Interpretable machine learning and mendelian randomization identify risk factors for lower extremity arterial embolism and thrombosis

Abstract

Background:

Lower extremity arterial embolism and thrombosis lead to significant morbidity, but their risk factors are not fully characterized.

Objectives:

To identify risk factors and develop an interpretable machine learning (ML) model for predicting lower extremity arterial embolism and thrombosis, with validation using mendelian randomization (MR).

Methods:

In this retrospective case-control study, data were collected from patients with lower extremity arterial embolism and thrombosis treated at our department of vascular surgery between January 2018 and November 2025. Predictors were selected using LASSO regression, and 11 ML models were developed using the selected variables. The optimal model was interpreted and implemented as a web-based prediction tool. Clinical utility and model calibration were assessed using decision curve analysis and calibration curves. Key predictors were further assessed using MR and multivariable logistic regression.

Results:

XGBoost achieved the highest discrimination, with an AUC of 0.951 (95% CI 0.925–0.972) in the test set. MR analyses indicated that genetically predicted cerebrovascular disease (CD) (OR 1.773; 95% CI 1.043–3.015; P = 0.035) and gamma-glutamyl transferase (GGT) (OR 1.327; 95% CI 1.105–1.595; P = 0.003) were risk factors, whereas mean platelet volume (MPV) was protective (OR 0.879; 95% CI 0.778–0.993; P = 0.038). Multivariable logistic regression confirmed CD (OR 10.19; 95% CI 4.36–23.82; P < 0.001) and MPV (OR 0.63; 95% CI 0.52–0.76; P < 0.001) as independent predictors.

Conclusions:

Interpretable ML combined with MR identified a history of CD and lower MPV as factors associated with risk of lower extremity arterial embolism and thrombosis.

1 Introduction

Lower extremity arterial embolism and thrombosis are major causes of acute limb ischemia (ALI). In the United States, the annual incidence of ALI is estimated at 15 to 26 cases per 100,000 persons per year (1), and the burden is increasing with population ageing. Although multiple revascularization options are available—including catheter-directed thrombolysis, mechanical thrombectomy, and surgical embolectomy—clinical outcomes remain unsatisfactory. U.S. data suggest that 1-year mortality and major amputation rates after intervention remain as high as 17% and 34%, respectively (2), attributable in part to delays in diagnosis and treatment (3). Early risk stratification is therefore essential for prevention, timely diagnosis, and optimized management (4); however, a validated clinical prediction model for early detection is still lacking.

Machine learning (ML) can capture nonlinear relationships and interaction effects from high-dimensional, complex data (5, 6); however, its black-box decision process can limit clinical adoption (7). SHapley Additive exPlanations (SHAP) quantify each feature's contribution to an individual prediction using a game-theoretic framework (8, 9), whereas local interpretable model-agnostic explanations (LIME) provide local explanations that are model-agnostic (10, 11). Together, these approaches improve model transparency and may increase clinician trust in ML-based decision support.

Mendelian randomization (MR) leverages genetic variants [single-nucleotide polymorphisms (SNPs)] as instrumental variables to estimate causal effects. Because these instruments are fixed at conception and are generally not influenced by postnatal environmental or behavioral factors, MR can reduce confounding and mitigate reverse causation, thereby strengthening causal inference in observational studies (12).

This study aimed to identify risk factors for lower extremity arterial embolism and thrombosis and to develop and validate an interpretable ML model for early, accurate risk prediction. SHAP and LIME were employed to elucidate feature importance and explain model predictions, with causal relationships further validated using MR and multivariable logistic regression.

