Baishideng Publishing Group Inc, 7041 Koll Center Parkway, Suite 160, Pleasanton, CA 94566, USA
Share the Article
Lv YW, Jiang AJ, Geng YM, Hu M, Wang YQ, Ma L, Shen N, Zhao JD. Integrating traditional Chinese medicine syndromes with clinical variables to predict lower extremity arterial disease in type 2 diabetes. World J Diabetes 2026; 17(11): 125698 [DOI: 10.4239/wjd.125698]
Ying-Wei Lv, Ai-Juan Jiang, Yu-Meng Geng, Min Hu, Yu-Qing Wang, Lin Ma, Nuo Shen, School of Integrated Chinese and Western Medicine, Anhui University of Chinese Medicine, Hefei 230012, Anhui Province, China
Jin-Dong Zhao, Department of Endocrinology, The First Affiliated Hospital of Anhui University of Chinese Medicine, Hefei 230031, Anhui Province, China
Author contributions: Lv YW, Jiang AJ, Geng YM, Hu M, Ma L, Shen N, Wang YQ and Zhao JD designed the research; Lv YW, Jiang AJ, Geng YM and Hu M performed the research; Ma L and Shen N contributed new analytic tools; Lv YW, Wang YQ, and Zhao JD analyzed the data; and Lv YW, Jiang AJ and Geng YM wrote the paper.
AI contribution statement: The authors take full responsibility and accountability for all content of this manuscript, including any portions for which AI tools were used as assistive technologies. All AI-assisted outputs were carefully reviewed, validated, and approved by the authors. AI tools were not used to generate original scientific data, perform independent scientific analyses, or draw scientific conclusions.
Supported by Anhui Provincial Natural Science Foundation Project, No. 2308085MH307; The "Basic-Clinical Integration" Special Project of Category A for 2025 of Anhui University of Chinese Medicine, No. JCLCA2025024; and National Natural Science Foundation of China, No. 82505372.
Institutional review board statement: This study was approved by Medical Ethics Committee of the First Affiliated Hospital of Anhui University of Chinese Medicine the (Approval No. 2021MCZQ11).
Informed consent statement: Patient consent was waived by the Medical Ethics Committee because this was a retrospective study using anonymized data.
Conflict-of-interest statement: The authors declare that they have no competing interests.
Data sharing statement: The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.
Corresponding author: Ai-Juan Jiang, PhD, Professor, School of Integrated Chinese and Western Medicine, Anhui University of Chinese Medicine, No. 350 Longzihu Road, Xinhua High-tech Zone, Hefei 230012, Anhui Province, China. jiangaijuan@ahtcm.edu.cn
Received: July 14, 2026 Revised: August 9, 2026 Accepted: October 9, 2026 Published online: November 15, 2026 Processing time: 115 Days and 11.6 Hours
Abstract
BACKGROUND
Lower extremity arterial disease (LEAD) is common macrovascular complication of type 2 diabetes mellitus (T2DM) that can lead to intermittent claudication, rest pain, and gangrene. Due to its insidious onset, LEAD is frequently diagnosed, missing the optimal window for intervention. Therefore, the construction of an accurate and convenient early risk prediction model is of significant clinical value.
AIM
To investigate the risk factors for T2DM-associated LEAD and to develop prediction models using machine learning algorithms based on integrated traditional Chinese medicine (TCM) and Western medicine features.
METHODS
Clinical data of 1018 inpatients with T2DM admitted to the Department of Endocrinology, The First Affiliated Hospital of Anhui University of Chinese Medicine, were retrospectively analyzed. The dataset was randomly divided into a training set (n = 712) and a validation set (n = 306) at a 7:3 ratio. LASSO regression was used for dimensionality reduction to identify core predictors. Five machine learning models, namely, logistic regression, XGBoost, random forest, gradient boosting decision tree (GBDT), and LightGBM, were constructed based on eight predictors: Age ≥ 60 years, hypertension, reduced hemoglobin, reduced red blood cell count, elevated blood urea nitrogen, emaciation, dark tongue, and fatigue. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, and F1 score. Ten-fold cross-validation was performed to assess model stability. A nomogram was constructed based on logistic regression coefficients for individualized risk visualization. Decision curve analysis (DCA) was used to evaluate clinical utility.
