This study protocol was registered with the PROSPERO International Prospective Register of Systematic Reviews (Registration: CRD420251046862). The research was conducted as per the PRISMA-DTA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses–Diagnostic Test Accuracy) guidelines (Checklist 1) [26].

This systematic search was independently designed and conducted by two researchers as per the PRISMA-DTA (Checklist 1) [26]. PubMed, the Cochrane Library, Embase, Web of Science, and IEEE Xplore were retrieved until September 24, 2025. The search strategy was structured around three core concepts: (1) the disease (“HS,” “non-alcoholic fatty liver disease,” and “MAFLD”); (2) the technology (“AI,” “machine learning [ML],” “DL”); and (3) the diagnostic context (“diagnosis,” “detection”).

Keywords within each conceptual category were combined via the OR operator (eg, AI OR DL OR ML), whereas keywords across different categories were linked using the AND operator (eg, AI AND MAFLD AND diagnosis).

The authors of the identified studies were not contacted. Reference lists of all included studies were manually reviewed to identify any additional eligible publications. No restrictions on language or publication date were applied at the database level to maximize search sensitivity. However, non-English records were excluded during subsequent screening. Gray literature, preprints, and unpublished studies were not systematically searched. This decision was made a priori to focus on peer-reviewed, full-text articles that had undergone editorial review, thereby ensuring baseline methodological quality and the availability of sufficient details for data extraction. The complete, reproducible search strings for all databases are provided in Table S1 in Multimedia Appendix 1.

Two independent reviewers initially screened all retrieved titles and abstracts. After removing duplicate records, studies were deemed eligible for inclusion if they met the following criteria:

Exclusion criteria were: (1) language: non-English publications; (2) study type: letters, conference abstracts, reviews, or academic papers lacking original data; (3) study subjects: animal or nonhuman research, bioinformatics-based analyses, and predictive modeling studies focused on indices, risks, or associations rather than diagnosis; and (4) data sufficiency: studies without key contingency data or insufficient information to calculate diagnostic performance metrics.

Two researchers independently extracted data based on the following domains: (1) study characteristics: first author, publication year, site of data collection, and duration of the study period; (2) study population: total sample size, and demographic characteristics (mean or median age); (3) methodological parameters: accessibility of clinical sample data, diagnostic reference standard, and validation strategy; (4) algorithmic architecture: type of algorithm, classifier employed, and application of transfer learning (TL); and (5) diagnostic efficacy: raw contingency table data, and aggregated diagnostic performance metrics.

Pooled estimates of sensitivity, specificity, and AUC, together with their 95% CIs, were presented using forest plots. Heterogeneity was quantified using the I² statistic. The AUC was designated as the primary indicator for overall diagnostic accuracy, as it integrates performance across all thresholds and remains unaffected by any single cut-off point. A SROC curve was constructed following an assessment of the threshold effect using the Spearman correlation coefficient between the logit of sensitivity and the logit of (1-specificity). Heterogeneity and its implications were further visualized via 95% PIs and bivariate boxplots. Potential small-study effects were evaluated using the Deeks funnel plot asymmetry test. Additional diagnostic indicators, including the diagnostic odds ratio (DOR), positive likelihood ratio (LRP), and negative likelihood ratio (LRN), were calculated. Clinical applicability was further examined using a Fagan nomogram, while the distribution of likelihood ratios across studies was illustrated via scatterplots.

Two investigators assessed the risk of bias via the QUADAS-2 (Quality Assessment of Diagnostic Accuracy Studies 2) checklist (Checklist 2) in Rev-Man in terms of PIRT framework [27]. The QUADAS-2 tool, recommended by the Cochrane collaboration, was used to assess the methodological quality and risk of bias of included diagnostic accuracy studies.

All screening, data extraction, and quality assessment procedures were independently conducted by 2 reviewers. Any discrepancies were first resolved through discussion to reach a consensus. When consensus was not achieved, a third senior investigator adjudicated the disagreement to make the final determination. This multilevel process ensured that all extracted data represented unanimous agreement within the research team.

The following independent meta-analyses were conducted:

Given the substantial heterogeneity observed among included studies with respect to patient populations, imaging devices, and AI algorithms, a bivariate mixed effects model was used to derive more accurate and reliable pooled estimates [25]. To ensure the robustness of the meta-analytic results, quantitative synthesis (eg, subgroup analysis) was performed only when at least 3 independent studies, defined as studies conducted by different authors, using distinct experimental protocols, or involving separate participant cohorts, were available. Multiple effect estimates from the same publication were included when they originated from distinct participant cohorts (eg, multicenter datasets or independent validation sets). When multiple model outputs were reported, only the best-performing model or that validated using an independent dataset was retained. Subgroup analyses were not conducted when fewer than three independent studies were available for a given subgroup. All statistical analyses and visualizations were performed via Stata MP 18 (StataCorp LLC). A 2-tailed P value <.05 denoted statistical significance.

