Machine learning models for predicting metabolic dysfunction-associated steatotic liver disease prevalence using basic demographic and clinical characteristics

Abstract

Background

Metabolic dysfunction-associated steatotic liver disease (MASLD) is a global health concern that necessitates early screening and timely intervention to improve prognosis. The current diagnostic protocols for MASLD involve complex procedures in specialised medical centres. This study aimed to explore the feasibility of utilising machine learning models to accurately screen for MASLD in large populations based on a combination of essential demographic and clinical characteristics.

Methods

A total of 10,007 outpatients who underwent transient elastography at the First Affiliated Hospital of Gannan Medical University were enrolled to form a derivation cohort. Using eight demographic and clinical characteristics (age, educational level, height, weight, waist and hip circumference, and history of hypertension and diabetes), we built predictive models for MASLD (classified as none or mild: controlled attenuation parameter (CAP) ≤ 269 dB/m; moderate: 269–296 dB/m; severe: CAP > 296 dB/m) employing 10 machine learning algorithms: logistic regression (LR), multilayer perceptron (MLP), extreme gradient boosting (XGBoost), bootstrap aggregating, decision tree, K-nearest neighbours, light gradient boosting machine, naive Bayes, random forest, and support vector machine. These models were externally validated using the National Health and Nutrition Examination Survey (NHANES) 2017–2023 datasets.

Results

In the hospital outpatient cohort, machine learning algorithms demonstrated robust predictive capabilities. Notably, LR achieved the highest accuracy (ACC) of 0.711 in the test cohort and 0.728 in the validation cohort, coupled with robust areas under the receiver operating characteristic curve (AUC) values of 0.798 and 0.806, respectively. Similarly, MLP and XGBoost showed promising results, with MLP achieving an ACC of 0.735 in the test cohort, and XGBoost registering an AUC of 0.798. External validation using the NHANES datasets yielded consistent AUC results, with LR (0.831), MLP (0.823), and XGBoost (0.784) performing robustly.

Conclusions

This study demonstrated that machine learning models constructed using a combination of essential demographic and clinical characteristics can accurately screen for MASLD in the general population. This approach significantly enhances the feasibility, accessibility, and compliance of MASLD screening and provides an effective tool for large-scale health assessments and early intervention strategies.

Background

Metabolic dysfunction-associated steatotic liver disease (MASLD), characterised by the accumulation of excess fat in the liver of individuals who consume little or no alcohol, has emerged as a global public health concern [1]. Its epidemiology reflects a troubling upward trend, mirroring the increasing prevalence of obesity and diabetes mellitus (DM) [2, 3]. As the most common chronic liver condition in the Western world, MASLD encompasses a spectrum of liver disorders, ranging from simple steatosis to non-alcoholic steatohepatitis, potentially progressing to cirrhosis and hepatocellular carcinoma [4, 5, 6]. The insidious nature of MASLD, of being often asymptomatic in its early stages, underscores the critical need for its early detection and intervention [7, 8, 9]. Furthermore, the multifactorial aetiology of the disease, involving complex interactions between genetic predisposition, lifestyle factors, and metabolic syndrome, poses significant challenges to its management and risk stratification [10, 11]. Thus, early and accurate risk assessment of MASLD is not only pivotal for timely therapeutic interventions but also paramount in preventing its progression to more advanced liver diseases [12, 13].

Contemporary diagnostic modalities for MASLD predominantly pivot around transient elastography (TE), a technique acclaimed for its diagnostic precision. TE, leveraging its ability to quantify liver stiffness, serves as a pivotal instrument in not only diagnosing MASLD but also in stratifying the disease severity [14, 15]. Despite its efficacy, TE relies on sophisticated equipment and the necessity of a clinical setting for implementation poses constraints, especially in terms of accessibility and patient compliance [16]. Similarly, traditional diagnostic approaches, including clinical evaluations, biochemical markers, and liver biopsies, underscore the dependency on patient cooperation and the need for specialised medical infrastructure [17]. Such methodologies, although integral, often fall short of convenience and patient adherence, impeding widespread screening and early detection efforts [18]. The inherent limitations of these approaches underscore the need for alternative and more accessible diagnostic strategies that can circumvent the aforementioned barriers and foster enhanced patient engagement in disease management [19].