2 Materials and methods

2.1 Study cohort

The study was approved by the Ethics Committee of the Affiliated Hospital of Southwest Medical University (No. KY2025688) in accordance with the Declaration of Helsinki. This study consecutively enrolled patients with lower extremity arterial embolism or thrombosis who underwent surgical treatment in the Department of Vascular Surgery between January 2018 and November 2025. The inclusion criteria were: (1) confirmed clinical diagnosis of lower extremity arterial embolism or thrombosis, established based on surgical findings, preoperative imaging, and clinical presentation; (2) receipt of surgical intervention (open or endovascular). Exclusion criteria included: (1) lower extremity diseases due to other causes (e.g., thromboangiitis obliterans, traumatic thrombosis) (n = 70); (2) iatrogenic thrombosis following revascularization procedures (n = 55); (3) missing key clinical or laboratory data (n = 49). Patients with uncomplicated varicose veins (without concomitant peripheral arterial disease) who were treated in the vascular surgery department of our hospital between January 2018 and November 2025 were recruited as controls. Given the practical difficulty of obtaining comprehensive clinical and laboratory data across various age groups from completely healthy individuals in a hospital-based setting, these patients were selected because they routinely undergo lower extremity vascular ultrasound during preoperative evaluation in our department, which helps exclude patients with concurrent peripheral arterial disease and ensures a reliable negative control status. After applying the predefined inclusion and exclusion criteria, all eligible patients within the predefined timeframe were included in the analysis, yielding a final study cohort of 238 patients with lower extremity arterial embolism and thrombosis and 753 control patients. The patient selection flowchart is presented in Supplementary Figure S1.

2.2 Clinical features and data processing

The data collected in this study were sourced from patients' inpatient electronic medical records (EMRs), including basic patient information, history of smoking and alcohol consumption, medical history, and laboratory parameters (first laboratory tests after admission, obtained prior to the administration of any anticoagulants or surgical/interventional procedures). Basic patient information includes age and sex. Medical history includes hypertension (HBP), diabetes mellitus (DM), cerebrovascular disease (CD), and coronary heart disease (CHD). Laboratory parameters included total cholesterol (TC, 2.9–5.18 mmol/L), triglycerides (TG, 0.4–1.7 mmol/L), high-density lipoprotein cholesterol (HDL-C, 1.04–2.08 mmol/L), low-density lipoprotein cholesterol (LDL-C, 1–3.37 mmol/L), apolipoprotein A1 (APOA1, 1.1–1.7 g/L), apolipoprotein B (APOB, 0.8–1.55 g/L), alanine aminotransferase (ALT, 9–50 U/L), aspartate aminotransferase (AST, 15–40 U/L), total protein (TP, 65–85 g/L), albumin (ALB, 40–55 g/L), gamma-glutamyl transferase (GGT, 10–60 U/L), urea (3.2–8.2 mmol/L), uric acid (UA, 208–428 μmol/L), creatinine (Crea, 57–111 μmol/L), glomerular filtration rate (GFR, 75–145 mL/min), fibrinogen (Fib, 2–4 g/L), prothrombin time (PT, 11.0–14.5 s), thrombin time (TT, 14–21 s), activated partial thromboplastin time (APTT, 26–40 s), mean platelet volume (MPV, 9.4–12.5 fl), and platelet count (PLT, 125–350 × 10⁹/L). Cases with any missing variables from the aforementioned list were excluded. Given the relatively small proportion of missing data, a complete-case analysis was adopted. Therefore, only patients with complete information for all study variables were included in the final analysis to preserve the integrity of the original dataset and avoid additional assumptions associated with data imputation. Outliers, defined as values exceeding 1.5 times the interquartile range above the upper quartile or below the lower quartile, were manually reviewed to rule out data entry errors. For patients with multiple records, only the first episode was retained for analysis. The disease and control cohorts were combined and then randomly split into a training set and a test set at a ratio of 7:3.

2.3 Selection of variables

Variable selection was performed using least absolute shrinkage and selection operator (LASSO) regression, which imposes an L1 penalty on regression coefficients, shrinking those of non-informative variables toward zero and thus improving model sparsity and stability. The optimal regularization parameter λ was chosen by 10-fold cross-validation with the glmnet package (version 4.1-8), using the value of λ within one standard error of the minimum cross-validation error.

2.4 Model development

Eleven ML models were developed using the LASSO-selected variables, including support vector machine (SVM), neural network, multilayer perceptron (MLP), Gaussian process, gradient boosting machine (GBM), logistic regression, naive Bayes, adaptive boosting (AdaBoost), extreme gradient boosting (XGBoost), random forest, and k-nearest neighbors (KNN). All models were implemented in R using the caret package (version 7.0-1) with 5-fold cross-validation and hyperparameter tuning.

2.5 Model evaluation

Model performance was evaluated using standard discrimination and accuracy metrics, including the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), accuracy, F1-score, and Brier score. Calibration curves were further employed to visualize the agreement between predicted and observed risks. Decision curve analysis (DCA) was applied to evaluate the net clinical benefit of the model across a range of threshold probabilities. Calibration curves were generated using the predtools package (version 0.0.3), and DCA was implemented in R using a custom function.