RESULTS
Among the 1018 patients with T2DM, 647 (63.56%) had LEAD. LASSO regression identified eight core predictors from 26 candidate variables (P < 0.05). In the validation set, the XGBoost model achieved the highest AUC of 0.763 (95% confidence interval: 0.707-0.815), with an accuracy of 0.712, sensitivity of 0.651, and F1 score of 0.743. The AUCs of the remaining models were: Random forest (0.762), logistic regression (0.758), LightGBM (0.754), and GBDT (0.744). Ten-fold cross-validation showed mean AUCs of 0.746 ± 0.006 and 0.736 ± 0.043 in the training and validation sets, respectively, with a minimal difference (ΔAUC = 0.010), indicating favourable stability without obvious overfitting. DCA indicated positive clinical benefit for all five models across a wide threshold probability range, significantly outperforming the “treat all” and “treat none” extreme strategies.
CONCLUSION
The XGBoost model built on dual-dimensional TCM and Western medicine features is effective for the early identification of LEAD in T2DM patients, providing support for individualized treatment strategies.
Core Tip: This study integrated Western medicine demographic and clinical characteristics, laboratory indicators, and traditional Chinese medicine syndrome features to construct prediction models for lower extremity arterial disease (LEAD) in type 2 diabetes mellitus using multiple machine learning algorithms. The XGBoost model based on integrated Western and Chinese medicine features exhibited the highest efficacy and has potential for the early identification and treatment of LEAD.
Citation: Lv YW, Jiang AJ, Geng YM, Hu M, Wang YQ, Ma L, Shen N, Zhao JD. Integrating traditional Chinese medicine syndromes with clinical variables to predict lower extremity arterial disease in type 2 diabetes. World J Diabetes 2026; 17(11): 125698
Lower extremity arterial disease (LEAD) is one of the most common macrovascular complications of type 2 diabetes mellitus (T2DM). LEAD can involve the entire lower extremity arterial tree and result in intermittent claudication, rest pain, and even gangrene, and is the leading cause of non-traumatic amputation, severely reducing patients’ quality of life and increasing the healthcare burden[1]. Diabetes is prevalent in China, and the likelihood of LEAD increases significantly as the disease progresses. As the early symptoms of LEAD are insidious, most patients present with moderate-to-severe vascular stenosis at the time of diagnosis, thereby missing the optimal window for intervention[2]. Therefore, the construction of an accurate and accessible model for early risk prediction has significant clinical value.
Traditional risk scores are typically based on linear regression and have limited ability to capture complex nonlinear relationships[3]. Machine learning algorithms are effective in identifying potential associations within high-dimensional clinical data and have demonstrated significant advantages in the prediction of disease risk. There has been a recent trend to integrate multi-dimensional data from traditional Chinese medicine (TCM) and Western medicine in the field of chronic disease prediction[4]. TCM syndrome features can reflect the overall pathological state of disease onset and progression, complementing the objective test results of Western medicine and aligning with the TCM principle of “preventive treatment before disease onset”.
This study aimed to integrate Western medicine demographic and clinical characteristics, laboratory indicators, and TCM syndrome features to construct models for predicting the risk of LEAD in patients with T2DM. Multiple machine learning algorithms were applied, and the efficacy of the resulting models was assessed. The findings provide a tool for the early identification of LEAD and the prevention and treatment of the disease using a combination of Chinese and Western medicine[5].
MATERIALS AND METHODS
Data sources
A total of 1724 patients with T2DM admitted to The First Affiliated Hospital of Anhui University of Chinese Medicine between September 2017 and August 2022 were selected as the study population. Of these, 1018 were ultimately enrolled after application of the inclusion and exclusion criteria. The study was reviewed and approved by the Medical Ethics Committee of the hospital.
Diagnostic criteria
T2DM: Patients were diagnosed according to the criteria outlined in the Clinical guidelines for prevention and treatment of type 2 diabetes mellitus in the elderly in China (2022 edition)[6].
LEAD: Diagnosis was based on the Guidelines for the Diagnosis and Treatment of Lower Extremity Arteriosclerosis Obliterans, combined with the presence of risk factors such as diabetes, hypertension, and hyperlipidemia. Patients were considered to have LEAD if ultrasound examination of the lower extremity arteries revealed intimal thickening, plaque formation, or lumen stenosis, and if they met the diagnostic criteria for diabetes.
