Applications of machine learning in the diagnosis of non-alcoholic fatty liver disease: a systematic review and meta-analysis Authors: Luying Qi, Baiwang Li, Jingnan Xue, Xiangpeng Wu, Hupo Bian, Jieqiong Chen, Hongxing Zhao Published: 2026 ABSTRACT ObjectiveTo systematically evaluate and quantify the diagnostic accuracy and performance of machine ML techniques for the detection of NAFLD, and to compare the performance of ML when assisting different diagnostic modalities, in order to provide the promising strategy for non-invasive and accurate diagnosis of NAFLD.MethodsA literature search was performed in the PubMed, Web of Science, Cochrane Library, and EMBASE databases from the establishment of the databases to March 2025. The quality of studies and risk of bias were assessed via the QUADAS-2 and the PROBAST. Studies reporting AUC, sensitivity, and specificity were included. Data analysis and plotting were conducted using R (version 4.3.1), Stata (version 17.0), and Origin 2022. This study adhered to the PRISMA guidelines.ResultsA systematic review and bias assessment were conducted on 24 articles, and a meta-analysis was performed on 21 sets of data from 18 articles. The sensitivity, specificity, and AUC values for the internal validation dataset were 0.82 [95% CI 0.77—0.86], 0.81 [95% CI 0.77—0.85], and 0.88 [95% CI 0.85—0.91], respectively. For the external validation dataset, the corresponding values were 0.80 [95% CI 0.70—0.87], 0.72 [95% CI 0.56—0.84], and 0.83 [95% CI 0.80—0.86]. Heterogeneity was observed among the studies, and subgroup analysis indicated that study design, reference standard, outcome, cross-validation and model type may be potential sources of heterogeneity. ML models assisting imaging achieved higher diagnostic accuracy than those assisting clinical variables/routine serological markers.ConclusionML-assisted imaging achieved promising diagnostic accuracy for NAFLD detection, offering the potential to assist clinicians in making accurate diagnoses and improving clinical decision-making. However, the lack of external validation limited the applicability and generalizability of the model. It is necessary to conduct multi-center studies and external validation in the future.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12876-026-04906-2. Introduction Non-alcoholic fatty liver disease (NAFLD) is a major global cause of chronic liver disease, with an adult prevalence of 32%, and its incidence is projected to rise worldwide [1, 2]. The annual direct medical cost of NAFLD exceeds 100 billion dollars, posing a substantial burden on public health [3]. NAFLD encompasses a spectrum from non-alcoholic fatty liver (NAFL, simple steatosis without inflammation) to non-alcoholic steatohepatitis (NASH, coexistence of steatosis, inflammation, and hepatocyte ballooning) [4]. NASH is a risk factor for fibrosis, indicating a poor prognosis and potential progression to cirrhosis and hepatocellular carcinoma [3, 5]. Without early diagnosis and intervention, NAFLD may advance to severe conditions such as inflammation, fibrosis, cirrhosis, and liver failure [6]. Recent nomenclature updates have replaced NAFLD with metabolic dysfunction-associated fatty liver disease (MAFLD) and metabolic dysfunction-associated steatotic liver disease (MASLD) [7]. At present, many studies still use NAFLD nomenclature. To ensure comprehensive inclusion of relevant literature, this study screened all three terms. Liver biopsy is the gold standard for diagnosing NAFLD, but it is invasive and affected by sampling variability and inter-observer differences, making it impractical to scale to the global population of up to 1 billion affected individuals, with significant limitations in real-world applications [8]. Currently, ultrasound is recommended as a first-line screening tool for NAFLD and remains the most widely used imaging modality for diagnosing hepatic steatosis [9, 10]. While it allows subjective estimation of fat infiltration in the liver, its performance in detecting mild steatosis is suboptimal [11]. Magnetic resonance imaging (MRI) techniques can accurately quantify steatosis or identify fibrosis stages. The proton density fat fraction (PDFF) has been proposed as a standardized MR biomarker for hepatic steatosis, demonstrating high specificity for fat deposition (unconfounded by other histological or clinical parameters), but its high cost limits routine clinical use [9, 12]. FibroScan (®) enables accurate, non-invasive assessment of hepatic steatosis in NAFLD through the Controlled Attenuation Parameter (CAP), which quantifies liver fat content independently of fibrosis [11, 13]. Several serum biomarkers—such as the fatty liver index (FLI), hepatic steatosis index (HIS), and NASH Index (ION)—are also available for diagnosing or grading hepatic steatosis [10]. Although non-invasive diagnostic methods have been proposed as alternatives to liver biopsy, their clinical validity remains debated [14]. Artificial intelligence (AI) is a broad term referring to systems designed to perform advanced reasoning and analyze vast datasets with minimal human intervention [15, 16]. Machine learning (ML), a subfield of AI, focuses on developing models that learn from data to make predictions, prioritizing maximal predictive performance [17, 18]. ML enables the creation of algorithms capable of automatically learning, identifying patterns, and generating outcome predictions [19]. Decharatanachart et al. [20] conducted a meta-analysis on the diagnostic performance of AI in chronic liver diseases, with a focus on cirrhosis and liver fibrosis. Sun et al. [21] compared the diagnostic accuracy of various non-invasive tests for NAFLD, but did not focus specifically on ML models. To date, a quantitative evaluation of ML algorithms for the diagnosis and detection of NAFLD has not been reported, and the effectiveness of these algorithms remains unassessed and unsummarized. The present systematic review and meta-analysis therefore aims to evaluate the diagnostic accuracy and performance of ML algorithms for NAFLD detection, to compare the performance of ML when integrated with different diagnostic modalities, and to offer insights that may guide future research endeavors. Methods This study followed the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines. The protocol was registered on PROSPERO (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) with the report number CRD420251049731. Ethical approval was not required. The primary aim of this study is twofold: (1) to systematically evaluate and quantify the diagnostic accuracy of ML algorithms for NAFLD detection; and (2) to compare the performance of ML