Machine learning fibrosis score for pediatric metabolic dysfunction-associated steatotic liver disease: Promising but premature

Abstract

Metabolic dysfunction-associated steatotic liver disease (MASLD) is now the leading cause of chronic liver disease in children, affecting up to 38% with obesity of children. With the global shift from non-alcoholic fatty liver disease (NAFLD) to MASLD using affirmative criteria (hepatic steatosis plus ≥ 1 cardiometabolic risk factor) and approximately 99% concordance in pediatrics, the development of non-invasive fibrosis tools is accelerating. Yao et al report a machine-learning “chronic MASLD with fibrosis (CH-MASLD-Fib)” score for advanced fibrosis with area under the receiver operating characteristic curve (AUROC) of 0.92. While timely, we urge caution. First, high accuracy from a single-center cohort signals overfitting: Complex models can learn cohort-specific noise and fail to generalize. Consistent with this, established pediatric scores (NAFLD fibrosis score, fibrosis-4, pediatric NAFLD fibrosis score) perform modestly (AUROC: Approximately 0.6-0.7), and aspartate aminotransferase-to-platelet ratio index is variable, raising concern that CH-MASLD-Fib’s result reflects a statistical artifact. Second, MASLD epidemiology varies by ethnicity (highest in Hispanic, lower in Black children); a model derived in a mono-ethnic Chinese cohort may misclassify other populations. Third, clinical utility and cost-effectiveness are unproven; dependence on specialized assays (e.g., bile acids, cholinesterase) would limit access and increase cost. We recommend external validation in multi-ethnic cohorts, head-to-head comparisons with simple serum indices and elastography, and formal economic analyses. Until then, clinical judgment anchored in readily available markers and judicious, targeted liver biopsy remains paramount.

Key Words: Liver fibrosis; Machine learning; Non-invasive biomarkers; Overfitting; Ethnic diversity; Cost-effectiveness; External validation; Health economics; Metabolic dysfunction-associated steatotic liver disease

Core Tip: Machine-learning models for detecting advanced fibrosis in pediatric metabolic dysfunction-associated steatotic liver disease routinely report striking accuracy, yet three recurring flaws limit clinical impact: Overfitting to single-center datasets, absence of multi-ethnic external validation, and dependence on costly, non-routine biomarkers. We outline a pragmatic roadmap: Prospective, multi-site cohorts with standardized liver histology, decision-curve and cost-utility analyses, and transparent model explainability to transform promising algorithms into trustworthy tools. Until these steps are fulfilled, the most reliable strategy combines simple serum tests (alanine aminotransferase, platelet-based indices), vibration-controlled transient elastography, and judicious, targeted liver biopsy for indeterminate or high-risk cases.

INTRODUCTION

We read with great interest the recent article by Yao et al[1] describing a machine learning-based fibrosis scoring system, the “chronic metabolic dysfunction-associated steatotic liver disease with fibrosis (CH-MASLD-Fib)” score, for non-invasive detection of advanced liver fibrosis in children with MASLD. The authors are to be commended for tackling one of the most significant challenges in modern pediatric hepatology. The escalating prevalence of pediatric MASLD, now the most common cause of chronic liver disease in youth, affects at least 10% of children in the United States and up to 38% of those with obesity[2]. Given the potential for progression to cirrhosis, end-stage liver disease, and even hepatocellular carcinoma, the development of accurate, non-invasive tools for risk stratification is a clinical and public health imperative[3].