TCM syndrome differentiation: Syndrome patterns were assessed with reference to the Guidelines for TCM Diagnosis and Treatment of Diabetic Peripheral Vascular Disease by two TCM physicians at or above the attending physician level. Four diagnostic methods, namely, inspection, auscultation and olfaction, inquiry, and pulse-taking, were used. In cases of disagreement, a senior physician was consulted.
Inclusion criteria
Patients were included if they: (1) Met the diagnostic criteria for T2DM; (2) Had complete clinical and laboratory data; and (3) Had provided written informed consent.
Exclusion criteria
Patients were not included if they had: (1) Type 1 diabetes or other non-T2DM diabetes types; (2) Acute diabetic complications (diabetic ketoacidosis, hyperosmolar hyperglycemic syndrome); (3) Lower extremity vascular lesions caused by other factors, such as trauma, infection, or thromboangiitis obliterans; (4) Concomitant malignant tumors, severe cardiac, hepatic, or renal failure, or autoimmune diseases; (5) Psychiatric disorders or cognitive impairment precluding cooperation with examination; and (6) More than 10% missing data.
The clinical data of all participants were extracted from the electronic medical record system. These data included three categories of candidate indicators.
Baseline demographic and clinical indicators and TCM syndrome features: Sex, age, hypertension, hyperlipidemia, diabetic nephropathy, coronary heart disease, fatty liver disease, emaciation, dark tongue, fatigue, thirst with preference for drinking, dry mouth and throat, and others, all coded as binary variables (yes/no).
Patients were randomly divided into a training set (70%) and a validation set (30%) at a ratio of 7:3, with stratification by LEAD status to preserve the disease prevalence in both subsets (seed = 4). The training set was used for model training and hyperparameter tuning, while the validation set was used for independent evaluation of model performance. Ten-fold cross-validation was performed on the optimal model to assess stability. The area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, and F1 score were used to evaluate model performance.
Statistical analysis was performed using Python 3.12 and R 4.5.1 software. Non-normally distributed continuous variables are expressed as medians (P25, P75), and between-group comparisons were conducted using the Mann-Whitney U test. Categorical variables are expressed as n (%) and were compared using the χ2 test. LASSO regression was used for dimensionality reduction of the candidate variables, with the optimal λ value selected via ten-fold cross-validation to identify core predictors. Five machine learning classification models were constructed using Python, and receiver operating characteristic (ROC) curves were plotted. Decision curve analysis (DCA) wasused to evaluate the clinical utility of the models. A nomogram was constructed for model visualization. The significance level was set at α = 0.05.
RESULTS
A total of 1018 patients with T2DM were included in the analysis, of whom 647 were in the LEAD group and 371 in the non-LEAD group. Univariate analysis revealed statistically significant differences between the two groups in terms of age, systolic blood pressure, diastolic blood pressure, hypertension, fatty liver disease, diabetic nephropathy, hemoglobin, RBC count, urine protein grade, FBG, TG, TC, LDL-C, serum creatinine, BUN, ALT, AST, abnormal complexion, obesity, emaciation, limb numbness, dark tongue, fatigue, limb pain, blurred vision, and dizziness with palpitation (P < 0.05). No significant association was found between HbA1c and LEAD (P = 0.611), possibly because both hemoglobin and RBC count were included in the model and HbA1c is physiologically linked with hemoglobin (it is lowered by anemia). Furthermore, fasting glucose levels also overlap with HbA1c in diabetic patients, providing little independent information. While the RBC count, hemoglobin, and BUN all fell within clinically normal reference ranges, all three were retained in the model after LASSO regression screening. The odds ratios (ORs) for RBC, BUN, and hemoglobin were 2.06, 1.72, and 1.75, respectively, with respective 95% confidence intervals (95%CIs) of 1.23-3.46, 1.19-2.49, and 1.32-2.32. These results demonstrate that despite the cutoff values for all three indicators remaining within the clinically normal reference ranges, their levels were associated with LEAD risk in the T2DM population and could serve as predictive features for predicting LEAD. Further details are shown in Tables 1 and 2.
Table 1 Baseline characteristics of included patients.
