This prospective, multicenter study recruited 506 maintenance hemodialysis patients from January 2024 to September 2025 across three branches of the Guangdong Provincial Hospital of Chinese Medicine, namely the University Town Hospital (n = 158), the Dadelu General Hospital (n = 240), and the Fangcun Hospital (n = 108). The study cohort consisted of 120 patients who had experienced cardiovascular events (CVE group) and 386 patients who had not (non-CVE group). The data from the Fangcun and Dadelu hospitals were randomly partitioned at a ratio of 7:3 into a training set (n = 243) and a validation set (n = 105). The data from the University Town Hospital (n = 158) were utilized as an independent external test set to assess the model’s generalizability.

Standardized tongue images were obtained using a specialized imaging device (Model DS01 B) during hemodialysis sessions, based on the traditional Chinese medicine tenet that “the heart opens to the tongue” (Huang and Yuan, 2020), which postulates that tongue morphology reflects cardiovascular health. Image acquisition was carried out under uniform lighting conditions to guarantee comparability among participants. In this research, tongue features are regarded as a non - invasive biomarker capable of detecting microcirculatory changes, such as those related to inflammation or fibrosis, which are associated with an increased cardiovascular risk in hemodialysis patients (Duan et al., 2024).

Simultaneously, 90 clinical and physicochemical variables were gathered, with laboratory values corresponding to the most recent test results available at the time of imaging. The study was approved by the hospital’s Ethics Committee (No. YE2024 022 01), and all participants provided written informed consent.

Inclusion criteria were as follows: (a) regular hemodialysis for ≥3 months; (b) age ≥18 years, conscious, and without major physical disabilities; (c) stable clinical condition with well - controlled comorbidities; and (d) ability to adhere to study procedures and provide reliable data. Exclusion criteria were: (a) active infection, aldosteronism, adrenal lesions, severe liver/brain/hematopoietic disorders, psychiatric conditions, or poor general health; and (b) history of kidney transplantation.

Two patients at the University Town Hospital were unable to cooperate with tongue image acquisition and were thus excluded from the study, and one patient at the Fangcun Hospital declined participation due to privacy concerns. No participants were lost to follow-up or withdrew from the study.

In this study, CVEs were defined as the initial occurrence of any of the following during the follow-up period, in accordance with the 2020 American Heart Association guidelines and international consensus (Merchant et al., 2020): (a) Myocardial infarction (MI) (National Expert Committee on Rational Drug Use NHaFPC, 2019), diagnosed by ischemic manifestations, dynamic electrocardiogram (ECG) alterations (e.g., ST - segment elevation or new left bundle - branch block), and the rise or fall of high-sensitivity troponin T (hs-TnT) with at least one value exceeding the 99th percentile; (b) Cardiovascular (Cardiovascular Disease Group SoP and Chinese Medical Association, 2022) mortality, ascribed to acute coronary syndrome, heart failure, documented malignant arrhythmia, or sudden cardiac death, without any non-cardiac cause, based on medical records, death certificates, or autopsy findings; (c) Hospitalization for heart failure (Juan et al., 2024), meeting the European Society of Cardiology (ESC) criteria and necessitating an admission of ≥24 h; (d) Hospitalized unstable angina, diagnosed in the absence of elevated troponin but with ischemic symptoms and objective evidence of ischemia; and (e) Clinically significant arrhythmias (Zhao et al., 2020)—sustained ventricular tachycardia, hemodynamically compromising atrial fibrillation (e.g., systolic blood pressure <90 mmHg or requiring urgent cardioversion), or high-grade atrioventricular (AV) block (Mobitz II or third-degree) requiring intervention. Cerebrovascular events such as stroke were excluded from the primary endpoint. All suspected CVEs were independently adjudicated by two blinded cardiologists using clinical data, ECGs, serial hs - TnT measurements, and imaging modalities (e.g., echocardiography, angiography), in line with the current ESC/American College of Cardiology (ACC) criteria; discrepancies were resolved through consensus by a third senior cardiologist.

All preprocessing operations were carried out subsequent to dataset splitting to preclude data leakage. Preprocessing, encompassing missing value imputation, outlier removal, and normalization, utilized statistics exclusively derived from the training set. The parameters were uniformly applied to the validation and test sets. Among the 90 physicochemical variables, those with a missing value proportion exceeding 30% were excluded (Sterne et al., 2009) (The threshold selection is based on Reference (Sterne et al., 2009) to guarantee data integrity). The remaining missing values were imputed via Multiple Imputation by Chained Equations (MICE) in R (mice package v4.4.0), resulting in the generation of five imputed datasets. The final values were obtained by averaging across the imputations. Outliers deviating by more than ±3 standard deviations from the training set mean were removed.

