Metabolic dysfunction–associated fatty liver disease (MAFLD) is a leading cause of chronic liver disease, affecting more than one-third of the global population [1,2] and resulting in annual, direct medical costs of US $103 billion in the United States and €35 billion (US $40 billion) in Europe [3]. The relabeling of nonalcoholic fatty liver disease as MAFLD reflects a deeper understanding of fatty liver disease [4]. It also helps identify adults at risk of serious prognoses [5] such as liver cirrhosis and hepatocellular carcinoma, which account for most liver-related deaths [6,7]. The exacerbation of comorbid conditions due to MAFLD amplifies its clinical significance; patients with chronic liver diseases often develop severe infections, chronic cardiovascular or kidney disease, cancer, and death [8]. Yet, therapeutic options for devastating MAFLD-induced liver diseases are limited. Liver transplantation is the optimal treatment [9,10] but is greatly restricted by organ availability and financial costs [11].

A diagnosis of MAFLD requires hepatic steatosis in the presence of excessive weight, type 2 diabetes mellitus, or metabolic dysregulation, manifested in the obese, diabetic, and lean phenotypes (subtypes) of MAFLD, respectively [12]. These phenotypes have distinct prognostic values [5], fibrotic burden [13,14], and complications [15,16]. For example, the diabetic phenotype is characterized by severe insulin resistance and is associated with the highest risk of any-cause and disease-specific mortality [17]. The obese phenotype is related to lifestyle factors (eg, diet and physical inactivity) and can lead to systemic inflammation and metabolic dysfunction. The lean phenotype involves ectopic fat deposition and genetic predispositions to MAFLD, although without obesity [18]. Because of the differences between the MAFLD phenotypes, accurate phenotype prediction is crucial for clinical decision-making, personalized care planning, and efficient resource allocation [19]. With relevant insights into the underlying etiology and pathology [20], effective phenotype prediction can facilitate patient stratification and treatment planning for streamlining diagnostic procedures, optimizing the use of laboratory tests or imaging, and specifying necessary lifestyle changes, all of which have cost-containment implications [21-23].

Physicians usually rely on liver biopsies [24,25] or score-based methods [26,27] that require contemporaneous clinical data, impose substantial costs, and misidentify at-risk adults. These constraints favor the potential of data-driven analytics for supporting timely identification of at-risk adults such that clinicians can formulate actionable risk reduction measures and effective patient stratification and researchers can design more appropriate clinical trials and treatment plans [28]. Despite the promise, data-driven analytics for MAFLD phenotyping face several challenges. First, MAFLD is a multisystem disease [29] because clinical, family genetics, lifestyle, and socioeconomic factors can influence fatty liver development and progression [30]. Incorporating such heterogeneous data in analytics methods, which typically are gathered from different sources, is difficult. For example, surveys designed to gather genetic family history data or lifestyle data tend to have small samples and often suffer from data incompleteness. Second, due to the complex nature of MAFLD, people with the same MAFLD phenotype may exhibit intraphenotype variability in etiology or pathology, which also should be considered for phenotype predictions. Third, both disease classification hierarchy and manifestations of MAFLD involve temporal complexity at the individual level.

In an effort to design a data-driven method to predict MAFLD phenotypes more accurately, we developed a novel, 2-stage, contrastive learning–based method. This method leverages graph representation learning, in combination with interindividual similarity, to process integrated individual-level data pertaining to genetic family history or lifestyle, which then can inform downstream predictions by complementing (incomplete) survey-based data with clinical data or vice versa. In addition, the proposed method incorporates multiview, contrastive pretraining that captures intraphenotype variability on the basis of clinical, genetic family history, and lifestyle data. By linking important data from different sources, it constructs individual-level representations for downstream tasks and predictions. Finally, its 2-stage estimation design accounts for disease hierarchy and temporal complexity, such that the proposed method can predict phenotypes among adults more accurately and explicitly than the existing analytics methods.

To demonstrate the predictive efficacy of the proposed method, we used clinical and survey-based data of 4408 adults in Taiwan [31] and included 8 prevalent methods as benchmarks. The results indicated that the proposed method consistently and significantly outperformed all the benchmarks in both F₁-score and area under the curve (AUC). This novel method can predict phenotypes accurately and can potentially contribute to medical informatics research and support personalized care for at-risk adults.