LASSO regression for identification of core predictors
The 26 candidate variables that showed statistical significance (P < 0.05) in the univariate analysis were entered into the LASSO regression model for dimensionality reduction. LASSO analysis was also used to integrate information on clinical history. The optimal λ value corresponding to the minimum binomial deviance under the one-standard-error criterion was selected using ten-fold cross-validation. Ultimately, eight core predictors with non-zero coefficients were identified from the 26 variables: Age, hypertension, reduced hemoglobin, reduced RBC count, elevated BUN, emaciation, dark tongue, and fatigue. Despite values for the RBC count, hemoglobin, and BUN falling within clinically normal reference ranges, all three were retained following LASSO regression (Figures 1 and 2).
Figure 2
Relationship between the tuning parameter (λ) and binomial deviance in the LASSO regression model with ten-fold cross-validation.
Further multivariate logistic regression analysis revealed that age (OR = 4.552, 95%CI: 3.339-6.207), hypertension (OR = 1.788, 95%CI: 1.337-2.391), BUN (OR = 1.678, 95%CI: 1.157-2.435), emaciation (OR = 1.900, 95%CI: 1.117-3.233), and fatigue (OR = 1.345, 95%CI: 1.012-1.787) remained significantly associated with LEAD (all P < 0.05). However, hemoglobin (OR = 1.512, 95%CI: 0.718-3.183), RBC (OR = 1.852, 95%CI: 0.853-4.020), and dark tongue (OR = 1.194, 95%CI: 0.859-1.658) did not reach statistical significance in the multivariate model (P > 0.05). Three LASSO-selected variables (hemoglobin, RBC count, dark tongue) were found to be non-significant in the multivariate analysis (P = 0.276, 0.119, 0.292). Nevertheless, they were retained as the models were constructed for prediction, rather than causal inference, and a feature can still improve discrimination through interaction and collinearity even when its standalone OR is unremarkable. All three were stably selected by LASSO across the 10 folds. Hemoglobin and RBC also showed effect sizes above 1.5, with P values that essentially reflected limited sample power. After application of the Benjamini-Hochberg procedure to control the false discovery rate, all eight variables retained statistical significance in the univariate analysis (P < 0.05). Details are shown in Table 3.
Table 3 Univariate and multivariate logistic regression models of predictors for lower extremity arterial disease.
Five machine learning prediction models were constructed using the eight selected core predictors. The ROC curves for the training and validation sets are shown in Figure 3. Comparison of performance in the validation set showed that the XGBoost model achieved the highest AUC (0.763), indicating the best comprehensive predictive performance. The AUCs of the remaining models were, in descending order: Random forest (0.762), logistic regression (0.758), LightGBM (0.754), and gradient boosting decision tree (GBDT; 0.744). Details are provided in Table 4. In the training set, LightGBM achieved the lowest calibration error [expected calibration error (ECE) = 0.019], narrowly ahead of XGBoost (ECE = 0.027); in the validation set, however, the logistic regression model exhibited the best generalization calibration (ECE = 0.043), followed closely by XGBoost (ECE = 0.044). The strong in-sample performance of the LightGBM model resulted from its leaf-wise boosting and built-in regularization, which yielded marginally less extreme probability estimates than XGBoost on the data used for fitting; however, this advantage was not retained in the validation set (ECE = 0.071). Logistic regression, by contrast, is inherently well-calibrated, and its smooth, bounded linear probabilities neither collapse toward 0 nor saturate near 1, resulting in generalization without overfitting. XGBoost avoids the severe overconfidence typical of deeper tree ensembles, resulting in a validation calibration similar to that of logistic regression models despite being tree-based (Figures 3 and 4). The XGBoost model showed acceptable calibration in the validation set, with a calibration slope of 1.18 and an intercept of -0.10 before recalibration, improving to 1.01 and -0.04 after Platt scaling (ECE: 0.044-0.034; Brier 0.186). These values indicate that the predicted risks were largely unbiased and could be reliably interpreted as probabilities. The Brier score was among the lowest of the five candidate models, reflecting both favorable calibration and discrimination. Two tuning-aware, leakage-free validation schemes yielded consistent estimates of true performance. Bootstrap optimism correction indicated a corrected AUC of 0.745 for XGBoost, while nested 5 × 5 cross-validation yielded an outer AUC of 0.746 (95%CI: 0.703-0.808). The close agreement between these independent methods supports their discriminative ability. Performance was stable across split ratios (AUC: 0.755-0.762 for 6:4-8:2). At a 0.5 decision threshold (validation prevalence, 63.7%), the model achieved a positive predictive value (PPV) of 0.74 and a negative predictive value (NPV) of 0.61. Raising the threshold to 0.7 increased the PPV to 0.86 at the cost of NPV (0.55). Due to its high baseline prevalence, the XGBoost model would be best used for screening increased risk to prioritize further vascular assessment rather than as a rule-out test, suggesting that the threshold should be adjusted according to the clinical context. After bias correction, the discriminative performance of the model was favorable with well-calibrated probabilities (Tables 4 and 5).