This work is particularly timely, arriving on the heels of the global consensus decision to transition the nomenclature from non-alcoholic fatty liver disease (NAFLD) to MASLD[4]. The name change was not superficial but a fundamental paradigm shift. It moves the field away from a diagnosis of exclusion defined by what it is not (“non-alcoholic”) to a positive, affirmative diagnosis based on the presence of hepatic steatosis plus at least one of five cardiometabolic risk factors[4]. This new framework rightly centers the disease’s pathophysiology on metabolic dysfunction, aims to reduce stigma, and enhances disease awareness. Fortunately, research has demonstrated a very high concordance between pediatric cohorts defined by legacy NAFLD criteria and the new MASLD definition[5], ensuring the continued relevance of historical data. However, the MASLD definition also reframes the research question for diagnostic models. A tool for MASLD is now applied to a population already identified as metabolically at-risk by definition, which may alter the predictive value of certain biomarkers and introduces a potential for circular reasoning if a model’s predictive power relies heavily on the very metabolic factors used for diagnosis.

While we applaud the innovative approach of Yao et al[1] and acknowledge the striking performance of their score [reported area under the receiver operating characteristic curve (AUROC): Approximately 0.92], we urge profound caution. Before this or any similar model can be considered for clinical adoption, its findings must be scrutinized through the lens of three critical challenges paramount in diagnostic model development: The high probability of statistical overfitting, the unproven generalizability across diverse populations, and the uncertain clinical and economic utility compared to established care pathways.

THE SPECTER OF OVERFITTING IN DIAGNOSTIC MODELS

In medical machine learning, an exceptionally high accuracy reported from a single developmental dataset is a classic hallmark of overfitting[6]. Overfitting occurs when a complex algorithm learns the training data too intimately, memorizing not only true underlying signals but also random noise and idiosyncrasies unique to that specific cohort. The consequence is a model that performs brilliantly in the laboratory setting but fails to generalize to new, unseen patients in the real world, leading to potentially harmful misclassifications and misguided clinical decisions.

The optimism for the CH-MASLD-Fib score must therefore be tempered by the sobering performance of existing non-invasive fibrosis scores in pediatric populations, as summarized in Table 1[7-10]. Scores developed for adults have consistently proven inadequate in children. The NAFLD fibrosis score (NFS), which incorporates age, body mass index, hyperglycemia, aspartate aminotransferase (AST)/alanine aminotransferase (ALT) ratio, platelets, and albumin, performs poorly in pediatric cohorts[7]; one large study even found it unable to significantly distinguish fibrosis stages (P = 0.14)[11]. Similarly, the fibrosis-4 (FIB-4) index (age, AST, ALT, platelets) has repeatedly demonstrated poor diagnostic accuracy in children, with reported AUROCs as low as approximately 0.32 for advanced fibrosis[7]. Even scores developed specifically for children have struggled. For example, the pediatric NFS has achieved only modest accuracy (AUROC: Approximately 0.74)[7]. This trend of modest performance holds true for more recent modeling efforts as well; for instance, a newly developed pediatric-specific model using variables including ethnicity, insulin, platelet count, and AST to predict significant liver stiffness (> 2.71 kPa, as measured by magnetic resonance elastography) achieved a similarly limited AUC of 0.70[12]. Critically, in the validation phase of that study, the model’s predictions showed a very weak correlation with measured liver stiffness (correlation coefficient of just 0.30), highlighting the challenge of accurately predicting fibrosis from lab data alone. The AST-to-platelet ratio index (APRI) has also shown highly variable performance fair accuracy for detecting any fibrosis, but poor utility for identifying the clinically significant advanced stages that demand intervention[7,11].

Table 1 Comparative performance and characteristics of non-invasive fibrosis scores in pediatric non-alcoholic fatty liver disease/metabolic dysfunction-associated steatotic liver disease.