We used 2-year longitudinal data of 4408 adults, obtained from a major healthcare organization in Taiwan, to evaluate the proposed method in comparison with 8 prevalent methods. No adults in the sample had MAFLD in year 1. For each person, the data include 2 demographic variables, 36 clinical variables, 32 lifestyle variables, and 42 genetic family history–related variables. Multimedia Appendix 1 provides the description and coding of variables. With these data, we evaluated the ability of each method to predict whether a person would develop MAFLD in year 2 and, if so, of which phenotype.

Of the 4408 individuals in our sample, 2999 (68.1%) were women, and 1409 (31.9%) were men, with an average age of 58.18 (SD 12.94) years. The outcome class distribution was imbalanced: 85.0% non-MAFLD (3747/4408), 11.5% nondiabetic MAFLD (507/4408), and 3.5% diabetic MAFLD (154/4408). We used class weights during model training to address the imbalance issue. Prior to making phenotype predictions, we applied z score standardization to numeric variables and one-hot encoding to categorical variables to prepare the data.

This study was approved by the Chang Gung Medical Foundation Institutional Review Board (201800270B0). All procedures were performed in accordance with relevant guidelines and regulations. Written informed consent was obtained from all participants. All patient information was anonymized prior to analysis, and the study complied with ethical standards for research involving deidentified healthcare data. Participants were informed that their involvement was voluntary and that they could withdraw from the study at any time without penalty. No financial compensation was provided.

Eight prevalent methods were included as benchmarks: DT [44], RF [45], XGBoost [46], MLP [48], autoencoder [49], GAT [57], GCN [56], and GraphSAGE [58]. These methods represent different analytics approaches and are frequently used for clinical prediction tasks; therefore, they are suitable for performance comparisons. Many of these benchmark methods are not designed to deal with incomplete data. Because the sample had missing values, we applied k-nearest neighbor (k=5) imputation [76-78] to the dataset and used one-hot encoding for categorical variables during data preprocessing to ensure consistency and comparability in the evaluations, that is, all methods used the same preprocessed data for fair comparisons. The only difference was that the proposed method also used the raw, nonimputed survey data (genetic family history and lifestyle data) as input for graph representation learning and contrastive learning, which are components capable of handling missing values. Moreover, we conducted an ablation study to examine the relative contribution of each key component to the proposed method’s overall performance.

To examine the prediction performance of each method, we randomly split the sample 10 times, using different random seeds to ensure robustness. In each trial, we used 80% of the data for model training and the remaining 20% for testing [76]. We also conducted 5-fold cross-validation on the training data prior to the evaluations and performed a series of analyses to fine-tune the key parameters of each method. Multimedia Appendix 2 summarizes important parameter values of the respective methods. Performance assessments relied on precision, recall, F₁-score, and AUC values. We did not consider accuracy, as it could not reflect prediction performance due to the imbalanced distribution of the outcome classes [79]. Compared with precision or recall, the F₁-score and AUC are arguably better indicators of a method’s efficacy of predicting MAFLD phenotypes. As reported by Docherty et al [80], we adopted a one-versus-rest strategy to assess each outcome class and compared the respective AUC values of all methods, which supports a fair, holistic analysis of their ability to predict MAFLD phenotypes.

Table 2 presents each method’s prediction performance across 10 trials. The proposed method has a 2-stage estimation design—stage 1 estimates whether an individual will develop MAFLD, and stage 2 predicts the likelihood of each MAFLD phenotype. Therefore, we report the results for each stage separately. As Table 2 shows, the proposed method attained higher AUC values in both stages, indicating its ability to distinguish patients with different outcomes. In stage 1, it accurately identified adults likely to develop MAFLD, with few false alarms, as signified by the relatively high precision and recall values. In stage 2, the proposed method generated effective predictions by consolidating the stage 1 results. The multiclass prediction results in stage 2 also allowed for direct comparisons with the benchmark methods. As seen in Table 2, the proposed method outperformed all benchmarks on both F₁-score and AUC. It exhibited a 7.2% improvement in AUC over the best-performing benchmark (0.898 vs 0.838) and had a 16.6% higher F₁-score than the best-performing benchmark (0.652 vs 0.559). Paired two-tailed t tests performed to examine differences in AUC indicated that the observed improvements were statistically significant (P<.001).