Based on a comprehensive consideration of the AUC and clinical indicators, the XGBoost model demonstrated the best overall performance of the five models. Ten-fold cross-validation was conducted on the XGBoost model, and the ROC curves are shown in Figure 5. The curves indicated AUC values of 0.746 ± 0.006 and 0.736 ± 0.043 in the training and validation sets, respectively. The closeness of these values suggests that the model was stable with no obvious overfitting.
Figure 5 Receiver operating characteristic curves of the XGBoost model based on ten-fold cross-validation.
ROC: Receiver operating characteristic; AUC: Area under the receiver operating characteristic curve; 95%CI: 95% confidence interval.
SHapley Additive exPlanations analysis of feature importance
SHapley Additive exPlanations (SHAP) analysis was used to rank the predictive factors in the model in terms of importance. The ranking showed: Age ≥ 60 years > hypertension > elevated BUN > dark tongue > emaciation > fatigue > decreased hemoglobin > decreased RBC count. As shown in the SHAP summary plot in Figure 6, age ≥ 60 years, hypertension, elevated BUN, dark tongue, emaciation, fatigue, decreased hemoglobin, and decreased RBC count were all positively associated with the development of LEAD.
Figure 6
SHapley Additive exPlanations analysis of the machine learning prediction model.
Clinical utility of the model
DCA curve analysis showed that at threshold probability values between 0.1 and 0.8, all five models demonstrated positive clinical net benefit, significantly outperforming the “treat all” and “treat none” extreme strategies in both the training and validation sets. In Figure 7, the X-axis represents the high-risk threshold probability and the Y-axis represents the net benefit. The orange "All" line and the black dashed "None" line represent the strategies of treating all patients and treating no patients, respectively. Machine learning models with curves above both reference lines provide positive clinical utility, with the results for the validation set demonstrating the generalizability of the findings (Figure 7). The curves represent the TCM-only model (AUC = 0.566), the Western medicine-only model (AUC = 0.696), the integrated Chinese-Western medicine model (AUC = 0.762), the treat-all strategy, and the treat-none strategy. The integrated model achieved the highest discrimination and the greatest net benefit across most high-risk thresholds, demonstrating that the combination of TCM and Western medicine features improves both predictive performance and clinical decision-making utility (Figure 8).
Figure 8 "Traditional Chinese medicine-only" and "Western-only" denote models built using traditional Chinese medicine features or Western medicine features alone, respectively.
"Integrated" denotes the model combining both feature sets. The "All" and "None" strategies represent treating all patients or no patients. Net benefit was computed across threshold probabilities from 0 to 1. TCM: Traditional Chinese medicine; AUC: Area under the receiver operating characteristic curve.
Nomogram for predicting LEAD risk
A nomogram for predicting the risk of LEAD in T2DM was constructed using the eight core predictors integrating Chinese and Western medicine features (Figure 9). Each predictor in the nomogram corresponds to a point scale at the top. The points assigned to each predictor are summed to obtain a total score, which is then mapped to the risk scale at the bottom to derive the individual probability of developing LEAD. The nomogram can be conveniently applied in clinical settings for individualized risk assessment. Age ≥ 60 years was selected as the demonstration variable, as older age was found to be the strongest predictor of LEAD and carries the highest point weighting in the nomogram. For example, consider a 72-year-old patient with T2DM (age ≥ 60 years, points: 72) with hypertension (points: 4.8), hemoglobin level of 125 g/L (points: 7), RBC count of 4.0 × 1012/L (points: 1.5), BUN level of 7 mmol/L (points: 0.5), no emaciation (points: 0), no dark tongue (points: 0), and exhibiting fatigue (points: 4.3). The total score is 72 + 4.8 + 7 + 1.5 + 0.5 + 0 + 0 + 4.3 = 90.1, corresponding to an approximate 89% probability of developing LEAD, indicating a high risk of LEAD in this older hypertensive patient.