across different diagnostic modalities — imaging, clinical variables, or routine serological markers — in order to identify the promising ML-assisted diagnostic strategy. Search strategy We conducted a comprehensive search of articles published from the establishment of the databases to March 2025, covering PubMed, Web of Science, Cochrane Library, and EMBASE. The search strategy combined theme words with free words, including "Machine Learning," "Artificial Intelligence," "Deep Learning," "Algorithms," "Data Mining," "Non-alcoholic Fatty Liver Disease," "NAFLD," "NASH," "NAFL," "MAFLD," "MASLD," and their synonyms. The detailed search strategies for all databases were provided in Supplementary Table 1. Inclusion and exclusion criteria The inclusion criteria were as follows: (1) Studies applying ML models for the detection or diagnosis of NAFLD; (2) Studies capable of extracting diagnostic performance metrics, such as sensitivity, specificity, accuracy, area under the curve (AUC), F1 score, recall, positive predictive value (PPV), and negative predictive value (NPV). Studies performed in the setting of “screening”, “identification”, or “prediction” were considered for inclusion only when their outcomes could be mapped to diagnostic accuracy measures; (3) Studies for which the reference standard comprised imaging features, clinical variables, and/or routine serological markers; (4) RCTs, cohort studies, case–control studies, or cross-sectional studies; (5) English-language articles; (6) Adult participants (general screening/non-NAFLD and NAFLD). The exclusion criteria were as follows: (1) Reviews, animal studies, conference abstracts, case reports, editorials, protocols, commentaries, or letters; (2) Studies focusing on topics such as disease progression, risk stratification, severity grading, or prognosis; (3) Insufficient data for analysis (e.g., inability to extract or pool effect sizes); (4) Lack of detailed data required to reconstruct 2 × 2 contingency tables for diagnostic performance evaluation; (5) Outcomes involving concurrent hepatic inflammation, fibrosis, cirrhosis, or hepatocellular carcinoma (HCC). Article screening All the articles were imported into EndNote X9. Duplicate records were removed through automated and manual deduplication. Screening was conducted based on titles and abstracts, followed by full-text review for eligibility. The screening process was performed independently by two researchers (Luying Qi and Jingnan Xue) and cross-checked. Discrepancies were resolved through consultation with a third researcher (Hupo Bian). Data extraction and risk of bias assessment From each study, the following information was extracted via predefined standardized forms: (1) Study characteristics: first author, publication year, country, study type, design, reference standard, sample size, and outcomes; (2) ML features: training set, validation set, cross-validation, external validation, ML algorithms, and algorithm performance evaluation metrics. For studies that did not report the exact values of true positives (TP), false positives (FP), false negatives (FN), or true negatives (TN), these values were manually calculated using formulas to reconstruct confusion matrices. Data extraction was conducted independently by two researchers (Baiwang Li and Xiangpeng Wu), followed by cross-checking. Discrepancies were resolved through consultation with a third researcher (Jieqiong Chen). Study quality and risk of bias were assessed via the Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2) for diagnostic accuracy studies, covering four domains: patient selection, index test, reference standard, and flow&timing. Each domain was rated as high, low, or unclear risk of bias. The Prediction Model Risk of Bias Assessment Tool (PROBAST) was used to evaluate bias risk and applicability across four domains: participants, predictors, outcome, and analysis. Deeks’ funnel plot was used to assess publication bias. Assessments were performed independently by two researchers (Jingnan Xue and Hupo Bian), with disagreements resolved by a third researcher (Luying Qi). Statistical analysis The sensitivity and specificity were calculated for each study. Pooled results were generated via bivariate mixed-effects models to produce forest plots for diagnostic performance evaluation. Summary receiver operating characteristic (SROC) curves were plotted, and the AUC with 95% confidence interval (CI) was calculated to assess overall diagnostic accuracy. Heterogeneity among studies was evaluated via the Cochrane Q statistic and Higgins I2 values, categorized as low, moderate, or high (25%, 50%, and 75%, respectively). Subgroup analyses were conducted to explore sources of heterogeneity. All analyses and plotting were performed via R (version 4.3.1), Stata (version 17.0) and Origin 2022 software. Results Process of article screening A total of 13,385 records were retrieved from four databases. After removing 2,618 duplicates, titles and abstracts of 10,767 articles were screened, leading to the exclusion of 10,668 records. Full texts of 99 articles were assessed for eligibility, and 24 studies [6, 22–44] were ultimately included in the systematic review, with 18 studies [22–30, 32, 33, 35–38, 41, 43, 44] included in the meta-analysis. The screening flowchart was presented in Fig. 1.Fig. 1PRISMA chart PRISMA chart Study characteristics A total of 24 studies were included. Over 90% of the studies were published in 2021 or later. The majority were conducted in China and the United States (12 in China [23, 25, 26, 29–31, 33, 34, 36, 37, 43, 44], 6 in the US [6, 22, 32, 40–42]). Eleven studies [6, 22, 24, 25, 30, 32, 33, 37, 40, 41, 44] were cross-sectional, eight [27, 28, 31, 34, 36, 39, 42, 43] were case–control studies, and five [23, 26, 29, 35, 38] were cohort studies. All studies had a training/testing set split. Sixteen studies (67%) [23–31, 35–37, 39–41, 44] employed cross-validation techniques, and seven studies (29%) [26, 29, 34, 36, 38, 41, 42] utilized external validation datasets for model evaluation. Among all the studies, boosting algorithms were the most frequently used in optimal predictive models (8 studies [22, 23, 25, 30, 33, 36, 41, 44]). Imaging diagnostics were used in 17 studies (ultrasound: 12 [22–24, 26, 29, 31, 35, 37, 38, 40, 41, 43]; CT: 1 [27]; MRI: 3 [28, 42, 44]; unspecified: 1 [36]), clinical parameters in 2 studies [6, 32], and combined diagnostics in 5 studies [25, 30, 33, 34, 39]. Characteristics of the included studies were summarized in Table 1.Table 1General characteristics of included studiesFirst AuthorYearCountryTypeDesignSamplesTraining set: Testing setCross validation(Y/N)External