	Key components	Target population	Reported AUROC for advanced fibrosis (F ≥ 3)	Ref.
CH-MASLD-Fib (machine learning model, hypothetical)	Multiple clinical and laboratory variables (machine learning-derived)	Pediatric (Chinese cohort)	Approximately 0.92 (derivation cohort)	Yao et al[1]
NAFLD fibrosis score	Age, BMI, impaired fasting glucose/diabetes, AST/ALT ratio, platelet count, albumin	Adult (developed in adults)	No diagnostic ability (AUROC: Approximately 0.50; P = 0.14)	Angulo et al[8]
FIB-4 index	Age, AST, ALT, platelet count	Adult (developed in adults)	Poor (AUROC range: Approximately 0.32-0.54)	Shah et al[9]
AST to platelet ratio index	AST level, platelet count	Adult (developed in adults)	Fair for any fibrosis (AUROC range: Approximately 0.70-0.80); poor for advanced fibrosis (AUROC range: Approximately 0.50-0.60)	Chrysanthos et al[10]
Pediatric NAFLD fibrosis score	ALT, alkaline phosphatase, platelet count, GGT	Pediatric (developed in children)	0.74 (95%CI: 0.66-0.82)	Alkhouri et al[7]

AUROC: Area under the receiver operating characteristic curve; CH-MASLD-Fib: Chronic metabolic dysfunction-associated steatotic liver disease with fibrosis; F: Fibrosis; NAFLD: Non-alcoholic fatty liver disease; FIB-4: Fibrosis-4; AST: Aspartate aminotransferase; BMI: Body mass index; ALT: Alanine aminotransferase; GGT: Gamma-glutamyl transpeptidase; CI: Confidence interval.

The vast discrepancy between the modest, real-world performance of these widely tested simpler scores and the near-perfect accuracy reported for the complex CH-MASLD-Fib model suggests a methodological artifact rather than a true breakthrough. It is highly probable that the model is over-parameterized for its development cohort size, a scenario that dramatically increases the risk of fitting to noise under the bias-variance tradeoff principle. This is the medical equivalent of the famous “husky in the snow” problem, where an algorithm learned to classify images of wolves vs dogs based on the presence of snow in the background instead of the animal itself[13]. The CH-MASLD-Fib model, trained on a single-center Chinese cohort, may have inadvertently learned spurious correlations unique to that hospital’s patient demographics, referral patterns, or even specific laboratory assay characteristics. Without rigorous external validation on a completely independent patient dataset, the impressive internal AUROC of 0.92 is likely an overestimation of real-world performance. In other words, the burden of proof for such a high-performing model must shift from demonstrating internal consistency to proving robust external validity in diverse settings.

THE IMPERATIVE OF GENERALIZABILITY ACROSS POPULATIONS

A second critical concern in proposing a score from a homogenous cohort is the profound heterogeneity of MASLD across different populations. This disease is far from monolithic; its prevalence, severity, and even dominant histologic patterns vary significantly by race and ethnicity, driven by a complex interplay of genetic predispositions, environmental exposures, and socioeconomic factors.

In the United States, data consistently show that children of Hispanic descent have the highest prevalence of MASLD and are at greater risk of progressing to its inflammatory form (metabolic dysfunction-associated steatohepatitis), compared to non-Hispanic white children[14]. Conversely, non-Hispanic Black children appear to have a lower prevalence of the disease[14]. Asian children are also recognized as a high-risk group, often developing MASLD at lower body mass index thresholds[15]. These disparities are not random; they are partly linked to underlying genetic factors such as population-specific frequencies of the PNPLA3 gene polymorphism, which strongly predisposes individuals to hepatic fat accumulation[16].

Therefore, a model developed exclusively in a cohort of Chinese children is not merely unvalidated for other groups it is biologically and demographically primed to fail elsewhere. An algorithm trained in one ethnic population will tend to weight predictive features according to that group’s unique genetic architecture and clinical context. Applying this model to, say, a Hispanic child in Los Angeles or a Black adolescent in Chicago would represent a dangerous extrapolation beyond its training data, risking systematic misclassification and potentially exacerbating healthcare disparities. The recent American Association for the Study of Liver Diseases (AASLD) pediatric practice guidance and other consensus statements have emphasized the need to understand and address such population differences in MASLD[17]. Consequently, any diagnostic tool intended for broad clinical use must demonstrate consistent, reliable performance across multiple, ethnically diverse pediatric populations as an absolute prerequisite for consideration.