The burgeoning field of machine learning (ML) presents a paradigm shift in medical diagnostics, offering a potent amalgam of precision and adaptability that traditional methodologies often lack [20, 21]. ML algorithms exhibit substantial potential in the realm of MASLD, primarily for refining risk stratification and prognostic assessment [22]. These algorithms can adeptly analyse complex multidimensional datasets and transcend the limitations of human cognition. In particular, methods such as random forest (RF), support vector machines (SVM), and neural networks have demonstrated remarkable capabilities in discerning intricate patterns and interactions among clinical variables [23]. This process not only enhances predictive accuracy but also facilitates the identification of novel risk factors and patient subgroups, thus enabling a more nuanced understanding of MASLD pathophysiology [24]. The adoption of ML in MASLD risk assessment could revolutionise the approach for early detection, allowing for timely intervention and personalised management strategies [25]. Moreover, the ability of ML models to process and interpret large-scale data swiftly aligns seamlessly with the demands of contemporary clinical practice, heralding a new era in the efficient and effective management of MASLD [24, 26].

Several recent studies have demonstrated promising applications of ML models in the screening and diagnosis of MASLD, yielding encouraging results. For instance, Chen et al. developed a risk prediction model based on multiple potential indicators, including homeostasis model assessment of insulin resistance (HOMA-IR), triglyceride glucose-waist circumference (TyG-WC), and achieved an impressive area under the receiver operating characteristic curve (AUC) of 0.960 [27]. Similarly, McTeer et al. used European MASLD registry data to construct various ML models using conventional clinical parameters, with AUC values ranging from 0.719 to 0.994, highlighting the utility of readily available clinical information in predicting MASLD stages and outcomes [28]. Nabrdalik et al. focused on ML-based recognition methods for MASLD in patients with diabetes, constructed a model using eight core parameters, and achieved an AUC of 0.84, demonstrating the potential of the model for cardiovascular risk prevention [29]. Despite the advancements presented by these studies, reliance on complex clinical or laboratory indicators, such as HOMA-IR and lipid-related parameters, limits the feasibility of MASLD screening, particularly in resource-limited settings or for large-scale population assessments. Although these advanced biomarkers are valuable in clinical practice, they are often not readily accessible for general screening. In contrast, this study adopted an innovative approach using a minimal set of easily accessible demographic and clinical factors, such as age, height, and weight, to develop a ML-based screening model. This strategy enhances the scalability and accessibility of MASLD detection, making it more applicable in diverse healthcare settings, improving early detection in underserved populations and facilitating timely stratified diagnosis, treatment, and intervention.

Leveraging ML models built on easily obtainable parameters, such as age, sex, body mass index (BMI), and chronic disease history, has emerged as a pivotal strategy to enhance accessibility and patient compliance in MASLD risk assessment [30]. This approach democratises the diagnostic process by transcending the limitations of clinical settings and specialised equipment. By integrating these readily available and self-measurable indicators within an ML framework, the potential to substantially increase diagnostic convenience and patient adherence can be realised. Furthermore, the adoption of such models in community settings amplifies screening ubiquity, making MASLD risk assessment more inclusive and accessible to the general populace [31, 32]. This paradigm fosters early detection through widespread accessibility and aligns with patient-centred care principles. The operational simplicity and non-reliance on invasive procedures inherent in this methodology substantially reduce barriers to screening, thereby paving the way for timely interventions and mitigating the progression of MASLD. Embracing this innovative approach could mark a significant milestone in MASLD management, aligning with the contemporary shifts towards preventative health strategies and personalised medicine.

This study investigated the efficacy of 10 ML algorithms in predicting MASLD prevalence using a combination of basic demographic characteristics and essential clinical indicators, including age, educational level, height, weight, waist and hip circumference, and history of hypertension and diabetes. Using a comprehensive outpatient cohort and external validation with National Health and Nutrition Examination Survey (NHANES) data, we sought to establish a feasible and accessible approach for MASLD screening in large populations.

Methods

Study design and participants

This prospective study, conducted at the First Affiliated Hospital of Gannan Medical University, enrolled 10,007 outpatients between October 2020 and February, 2024. Eligible participants were aged 15 years or older and of any sex. The participants provided informed consent after understanding the study’s objectives, methodology, and importance. Individuals were excluded if they had severe cardiovascular or cerebrovascular diseases or malignant tumours, were pregnant, or were unable to adhere to the study protocol, including completing questionnaires and undergoing TE (Fig. 1). The study was approved by the relevant medical ethics committee and all participants provided written informed consent.

Fig. 1

Flowchart of this study

Demographic data collection

Comprehensive data collection involved direct interviews and questionnaires, capturing demographic details (age, sex, educational level, and occupation); a thorough review of diabetes diagnoses, including type, duration, and treatment modalities; and medical history with a focus on hepatitis B virus (HBV) and hepatitis C virus (HCV) infections and other liver diseases. Physical assessments included height, weight, BMI and hypertension status. Detailed definitions and descriptions of all study variables are provided in Table S1.