Figure 5 presents the respective receiver operating characteristic curves of all methods. The proposed method’s AUC curve was notably better than that of any benchmark method. This result further affirms its superior efficacy in estimating MAFLD phenotypes among adults compared with many prevalent methods.

In addition to overall performance, we examined the respective methods’ performance for each outcome class. As shown in Table 3, the proposed method achieved the highest F₁-score and AUC values for each outcome class, reaffirming its superior prediction ability. It attained a higher F₁-score (0.913) and AUC (0.859) for non-MAFLD than the respective best-performing benchmarks (DT: F₁-score=0.908; GraphSAGE: AUC=0.801). The performance improvements were especially prominent for the MAFLD phenotypes. For nondiabetic MAFLD, our method achieved an F₁-score of 0.531, much higher than that of the best-performing benchmark (MLP: F₁-score=0.400), exhibiting a 32.8% improvement. It also attained the highest AUC (0.878), higher than that of the best-performing benchmark (GraphSAGE: AUC=0.804). For diabetic MAFLD, the proposed method’s F₁-score (0.519) was 30.4% higher than that of the best-performing benchmark (GraphSAGE: F₁-score=0.398). Moreover, its precision value was superior to that of other methods, suggesting that it can identify adults who are likely to develop diabetic MAFLD with fewer false alarms.

The box plots in Figure 6 indicate the proposed method’s robust performance for each outcome class across 10 trials. It attained high F₁-scores for each outcome class, especially nondiabetic MAFLD and diabetic MAFLD, while the benchmark methods exhibited notably greater variance and occasional outliers. Together, these plots provide further evidence of the proposed method’s efficacy and value for clinical decision-making and patient management.

We also performed an ablation study to examine the relative contribution of each key component of the proposed method. We considered MLP, Graph, Graph + contrastive learning, and the (complete) proposed method. In essence, MLP serves as a baseline because it only uses the preprocessed data, without any key components of the proposed method. Graph builds on MLP and includes the graph representation learning of genetic family history and lifestyle data, together with the learned embeddings concatenated to the preprocessed dataset to train the MLP classifier. Graph + contrastive learning further extends Graph by incorporating contrastive learning after graph representation learning. The complete proposed method included all 3 key components. The results of the ablation study (Table 4) revealed how each component contributed to the method’s performance. They jointly produced the best predictions, indicating that MAFLD phenotype prediction can benefit from graph representation, multiview contrastive pretraining, and 2-stage estimation design.

To gain clinical insights into the proposed method’s learned representations, we examined its interpretability by depicting the embeddings visually. Specifically, we applied t-distributed stochastic neighbor embedding (t-SNE) [81] to visualize the contrastive pretraining embeddings and performed a Shapley additive explanation (SHAP) analysis [82] to reveal feature importance. Figure 7A presents a visualization of the original lifestyle and genetic features, and Figure 7B provides a visualization of the features obtained by concatenating the contrastive pretraining embeddings with the original lifestyle and genetic features. The original lifestyle and genetic features exhibited a scattered distribution, without any clear patterns. With contrastive pretraining embeddings, more distinctive clusters emerged, suggesting that patients with similar characteristics tend to cluster more closely than those with dissimilar characteristics. While these visual plots are exploratory without formal proof of class separability, they still illustrate that incorporating contrastive pretraining embeddings can potentially create a more structured, distinguishable representation of patient outcomes for effective MAFLD phenotype prediction.

We further examined the feature importance for each outcome class, as depicted by the SHAP summary plots in Figure 8. Because the proposed method adopted a 2-stage estimation (architecture) design, the model-agnostic explainer KernelSHAP was used with a background dataset of 100 training instances. For all test instances, SHAP values were computed on a representative model instance (ie, median test AUC across 10 trials). As seen in Figure 8, several metabolic indicators were important predictors consistently across different outcome classes. For example, BMI and waist circumference were highly influential. As Figure 8A shows, high BMI values (marked as red points) greatly reduced the likelihood of non-MAFLD predictions; particularly, high BMI and waist circumference values were associated with a greater likelihood of nondiabetic MAFLD or diabetic MAFLD, as shown in Figure 8B and C. Predictions of nondiabetic MAFLD were influenced by a combination of general metabolic indicators (eg, BMI and waist circumference) and lifestyle factors (eg, smoking and sleep disturbance). For diabetic MAFLD, definitive disease markers and factors related to disease consequence and management, such as self-care status and nutritional status (Mini Nutritional Assessment), appeared to be essential auxiliary predictors. These results align with clinical knowledge and reveal the proposed method’s ability to capture phenotype-specific patterns from patient data, with desirable interpretability.

