLD patients, characterized by increased genetic susceptibility, enhanced insulin resistance despite normal body mass index (BMI), and more pronounced mitochondrial dysfunction[45]. These pathophysiological differences may be reflected in unique EV signatures. Lean MASLD patients may release EVs with distinct compositions, including specific microRNA profiles (especially miR-122)[46] and altered lipid compositions that reflect their unique metabolic dysregulation[47]. The size distribution of EVs in lean MASLD may also differ, potentially showing a higher proportion of smaller exosomes due to enhanced ESCRT-dependent secretion driven by genetic and metabolic factors independent of obesity-related inflammation.

The metabolic dysregulation in MASLD, particularly insulin resistance and dyslipidemia, directly impacts EV biogenesis and composition. Insulin resistance enhances hepatocyte EV secretion through activation of specific signaling pathways, while simultaneously altering EVs to include pro-inflammatory and pro-fibrotic mediators[43]. Dyslipidemia, especially elevated free fatty acids and ceramides, modulates membrane lipid composition, affecting EV budding efficiency and size distribution[48]. These mechanistic insights support the rationale for using EV physical characteristics as biomarkers and highlight the need for future studies to investigate EV to enhance diagnostic specificity and provide insights into disease pathogenesis.

Collectively, these findings suggest that EV alterations are not merely epiphenomena but reflect key pathophysiological processes in MASLD progression.

ADVANCES IN EV-BASED DIAGNOSTIC MODELS FOR MASLD

Recent studies have applied machine learning models to improve diagnostic accuracy in MASLD by integrating clinical and biomarker data[3]. Kornek et al[49] confirmed that circulating EVs correlate with disease severity and inflammatory activity in liver diseases. Trifylli et al[5] conducted a study with 76 patients from an initial 798 individuals with metabolic dysfunction. The study used nanoparticle tracking analysis to assess plasma EVs characteristics and the ultrasound attenuation parameter from transient elastography for steatosis staging (S0-S3). Twenty machine learning models were developed using the CatBoost algorithm, focusing on two cases: Case 1 distinguished no steatosis (S0) from steatosis (S1-S3), and Case 2 identified severe steatosis (S3). The models incorporated EV features, anthropometric data, and clinical features (diabetes and fibrosis), with performance assessed via cross-validation and feature importance analyzed using SHapley Additive exPlanations (SHAP). In Case 1, the CatBoost C1a model achieved a receiver operating characteristic area under the curve (ROC-AUC) of 0.71/0.86 (training/test sets) using only EV features. In Case 2, the CatBoost C2h-21 model, which combined multidimensional features, achieved an ROC-AUC of 0.81/1.00 with a sensitivity of 0.92 and specificity of 0.87. These findings suggest that physical characteristics of circulating EVs can serve as novel non-invasive biomarkers for MASLD.

This study is among the few that integrate the ultrasound attenuation parameter, machine learning, and explainable AI, thereby surpassing the constraints of traditional linear statistical methods by uncovering non-linear relationships among MASLD disease features. In the early diagnosis of MASLD, when hepatic steatosis occurs (stages S1-S3), lipid infiltration induces enhanced intercellular communication, and stress-induced cells release an increased number of EVs that convey pathological information. This also results in a regular increase in EV release volume with disease progression, while the size distribution shifts alongside the degree of inflammation and fibrosis. Based on this principle, the researchers employed the CatBoost algorithm and SHAP analysis, comprehensively incorporating EV mean size, concentrations of different size subpopulations, clinical features, and anthropometric data as inclusion factors. Through multi-threaded analysis, these factors were integrated into the CB-C2h-21 model, which significantly improved the diagnostic performance for severe steatosis (S3). The model achieved an area under the curve (AUC) of 0.81 in the training set and 1.00 in the test set, demonstrating substantial translational application potential. The CB-C2h-21 model, which incorporates EVs, clinical, and anthropometric data, demonstrates a significant improvement in diagnostic efficacy for severe steatosis (S3), with an AUC of 0.81 for training and 1.00 for testing, indicating substantial translational potential.

STRENGTHS AND LIMITATIONS

Trifylli et al[5] exhibited methodological rigor by implementing a stringent standardized preanalytical protocol, including fixed blood collection timing, consistent centrifugation parameters, and controlled freeze-thaw cycles, which minimized variability in EVs analysis. The assessment of interday variability on different dates further ensured data reliability. The application of ten random five-fold cross-validations (C1) and three-fold cross-validations (C2) effectively reduced the impact of splitting randomness in small datasets. The CB-C2h-21 model achieved a perfect ROC-AUC of 1.00 on the test set, while maintaining a low standard deviation in cross-validation, indicating excellent generalizability. Furthermore, the study demonstrated that EV features alone could effectively differentiate between S0 and S1-S3 stages, with a ROC-AUC of 0.71. This performance is comparable to that of currently utilized non-invasive tests, including the controlled attenuation parameter (ROC-AUC of 0.70-0.75)[50] and Fibrosis-4 score (ROC-AUC of 0.70-0.80)[51], while advanced imaging techniques, including MRI-PDFF offer higher accuracy but remain costly and less accessible in routine practice. These comparative results indicate the potential clinical utility of EVs as independent biomarkers.