MASLD assessment

Liver fat alterations were evaluated using FibroScan (Echosens, France) for TE. This non-invasive technique measures the controlled attenuation parameter (CAP) to quantitatively evaluate liver fat content [33]. Participants fasted for at least 3 h prior to the examination. Trained and certified medical technicians operated the FibroScan TE device to measure the CAP [34]. Measurements were performed with the participants in the supine position, with their right arm behind the head, and a slight leftward tilt of the upper body to maximally expand the intercostal space. The measurement area ranged from the right anterior axillary line to the mid-axillary line between the seventh and eighth, or eighth and ninth ribs, maintaining a probe orientation perpendicular to the skin surface. Quality control standards for each participant required at least 10 valid CAP measurements, with a quartile-to-median ratio ≤ 30%, taking the median value as the final measurement.

CAP was used to evaluate liver cell fat content. MASLD was classified as follows: none or mild: CAP ≤ 269 dB/m; moderate: 269–296 dB/m; severe: CAP > 296 dB/m.

Data preprocessing and feature selection

In this study, the missForest algorithm was employed to handle the missing data (< 25%) in both the original and NHANES datasets. This approach efficiently handles both continuous and categorical variables by leveraging the observed values from other variables to predict the missing variables. By imputing missing data, missForest preserves data completeness and ensures the reliability of subsequent analyses. Table S2 summarises the proportion of missing data for each variable, and the corresponding imputation methods applied to both datasets.

For feature selection, we employed a two-step approach combining least absolute shrinkage and selection operator (LASSO) regression and the Boruta algorithm. LASSO regression, which is effective in handling multicollinearity and enhancing model interpretability, was applied to refine the variables by reducing the coefficients of less informative predictors to zero. The selection process was guided by the mean squared error and alpha values to optimise the model performance (Figure S1A–B). Subsequently, the Boruta algorithm, a robust and stable method, was used to identify relevant variables by comparing the importance of predictive variables with that of randomly permuted ‘shadow variables’ through multiple iterations of the RF algorithm, systematically excluding non-essential variables (Figure S1C). The intersection of the variables selected using both methods was used as the foundation for subsequent analyses.

ML algorithms and model development

To predict MASLD risk stratification, we employed 10 state-of-the-art ML tools for developing predictive models: logistic regression (LR), multilayer perceptron (MLP), extreme gradient boosting (XGBoost), bootstrap aggregating (Bagging), decision tree (DT), K-nearest neighbours (KNN), light gradient boosting machine (LightGBM), naive Bayes (NB), RF, and SVM. The dataset was divided into training (70%), validation (20%), and test (10%) sets. All analyses were conducted using Python 3.8.10 and R 3.3.2. LR, a multilinear regression method, utilises maximum likelihood estimation to gauge predictive probabilities. MLP, a neural network-based algorithm, captures complex non-linear relationships through interconnected layers and optimises weights via backpropagation. Bagging improves stability and accuracy by generating multiple training subsets through resampling and aggregating predictions from individual models. KNN classifies instances by measuring the proximity in the feature space, relying on majority voting among the nearest neighbours to make predictions. NB, which is a probabilistic model based on Bayes’ theorem, assumes feature independence and demonstrates robust performance with categorical and continuous data. SVM classifies data points by maximising the margin in a transformed dimensional space. DT is a structured tree predictive model that selects optimal features and splits points based on the Gini index. RF, which employs DTs as base classifiers, recursively builds multiple trees, with the final prediction determined by a voting mechanism. XGBoost, another tree-based ensemble classifier, minimises the current loss function by generating new decision trees until the residual of the loss function approaches zero, by employing techniques such as second-order derivatives, Taylor series, and parallel learning. LightGBM, an advanced gradient boosting DT algorithm, enhances the performance through parameter optimisation, including adjustments in iteration numbers, maximum leaves, and tree depth. After the initial development of the MASLD risk prediction models, hyperparameter tuning was conducted for all 10 ML algorithms to optimise model performance. Specifically, Bayesian optimisation was selectively applied to models with a higher number of hyperparameters and more complex tuning requirements, including XGBoost, LightGBM, RF, SVM, and MLP, owing to its efficiency in navigating high-dimensional parameter spaces. For simpler models, such as LR, NB, DT, KNN, and Bagging, hyperparameters were optimised using grid search or random search because their parameter spaces were relatively limited and computationally less intensive. The final hyperparameters adopted for each model are listed in Table S3.