Notably, the very fact that an ethnicity-specific model was developed highlights the possibility that a single universal fibrosis score for all children may be unattainable. The future of non-invasive diagnostics in pediatric MASLD might lie in population-adjusted or even ethnicity-specific models. However, pursuing that path presents immense logistical, financial, and ethical challenges. It would require assembling large international consortia with sufficient biopsy-proven data from every major demographic group. It also raises difficult questions: How would we validate and implement different cut-offs or algorithms for children of mixed heritage or in multi-ethnic societies? Yao et al’s work, by its nature, forces the MASLD research community to confront the true complexity of developing globally applicable diagnostics[1].

One potential avenue to improve the model’s generalizability is through advanced machine learning strategies. For example, transfer learning or federated learning could be employed to adapt the CH-MASLD-Fib model using multi-ethnic datasets, mitigating the need to build separate models from scratch for each population. In addition, alternative technological pathways merit exploration: Multimodal imaging “radiomics” features [e.g., from ultrasound or magnetic resonance imaging (MRI)] could be integrated with clinical and laboratory data to capture fibrosis in a more universal manner, reducing reliance on cohort-specific correlations. To address the ethical and practical challenges of “population-specific” models particularly for children of mixed ancestry a pragmatic approach might include incorporating genetic risk markers (such as PNPLA3 or TM6SF2 polymorphisms) into the assessment algorithm. In cases where a patient’s risk stratification remains ambiguous, implementing targeted genetic testing and convening a multidisciplinary team (including genetic counselors and hepatologists) could guide personalized management decisions.

CLINICAL UTILITY AND COST-EFFECTIVENESS: THE FINAL HURDLES

Beyond statistical accuracy and generalizability, a diagnostic score’s ultimate value depends on its feasibility, utility, and cost-effectiveness in real-world practice. The practical utility of the CH-MASLD-Fib score will hinge on which variables it uses. If the score relies only on simple, inexpensive, universally available lab tests, like those in NFS or FIB-4 its potential for widespread adoption is high, provided its accuracy is validated externally[3]. However, if the machine learning algorithm identified more esoteric or costly biomarkers as key predictors such as total bile acids (which can cost approximately 70-80 US dollars for a basic serum test and over 300 US dollars for a fractionated panel) or serum cholinesterase activity its clinical utility would plummet. Such tests are not part of routine pediatric work-ups and are not available in all laboratories, creating major barriers to access (especially in primary care settings) and adding expense[18-20].

Furthermore, any new tool must demonstrate its value within, or superiority to, existing clinical pathways. For example, the North America Society for Pediatric Gastroenterology, Hepatology and Nutrition currently recommends screening at-risk children (those with obesity, starting around age 9-11) using ALT levels[21]. The 2024 AASLD pediatric practice statement similarly outlines a tiered evaluation framework that acknowledges the limitations of current non-invasive tools and reserves liver biopsy as the gold standard for staging[17]. Of course, liver biopsy is an invasive procedure with a non-negligible risk of complications, so it is reserved as a last resort despite its diagnostic value. In practice, the typical care pathway for suspected pediatric MASLD is stepwise: An elevated ALT triggers liver ultrasound or elastography, followed by referral to a specialist for further management. Any complex new scoring system must prove that it can improve on this pathway by identifying high-risk patients more accurately or earlier without adding prohibitive cost or complexity. If implementation of the score requires expensive send-out assays or sophisticated computing infrastructure, it may simply be impractical outside major tertiary centers.