Figure 9 Nomogram for the prediction model of lower extremity arterial disease in type 2 diabetes mellitus.
RBC: Red blood cell; BUN: Blood urea nitrogen.
DISCUSSION
This study constructed and compared five machine learning models for predicting the risk of LEAD in patients with T2DM. The models were based on eight core predictors identified by LASSO regression. In the validation set, the XGBoost model achieved the highest AUC of 0.763, higher than those of the random forest (0.762), logistic regression (0.758), LightGBM (0.754), and GBDT (0.744) models, and demonstrated the best comprehensive predictive performance with an accuracy of 0.712, sensitivity of 0.651, and F1 score of 0.743. XGBoost is an optimized gradient boosting tree algorithm that generates multiple decision trees through iterative learning for joint decision-making, effectively capturing nonlinear interactions among variables[7,8]. Compared with the findings of Kuang et al[2] based on a national database of 3000 diabetic patients (LEAD prevalence, 15.9%; logistic regression and multilayer perceptron models, AUCs of 0.73 and 0.72, respectively), the present single-center model achieved higher discrimination. The model was further evaluated for calibration and clinical utility, despite the smaller cohort and higher LEAD prevalence in the current sample restricting direct comparison. The built-in regularization terms in XGBoost reduce the risk of overfitting, and these models have been widely validated in chronic disease risk prediction. The development of LEAD in patients with T2DM results from long-term synergistic interactions among multiple factors. XGBoost is more effective in handling nonlinear relationships than logistic regression, and the results in the validation set showed stable performance with no obvious overfitting. Moreover, as shown by the calibration curves, the XGBoost model achieved an ECE of 0.027 in the training set, indicating significant consistency between predicted probabilities and actual outcomes, which is a critical requirement for clinical decision-making. The Brier score of 0.191 further confirmed the reliability of the model in probability estimation. Notably, although the AUC values of the five models were relatively close (between 0.744 and 0.763), the superior sensitivity (0.651) and F1 score (0.743) of the XGBoost model are particularly important in clinical screening contexts, where the identification of high-risk patients who may benefit from early intervention takes priority over minimizing false positives. The relatively narrow AUC range across the models also suggests that the eight selected core predictors represent significant discriminative indicators, and the choice of algorithm primarily influences calibration and sensitivity rather than overall discrimination.
The results of the SHAP analysis further enhanced the interpretability of the XGBoost model by quantifying the contribution of each predictor to individual predictions. The ranking of features in terms of importance revealed that age ≥ 60 years, hypertension, and elevated BUN were the top three contributors, consistent with the established clinical understanding of LEAD risk factors. This alignment between data-driven assessment of feature importance and clinical knowledge increases confidence in the validity of the model and would contribute to its acceptance by clinicians. Unlike traditional logistic regression coefficients, which assume linear additive effects, SHAP values capture both the magnitude and direction of the effects of individual features, enabling explanations of personalized risk. This is particularly valuable in heterogeneous patient populations where the same risk factor may vary in its effects depending on the patient’s overall clinical profile. The number-four ranking of the dark tongue feature is noteworthy, as this is a reflection of blood stasis in TCM theory and may correspond to microcirculatory disturbances detectable through non-invasive observation, highlighting the complementary value of TCM features in the predictive framework.
The high prevalence of LEAD (63.56%) observed in the present study underscores the substantial burden of lower extremity vascular complications among hospitalized T2DM patients and highlights the urgent need for effective screening tools. This prevalence is consistent with epidemiological reports indicating that LEAD affects approximately 50%-70% of patients with long-term diabetes. The LASSO regression approach used here for feature selection was effective in addressing the challenge of high-dimensional clinical data by applying L1 regularization, which simultaneously shrinks less informative coefficients to zero and retains only the most reliable predictors. Eight core predictors were identified from the original 26 candidate variables, achieving a balance between model simplicity and predictive performance. This parsimonious model is advantageous in clinical practice, as it requires only readily available clinical and laboratory parameters with no need for additional specialized testing, facilitating implementation in routine clinical workflows.