Prediction model evaluation

To comprehensively assess the performance of the ML models in predicting MASLD risk, a suite of evaluation metrics was employed. The primary metric used was the AUC, which provides a measure of the model’s ability to distinguish between different risk categories. A higher AUC value indicates a greater discriminatory power of a model. Additional metrics included accuracy (ACC, the proportion of true results among the total number of cases examined), sensitivity (ability of the model to correctly identify individuals with MASLD), specificity (ability of the model to correctly identify individuals without MASLD), precision (the proportion of true positive predictions among all positive predictions, reflecting the reliability of the model in identifying individuals with MASLD), and F1 score (a harmonic mean of precision and recall, providing a balance between sensitivity and specificity). These metrics collectively offer a multifaceted view of the effectiveness of the models, ensuring a thorough and nuanced evaluation of their predictive capabilities in the context of MASLD risk stratification.

External validation using NHANES data

NHANES data from a comprehensive cross-sectional population-based study were utilised to supplement our findings. By covering the cycles from 2017 to 2023, the NHANES provides crucial insights into the health and nutritional status of the U.S. population. The NHANES protocol, approved by the National Center for Health Statistics Research Ethics Review Board, ensured rigorous ethical standards, and all participants provided written informed consent. The NHANES data from 2017 to 2023 were used for external validation. We extracted demographic data (age, sex, and educational level) and disease history (hypertension and DM) from the NHANES data. The physical examination metrics included height, weight, BMI, waist and hip circumference, liver stiffness measurement, and CAP. Detailed mapping and descriptions of the variables used in this study are provided in Table S4. This table outlines the variable codes, predictors, descriptions, types, and value ranges, ensuring traceability and comparability of the data across the NHANES datasets.

Statistical analysis

All statistical analyses were conducted using SPSS software (version 23.0), R software (version 3.3.2), and Python software (version 3.8.10). Continuous variables were presented as mean ± standard deviation for normally distributed data and as median with interquartile range for non-normally distributed data. Categorical variables were summarised as percentages and frequencies. Comparative analyses between the groups were performed using non-parametric tests for continuous variables (Mann–Whitney U test), as all continuous data in this study were non-normally distributed. For categorical variables, the chi-squared test was used to compare group differences. Statistical significance was set at P < 0.05.

Results

Demographic characteristics of participants

This study enrolled 10, 007 outpatients who were categorised into distinct groups: a training cohort comprising 7,003 individuals, an internal validation cohort with 2,002 participants, and an internal testing cohort containing 1,002 participants. The allocation followed a stratified randomisation approach, at a ratio of 7:2:1, ensuring that each subgroup retained a similar distribution of relevant features. Detailed descriptions of the baseline demographics, physical evaluations, and clinical diagnoses of each cohort are summarised in Table 1. In the evaluation of the demographic variables, significant disparities emerged between the MASLD and non-MASLD groups. For ease of reference, detailed data are presented in Table S5. The median age was higher in the MASLD group (52 [40–58] years) than in the non-MASLD group (49 [37–57] years; P < 0.001). Sex distribution also differed significantly, with female individuals comprising 56.21% (n = 3235) and 68.72% (n = 2922) of the MASLD and non-MASLD groups, respectively (P < 0.001). Education level showed a notable variation between the groups, with a higher proportion of individuals with low education (12.27% vs. 9.05%) and a lower proportion of those with higher education (25.49% vs. 29.70%, P < 0.001) in the MASLD group compared with the non-MASLD group. Occupational distribution further highlighted the intergroup differences, with a slightly higher percentage of the labour force in the MASLD group than in the non-MASLD group (77.15% vs. 75.16%, P = 0.008). Physical examination data demonstrated that the MASLD group exhibited significantly higher median values for height (164 [158–170] cm), weight (67 [60–75] kg), BMI (24.91 [23.23–27.06] kg/m²), waist circumference (85 [80–91] cm), and hip circumference (97 [90–101] cm), compared with the non-MASLD group (all P < 0.001). Additionally, the MASLD group exhibited a higher prevalence of DM (4.07%) than the non-MASLD group did (3.10%, P = 0.013), further highlighting the association between MASLD and DM. Similarly, hypertension was noted significantly more frequently in the MASLD group (27.72%) than in the non-MASLD group (17.03%, P < 0.001), highlighting the metabolic and cardiovascular risk profile associated with MASLD. Notably, the prevalence of HBV and HCV infections did not differ significantly between the two groups; HBV was observed in 1.60% of the MASLD cohort and 1.70% of the non-MASLD cohort (P = 0.828) and HCV in 0.10% and 0.20%, respectively (P = 0.528). These findings collectively underscore the distinctive demographic and clinical profiles of MASLD and non-MASLD populations.