Tongue images were visually examined by trained professionals. Images of poor quality (blurry, over/under - exposed, contaminated) were excluded. High - quality images were manually annotated using LabelMe (v5.2.1) to demarcate the tongue region. The JSON annotations were transformed into binary masks (using labelme_json_to_dataset) to segment the tongue and mask non - tongue areas (lips, skin, background), thereby minimizing visual noise during the training process.

Finally, all numerical features, including those derived from physicochemical data and tongue images, were standardized through Z - score normalization based on the training set means (μ) and standard deviations (σ), as shown in Equation 1. The same parameters were applied to the validation and test sets to ensure consistent evaluation and reproducibility. z = x − μ t r a i n σ t r a i n (1)

We conducted an evaluation of four algorithms: Logistic Regression (LR) with L2 regularization, LightGBM featuring moderate depth and strong L1/L2 regularization, AdaBoost employing shallow decision stumps, and a Multilayer Perceptron (MLP) equipped with a single hidden layer. Hyperparameters were adjusted in accordance with preliminary experiments to mitigate overfitting. All models were trained under three configurations: using only clinical variables, utilizing only tongue features, and incorporating fused features.

This research employed both hand - crafted and deep - learning techniques to extract multi - dimensional features from segmented tongue images. Hand - crafted features (n = 1,354) were calculated using OpenCV (v4.12.0) and PyRadiomics (v3.1.0), which included: (a) color descriptors (mean and standard deviation across RGB, CIELAB, HSV; 18 features); (b) texture metrics (GLCM, LBP, Gabor filters at 4 scales and 6 orientations); (c) shape descriptors (area, perimeter, eccentricity, and morphological measures from edge contours); and (d) Shannon entropy for textural complexity.

Regarding deep features, a ResNet - 50 model pre - trained on ImageNet was fine - tuned. Stochastic Gradient Descent (SGD) was utilized with a learning rate of 0.01, a momentum of 0.9, and a batch size of 32 over 30 epochs. Data augmentation (±15° rotation, horizontal flip, ±10% scaling) and cross - entropy loss were applied on an NVIDIA RTX 4090 GPU. The 2,048 - dimensional output from the penultimate layer was reduced to 8 dimensions through Principal Component Analysis (PCA), retaining components that accounted for ≥95% of the cumulative variance.

The final input consisted of 90 clinical/laboratory variables, 1,354 hand - crafted features, and 8 deep features (totaling 1,460 dimensions). To mitigate redundancy, LASSO regression with L1 regularization was employed. The optimal λ was selected via 10 - fold cross - validation using the one - standard - error rule. Only features with non - zero coefficients were retained. All features were Z - score normalized based on the training - set statistics prior to fusion. The entire pipeline was implemented in Python 3.9, leveraging opencv - python, PyRadiomics, scikit - learn, and PyTorch (v1.12.1).

All statistical analyses were carried out in Python 3.9, leveraging scipy (v1.13.1), statsmodels (v0.13.2), and scikit - learn (v1.0.2), with a two - sided significance level set at α = 0.05. Baseline characteristics were compared between patients with and without cardiovascular events (CVEs). Normality was evaluated through the Shapiro–Wilk test and Q–Q plots. For normally distributed continuous variables (characterized by homogeneous variance), they were summarized as the mean ± standard deviation (SD) and compared using the independent - samples t - test. In contrast, non - normal variables were reported as the median (inter - quartile range, IQR) and analyzed via the Mann–Whitney U test. Categorical variables were presented as counts (percentages) and compared using the chi - squared test; when the expected cell counts were <less than 5, Fisher’s exact test was employed.

To augment the robustness and validation of the model, sensitivity analyses were conducted. Specifically, 10 - fold cross - validation was implemented during LASSO hyperparameter tuning to alleviate overfitting. Uncertainty was quantified by calculating 95% confidence intervals for the area AUC using DeLong’s method, and further validation was achieved through 1,000 bootstrap resampling iterations.

The performance of the model on the independent test set was evaluated using multiple metrics, including the AUC, accuracy, sensitivity, specificity, positive and negative predictive values (PPV/NPV), and F1 - score. Calibration was assessed visually through calibration plots and statistically via the Hosmer–Lemeshow test (where P > 0.05 indicates an adequate fit). Clinical utility was determined through decision curve analysis (DCA), which measured the net benefit across risk thresholds ranging from 5% to 30%. A model was considered clinically useful if it outperformed both the “treat all” and “treat none” strategies.

This study encompassed 506 patients, with 348 patients from Fangcun and Dade Road General Hospitals constituting the model development cohort, and 158 patients from University Town Hospital serving as a completely independent external test cohort, which was not utilized at all during model development, hyper - parameter tuning, or data splitting.