As of September 24, 2025, 2536 articles were retrieved. After removing 864 duplicates, the titles and abstracts of the rest were screened as per the predefined eligibility criteria, resulting in the exclusion of 1596 articles. Specifically, 9 were non-English publications, 884 were of other types, 673 involved inappropriate study subjects, and 30 used unsuitable research methods. The full texts of the remaining 76 articles were subsequently reviewed. Seventeen studies were excluded for incomplete data, 7 for being of other types, 7 for inappropriate methodologies, and 9 for being inaccessible. Ultimately, 36 studies were included in the final analysis (Figure 1). The characteristics of the included studies are summarized in Table 1, and the results of the subgroup analyses are presented in Table 2.

Of the 36 included studies, 33 (comprising 36 datasets) satisfied the criteria for subgroup analysis. The pooled results (Figure 2) demonstrated a summary sensitivity of 0.96 (95% CI 0.93‐0.97), a specificity of 0.93 (95% CI 0.91‐0.95), and an AUC of 0.98 (95% CI 0.96‐0.99), indicating excellent diagnostic discrimination by the AI models. Substantial heterogeneity was observed across studies (I²>75%). The Spearman correlation coefficient (0.21, P=.05) suggested that threshold effects contributed minimally to overall heterogeneity. The broad 95% PI, however, indicated that differences in diagnostic thresholds were a major source of variability. No significant small-study effects were identified (P=.65). In terms of clinical applicability, at a pretest probability of 50%, a positive AI result increased the posttest probability to 93%, whereas a negative result reduced it to 4%. Likelihood ratio scattergram analysis confirmed that the pooled estimates were located within the “confirm and exclude” quadrant (LRP >10 and LRN <0.1), underscoring the strong clinical value of AI for both confirming and excluding HS.

The QUADAS-2 quality assessment (Figure 3) revealed that 44% (16/36) of studies exhibited a high risk of bias in the patient selection domain, primarily due to selection bias, limited representativeness of study populations, and incomplete reporting of key clinical parameters. In the Index test domain, 6% (2/36) of studies were rated as high risk, largely attributable to the absence of image quality control, subjective elements during image processing, and nonstandardized training or validation procedures. In the reference standard domain, 17% (6/36) of studies demonstrated a high risk of bias, most commonly due to deviations from gold-standard reference methods, unclear blinding procedures, or incomplete pathological sampling information. The flow and timing domain exhibited unclear risk in 36% (14/36) of studies, often due to insufficient reporting on patient inclusion pathways and the interval between image acquisition and diagnostic confirmation. These methodological limitations may contribute to an overestimation of AI model performance in real-world clinical practice.

Subgroup analyses were conducted for 7 key variables (Figures S1-S16 in Multimedia Appendix 1), with diagnostic performance interpreted relative to clinical applicability thresholds (LRP >10 for strong rule-in capability; LRN <0.1 for strong rule-out capability).

This meta-analysis of 36 studies demonstrates the superior diagnostic performance of AI in identifying HS, yielding a pooled AUC of 0.98, surpassing that of conventional ultrasound, CT, and MRI, whose pooled AUCs were 0.93, 0.975, and 0.97, respectively [64-66]. These findings underscore AI’s potential to overcome the inherent physical constraints of individual imaging techniques, thereby establishing it as a versatile and adaptive diagnostic approach. This capability carries considerable clinical significance, providing a strong rationale for further meta-analyses dedicated to AI-based diagnostic technologies and informing the development of flexible, context-specific clinical applications. By optimizing the use of the most accessible and cost-effective diagnostic resources, AI could markedly broaden the availability of early HS screening across diverse health care settings.

Although the encouraging performance of AI in diagnosing HS is promising, interpretation of these findings must be tempered by a critical appraisal of the methodological limitations underlying this research. Our analysis revealed substantial heterogeneity (I²>75%) and a high overall risk of bias among included studies, particularly within the patient selection domain, where 44% (16/36) were judged to be at high risk. A major limitation lies in the predominance of retrospective, single-center designs (25/36, 69%). Such studies typically develop and validate models within controlled, idealized data environments, meaning that reported metrics may reflect “best-case scenarios” rather than true clinical performance across diverse devices, operators, and patient populations in routine practice. Moreover, independent external validation and multicenter prospective trials remain notably scarce, severely limiting the assessment of these models’ generalizability. Therefore, while existing evidence underscores AI’s considerable potential in HS diagnosis, the current body of research remains insufficient to justify its widespread clinical adoption. Bridging the translational gap from high-performing algorithms to reliable, universally applicable clinical tools thus remains a substantial challenge.