validation(Y/N)Option modelAUCReference standardOutcomeAlizargar, A [6]2024USACross-sectionalMulti-center25058:2NNEnsemble model incorporating RF, XGBoost, and SVMTraining set: 0.981HIS index diagnosis: HSI = 8 alanine aminotransferase/aspartate aminotransferase ratio + body mass index (BMI) + 2 if diabetes + 2 if femaleNAFLDAtsawarungruangkit, A [22]2021USACross-sectionalMulti-center32357:3NNEnsemble model of RUS boosted treesTraining set: 0.76Testing set: 0.79Ultrasound diagnosisNAFLDTengrui Cao [23]2024ChinaCohortSingle-Center22,1407:3YNXGBoost modelTraining set: 0.967Testing set: 0.914Ultrasound diagnosis: (1)enhanced echogenicity in the near field and attenuated echogenicity in the far field of the liver parenchyma; (2)intrahepatic ducts that are relatively obscure in structure and rounded liver margins with hepatomegalyNAFLDChen Chen [24]2024JapanCross-sectionalSingle-Center47227:3YNRF modelTesting set: 0.889Ultrasound diagnosis: (1)the total score of blurred blood vessels (0—1 points); (2)liver brightness (0—4 points); (3)deep attenuation (0—2 points), (4)liver and kidney echo contrast (0—4 points) exceeds 2MASLDJiale Deng [25]2024ChinaCross-sectionalMulti-center5,171,3928:2YNCatBoost modelTraining set: 0.878Testing set: 0.862Any of the following three criteria diagnosis: excessive overweight or obesity/type II diabetes mellitus/radiological imaging-confirmed hepatic steatosisMAFLDYuhan Deng [26]2023ChinaCohortMulti-center12,2218:2YYLSTM modelInternal validation set: 0.818Ultrasound diagnosisNAFLDFujii, T [27]2024JapanCase–controlSingle-Center1867:3YN"clinicoradiomic" modelTraining set: 0.79Testing set: 0.80CT diagnosis: median CT attenuation values of the liver parenchyma of < 40 Hounsfield Units (HU) without a contrast agentNAFLDAiguo Han [28]2020PolandCase–controlSingle-Center2041:1YNCNN modelTesting set: 0.98MRI diagnosis: MRI PDFF = 5%NAFLDGuoqing Huang [29]2023ChinaCohortSingle-Center61967:3YYLR modelInternal set: 0.778External set: 0.806Ultrasound diagnosisNAFLDWeidong Ji [30]2022ChinaCross-sectionalMulti-center490,9807:3YNXGBoost modelInternal validation set:0.951Three criteria diagnosis: (1)subjects without drinking or drinking history; (2)no specific diseases leading to fatty liver such as viral hepatitis, liver disease induced by drug, total parenteral nutrition, hepatolenticular degeneration, and autoimmune liver disease; (3)liver imaging of subjects was consistent with the diagnostic criteria for diffuse fatty liverNAFLDQingqun Li [31]2022ChinaCase–controlSingle-Center2007:3YNRF modelTraining set: 0.945Ultrasound diagnosis (two of the following four abdominal ultrasounds): (1)the near-field echo of the liver area is diffusely enhanced, the echo is stronger than that of a kidney; (2)the far-field echo gradually attenuates; (3)the intrahepatic duct is not displayed; (4)the liver may be slightly or moderatelyNAFLDCan Liu [32]2023USACross-sectionalMulti-center66137:3NNANN modelTraining set: 0.874Validation set: 0.883USFLI diagnosis: USFLI cutoff value of = 30 (USFLI was calculated based on age, race, waist circumference, fasting glucose, gamma glutamyl transferase, and fasting insulin level)NAFLDYuanxing Liu [33]2021ChinaCross-sectionalSingle-Center15,3157:3NNXGBoost modelTraining set: 0.926Testing set: 0.873Imaging/histological examination diagnosisNAFLDXuefeng Ma [34]2021ChinaCase–controlMulti-center2667:12NYRF modelTraining set: 0.837Validation set: 0.732Pathology/B ultrasound/CT diagnosisNAFLDMasaebi, F [35]2024IranCohortMulti-center61807:3YNLR modelTraining set: 0.92Ultrasound diagnosisMASLDHongye Peng [36]2023ChinaCase–controlMulti-center709578:131YYXGBoost modelTesting set: 0.938Ultrasound or MRI diagnosis: ultrasound or MRI-PDFF = 5% according to the guideline of American Association for the Study of Liver DiseasesNAFLDShenghua Qin [37]2023ChinaCross-sectionalSingle-Center14,4397:3YNSVM modelTesting set: 0.85Color Doppler ultrasound diagnosisNAFLDSorino, P [38]2021ItalyCohortMulti-center29698:2NYNN modelTraining set: 0.84Testing set: 0.82Ultrasound diagnosisNAFLDÜlger, Y [39]2022TurkeyCase–controlSingle-Center4008:2YNANN modelTraining data: 0.96Testing data: 0.95Clinical parameters, imaging techniques (USG/MRI) and liver biopsy diagnosisNAFLDYazhi Wang [40]2025USACross-sectionalMulti-center24287:3YNLR modelTraining set: 0.856Validation set: 0.823CAP diagnosis: CAP = 268(57–284)dBNAFLDBo Yang [41]2024USACross-sectionalMulti-center14,9138:2YYLightGBM modelTraining set: 0.98Internal validation set: 0.90External validation set: 0.81Ultrasound diagnosis/CAP and liver stiffness measurement (LSM) diagnosisNAFLDTahmasebi, A [42]2023USACase–controlSingle-Center1308:1:1NYAutoML model/MRI diagnosis: MRI PDFF > 6.4%NAFLDShang-Yu Chiang [43]2025ChinaCase–controlSingle-Center4956:2:2YNPBCS-ConvNeXt modelTraining set: 0.88ATI images diagnosisNAFLDYip, T. C [44]2017ChinaCross-sectionalSingle-Center922500:422NNAdaBoost modelTraining set: 0.90Validation set: 0.831H-MRS (proton-magnetic resonance spectroscopy): IHTG content = 5%NAFLDY:yes N:no Table 1General characteristics of included studiesFirst AuthorYearCountryTypeDesignSamplesTraining set: Testing setCross validation(Y/N)External validation(Y/N)Option modelAUCReference standardOutcomeAlizargar, A [6]2024USACross-sectionalMulti-center25058:2NNEnsemble model incorporating RF, XGBoost, and SVMTraining set: 0.981HIS index diagnosis: HSI = 8 alanine aminotransferase/aspartate aminotransferase ratio + body mass index (BMI) + 2 if diabetes + 2 if femaleNAFLDAtsawarungruangkit, A [22]2021USACross-sectionalMulti-center32357:3NNEnsemble model of RUS boosted treesTraining set: 0.76Testing set: 0.79Ultrasound diagnosisNAFLDTengrui Cao [23]2024ChinaCohortSingle-Center22,1407:3YNXGBoost modelTraining set: 0.967Testing set: 0.914Ultrasound diagnosis: (1)enhanced echogenicity in the near field and attenuated echogenicity in the far field of the liver parenchyma; (2)intrahepatic ducts that are relatively obscure in structure and rounded liver margins with hepatomegalyNAFLDChen Chen [24]2024JapanCross-sectionalSingle-Center47227:3YNRF modelTesting set: 0.889Ultrasound diagnosis: (1)the total score of blurred blood vessels (0—1 points); (2)liver brightness (0—4 points); (3)deep attenuation (0—2 points), (4)liver and kidney echo contrast (0—4 points) exceeds 2MASLDJiale Deng [25]2024ChinaCross-sectionalMulti-center5,171,3928:2YNCatBoost modelTraining set: 0.878Testing set: 0.862Any of the following three criteria diagnosis: excessive overweight or obesity/type II diabetes mellitus/radiological imaging-confirmed hepatic steatosisMAFLDYuhan Deng [26]2023ChinaCohortMulti-center12,2218:2YYLSTM modelInternal validation set: 0.818Ultrasound