Within the development cohort, stratified random splitting based on cardiovascular event status was carried out 10 times at a 7:3 ratio to maintain event prevalence and evaluate performance variability. For each iteration, models were trained on the training subset and evaluated on the validation subset, with the AUC serving as the primary evaluation metric. Although selecting the split with the highest performance may introduce a slight optimistic bias, this approach ensured consistency across analyses. The final conclusions were validated on the reserved external test set to minimize the risk of overfitting.

The split with the highest validation AUC was chosen as the fixed partition for all subsequent steps. It consisted of a training set of 243 patients (57 events, 23.5%) and a validation set of 105 patients (20 events, 19.0%). All subsequent procedures, including feature selection, hyper - parameter tuning, model comparison, and performance reporting, were strictly based on this fixed split to ensure reproducibility.

The external test cohort included 158 patients, 43 of whom experienced cardiovascular events (CVEs, (27.2%). The baseline characteristics were generally comparable to those of the development cohort (Table 1), indicating its suitability for unbiased external validation.

To conduct a systematic evaluation of the predictive value of diverse data modalities, a comparison was made regarding the performance of four machine - learning models, namely LR, LightGBM, AdaBoost, and MLP, across three feature configurations. These configurations included clinical variables solely, tongue image features solely, and a fused set integrating both. The performance was evaluated on an independent external test cohort, and the results are presented in Table 2.

Models relying on tongue image features significantly outperformed those utilizing only clinical variables. The optimal clinical - only model (AdaBoost) attained a moderate AUC of 0.682, with a balanced sensitivity of 52.4% and specificity of 75.0%. However, it exhibited limited overall accuracy (69.6%) and F1-score (0.452). In contrast, the tongue - only AdaBoost model achieved a substantially higher AUC of 0.786, accompanied by enhanced accuracy (81.0%), sensitivity (71.4%), specificity (84.0%), and F1 - score (0.643), which demonstrated the robust discriminative capacity of non - invasive tongue imaging features.

The fusion of clinical and tongue features led to further, albeit marginal, improvements. The fused AdaBoost model achieved the highest AUC in this study (0.787), indicating a slight incremental value from multimodal integration. Nevertheless, this was accompanied by a trade - off in accuracy (74.5%) and F1 - score (0.579), primarily attributable to an increase in false positives. Other fused models (LR, MLP) showed consistent improvements over clinical-only models and modest enhancements in overall discrimination, albeit with a slight trade-off in specificity compared to tongue-only counterparts. For instance, the specificity of LR decreased from 85.0% (tongue-only) to 80.0% (fused), while the AUC increased from 0.750 to 0.776.

Notably, while the training performance was high across all models (e.g., training AUCs >0.85 for tongue and fused models), the test performance declined, particularly for MLP, suggesting some degree of overfitting. This is likely due to the high feature dimensionality relative to the sample size. Despite this, DCA indicated a positive net benefit across clinically relevant risk thresholds (5%–30%) for all tongue - based and fused models, supporting their potential clinical applicability.

In conclusion, tongue image features alone outperformed conventional clinical variables in predicting cardiovascular events in hemodialysis patients. Multimodal fusion provided a small yet consistent improvement in discrimination, with AdaBoost emerging as the most balanced and effective algorithm across all configurations. As depicted in the ROC curves (Figures 2A,B), the performance gap between the training and test sets suggests potential overfitting, which was alleviated through cross - validation.

As presented in Table 2, the tongue image–based model (Tongue_all) attained a notably higher AUC value of 0.786 in comparison to the clinical model, which had an AUC of 0.682. This finding suggests that tongue features offer incremental predictive value. For example, when combined with tongue features, the AdaBoost model achieved a sensitivity of 71.4%, outperforming the model relying solely on clinical data, which had a sensitivity of 52.4%. This advantage is likely attributable to the tongue’s capacity to capture microvascular alterations that cannot be directly measured by conventional clinical variables.

The Tongue_all model, which combines deep learning and radiomic features, achieved the highest accuracy of 81.0% and an F1 - score of 0.643, with an AUC of 0.786. The Fused model exhibited a slight increase in AUC (0.787), yet demonstrated lower accuracy (74.5%) and specificity (75.0%), indicating a trade - off with a higher rate of false positives. The clinic - only model performed the least effectively, with an AUC of 0.682. DCA and Integrated Discrimination Improvement (IDI) analyses confirmed the superior clinical utility and incremental value of tongue - derived features (Table 3; Figures 2–5).

SHAP analysis demonstrated that both tongue texture features (such as wavelet - derived metrics) and clinical biomarkers (especially PT%) were the key determinants of prediction. Elevated values of specific texture features were correlated with an augmented risk, possibly mirroring underlying inflammation or microvascular alterations. In contrast, a higher PT% (signifying superior coagulation function) exerted a protective effect. The substantial influence of texture metrics corroborates the TCM notion that tongue morphology serves as an indicator of systemic health. This interpretability narrows the disparity between model output and clinical comprehension.