Subgroup analyses provide valuable insights for optimizing AI model design and informing clinical integration. DL-based models demonstrate exceptionally high specificity in diagnosing HS, offering a distinct clinical advantage by reducing unnecessary liver biopsies. However, these models require higher-quality data annotation and greater computational resources. Model performance was also closely associated with the reference standard and imaging modality used. Systems using histopathological images as input achieved the highest diagnostic accuracy. Nevertheless, their clinical applicability is restricted by procedural invasiveness and sampling error [67]. The AI-assisted whole-slide analysis model proposed by Roy et al [37] improves quantitative consistency but faces practical barriers related to cost and patient acceptance.

The choice of imaging modality inherently involves trade-offs between diagnostic accuracy, accessibility, and cost. MRI-PDFF, while providing precise, noninvasive quantification, is affected by confounders such as iron overload, edema, and concurrent pathologies [7], and its high cost limits use in primary care. Ultrasound remains the most accessible and economical option but suffers from operator dependency, reduced sensitivity for mild steatosis, limited penetration in obese individuals [68,69], and suboptimal accuracy in detecting fibrosis [70]. AI integration could mitigate these limitations by standardizing acquisition and interpretation, though at the cost of increased system complexity and computational demand. CT achieves a sensitivity and specificity of 0.93 but is constrained by ionizing radiation exposure and potential interference from iodine-based contrast agents [71]. The dual-energy CT 3D nnU-Net model developed by Yoo et al [54] achieved an AUC of 0.97 for distinguishing steatotic from normal tissue, yet its clinical application is constrained by limited equipment availability.

To optimize model performance and data use, TL has been widely adopted, an especially valuable strategy given the substantial costs associated with medical data annotation. Nonetheless, the effectiveness of TL depends critically on the degree of similarity between the source and target domains; substantial domain discrepancies may lead to “negative transfer,” as illustrated by sensitivity variations of up to 10% in the Inception-v3 model reported by Constantinescu et al [40]. To address this limitation, emerging approaches such as adversarial domain adaptation frameworks have achieved near-human classification accuracy on heterogeneous MRI datasets [72]. Similarly, hybrid pretraining strategies [73] and federated learning techniques have reached up to 99% of the performance attained through centralized training [74]. These approaches enhance model robustness while effectively addressing data privacy and heterogeneity.

Beyond algorithmic optimization, the real-world implementation of AI is profoundly influenced by study design and data governance. Retrospective studies, which constituted the majority (25/36, 69%) of the included reports, demonstrated significantly higher performance than prospective studies (AUC: 0.98 vs 0.94), likely reflecting the high-quality and well-curated imaging data typically available in retrospective cohorts. In contrast, prospective designs more faithfully capture real-world clinical workflows but are inherently subject to operational variability, such as inconsistent imaging protocols and unpredictable patient factors, thereby leading to attenuated performance.

Furthermore, data governance and accessibility are pivotal in determining model generalizability. Multicenter collaborations and data sharing can improve generalizability and reproducibility, though they require standardized imaging protocols, increased logistical coordination, and greater resource investment, posing feasibility challenges in resource-limited settings. Moreover, access to medical data for AI development remains hindered by privacy regulations, institutional policies, and technical interoperability barriers. Privacy-preserving strategies, such as federated learning, offer promising solutions by enabling multi-institutional collaboration without direct data exchange, albeit at the cost of increased computational demands and system complexity. It should also be noted that publicly available datasets may not fully represent the clinical heterogeneity encountered in real-world practice, thereby introducing potential selection bias. These factors, while critical for improving AI performance, also contribute substantially to heterogeneity, underscoring the necessity of comprehensive external validation and context-specific adaptation before large-scale clinical implementation.

The use of AI extends beyond diagnostic precision to encompass the comprehensive management of HS. Accumulating evidence indicates that AI not only enables accurate quantification of hepatic fat but also integrates radiomic, pathological, and clinical data to facilitate fibrosis staging, predict HCC risk, assess posttransplant survival, and stratify cardiovascular complications. For instance, a VGG16-based ultrasound model outperformed human interpretation in classifying borderline cases [75]. The integration of macrogenomic sequencing with ML has proven effective for the differential diagnosis of HS in obese pediatric populations [76]. Similarly, an ML model based on MRI-derived liver fat quantification markedly improved diagnostic accuracy for liver fibrosis [77]. AI-powered digital pathology platforms reduce the inherent subjectivity of conventional histological assessment [78], while DL-based radiomics facilitates the identification of critical pathological features such as microvascular invasion [79]. A DL algorithm demonstrated 99% accuracy in predicting postliver transplantation survival [80]. In the context of MAFLD-related complications, AI algorithms have been employed to accurately identify affected patients from electronic health records, revealing type 2 diabetes mellitus as a significant predictor of all-cause mortality (hazard ratio: 1.36) [81]. Moreover, a dual model combining tongue imaging with clinical indicators achieved precise prediction of coronary heart disease risk among patients with fatty liver [82]. The foregoing advances signal a diagnostic paradigm shift in HS management from a traditional “liver-centric” approach towards a “patient-centric” model of multi-system risk management, paving the way for early intervention and personalized therapy.