Additionally, SHAP analyses allow for reasoning at the individual level. Figure 9 provides a visualization of SHAP values for 10 patients who were predicted to develop diabetic MAFLD. The heat map shows that diabetes mellitus and high hemoglobin A_1c diagnoses were consistently important predictors for most patients in this group, including patients B, G, and H. We also observed significant intraphenotype variability among patients. For example, the prediction for patient J was also significantly influenced by BMI and waist circumference, whereas triglycerides were a more important factor for patient C. The interpatient variability can help physicians better understand the impact of different factors at the individual level and thereby support personalized care and treatment planning.

The proposed method leverages deep learning to estimate MAFLD phenotypes among adults, using graph representation learning and contrastive learning. It provides several methodological novelties that can advance medical informatics research and enhance clinical decision-making for improved patient management. The evaluation results establish its predictive efficacy, demonstrate the value of combining clinical and survey-based data, and underscore the importance of intraphenotype variability and disease dynamics for MAFLD phenotype prediction. Furthermore, this method is generalizable and can be applied to other prediction tasks in similar clinical scenarios (eg, gauging the risk of diabetes or CVD) that feature multisource data, individual heterogeneities, intraclass variance, and intervariable relationships.

Using the proposed method, physicians will be able to identify individuals at higher risk of fibrosis and generate timely alerts for effective patient-centric care [83], which can mitigate the likelihood of significant disease progression and serious patient outcomes. Accurate prediction of MAFLD phenotypes also helps reduce hepatic complications such as CVD, chronic kidney disease [16], hepatocellular carcinoma [6], osteoporosis, endocrine disorders, and cognitive impairment [84]. The proposed method is capable of distinguishing high-risk versus low-risk adults on the basis of pathogenesis, spanning lifestyle, genetic, and metabolic factors; as a result, the likelihood of fibrosis or cirrhosis can be reduced, with broad implications for precision medicine and drug development [85]. In a related sense, its ability to predict phenotypes in an accurate and timely manner also enables personalized surveillance, treatment choice assessments, lifestyle changes, and treatment planning.

Although the proposed method does not achieve an objectively high F₁-score for MAFLD phenotypes, it still offers meaningful improvements over prevalent methods, even in the presence of the inherent challenges created by highly imbalanced patient clinical data. In our sample, most adults were in the non-MAFLD category, and few had MAFLD phenotypes, which made model training difficult for every method we investigated. This challenge is common to many clinical settings and has been documented across different patient outcome or risk prediction tasks. For example, recent related studies report F₁-scores in the range between 0.10 and 0.51 for minority classes [86,87]. Despite this persistent difficulty, the proposed method consistently outperformed all the benchmarks on MAFLD phenotypes (minority classes), which are clinically important. Hence, the observed improvements with our method represent valuable advances [22,88].

We illustrate the clinical use of the proposed method as a proactive risk stratification approach for clinical decision support and patient management. In stage 1, it estimates the probability of a person developing MAFLD within 1 year. To flag individuals as high risk, a physician can use the probability to select a decision threshold for balancing the trade-off between precision (the proportion of flagged individuals who are truly at high risk) and recall (the proportion of all true positive individuals who are truly flagged as high risk). If the physician prefers high certainty, they can choose a high threshold value. For example, our post hoc analysis showed that by setting the threshold to 0.60, the proposed method’s precision increased to 0.777, that is, approximately 78% of flagged patients indeed developed MAFLD. By choosing an even higher threshold value of 0.70, its precision further increased to 0.820, although at the cost of reduced sensitivity (0.59 in recall). As a result, the physician can identify adults who should be monitored more closely (for example, a semiannual follow-up instead of an annual follow-up), need immediate lifestyle counseling, or require proactive baseline liver function tests to track changes over time.