Among the eight core predictors, age and hypertension are established risk factors for LEAD. In diabetes, older age compounds the cumulative endothelial injury inflicted by chronic hyperglycemia, including the accumulation of advanced glycation end-products, oxidative stress, and diminished nitric oxide (NO) bioavailability, which accelerate vascular senescence and intimal thickening[9,10]. Long-term hypertension adds hemodynamic insult to this compromised vasculature, while increased shear stress disrupts endothelial integrity and promotes plaque instability in lower-extremity arteries[11,12]. Reduced hemoglobin levels and RBC counts indicate impaired oxygen-carrying capacity and microcirculatory dysfunction[13-15]. The diabetic vasculature is peculiarly susceptible to hypoxic injury, as microangiopathy narrows the nutritive arteriolar bed, and concurrent macrovascular stenosis reduces bulk blood flow[16]. Hemoglobin transports NO as the S-nitrosohemoglobin complex, which regulates vascular tone and platelet aggregation; its reduction compromises NO-mediated vasodilation and shifts the vascular milieu toward thrombosis[17,18]. Low RBC counts alter blood viscosity and rheology, further impairing microvascular perfusion[19]. These changes in hematological parameters are consistent with evidence that anemia is predictive of cardiovascular events in diabetic populations. Elevated BUN values indicate compromised renal function and are closely linked to the kidney-vascular axis in diabetes[20]. Diabetic nephropathy induces systemic inflammation, oxidative stress, and calcium-phosphate dysregulation, each of which accelerates medial arterial calcification and reduces vessel compliance[21]. Arterial calcification is of particular clinical importance in the lower extremities, as it hardens vessel walls, inhibits adaptive vasodilation, and increases susceptibility to ischemia. BUN is a readily available renal marker and represents an indicator of the systemic metabolic milieu predisposing to vascular injury[22-24]. Emaciation, dark tongue, and fatigue reflect the constitutional and symptomatic dimensions of diabetic vascular disease. In TCM theory, diabetes (Xiao-Ke) originates from yin deficiency with dryness-heat; prolonged illness consumes qi and yin, and emaciation indicates this consumptive trajectory of chronic metabolic depletion with impaired tissue nourishment[25]. Dark tongue is a reflection of blood stasis, the central pathological mechanism by which TCM conceptualizes diabetic complications, namely, microcirculatory disturbance, hyperviscosity, and tissue hypoperfusion, which correspond directly to the pathophysiology of arterial occlusive disease[26,27]. Fatigue indicates qi deficiency in the spleen and is associated with reduced exercise tolerance and impaired peripheral perfusion[28,29]. These TCM features are reflective of systemic dysfunction that individual laboratory indicators may not fully capture. The convergence of TCM constitutional assessment with Western biomarkers represents the principle of integrating macroscopic pattern differentiation with microscopic objective indicators; specifically, markers used in Western medicine quantify the biochemical and structural consequences of diabetic vascular injury, while TCM features characterize the functional and systemic phenotype of the same disease process[30-32]. The integration of TCM with Western medicine enabled the model to detect risk signals beyond individual parameters, as the DCA results demonstrated positive net clinical benefit across a wide threshold of probabilities[33,34].
The incorporation of TCM syndrome features into a Western medicine-based machine learning framework represents a methodological innovation in this study. In contrast to earlier prediction models that relied exclusively on demographic, clinical, and laboratory variables, the present model includes three TCM features, emaciation, dark tongue, and fatigue, which capture aspects of the patient’s functional and constitutional status that are not directly measurable through standard biochemical tests. From the perspective of integrative medicine, these TCM features represent phenotypic markers of underlying pathophysiological processes. Specifically, dark tongue may reflect chronic hypoxia and microcirculatory stasis, emaciation may indicate elevated catabolism and protein-energy wasting, and fatigue may signal mitochondrial dysfunction or impaired tissue perfusion[35,36]. The finding that all three TCM features were positively associated with LEAD in the SHAP analysis supports the hypothesis that TCM pattern differentiation can provide additional prognostic information beyond the assessment of conventional risk factors, confirming the value of a dual-dimensional feature framework in chronic disease prediction.