diagnosisNAFLDFujii, T [27]2024JapanCase–controlSingle-Center1867:3YN"clinicoradiomic" modelTraining set: 0.79Testing set: 0.80CT diagnosis: median CT attenuation values of the liver parenchyma of < 40 Hounsfield Units (HU) without a contrast agentNAFLDAiguo Han [28]2020PolandCase–controlSingle-Center2041:1YNCNN modelTesting set: 0.98MRI diagnosis: MRI PDFF = 5%NAFLDGuoqing Huang [29]2023ChinaCohortSingle-Center61967:3YYLR modelInternal set: 0.778External set: 0.806Ultrasound diagnosisNAFLDWeidong Ji [30]2022ChinaCross-sectionalMulti-center490,9807:3YNXGBoost modelInternal validation set:0.951Three criteria diagnosis: (1)subjects without drinking or drinking history; (2)no specific diseases leading to fatty liver such as viral hepatitis, liver disease induced by drug, total parenteral nutrition, hepatolenticular degeneration, and autoimmune liver disease; (3)liver imaging of subjects was consistent with the diagnostic criteria for diffuse fatty liverNAFLDQingqun Li [31]2022ChinaCase–controlSingle-Center2007:3YNRF modelTraining set: 0.945Ultrasound diagnosis (two of the following four abdominal ultrasounds): (1)the near-field echo of the liver area is diffusely enhanced, the echo is stronger than that of a kidney; (2)the far-field echo gradually attenuates; (3)the intrahepatic duct is not displayed; (4)the liver may be slightly or moderatelyNAFLDCan Liu [32]2023USACross-sectionalMulti-center66137:3NNANN modelTraining set: 0.874Validation set: 0.883USFLI diagnosis: USFLI cutoff value of = 30 (USFLI was calculated based on age, race, waist circumference, fasting glucose, gamma glutamyl transferase, and fasting insulin level)NAFLDYuanxing Liu [33]2021ChinaCross-sectionalSingle-Center15,3157:3NNXGBoost modelTraining set: 0.926Testing set: 0.873Imaging/histological examination diagnosisNAFLDXuefeng Ma [34]2021ChinaCase–controlMulti-center2667:12NYRF modelTraining set: 0.837Validation set: 0.732Pathology/B ultrasound/CT diagnosisNAFLDMasaebi, F [35]2024IranCohortMulti-center61807:3YNLR modelTraining set: 0.92Ultrasound diagnosisMASLDHongye Peng [36]2023ChinaCase–controlMulti-center709578:131YYXGBoost modelTesting set: 0.938Ultrasound or MRI diagnosis: ultrasound or MRI-PDFF = 5% according to the guideline of American Association for the Study of Liver DiseasesNAFLDShenghua Qin [37]2023ChinaCross-sectionalSingle-Center14,4397:3YNSVM modelTesting set: 0.85Color Doppler ultrasound diagnosisNAFLDSorino, P [38]2021ItalyCohortMulti-center29698:2NYNN modelTraining set: 0.84Testing set: 0.82Ultrasound diagnosisNAFLDÜlger, Y [39]2022TurkeyCase–controlSingle-Center4008:2YNANN modelTraining data: 0.96Testing data: 0.95Clinical parameters, imaging techniques (USG/MRI) and liver biopsy diagnosisNAFLDYazhi Wang [40]2025USACross-sectionalMulti-center24287:3YNLR modelTraining set: 0.856Validation set: 0.823CAP diagnosis: CAP = 268(57–284)dBNAFLDBo Yang [41]2024USACross-sectionalMulti-center14,9138:2YYLightGBM modelTraining set: 0.98Internal validation set: 0.90External validation set: 0.81Ultrasound diagnosis/CAP and liver stiffness measurement (LSM) diagnosisNAFLDTahmasebi, A [42]2023USACase–controlSingle-Center1308:1:1NYAutoML model/MRI diagnosis: MRI PDFF > 6.4%NAFLDShang-Yu Chiang [43]2025ChinaCase–controlSingle-Center4956:2:2YNPBCS-ConvNeXt modelTraining set: 0.88ATI images diagnosisNAFLDYip, T. C [44]2017ChinaCross-sectionalSingle-Center922500:422NNAdaBoost modelTraining set: 0.90Validation set: 0.831H-MRS (proton-magnetic resonance spectroscopy): IHTG content = 5%NAFLDY:yes N:no General characteristics of included studies Training set: 0.76 Testing set: 0.79 Training set: 0.967 Testing set: 0.914 Training set: 0.878 Testing set: 0.862 Training set: 0.79 Testing set: 0.80 Internal set: 0.778 External set: 0.806 Internal validation set: 0.951 Training set: 0.874 Validation set: 0.883 Training set: 0.926 Testing set: 0.873 Training set: 0.837 Validation set: 0.732 Training set: 0.84 Testing set: 0.82 Training data: 0.96 Testing data: 0.95 Training set: 0.856 Validation set: 0.823 Training set: 0.98 Internal validation set: 0.90 External validation set: 0.81 Training set: 0.90 Validation set: 0.83 Y:yes N:no The primary input features (top 30 based on the number of studies) for all ML models encompassed demographic characteristics, chronic disease conditions, lifestyle habits, and laboratory indicators. Age and sex were the most prevalent input features, followed by BMI and ALT. Detailed information can be found in Fig. 2 and Supplementary Table 2.Fig. 2Input features Input features Quality assessment and publication bias QUADAS-2 tool For QUADAS-2, the results of the quality assessment were shown in Fig. 3. The primary sources of high risk of bias in the included studies stemmed from reference standard, with additional high risks observed in patient selection, index test, and flow & timing. In terms of patient selection, the most common high-risk sign was the inclusion of patients (non-consecutive or non-random samples); Case–control studies also suggested high risk. In the index test, diagnostic methods based on subjective indicators were considered to have a high risk of bias, while most studies relied on objective indicators. Regarding the reference standard, high risk of bias was most common, primarily due to the lack of blinding in the studies. For the domain of flow & timing, unclear risk of bias was most commonly observed, primarily due to a lack of clarification regarding whether an appropriate interval existed between index tests and reference standard; High risk of bias was mainly attributed to the fact that only a subset of participants underwent the same reference standard. Regarding applicability concerns, the proportion of low concern was high. Five studies [24, 25, 27, 35, 39] were rated as having high concern in patient selection domain, as their study populations differed substantially from our systematic review's target population in terms of demographic characteristics. In index test domain, the diagnostic methods employed in the two studies [33, 34] were not clearly reported, so it was rated as unclear. One study [25] was rated as having high concern because the diagnostic methods used in the study had low agreement with the reference standard specified in the systematic review. No studies were rated as high concern in reference standard domain. Two articles [33, 34] were rated as unclear concern due to ambiguous descriptions of the reference standard. Detailed information was provided in Supplementary Table 3.Fig. 3The risk of bias and applicability concerns of included studies The risk of bias and applicability concerns of included studies PROBAST tool Regarding the PROBAST tool, two studies [23, 