Furthermore, the proposed method provides additional insights based on the stage 2 estimate, which can support personalized planning and care. In general, obtaining clinically meaningful precision requires a higher threshold value. For example, with a threshold value of 0.50, the proposed method’s precision reached 0.506 for nondiabetic MAFLD and 0.500 for diabetic MAFLD. Emphasizing high-probability instances with a threshold value of 0.70 increased the precision to 0.762 and 0.778, respectively, which would allow physicians to tailor management strategies for adults whose phenotype can be predicted with higher confidence. Additionally, physicians can leverage the instance-level SHAP analysis, as depicted in Figure 9, to identify the specific factors that drive patient risk. While these insights do not directly indicate a definitive diagnosis, they can still guide physicians to engage in preventive care through patient risk stratification, while coping with the challenge of precise phenotype classification. Overall, physicians can adopt an appropriate threshold value to balance precision and recall while minimizing the likelihood of missing at-risk individuals for proactive stratification.

In summary, a multiview architecture leverages complementary information from lifestyle, genetic, and clinical data perspectives for richer representations that help distinguish infrequent yet clinically important MAFLD phenotypes, without sacrificing interpretability. The 2-stage design offers flexibility and additional utility. Accurate and robust estimates in stage 1 help physicians assess whether or not an individual is likely to develop MAFLD for initial screening purposes. In addition to that determination, even a moderate improvement in the F₁-score in stage 2 can facilitate physicians’ decision-making by providing additional information and clinical insights. These valuable risk stratification capabilities enable physicians to identify high-risk adults who may need close monitoring or alternative treatments. The 2-stage design also offers beneficial flexibility. Physicians can adjust their focus across the first or second stage, depending on their objective (eg, early screening, risk stratification, or intervention planning). According to 2 experienced hepatologists (who wish to remain anonymous), “Early, better estimates of individuals’ likelihood of MAFLD is valuable clinically,” and “The use of data-driven analytics methods to predict MAFLD phenotypes can enhance clinical decision-making and personalized patient management” (September 2, 2025). These expert inputs affirm the clinical value and practicality of our proposed method.

This study has several limitations, and it can be extended by further research. First, we used a sample from a single healthcare organization, which offered relatively limited diversity in terms of data sources and patient populations. In a related sense, our sample was imbalanced in the outcome class distribution, which constrained the prediction performance for minority classes, as reflected by the relatively low F₁-scores, which is in line with previous research [87]. Future studies should consider additional data sources and types such as image and text [89] to extend the proposed method, use different patient cohorts to affirm its efficacy, and apply synthetic data augmentation or multimodal foundation models to better address the issue of imbalanced outcome class distribution with cross-modal learning capabilities [62]. Second, because intraphenotype variability introduces complexity with regard to achieving compact clusters in the embedding space, a trade-off arises between variability and compactness, which could restrict the predictive utility for large datasets or different diseases. Therefore, we call for efforts to explore an optimal balance of variability and compactness for both accuracy and generalizability, such as clustering-based contrastive learning [90]. Third, the proposed 2-stage method provides some limited interpretability, through a feature attribution–based approach (ie, SHAP); its contrastive pretraining component deserves further exploration for greater transferability and interpretability. Ongoing efforts could facilitate and interpret embeddings in focal clinical contexts. Fourth, an international, multisociety Delphi process led to the proposal of metabolic dysfunction–associated steatotic liver disease (MASLD) in 2023 [91]. Although our findings might be extrapolated to adults with MASLD [92], the proposed method should be extended with research that tests for differences between MAFLD and MASLD and refines the proposed method to ensure robustness and prediction performance.

Predicting MAFLD phenotypes among adults is crucial, but existing analytic methods overlook its multisystem nature and phenotypic heterogeneity. As a solution, we developed a novel method that leverages graph representation learning, multiview contrastive pretraining, and a 2-stage estimation design to produce effective predictions that reflect phenotypic heterogeneity, complex relationships, and disease dynamics. It is effective in identifying at-risk adults and thus offers support for clinical decision-making and personalized care. This study reveals a promising pathway to advance health informatics research and clinical practice by leveraging rich, detailed clinical data in electronic health records and survey-based data to predict MAFLD phenotypes.