2.6 Model explanation

To address the “black-box” nature of ML models and enhance their clinical acceptability, we employed both SHAP and LIME for post hoc model interpretation. SHAP, grounded in cooperative game theory, provides both global feature importance and local explanations for individual predictions. LIME was used as a complementary approach to further improve local interpretability. SHAP values were computed using the fastshap package (version 0.1.1) and visualized with the shapviz package (version 0.10.2), whereas LIME analyses were performed with the lime package (version 0.5.3).

2.7 Network calculator

The final prediction model was deployed as an R Shiny web application hosted on shinyapps.io (https://www.shinyapps.io), enabling individualized estimation of the risk of lower extremity arterial embolism or thrombosis.

2.8 MR analysis

MR analysis is based on three core assumptions: (1) strong association between genetic instruments and the exposure; (2) no direct association with the outcome; and (3) independence from confounders (13). Instrumental variables were selected using a genome-wide significance threshold (P < 5 × 10⁻⁸). Linkage disequilibrium was controlled for (clumping distance = 10,000 kb, r² < 0.001). The instrument strength was assessed using the F-statistic, calculated as F = [R²/(1 − R²)] × (N − 2), where R² = 2 × EAF × (1 − EAF) × β². Variants with F < 10, indicating weak instruments, were excluded (14). MR analyses were primarily performed using the inverse variance weighted (IVW) method, with a P value < 0.05 indicating significant evidence for a causal effect. Sensitivity analyses were conducted using MR Egger and weighted median methods. The robustness of the findings was further assessed by evaluating heterogeneity (Cochran's Q test) and horizontal pleiotropy (MR-Egger intercept test). All analyses were implemented with the “TwoSampleMR” R package (version 0.6.8). Genome-Wide Association Study (GWAS) summary data for lower extremity arterial embolism and thrombosis (1,297 cases; 463,106 controls) were sourced from the FinnGen R12. Data for CD, smoking, DM, MPV, AST, urea, GFR, and GGT were obtained from the IEU OpenGWAS and GWAS Catalog. CHD was excluded from MR analysis as it provided no predictive contribution in the model. Complete GWAS data are detailed in Supplementary Materials Table S1, with beta coefficients for CD derived from log (OR).

2.9 Statistics

All analyses were conducted using R (version 4.4.2). Continuous variables were expressed as mean ± standard deviation (SD) or median (interquartile range, IQR). For comparisons between two groups, the student's t test was used for normally distributed data, and the Mann–Whitney U test was used for non-normally distributed data. The normality of distribution was assessed using the Shapiro–Wilk test. Categorical variables were presented as n (%) and compared by chi-square tests. Dose-response relationships were assessed using Restricted Cubic Spline (RCS). Statistical significance was set at a two-sided P value < 0.05; P values ≥ 0.001 were reported as exact values, whereas P values < 0.001 were reported as P < 0.001. Multivariable logistic regression was used to assess independent risk factors, with female sex as the reference category for gender and “no” as the reference for all other categorical variables. Key predictors were identified as the intersection of variables demonstrating significant associations in both multivariable logistic regression and MR analysis.

3 Results

3.1 Baseline clinical information

The disease group (lower extremity arterial embolism and thrombosis) comprised 238 patients, while the control group (patients with simple varicose veins) included 753 patients. In the analysis of lower extremity arterial embolism, the training set contained 695 samples and the test set contained 296 samples. Compared with the control group, the disease group had a higher proportion of males (66% vs. 48%), was older (76 vs. 60 years), and demonstrated a higher prevalence of diabetes (24% vs. 6%), hypertension (48% vs. 22%), cerebrovascular disease (23% vs. 3%), and coronary heart disease (17% vs. 2%). Additionally, the disease group contained more patients with a history of smoking (37% vs. 21%) and alcohol consumption (33% vs. 21%). Regarding lipid profiles, the disease group showed lower levels of TC (4.51 vs. 4.67 mmol/L), LDL-C (2.68 vs. 2.82 mmol/L), and ApoA1 (1.39 vs. 1.47 g/L). Patients in the disease group showed impaired hepatic and renal function, as evidenced by elevated levels of ALT (22.95 vs. 18.80 U/L), AST (26.35 vs. 21.00 U/L), GGT (34.35 vs. 20.00 U/L), TP (68.65 vs. 67.50 g/L), urea (6.52 vs. 5.76 mmol/L), UA (339.60 vs. 327.60 μmol/L), and Crea (75.85 vs. 64.00 μmol/L), along with decreased ALB (40.20 vs. 40.80 g/L) and GFR (82.15 vs. 98.10 mL/min). In terms of coagulation function, the disease group exhibited prolonged PT (13.30 vs. 12.90 s), shortened APTT (32.90 vs. 34.70 s) and TT (17.40 vs. 18.00 s), and elevated Fib (3.86 vs. 2.81 g/L). Meanwhile, MPV was lower in the disease group (9.80 vs. 10.80 fl). Variables were generally well-balanced between the training and test sets, except for slightly older age in the training set (64 vs. 61 years) (Table 1).