Despite the methodological rigor, the high attrition rate of the study warrants attention. Although 798 patients were initially enrolled, only 76 completed all procedures, resulting in a substantial dropout rate of 90.5%, which may introduce selection bias. The limited sample size, predominantly comprising overweight patients without a healthy control group, may constrain the effectiveness of the model in evaluating lean patients with MASLD, thereby limiting statistical power, particularly in the S3 group and the advanced fibrosis group. The absence of a healthy control group is particularly concerning for the Case 1 model, which aims to distinguish S0 from S1-S3 stages. Without healthy controls, the model’s ability to differentiate true absence of steatosis from early-stage disease may be compromised, potentially affecting its specificity and clinical utility in screening asymptomatic populations. This limitation underscores the necessity for future studies to include well-characterized healthy controls matched for age, sex, and metabolic parameters to establish robust baseline EV profiles and improve the discriminatory power of EV-based diagnostic models[52]. The authors acknowledge this limitation and emphasize the necessity for large-scale multicenter validation.

While the CB-C2h-21 model achieved a perfect ROC-AUC of 1.00 on the test set for S3 identification, this result warrants cautious interpretation. A perfect AUC, particularly in a small test cohort (n = 21 for S3), may indicate potential overfitting despite cross-validation efforts. The substantial difference between training AUC (0.81) and test AUC (1.00) suggests that the model may have captured dataset-specific patterns rather than generalizable biological signals. This concern is amplified by the limited sample size and the high dimensionality of input features. Independent external validation in larger, diverse cohorts is essential to confirm whether this exceptional performance reflects true biological discriminatory capacity or statistical artifact[53].

With respect to EVs characterization standards, the study focused exclusively on the physical features of EVs without validating EVs markers (CD9, CD63, CD81, TSG101) as recommended by the minimal information for studies of EVs 2023 guidelines[54], and lacked assessment of EVs purity and content. Recent research indicates that specific EV-carried microRNAs, including miR-122-5p, constitute approximately 72% of circulating extracellular RNA in patients with MASLD and are correlated with the severity of steatosis[46,52]. The integration of EVs proteomic and transcriptomic data holds the potential to substantially enhance diagnostic accuracy.

FUTURE RESEARCH DIRECTIONS

It is noteworthy that, given the prevalence of obesity among MASLD patients, BMI was utilized as a critical reference indicator during subject enrollment, with a significant proportion of participants classified as overweight or obese. Patient weight emerged as a significant predictor within the model.

However, lean MASLD, characterized by greater insidiousness and poorer prognosis[47], may affect the weight assigned to BMI in this evaluation criterion. We propose that subsequent studies should expand the sample scope, incorporate multicenter, multi-ethnic, and multi-country and regional comparisons, and adopt rigorous statistical evaluation[55]. Critically, future validation cohorts should be stratified by BMI categories, particularly distinguishing lean MASLD patients (BMI < 23 kg/m² in Asian populations or < 25 kg/m² in Western populations) from overweight and obese patients[56]. This stratification is essential because lean MASLD patients may exhibit distinct EVs profiles driven by different pathogenic mechanisms, including genetic predisposition, visceral adiposity despite normal BMI, and unique metabolic signatures[45,57]. Establishing BMI-specific diagnostic thresholds and potentially developing separate predictive models for lean vs non-lean MASLD populations would enhance the clinical applicability and accuracy of EV-based diagnostics across diverse patient phenotypes. We also anticipate the performance in evaluating lean MASLD in future research.

Beyond EVs physical characteristics, integrating EVs lipidomics may provide additional diagnostic value. EVs carry distinct lipid signatures that reflect the metabolic state of their cells of origin. In MASLD, EVs lipid composition is altered, with increased ceramides, sphingomyelins, and oxidized lipids that correlate with disease severity[20]. Combining EVs lipidomic profiles with size and concentration measurements could enhance the discriminatory power of diagnostic models and provide mechanistic insights into lipotoxicity-driven disease progression.

To mitigate potential EVs contamination, methods including size-exclusion chromatography or density gradient ultracentrifugation have demonstrated superiority in EVs separation and the reduction of lipoprotein contamination[58,59]. Transmission electron microscopy could be utilized for the observation of EVs morphology, validation of EVs markers, and nano-flow cytometry for EVs subpopulation analysis[60], enabling further molecular characterization of EVs for subsequent mechanistic investigations. Adherence to MISEV2023 guidelines is paramount for ensuring reproducibility and comparability across studies. Future investigations should incorporate comprehensive EVs characterization, including validation of classical EV markers (CD9, CD63, CD81, TSG101), assessment of EVs purity through negative markers (apolipoprotein B, albumin), quantification of particle-to-protein ratios, and morphological confirmation via electron microscopy[54]. Rigorous characterization will not only validate the biological relevance of observed EVs alterations but also facilitate standardization of EV-based diagnostics for clinical translation.

CONCLUSION

The study conducted by Trifylli et al[5] offers novel insights into the non-invasive staging of MASLD, illustrating that the integration of the physical characteristics of EVs with explainable AI has the potential to capture the non-linear biological information associated with the disease. Nonetheless, several challenges remain, including a limited sample size, a high attrition rate, variability in reference standards, insufficient molecular characterization of EVs, and a lack of external validation. These issues necessitate resolution through large-scale, multicenter cohorts with histological confirmation. Future prospective studies are warranted to validate the generalizability of these models and to clarify the causal relationship between EVs and MASLD pathogenesis.