26] were classified as having low risk overall in the domain of risk of bias. No high-risk indicators were identified in the domain of participants. Regarding predictors, 15 studies [6, 22, 25, 27, 28, 31, 32, 34, 36, 39–44] evaluated predictors with clear outcome information, with the absence of blinding identified as the primary reason for high risk of bias. Regarding outcome, 4 studies [6, 25, 32, 42] incorporated at least one predictor into the definition of the outcome, and 15 studies [6, 22, 25, 27, 28, 31, 32, 34, 36, 39–44] had clear knowledge of predictor information when determining the outcome, both of which were rated as high risk of bias. For analysis, the principal factor leading to high risk of bias was insufficient sample size; in addition, high risk was also attributed to selecting predictors based on univariate analysis. In terms of applicability, 13 studies [22, 23, 26, 29–32, 36–38, 40, 41, 44] were generally rated as low concern. The demographic characteristics of the participants in 5 articles [24, 25, 27, 35, 39] were inconsistent with the research question of the systematic review; one study [43] did not include clinical variables or serological markers in the predictors; three studies [24, 25, 35] used MAFLD or MASLD as outcome indicators. As these aspects deviated substantially from the systematic review, all were rated as having high concern. A summary of results was presented in Table 2, and detailed responses to all 20 signaling questions were available in Supplementary Table S4.Table 2Summary of the risk of bias and applicability concerns for each study by PROBASTStudy(Year)Risk of BiasApplicabilityOverallParticipantsPredictorsOutcomeAnalysisParticipantsPredictorsOutcomeRisk of BiasApplicabilityAlizargar, A (2024) [6]?---? + + -?Atsawarungruangkit, A (2021) [22] + --? + + + - + Tengrui Cao (2024) [23] + + + + + + + + + Chen Chen (2024) [24] + + ? + - + -?-Jiale Deng (2024) [25] + ---- + ---Yuhan Deng (2023) [26] + + + + + + + + + Fujii, T (2024) [27]?---- + + --Aiguo Han (2020) [28] + --? + + + --Guoqing Huang (2023) [29] + + + - + + + - + Weidong Ji (2022) [30] + ??? + + + ? + Qingqun Li (2022) [31] + --- + + + - + Can Liu (2023) [32] + --- + + + - + Yuanxing Liu (2021) [33] + ??? + + ???Xuefeng Ma (2021) [34] + --- + + ?-?Masaebi, F (2024) [35]?? + -- + ---Hongye Peng (2023) [36] + --- + + + - + Shenghua Qin (2023) [37] + ??? + + + ? + Sorino, P (2021) [38]? + ?? + + + ? + Ülger, Y (2022) [39] + ---- + + --Yazhi Wang (2025) [40] + --- + + + - + Bo Yang (2024) [41] + -- + + + + - + Tahmasebi, A (2023) [42] + --- + + + --Shang-Yu Chiang (2025) [43]?---?- + --T.C.-F. Yip (2017) [44] + --- + + + - + " + ": low risk of bias/low applicability concern; "?": unclear risk of bias/unclear applicability concern; "-": high risk of bias/high applicability concern" + ": low risk of bias/low applicability concern; "?": unclear risk of bias/unclear applicability concern; "-": high risk of bias/high applicability concern Table 2Summary of the risk of bias and applicability concerns for each study by PROBASTStudy(Year)Risk of BiasApplicabilityOverallParticipantsPredictorsOutcomeAnalysisParticipantsPredictorsOutcomeRisk of BiasApplicabilityAlizargar, A (2024) [6]?---? + + -?Atsawarungruangkit, A (2021) [22] + --? + + + - + Tengrui Cao (2024) [23] + + + + + + + + + Chen Chen (2024) [24] + + ? + - + -?-Jiale Deng (2024) [25] + ---- + ---Yuhan Deng (2023) [26] + + + + + + + + + Fujii, T (2024) [27]?---- + + --Aiguo Han (2020) [28] + --? + + + --Guoqing Huang (2023) [29] + + + - + + + - + Weidong Ji (2022) [30] + ??? + + + ? + Qingqun Li (2022) [31] + --- + + + - + Can Liu (2023) [32] + --- + + + - + Yuanxing Liu (2021) [33] + ??? + + ???Xuefeng Ma (2021) [34] + --- + + ?-?Masaebi, F (2024) [35]?? + -- + ---Hongye Peng (2023) [36] + --- + + + - + Shenghua Qin (2023) [37] + ??? + + + ? + Sorino, P (2021) [38]? + ?? + + + ? + Ülger, Y (2022) [39] + ---- + + --Yazhi Wang (2025) [40] + --- + + + - + Bo Yang (2024) [41] + -- + + + + - + Tahmasebi, A (2023) [42] + --- + + + --Shang-Yu Chiang (2025) [43]?---?- + --T.C.-F. Yip (2017) [44] + --- + + + - + " + ": low risk of bias/low applicability concern; "?": unclear risk of bias/unclear applicability concern; "-": high risk of bias/high applicability concern" + ": low risk of bias/low applicability concern; "?": unclear risk of bias/unclear applicability concern; "-": high risk of bias/high applicability concern Summary of the risk of bias and applicability concerns for each study by PROBAST " + ": low risk of bias/low applicability concern; "?": unclear risk of bias/unclear applicability concern; "-": high risk of bias/high applicability concern " + ": low risk of bias/low applicability concern; "?": unclear risk of bias/unclear applicability concern; "-": high risk of bias/high applicability concern Deeks’ funnel plot In the results of the Deeks’ funnel plot, the P value was greater than 0.05, indicating no significant publication bias in the internal and external validation set, as shown in Fig. 4.Fig. 4Deeks’ funnel plot to evaluate publication bias (a internal validation set b external validation set) Deeks’ funnel plot to evaluate publication bias (a internal validation set b external validation set) Predictive performance and subgroup analysis After excluding studies with incomplete model evaluation metrics (AUC, TP, FP, TN, FN) and duplicate cohorts, a total of 18 studies were included. For studies that provided external validation, these were extracted and analyzed as independent datasets. Ultimately, a meta-analysis was performed on 21 datasets containing the best-performing models, with detailed characteristics presented in Table 3. Of these, 17 datasets were based on internal validation. The pooled sensitivity and specificity for internal validation sets were 0.82 (95% CI: 0.77–0.86) and 0.81 (95% CI: 0.77–0.85), respectively, with an AUC of 0.88 (95% CI: 0.85–0.91). The coupled forest plot was shown in Fig. 5. The remaining 4 datasets were based on external validation. The pooled sensitivity and specificity for external validation sets were 0.80 (95% CI: 0.70–0.87) and 0.72 (95% CI: 0.56–0.84), respectively, with an AUC of 0.83 (95% CI: 0.80–0.86). The coupled forest plot was shown in Fig. 6. The SROC curves were presented in Fig. 7.Table 3Summary of meta analysis characteristics including internal and external validationAuthor(Year)Type of validation setSamplesOptimal algorithmAUCTPFNTNFPAtsawarungruangkit, A (2021) [22]internal970Ensemble of RUS boosted trees0.7917867512213Tengrui Cao (2024) [23]internal6641XGBoost0.9127425412801557Chen Chen (2024) [24]internal1417RF0.8934678815178Jiale Deng (2024) [25]internal1,034,279CatBoost model0.8684,95319,412700,225229,689Yuhan Deng (2023) [26]internal2098LSTM0.8216083926830Yuhan Deng (2023) [26]external1728LSTM0.79791209506222Fujii, T (2024) [27]internal56"clinicoradiomic"0.80168266Aiguo Han (2020) [28]internal102CNN0.98682302Guoqing Huang (2023) [29]internal1859LR0.78564236737322Guoqing Huang (2023) [29]external4337LR0.819615692513294Weidong Ji (2022) [30]internal147,294XGBoost0.9565,729780463,9509811Can Liu (2023) [32]internal2004ANN0.88128033432367Yuanxing Liu (2021) [33]internal4942XGBoost0.8711367242802280Masaebi, F (2024) [35]internal1854LR0.92351571280166Hongye Peng (2023) [36]external131XGBoost0.94889268Shenghua Qin (2023) [37]internal4332SVM0.8516724311798431Sorino, P (2021) [38]internal574NN0.821025634076Bo Yang (2024) [41]internal2983LightGBM0.902246172345220Bo Yang (2024) [41]external1798LightGBM0.8155177661509Shang-Yu Chiang (2025) [43]internal99PBCS-ConvNeXt0.88409419T.C.