In summary, the advantages of AI in HS diagnosis are threefold as follows:

Despite its promising outlook, the widespread clinical adoption of AI in HS management faces multiple challenges. Technically, data heterogeneity, stemming from variations in imaging quality [84], scanner types, and reference standard thresholds, impedes the development of universally robust and generalizable models. Many high-performing algorithms are derived from single-center, retrospective datasets (eg, Yang et al [22], n=50, Beijing) with limited demographic diversity, thereby compromising their external validity and real-world applicability. Moreover, most existing models primarily focus on imaging biomarkers for fat quantification without adequately elucidating the complex pathophysiological interplay among steatosis, metabolic comorbidities, and fibrosis, limiting both clinical interpretability and holistic disease assessment.

From a clinical integration perspective, the transition from algorithmic development to real-world deployment necessitates careful consideration of workflow compatibility, device dependency, and cost-effectiveness. Lightweight AI models hold promise for incorporation into primary care ultrasound systems, facilitating large-scale population screening, whereas more advanced MRI- or CT-based models may be more appropriately implemented in tertiary medical centers. The overarching objective is seamless integration into existing clinical workflows, ensuring that AI serves as an assistive, rather than disruptive, technology that streamlines radiological practice, conserves clinician time, and enhances diagnostic efficiency [85]. Furthermore, issues concerning data privacy [86], algorithmic bias [87], and accountability [88] lack clear regulatory frameworks.

From a global health perspective, the clinical use of AI in HS diagnosis varies according to resource availability. To promote both efficiency and equity in HS diagnosis and management, a phased implementation framework is proposed:

Several limitations warrant cautious interpretation. First, considerable methodological and clinical heterogeneity was observed across the included studies, constraining the reliability of the conclusions. Despite extensive subgroup analyses, variability arising from differences in patient characteristics, imaging equipment, and diagnostic thresholds could not be fully addressed. This residual heterogeneity undermines the robustness of pooled estimates and suggests the influence of unmeasured factors affecting AI performance.

Second, the analysis was limited by methodological shortcomings inherent in the primary studies. Inadequate reporting of key patient characteristics hindered subgroup analyses by disease etiology, particularly distinguishing pure MAFLD from mixed forms, a critical gap given the potential impact of comorbidities on diagnostic accuracy. Furthermore, wide variation in AI architectures and the limited number of comparable models precluded meaningful comparisons across technical approaches, leaving the effect of architectural design on diagnostic performance unclear.

Third, the generalizability and real-world applicability of the findings remain limited. Most studies were retrospective, single-center designs prone to selection bias, with scarce external or temporal validation. Thus, the high-performance metrics reported may represent an idealized best-case scenario rather than outcomes achievable in prospective clinical settings.

Additionally, although our restriction to peer-reviewed full-text publications ensured a baseline level of methodological rigor, the exclusion of relevant preprints and gray literature may have introduced publication bias. Such selective inclusion likely favored studies reporting positive outcomes, potentially leading to overestimated performance measures. Moreover, key practical factors, such as computational burden, workflow integration, and technical expertise, could not be quantitatively evaluated, despite their importance for real-world implementation.

This meta-analysis highlights the substantial diagnostic potential of AI, particularly DL, in assessing HS. Its key contribution lies in establishing a unified, imaging-modality-independent analytical framework that provides comprehensive evidence beyond the constraints of individual imaging techniques. Nonetheless, these results reflect technical promise rather than confirmed clinical use. The translation from high-performing algorithms to reliable clinical tools remains hindered by performance heterogeneity, retrospective study designs, and insufficient external validation. While the technological foundation of AI in HS is encouraging, clinical maturity has yet to be achieved. Bridging this translational gap will require prospective multicenter studies, standardized reporting protocols, and rigorous external validation. Ultimately, successful clinical adoption will depend on demonstrating not only algorithmic robustness but also tangible improvements in patient outcomes and workflow efficiency across real-world health care settings.