Table 1 Baseline characteristics and outcomes of MASLD patients in the training, internal validation and internal testing cohort

Feature selection

The two-step feature selection approach identified a robust predictor set for the MASLD. LASSO regression and the Boruta algorithm yielded a combined set of statistically significant variables, including age, educational level, height, weight, BMI, waist and hip circumference, and history of hypertension and DM (Figure S1D). Notably, although BMI was initially identified as relevant, it was excluded from the final model because of its high correlation with height and weight, which could lead to multicollinearity issues in ML modelling.

Development of ML models for MASLD prediction

We employed 10 ML algorithms based on demographic characteristics to construct diagnostic models for MASLD. These include LR, MLP, XGBoost, Bagging, RF, NB, LightGBM, SVM, KNN, and DT. To further investigate the contribution of each feature across the different models, we assessed the variable importance rankings derived from all 10 ML algorithms (Figure S2). LR emerged as the most effective, achieving an ACC of 0.728, a sensitivity of 0.749, and an AUC of 0.806 in the internal validation cohort (Fig. 2A; Table 2). XGBoost and MLP also exhibited commendable performances, with XGBoost slightly trailing LR in terms of sensitivity. The Bagging model, although comparable in ACC to the LR model, had a lower sensitivity. Conversely, DT demonstrated the least effectiveness, with the lowest ACC and sensitivity, underscoring its limitations. In the internal testing cohort, LR maintained its lead, with an ACC of 0.711 and an AUC of 0.798. Other models, such as XGBoost and MLP, also demonstrated robust performance, with AUCs of 0.798 and 0.796, respectively (Fig. 2B). To further assess the classification efficacy of the LR model, confusion matrices were generated for both the validation and test cohorts (Figures S3A–B). These matrices provide a detailed breakdown of true positives, false positives, true negatives, and false negatives for the model. The findings indicate that the LR model effectively identified individuals with MASLD (true positives) and those without MASLD (true negatives). Notably, the relatively low incidence of false positives and false negatives underscore the reliability of the model’s clinical risk prediction, minimising the likelihood of both over-diagnosis and missed cases. These results align with the model’s strong performance metrics, including high AUC and F1 scores, further reinforcing its potential for clinical application. Additionally, pairwise comparisons of the AUCs between LR and alternative classification models, as detailed in Tables S6–S8, substantiate the statistical significance of performance differences, underscoring the robustness of the LR model in predictive accuracy.

Fig. 2

Prediction performance by AUCs for the ten machine-learning models in the internal validation cohort (A), internal testing cohort (B) and external validation cohort (C)

Table 2 Performance of various machine learning models for predicting MASLD across internal validation, internal testing, and external validation cohorts

A comprehensive decision curve analysis (DCA) conducted on the internal validation cohort demonstrated the superior clinical utility of the LR model among various ML approaches, reinforcing its applicability in real-world settings (Figure S4A). Similarly, DCA results for the internal testing cohort confirmed that using the LR model for risk prediction yielded substantial net benefits (Figures S4B). Figure S5A and B depict the calibration curves for different models in the internal validation and testing cohorts. The LR model exhibited strong calibration for all datasets, with its calibration curves closely aligning with the ideal 45-° reference line, indicating high concordance between predicted and actual event probabilities. In contrast, the DT and NB models displayed pronounced calibration biases, particularly in higher probability ranges, where their predictions deviated markedly from observed outcomes.

Furthermore, the analysis of the precision-recall (PR) curves across the models (Figure S6A–B) further highlights the superior performance of the LR model. PR curves are particularly useful for assessing models on imbalanced datasets, as they emphasize the trade-off between precision (positive predictive value) and recall (sensitivity). The LR model demonstrated high precision while maintaining strong recall, effectively minimizing false positives without compromising sensitivity. These findings underscore the robust discriminatory capability and reliability of the LR model for clinical risk prediction in imbalanced clinical datasets, making it the optimal choice for this application.