-F. Yip (2017) [44]internal922AdaBoost0.83105138816Fig. 5Coupled forest plot of pooled sensitivity and specificity (internal validation set)Fig. 6Coupled forest plot of pooled sensitivity and specificity (external validation set)Fig. 7Summary receiver operating characteristic plots (a internal validation set b external validation set) Table 3Summary of meta analysis characteristics including internal and external validationAuthor(Year)Type of validation setSamplesOptimal algorithmAUCTPFNTNFPAtsawarungruangkit, A (2021) [22]internal970Ensemble of RUS boosted trees0.7917867512213Tengrui Cao (2024) [23]internal6641XGBoost0.9127425412801557Chen Chen (2024) [24]internal1417RF0.8934678815178Jiale Deng (2024) [25]internal1,034,279CatBoost model0.8684,95319,412700,225229,689Yuhan Deng (2023) [26]internal2098LSTM0.8216083926830Yuhan Deng (2023) [26]external1728LSTM0.79791209506222Fujii, T (2024) [27]internal56"clinicoradiomic"0.80168266Aiguo Han (2020) [28]internal102CNN0.98682302Guoqing Huang (2023) [29]internal1859LR0.78564236737322Guoqing Huang (2023) [29]external4337LR0.819615692513294Weidong Ji (2022) [30]internal147,294XGBoost0.9565,729780463,9509811Can Liu (2023) [32]internal2004ANN0.88128033432367Yuanxing Liu (2021) [33]internal4942XGBoost0.8711367242802280Masaebi, F (2024) [35]internal1854LR0.92351571280166Hongye Peng (2023) [36]external131XGBoost0.94889268Shenghua Qin (2023) [37]internal4332SVM0.8516724311798431Sorino, P (2021) [38]internal574NN0.821025634076Bo Yang (2024) [41]internal2983LightGBM0.902246172345220Bo Yang (2024) [41]external1798LightGBM0.8155177661509Shang-Yu Chiang (2025) [43]internal99PBCS-ConvNeXt0.88409419T.C.-F. Yip (2017) [44]internal922AdaBoost0.83105138816 Summary of meta analysis characteristics including internal and external validation Coupled forest plot of pooled sensitivity and specificity (internal validation set) Coupled forest plot of pooled sensitivity and specificity (external validation set) Summary receiver operating characteristic plots (a internal validation set b external validation set) Leave-one-out sensitivity analyses revealed that the overall AUC, sensitivity, and specificity did not change substantially when any single dataset was omitted, indicating that no individual dataset dominated the pooled results and that the findings were robust. Subgroup analyses were performed on the 17 internally validated datasets. The results showed that study design, reference standard, outcome, and the use of cross-validation were all statistically significant for both sensitivity and specificity (P < 0.05), indicating significant differences among studies, which may be sources of data heterogeneity. Regarding the participants, the difference between local population and public database did not reach statistical significance; however, numerical differences were observed, with lower sensitivity and specificity in the public database group (0.76/0.77) compared to the local population group (0.82/0.82). Cross-sectional studies yielded slightly higher sensitivity and specificity (0.82/0.81) than non-cross-sectional studies (0.80/0.80). ML technologies combined with imaging(ultrasound, CT, or MRI) demonstrated higher sensitivity and specificity (0.83/0.80) than combined with clinical variables and routine serological markers (0.80/0.79). Studies using MASLD as the outcome measure showed higher sensitivity and specificity (0.84/0.86) than those using NAFLD (0.81/0.81). Studies that employed cross-validation had higher sensitivity (0.84 vs. 0.74) but lower specificity (0.80 vs. 0.83) than those without cross-validation. Across different models, the deep learning (DL) model (0.82) demonstrated statistically significantly higher specificity compared to the boosting tree model (0.80). The detailed results of the subgroup analyses for the internal validation sets were presented in Table 4.Table 4Subgroup analysis resultsSubgroupNumbersSensitivity(95%CI)P1Specificity(95%CI)P2Participants0.310.07 Local population150.82 [0.78—0.87]0.82 [0.78—0.85] Public database20.76 [0.61—0.92]0.77 [0.65—0.89]Study Design< 0.01< 0.01 Cross-sectional study90.82 [0.76—0.88]0.81 [0.76—0.86] Non-cross-sectional study70.80 [0.72—0.88]0.80 [0.73—0.86]Reference Standard 0.020.01 Imaging(ultrasound, CT or MRI)140.83 [0.79—0.87]0.80 [0.76—0.84] Clinical variables or routine serological markers20.80 [0.68—0.93]0.79 [0.69—0.90]Outcome0.05< 0.01 NAFLD140.81 [0.76—0.87]0.81 [0.76—0.85] MASLD20.84 [0.72—0.96]0.86 [0.77—0.94]Cross-validation< 0.01< 0.01 Yes120.84 [0.80—0.88]0.80 [0.75—0.85] No50.74 [0.65—0.84]0.83 [0.77—0.89]Model0.05< 0.01 Boosting tree70.83 [0.76—0.90]0.80 [0.74—0.87] DL50.82 [0.73—0.92]0.82 [0.74—0.90] Table 4Subgroup analysis resultsSubgroupNumbersSensitivity(95%CI)P1Specificity(95%CI)P2Participants0.310.07 Local population150.82 [0.78—0.87]0.82 [0.78—0.85] Public database20.76 [0.61—0.92]0.77 [0.65—0.89]Study Design< 0.01< 0.01 Cross-sectional study90.82 [0.76—0.88]0.81 [0.76—0.86] Non-cross-sectional study70.80 [0.72—0.88]0.80 [0.73—0.86]Reference Standard 0.020.01 Imaging(ultrasound, CT or MRI)140.83 [0.79—0.87]0.80 [0.76—0.84] Clinical variables or routine serological markers20.80 [0.68—0.93]0.79 [0.69—0.90]Outcome0.05< 0.01 NAFLD140.81 [0.76—0.87]0.81 [0.76—0.85] MASLD20.84 [0.72—0.96]0.86 [0.77—0.94]Cross-validation< 0.01< 0.01 Yes120.84 [0.80—0.88]0.80 [0.75—0.85] No50.74 [0.65—0.84]0.83 [0.77—0.89]Model0.05< 0.01 Boosting tree70.83 [0.76—0.90]0.80 [0.74—0.87] DL50.82 [0.73—0.92]0.82 [0.74—0.90] Subgroup analysis results Discussion Comparison with previous reviews In recent years, the widespread application of AI in medical imaging has provided new possibilities for disease diagnosis and treatment. ML models are increasingly becoming important tools for the early prediction