The nomogram constructed from the eight core predictors translates the complex machine learning model into a practical bedside tool[37,38]. The nomogram assigns point values to each predictor and uses the total score to calculate the likelihood of LEAD development, thereby enabling clinicians to perform rapid, individualized risk assessments without the need for computational resources[39]. The illustrative case of a 72-year-old hypertensive patient with moderate anemia and fatigue, who showed an approximate 89% probability of developing LEAD, demonstrates the potential of the model in identifying high-risk individuals who warrant further vascular evaluation. The clinical utility of the model was further supported by the DCA results, which showed positive net clinical benefit for all five models over a threshold probability range of 0.1-0.8, significantly outperforming both the “treat all” and “treat none” strategies[40]. This indicates that the application of the model in guiding clinical decisions would confer a net benefit to patients across a wide range of risk thresholds, which is essential for a screening tool intended for broad clinical application.
CONCLUSION
The proposed model integrating TCM syndrome features with clinical variables offers a practical tool for predicting the likelihood of LEAD development in T2DM. This study has several limitations. First, the data were obtained from a single institution, potentially limiting generalizability of the findings. Second, the range of variables was not fully comprehensive, and some risk factors were not included. In future work, we will expand the sample to include larger multi-center cohorts, enrich the feature set, include standardized, instrument-assisted TCM assessment, and validate the model externally to support its clinical use.
ACKNOWLEDGEMENTS
We gratefully acknowledge all study participants and the First Affiliated Hospital of Anhui University of Chinese Medicine.
Writing Group of the Clinical Guidelines for the Prevention and Treatment of Type 2 Diabetes in the Elderly in China; Geriatric Endocrinology and Metabolism Branch of the Chinese Geriatrics Society; Geriatric Endocrinology and Metabolism Branch of the Chinese Association of Geriatric Health Care Medicine; Geriatric Medicine Professional Committee of the Beijing Medical Award Foundation; National Clinical Research Center for Geriatric Diseases. [Clinical guidelines for prevention and treatment of type 2 diabetes mellitus in the elderly in China (2022 edition)].Zhonghua Neike Zazhi. 2022;61:12-50.
[RCA] [PubMed] [DOI] [Full Text][Cited by in RCA: 128][Reference Citation Analysis (1)]
Wu Q, Zhang YN, Zhang NN, Liu QY, Cai JR, Chen HS. Age affects the association of red blood cell indices with efficacy of remote ischemic conditioning in patients with acute moderate ischemic stroke.Sci Rep. 2024;14:22561.
[RCA] [PubMed] [DOI] [Full Text][Cited by in RCA: 3][Reference Citation Analysis (0)]
Meng X, Liu X, Tan J, Sheng Q, Zhang D, Li B, Zhang J, Zhang F, Chen H, Cui T, Li M, Zhang S. From Xiaoke to diabetes mellitus: a review of the research progress in traditional Chinese medicine for diabetes mellitus treatment.Chin Med. 2023;18:75.
[RCA] [PubMed] [DOI] [Full Text][Cited by in RCA: 44][Reference Citation Analysis (0)]
Shenyi J, Yahua L, Xu H, Mengjie C, Jiatuo XU, Hao LU, Qingguang C. Dark red tongue color formation caused by hyperglycemia is attributed to decreased blood flow of tongue tissue partially due to nuclear factor-kappa B pathway activation.J Tradit Chin Med. 2023;43:1118-1125.
[RCA] [PubMed] [DOI] [Full Text][Cited by in RCA: 2][Reference Citation Analysis (0)]
Liu X, Wang X, Xie R, Liu Z, Wang B, Zhai H. Explainable machine learning for predicting mechanical ventilation in Alzheimer's patients with pneumonia: a SHAP-guided XGBoost nomogram based on MIMIC-IV.Sci Rep. 2026;16:17250.
[RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)][Cited by in Crossref: 1][Reference Citation Analysis (0)]
Creativity or innovation: Grade A, Grade A, Grade A, Grade B
Scientific significance: Grade A, Grade B, Grade B, Grade B
P-Reviewer: Gao M, Associate Chief Nurse, Chief Nurse, Lecturer, China; Yan J, Chief Physician, Head, MD, Professor, China S-Editor: Lin C L-Editor: A P-Editor: Wang WB