External validation of the predictive models

We used the NHANES dataset (2017–2023) as an external validation cohort to evaluate the predictive performance of the models. A comprehensive baseline analysis was performed to compare demographic, clinical, and liver-related characteristics between the MASLD and non-MASLD groups (Table S9). Key findings indicated that patients with MASLD exhibited significant differences across multiple dimensions, compared with individuals without MASLD. Demographically, patients with MASLD were older (median age: 55 vs. 40 years, P < 0.001), more likely to be male (50.73% vs. 45.16%, P < 0.001), and had higher rates of low or secondary education (P < 0.001), compared with individuals without MASLD. Anthropometric measurements further highlighted the substantial disparities. Patients with MASLD had significantly elevated BMI (median: 31.10 vs. 24.50 kg/m², P < 0.001), waist circumference (105.90 vs. 86.90 cm, P < 0.001), and hip circumference (109.10 vs. 98.30 cm, P < 0.001), compared with individuals without MASLD. These differences were accompanied by higher weight and height, underscoring the metabolic profile of MASLD. Clinically, patients with MASLD demonstrated a markedly higher prevalence of hypertension (42.20% vs. 22.94%, P < 0.001) and DM (17.94% vs. 5.87%, P < 0.001).

Post-imputation analysis (Table S10) corroborated these findings, with similar trends observed for demographics and metabolism. Patients with MASLD were significantly older, had a higher BMI and waist circumference, and had a greater prevalence of related conditions, compared with individuals without MASLD. These results provide strong evidence of the distinct demographic and clinical profiles of patients with MASLD, emphasising the importance of these variables in predictive modelling and their relevance for external validation using real-world datasets.

External validation confirmed the robustness and generalisability of our predictive models, with LR emerging as the best-performing model. The LR model achieved an AUC of 0.831, outperforming the other models, as demonstrated by the receiver operating characteristic curves (Fig. 2C). This indicated a strong ability to distinguish patients with MASLD from individuals without MASLD in the NHANES dataset. We conducted a series of complementary analyses to further validate the performance of the model. The confusion matrix (Figure S3C) demonstrated high ACC, whereas DCA (Figure S4C) confirmed its clinical utility with a favourable net benefit across various thresholds. Moreover, the calibration curve (Figure S5C) showed excellent agreement between the predicted and observed outcomes, and the PR curve (Figure S6C) indicated robust performance, even in the presence of class imbalance. These results collectively validated the LR model as a reliable tool for external datasets. Collectively, these results highlight the robust performance and clinical applicability of the LR model, making it a reliable tool for identifying MASLD in external real-world datasets.

Visualization by SHapley additive explanations (SHAP)

Figure 3B illustrates the importance of the SHAP features for the LR model. The features under scrutiny were arranged in the descending order of their influence on the projected outcomes, as indicated by the mean absolute value of SHAP. The top five pivotal features were weight, height, waist circumference, age, and educational level. The SHAP summary plot further illustrates the impact of these variables on model predictions (Fig. 3A). Notably, higher SHAP values associated with certain features indicated an increased predisposition to developing MASLD. For instance, patients exhibiting aberrant weight were found to be predisposed to a higher incidence of MASLD, compared with those with normal weight.

Fig. 3

(A) SHAP summary plot showing the distribution of the SHAP values of each feature; (B) SHAP feature importance shown according to the mean absolute SHAP value of each feature

Interpreting ML models at the patient level

We utilized the SHAP method for individual patient predictions and assess the influence of the LR model on specific patient features. Feature contributions are visually represented by colour coding, where red signifies a positive contribution and blue denotes a negative contribution. For true-positive patients (Figure S7A), the LR model inferred a relatively high probability of developing MASLD, with key features such as weight, height, and educational level exerting a predominant positive influence. Conversely, for true-negative individuals (Figure S7B), the model predicted a lower probability of MASLD, where features such as weight, waist circumference, age, and height exhibited a contrasting contribution pattern.

To further elucidate the interpretability of the LR model, we performed a local interpretable model-agnostic explanation (LIME) analysis to evaluate the contributions of individual features to true-positive (Figure S8A) and true-negative (Figure S8B) predictions. In true-positive predictions, the LR model assigned a high probability (0.87) of developing MASLD (Figure S9A). Features such as height, weight, and waist circumference were identified as the most influential contributors to positive predictions. Negative influences, such as DM and age, had relatively smaller effects on the model output. For true-negative predictions, the model assigned a low probability (0.10) for MASLD (Figure S9B), which was primarily driven by negative contributions from weight and waist circumference. Minimal positive contributions, such as from height and educational level, were insufficient to offset the dominant negative factors. The combined LIME and SHAP analyses consistently highlighted weight and waist circumference as key predictive factors, with the results cross-validating each other and demonstrating the robustness and reliability of the LR model for personalised MASLD risk prediction.