and identification of NAFLD. Zamanian et al. [45] systematically reviewed 64 studies that employed ML or DL models using ultrasound (including elastography) or clinical (including serological) approaches to detect the presence of NAFLD, and found these models to be effective, accurate, accessible, and capable of reducing diagnostic costs. However, their study only summarized the algorithms and optimal accuracy values from the literature without quantitatively pooling the results. In contrast, our study conducted a quantitative meta-analysis of ML-assisted NAFLD diagnosis, pooling diagnostic performance metrics such as AUC, sensitivity, specificity, PPV, and NPV, thereby providing more precise information for clinical decision-making. Decharatanachart et al. [46] performed a meta-analysis of 13 studies published up to 2021, focusing primarily on AI-assisted ultrasound and clinical diagnosis of NAFLD, with pooled AUC values of 0.98 and 0.85, respectively, demonstrating promising performance improvements. Our study analyzed 21 datasets, focusing on rapidly evolving ML technologies, most of which were published in or after 2021. In addition to ultrasound and clinical data, our study also included ML-assisted diagnosis and prediction using CT and MRI, serving as an update and complement to the work by Decharatanachart et al. Li et al. [47] reviewed the potential of DL in the diagnosis of MASLD, but focused only on ultrasound-based diagnosis without involving other imaging modalities, clinical data, or laboratory tests, and only reported AUC/ACC values. Our study aggregated and quantitatively analyzed multiple diagnostic methods and various performance metrics, aiming to provide a convenient and cost-effective diagnostic tool for future NAFLD screening and diagnosis. Li et al. [2] summarized the applications of AI in the diagnosis and treatment of NAFLD, including diagnostic models based on electronic health records, laboratory tests, ultrasound, radiological imaging, and liver histopathology data, yet noted a lack of high-quality machine learning development datasets. Our study synthesized multiple studies that developed and validated ML models for NAFLD, and presents model performance more intuitively and accurately through forest plots and SROC curves, thereby addressing this gap. Quality and bias assessment The high risk of bias and concern identified by the QUADAS-2 and PROBAST tools may be related to the following factors. The enrolled patients were not drawn from a continuous or random sample, but were selected from public databases or retrieved from hospital electronic medical record systems. The case–control design may have led researchers to be influenced by their awareness of patient status, thereby affecting the randomness of participant selection and the interpretation of index test results. Diagnostic methods relying on subjective criteria, combined with researchers' prior knowledge of the patients' final diagnoses, may lead to an overestimation of diagnostic performance. The absence of blinding, whereby researchers had prior knowledge of the index test results, may have led to biased interpretation of the reference standard outcomes. Additionally, failure to implement blinding may create a risk that knowledge of outcome data affected the process of predictor assessment. Failure to apply the same reference standard to all participants in the study may have introduced verification bias. Because the predictors incorporated in the study model were also components of the outcome definition, the association between predictors and the outcome may have been overestimated. Among predictive models, an EPV < 20 in the development study or a sample size < 100 in the validation study indicated an insufficient sample size. Selecting predictors based on univariate analysis may lead to overfitting or an overestimation of model performance. High concern was potentially associated with the enrollment of patients who had other underlying diseases at baseline. Additionally, cases with MAFLD or MASLD as outcome were also classified as high-concern. Multiple studies have demonstrated that over 95% of NAFLD cases met the diagnostic criteria for MAFLD and MASLD. The evolution of terminology from NAFLD to MAFLD and MASLD reflected a highly overlapping relationship between these terms, which can be considered clinically equivalent [48–52]. However, their diagnostic methods and definitions were not entirely consistent, and we regarded this as high concern. Heterogeneity and model robustness Through subgroup analyses, we found that study design, reference standard, outcome, and the use of cross-validation all had statistically significant effects on sensitivity and specificity (P < 0.05). The specificity between different models was also statistically significant, suggesting that these factors may be sources of heterogeneity. No statistically significant differences in sensitivity or specificity were observed between studies of local population and of public database. However, only two studies utilized public database; the small sample size may have resulted in insufficient statistical power to detect true differences. Cross-sectional studies demonstrated slightly higher diagnostic accuracy than non-cross-sectional studies, along with more reliable estimates (narrower confidence intervals). Non-cross-sectional studies (predominantly cohort studies) may be subject to selection bias and verification bias, leading to greater uncertainty in diagnostic accuracy estimates. Sun et al. [21] confirmed that differences in study design were important sources of heterogeneity in the accuracy of non-invasive diagnostic tests for NAFLD, supporting the findings of our heterogeneity analysis. Compared to ML algorithm-based clinical variables/routine serological markers, ML algorithm-based imaging (ultrasound, CT, or MRI) showed higher diagnostic accuracy and narrower confidence intervals, indicating they were more objective and reproducible diagnostic standards, the finding consistent with the results of Decharatanachart et al. [46]. Furthermore, a meta-analysis by Wang et al. [53] reported that the sensitivity and specificity of MR-based techniques for diagnosing NAFLD were 0.82 and 0.87. While Khanmohammadi et al. [54] reported lower diagnostic accuracy for the body roundness index (BRI) in identifying NAFLD, with a sensitivity of 0.81 and specificity of 0.69. These findings suggested a potential priority for ML-assisted imaging in the diagnosis of NAFLD. The diagnostic accuracy was higher in the MASLD subgroup. This may be related to the fact that MASLD is an updated definition of NAFLD, with a stronger emphasis on