This leads to the unsettled economics of non-invasive testing in MASLD. Unlike other chronic liver diseases such as hepatitis C, where non-invasive fibrosis tests have been proven as cost-effective alternatives to biopsy and upfront treatment is available[22], the economic case for routine fibrosis testing in pediatric MASLD remains weak. Modeling studies for NAFLD/MASLD have yielded inconclusive or highly uncertain results, with cost-effectiveness often hinging on optimistic assumptions about test costs and, critically, whether a diagnosis leads to interventions that improve long-term outcomes[23]. This uncertainty is compounded by the fact that there are currently no Food and Drug Administration (FDA)-approved pharmacotherapies for pediatric MASLD. The cornerstone of management remains lifestyle modification an intervention with notoriously variable adherence and success. Introducing a new, potentially expensive diagnostic test that ultimately results in the same recommendation (“eat a healthier diet and exercise more”) is unlikely to be cost-effective[14,18]. In essence, the clinical and economic value of advanced diagnostics for MASLD is tightly linked to the therapeutic landscape. As effective treatments for pediatric metabolic dysfunction-associated steatohepatitis are developed and approved, the need to accurately identify high-risk patients will become critical; until then, the utility of such tests remains limited.

Finally, we note that machine learning’s potential is rapidly evolving, and it need not be considered in isolation from other diagnostic technologies. For example, advanced imaging techniques like multiparametric MRI have emerged as powerful non-invasive tools for liver fibrosis assessment. The corrected T1 mapping metric now FDA-cleared for clinical use has shown in a recent multinational trial that adding MRI-based tissue characterization to MASLD management can avoid roughly 50% of unnecessary biopsies while improving diagnostic rates, all in a cost-effective manner[24]. Such imaging biomarkers provide a complementary approach to fibrosis detection alongside serum-based models. Looking ahead, it is conceivable that future artificial intelligence (AI) algorithms will integrate these radiologic markers (through imaging-based data, i.e. radiomics) with traditional clinical and laboratory inputs, creating a multimodal diagnostic paradigm. Far from being static, machine learning models can undergo rapid iterative improvement; thus, initial limitations might be overcome through continual refinement and incorporation of new data types. By acknowledging these parallel advances in imaging and AI, we present a more balanced perspective on the evolving landscape of non-invasive fibrosis assessment.

CONCLUSION

The CH-MASLD-Fib score proposed by Yao et al[1] represents an intriguing application of machine learning to a pressing clinical problem. However, the journey from a promising algorithm to a trustworthy clinical tool is long, and it requires navigating significant pitfalls. Our analysis highlights three major reasons for skepticism at this early stage: The high probability of statistical overfitting, the unaddressed challenge of ethnic heterogeneity (making broad generalizability doubtful), and an undefined clinical and economic utility. We strongly caution against any premature enthusiasm for, or adoption of, this score in routine practice. Instead, we advocate for a clear and rigorous validation pathway that must be completed before this tool can be considered reliable for clinical use: (1) Independent external validation: The model should be tested on large, independent cohorts from multiple centers (outside the original institution) to check for overfitting and confirm its accuracy; (2) Multi-ethnic cohort studies: Prospective validation in ethnically diverse pediatric populations (including Hispanic, non-Hispanic white, Black, and other Asian groups) is essential to establish generalizability; (3) Head-to-head comparison: The score’s performance should be directly compared, in parallel, with existing simple serum scores (APRI, FIB-4) and non-invasive imaging modalities (e.g., transient elastography) in the pediatric setting; and (4) Cost-effectiveness analysis: A formal economic evaluation should determine if using the score provides value over the current standard of care, in terms of improved outcomes per cost. This should consider different healthcare settings and the evolving availability of effective therapies. The quest for a non-invasive “holy grail” of fibrosis assessment in pediatric MASLD is undeniably important. Sophisticated models like that of Yao et al[1] are valuable scientific contributions and may indeed form part of the solution. Yet we must not lose sight of the ultimate benchmark: Proven clinical value. A truly useful test must not only be accurate, but also generalizable, accessible, and cost-effective. Until the CH-MASLD-Fib score has met this high bar through robust, multi-faceted validation, clinical judgment guided by established simple non-invasive tests and, when necessary, confirmatory liver biopsy will remain our most trusted approach for managing children with MASLD.