Association between predictors and predicted probability

We utilised restricted cubic splines (RCS) to assess the non-linear relationships between the predictors and the predicted probability of MASLD, adjusting for potential confounders. The analyses demonstrated significant overall associations between the predictors (age, height, weight, BMI, and waist and hip circumference) and the outcomes, as evidenced by p-values of < 0.001 for the overall effects across all variables (Figure S10).

To further explore the association between the predictors and MASLD risk, continuous variables identified in the RCS analysis were dichotomised using clinically relevant cutoff values. A forest plot was generated to present the odds ratios (ORs) and 95% confidence intervals (CIs) for each subgroup (Figure S11). When age was dichotomised using a cut-off of 28 years, individuals aged > 28 years showed a significantly higher risk of MASLD, compared with those aged ≤ 28 years (OR: 1.409, 95% CI: 1.174–1.691, P < 0.001). Similarly, weight > 56.5 kg was associated with a substantially increased risk of MASLD (OR: 4.887, 95% CI: 4.332–5.512, P < 0.001). Height, dichotomized using a cut-off of 158.5 cm, showed an inverse relationship, with individuals taller than the cut-off exhibiting a significantly reduced risk of MASLD (OR, 0.621; 95% CI, 0.551–0.701; P < 0.001), compared with those shorter than the cut-off. BMI > 29.4 kg/m² demonstrated the strongest association, conferring an eight-fold increased risk of MASLD (OR, 8.742; 95% CI, 5.541–13.792; P < 0.001). In addition, larger waist circumference (> 99 cm) and hip circumference (> 85 cm) were both significantly associated with a higher MASLD risk, with ORs of 3.217 (95% CI: 2.364–4.378, P < 0.001) and 1.915 (95% CI: 1.609–2.279, P < 0.001), respectively.

Construction and clinical application of an online prediction tool

Building upon the previously established LR model, we developed an online calculator (https://lalalaanjila.shinyapps.io/MASLD_Risk_Prediction_Calculator/) to estimate the probability of MASLD. The tool integrates multiple clinical variables and applies the logit transformation to individual input data, generating a predicted risk probability for MASLD. By offering a user-friendly and precise platform for personalized risk assessment (Figure S12). Compared with conventional scoring systems, this calculator not only improves the predictive accuracy but also streamlines the assessment process, facilitating more efficient individualized treatment strategies. Preliminary validation demonstrated that its performance on the validation set was in line with the original expectations of the model, further supporting its potential for clinical implementation.

Discussion

This study highlighted the application of various ML algorithms to accurately predict the prevalence of MASLD using basic demographic characteristics. The novel approach of utilising 10 distinct algorithms—SVM, MLP, XGBoost, Bagging, KNN, LightGBM, NB, RF, DT, and LR—allowed us to identify the most effective model for MASLD screening.

Each of the tested algorithms demonstrated unique strengths suited to specific application scenarios. LR is widely used due to its simplicity and interpretability; DT provides a transparent decision-making process; RF enhances predictive stability and reduces variance by aggregating multiple DT; SVM excels in handling high-dimensional spaces and capturing complex decision boundaries; KNN performs well in non-parametric classification with well-separated data distributions; MLP, as a neural network approach, captures non-linear patterns effectively; XGBoost and LightGBM demonstrate strong predictive power and computational efficiency, especially in large-scale datasets; Bagging improves model robustness through ensemble learning, and NB is advantageous for probabilistic classification tasks, particularly when feature independence assumptions hold. By comparing the performance metrics of these models, including the ACC, sensitivity, specificity, and AUC, we were able to determine the optimal algorithm for this task. The LR model emerged as the most effective, achieving an impressive ACC of 0.728, a sensitivity of 0.749, and an AUC value of 0.806. This superior performance underscores the capability of LR to balance the classification of positives and negatives, making it a robust tool for early MASLD screening and a crucial reference for future studies.

Furthermore, this study introduced an innovative home-based self-assessment tool for early MASLD detection. This tool, designed to be user-friendly and suitable for domestic use, employs a straightforward questionnaire combined with non-invasive indicators such as age, educational level, height, weight, waist and hip circumference, and history of hypertension and diabetes. Validation in large-scale population studies has demonstrated its efficacy and reliability, offering a practical solution to reduce healthcare costs and enhance patient adherence through early screening and intervention.