metabolic dysfunction and stricter diagnostic criteria. However, only two studies were included in the MASLD subgroup, and the 95% confidence intervals were wide; therefore, these results should be interpreted with caution. Cross-validation plays a critical role in the development and evaluation of ML models by preventing overfitting [55], selecting optimal models [56], assessing model generalizability [57], and performing hyperparameter tuning and calibration [58]. In our study, studies that employed cross-validation had higher sensitivity (0.84) than those without (0.74), but lower specificity (0.80) compared to those without cross-validation (0.83). This may be because studies without cross-validation were at higher risk of overfitting and may suffer from class imbalance. The difference in sensitivity between the boosting tree model (0.83) and the DL model (0.82) was of borderline statistical significance (P = 0.05), suggesting that the two models may exhibit comparable performance in patient identification, albeit with cautious interpretation warranted. Regarding specificity, the DL model (0.82) demonstrated superiority over the boosting tree model (0.80), indicating a potential advantage of the DL approach in ruling out non-NAFLD cases. Previous studies have yielded inconsistent findings regarding the comparative advantage of DL models versus boosting tree models [59–61]. Therefore, further investigations are warranted to validate these observations. A total of 21 datasets from 18 studies included in this meta-analysis were analyzed. Among these, 17 datasets were based on internal validation, with pooled sensitivity, specificity, and AUC of 0.82, 0.81, and 0.88, respectively. The remaining four datasets were based on external validation, yielding pooled sensitivity, specificity, and AUC of 0.80, 0.72, and 0.83, respectively, indicating that the models performed robustly in identifying patients. The external validation primarily involved spatially independent datasets. While the models demonstrated good sensitivity, a decrease in specificity was observed, which may be attributable to differences in disease spectrum across populations, inconsistencies in reference standards, or model overfitting. These findings suggested that model calibration or threshold adjustment may be necessary before applying the models to new populations. Furthermore, the number of external validation sets in our study was limited. The lack of external validation in ML-based studies may hinder the accurate reflection of their applicability and reliability in this health domain, thereby limiting model stability and generalizability [62]. Future research should focus on further optimizing model design, expanding sample sizes, and conducting additional external validation to enhance the reliability and generalizability of ML models in clinical practice. Clinical value Clinicians typically diagnosed NAFLD by synthesizing symptoms, laboratory tests, and imaging findings. For example, patients presenting with obesity, persistent fatigue, or right upper quadrant discomfort, combined with elevated AST/ALT levels and imaging evidence of hepatic steatosis or fibrosis, might indicate NASH or fibrosis. However, early-stage NAFLD often lacked overt symptoms or significant laboratory markers of liver injury, and was frequently incidentally detected during routine screenings. US is widely used for NAFLD screening; however, its sensitivity for detecting mild steatosis is limited, reaching only 0.65, and it is affected by the variability of operators and examinees [63, 64]. Cao et al. [65] found that AI-assisted ultrasound has substantial potential in diagnosing hepatic steatosis, with significantly improved diagnostic accuracy compared to conventional ultrasound. Muhammad et al. [66] reported that AI-assisted ultrasound achieved an AUC of 0.98 for detecting NAFLD, and AI-assisted CT achieved an AUC of 0.97 for staging fibrosis (= F3), which were markedly higher than those of electronic health record-based tools (AUC 0.76). Nivethitha V et al. [67] observed that AI models using imaging-based reference standards (e.g., MRI-derived fat fraction or expert-graded ultrasound) demonstrated higher diagnostic accuracy than those using liver biopsy as the reference standard. Our study further showed that ML models assisting imaging achieved higher diagnostic accuracy than those assisting clinical variables/routine serological markers. Collectively, these findings suggest that AI/ML-assisted imaging may serve as a non-invasive alternative for NAFLD diagnosis. Regarding cost-effectiveness, Zhang and Njei et al. [68, 69] demonstrated that AI/ML-based diagnostic models offer advantages in improving NAFLD/MASLD screening efficiency and reducing costs. Limitations Firstly, the small number of studies in certain subgroup analyses hindered the assessment of their diagnostic accuracy. Secondly, the included studies varied in clinical application scenarios (e.g., screening, prediction, and identification). Although these results were integrated into a diagnostic accuracy framework, this may have introduced additional heterogeneity. Furthermore, the predominance of internally validated models raised concerns about potential overfitting and limited external generalizability. The relatively small number of studies employing external validation hindered their applicability and generalizability in this field. Conclusion This systematic review and meta-analysis provided the first quantitative evaluation of ML algorithms for detection and diagnosis of NAFLD. The results indicated that ML-assisted imaging demonstrated promising diagnostic accuracy for NAFLD detection, suggesting that ML algorithms may assist clinicians with early screening and optimize diagnostic and therapeutic decision-making. Patients with underlying comorbidities, including hypertension, diabetes mellitus, and cardiovascular disease, should be prioritized as key screening targets. A standardized management pathway can be established through the integration of electronic health record systems, the development of quality control indicators, coordinated comorbidity management and risk stratification, as well as patient education and lifestyle interventions. In future studies, we will further investigate disease progression (e.g., from NAFL to NASH, and from NASH to fibrosis) with the aim of delaying disease progression and reducing the public health burden. In addition, it is hoped that through multicenter research and external validation, the predictive model of ML algorithm can be applied to clinical practice. Supplementary Information Supplementary Material 1. Supplementary Material 1.