The study’s utilisation of a large and diverse cohort of 10,007 participants, along with external validation using NHANES data from 2017 to 2023, underscores the robustness and generalisability of our findings. The dataset was systematically partitioned into training, internal testing, and internal validation cohorts, whereas external validations confirmed the performance of the model across various metrics. This extensive cohort not only bolsters the reliability of the results but also enhances the representativeness of the study, providing a dependable tool for early MASLD screening and serving as a significant reference for future research.

In the landscape of escalating global incidence of MASLD, exacerbated by demographic shifts towards an aging population and changing dietary habits [1], our study addresses a critical gap in early screening and intervention strategies. The current diagnostic and monitoring protocols for MASLD, which are predominantly centralised in specialised medical centres, involve intricate, often costly, and sometimes invasive procedures [35]. This complexity hinders widespread compliance and proactive engagement in MASLD screening, especially when asymptomatic individuals are considered [36]. Therefore, there is an urgent need to develop an accessible and accurate predictive mechanism for MASLD prevalence based solely on fundamental demographic characteristics. Simplifying the diagnostic paradigm to leverage basic baseline features, such as age, educational level, height, weight, and waist and hip circumference, without the need for specialised medical consultations could revolutionise public adherence to liver health screening. As evidenced by our research, this technology has the potential to significantly enhance the ubiquity and compliance of MASLD screening across populations, thereby mitigating the burgeoning burden of liver health issues globally.

Our research, harnessing a selection of robust ML algorithms, demonstrates the feasibility of employing basic demographic and clinical characteristics to predict MASLD in the general population. The significant influences of age, educational level, height, weight, waist and hip circumference, and history of hypertension and diabetes on the development and progression of MASLD were empirically validated. The power of ML to distil these common demographic attributes into actionable insights enables individuals to manage their liver health proactively with minimal intrusion into daily life. Our findings confirmed that ML models can effectively classify the severity of MASLD, with impressive and robust AUC results. We stratified the MASLD group into three categories, offering tailored intervention strategies: (1) individuals with mild or no MASLD, advised to maintain healthy dietary habits and engage in regular physical activity; (2) individuals with moderate MASLD, necessitating more stringent lifestyle and dietary changes, potentially including significant weight management to reduce overall body weight by more than 10%, a crucial factor for improving hepatic fat accumulation [37, 38], along with regular monitoring of liver function and other related parameters; and (3) individuals with severe MASLD, who should seek comprehensive evaluation at specialized liver centres, including liver imaging and biochemical monitoring, with potential recommendations for medications to regulate blood sugar, fat, and cholesterol levels [39]. This stratification not only streamlines the screening process but also equips healthcare providers with a nuanced approach to effectively manage MASLD at various stages [40].

The validation process revealed the high accuracy and stability of the LR model, which were further corroborated by internal and external validations. The LR model excelled across key performance indicators, consistently outperforming other models in terms of ACC and AUC values, and demonstrating robust generalization capabilities across diverse datasets. These findings provide strong support for the potential of the model for early MASLD detection and intervention, paving the way for future research and clinical applications.

Although our study presents significant advancements in predicting MASLD using basic demographic characteristics, it is important to acknowledge its limitations and scope for further refinement. First, as this was a retrospective study, the selection of outpatient participants may have introduced an inherent bias. Future research should aim for larger-scale prospective studies to validate and replicate these findings. Second, the absence of venous blood sampling for all participants resulted in a lack of comprehensive assessment of relevant liver functions and lipid profiles, which is crucial for a holistic understanding of MASLD [41]. Thirdly, we relied on TE for diagnosing MASLD, as opposed to histopathological biopsy; although practical and non-invasive, TE may not provide the complete picture of the liver condition [15, 42]. This approach is particularly justified in outpatient settings where symptoms are not pronounced and invasive liver biopsies are impractical. Fourth, the potential presence of unrecorded or undiagnosed cases of DM among the screened individuals may have introduced a degree of selection bias, affecting the study results. Finally, the black-box nature of some algorithms can pose challenges in terms of interpretability and transparency, which are critical in medical decision-making. Despite these limitations, the strengths of this study cannot be overlooked. By leveraging simple demographic features, we successfully developed ML models that not only accurately predict MASLD but also stratify patients based on disease severity. This stratification facilitates more personalized intervention and monitoring [43, 44, 45]. Our findings significantly enhance the feasibility and compliance of MASLD screening in the general population, thereby strengthening the public focus on liver health.

Conclusions

This study conclusively demonstrated that ML models, particularly the LR algorithm, effectively utilised a limited number of demographic and readily available clinical characteristics to accurately screen for MASLD in a broad population. This approach significantly enhances the feasibility and compliance with MASLD screening in general healthcare settings.