To assess the potential impact of complete-case analysis on selection bias, baseline characteristics were compared between included patients and those excluded because of missing data (Supplementary Materials Table S2). Except for ALT (22.95 vs. 17.00 U/L) and TT (17.40 vs. 16.90 s), no statistically significant differences were observed for most baseline characteristics, including age, sex, major comorbidities, lipid profiles, renal function parameters, and MPV. These findings indicate that the baseline characteristics of the included and excluded patients were highly similar, suggesting a minimal risk of substantial selection bias resulting from the complete-case analysis.

3.2 Selection of predictor variables

LASSO regression retained 13 predictors with non-zero coefficients for model development, including CD, Fib, smoking, DM, MPV, CHD, year, PT, urea, GFR, AST, GGT, and TT (Figures 1A–C).

Figure 1

3.3 Model development, performance comparison and sensitivity analysis

The eleven ML models showed good discriminative performance for predicting lower extremity arterial embolism and thrombosis, with training AUCs ranging from 0.919 to 1.000 and test AUCs ranging from 0.905 to 0.957 (Figures 1D,E). After excluding overfitted models (training AUC = 1), the XGBoost algorithm demonstrated the most favorable balance of discrimination and generalization, achieving a training AUC of 0.978 and a test AUC of 0.951. The model also demonstrated strong performance in accuracy (0.922 training; 0.899 test), specificity (0.975 training; 0.973 test), and precision (0.906 training; 0.887 test), although sensitivity was comparatively lower (0.754 training; 0.662 test). A comprehensive summary of all performance metrics is provided in Table 2. Calibration curves showed that XGBoost followed the ideal diagonal line reasonably well across most probability ranges in both the training (Figure 1F) and test sets (Figure 1G), indicating generally good agreement between predicted and observed risks. DCA in the training (Figure 1H) and test sets (Figure 1I) demonstrated that the XGBoost model yielded a higher net clinical benefit across a wide range of threshold probabilities.

3.4 Model explanation

SHAP analysis of the XGBoost model, developed using LASSO-selected features, revealed the order of feature importance based on mean absolute SHAP values as Fib, year, GFR, GGT, MPV, PT, TT, AST, CD, urea, smoking, and DM, while CHD exhibited a SHAP value of zero (Figure 2A). The SHAP beeswarm plot illustrates the directional impact of feature variations on model predictions, where each point represents an individual sample. In the XGBoost model, advanced age, elevated Fib, GGT, PT, TT, AST, and urea levels, decreased GFR and MPV levels, and positive history of CD, DM, or smoking were identified as predictive factors for lower extremity arterial embolism and thrombosis. CHD demonstrated no predictive influence (Figure 2B).

Figure 2

SHAP dependence plots further delineated the specific relationship patterns between individual features and the risk of lower extremity arterial embolism and thrombosis. These plots illustrate how changes in each continuous variable affect the predicted probability, while for categorical variables, they reveal which category elevates risk. For instance, a history of DM significantly increased the predicted probability of embolism and thrombosis (Figure 2C).

SHAP waterfall plots (Figure 2D) and force plots (Figure 2E) enabled model interpretation at the individual level. The predicted probability of embolism and thrombosis reached 0.919 in high-risk Sample 50 with lower extremity arterial embolism and thrombosis, with the most significant positive contributions from elevated fib (7.72, +0.197), decreased MPV (8.4, +0.1286), advanced age (86 years, +0.1041), reduced GFR (45.9, +0.0883), and prolonged PT (14.1, +0.0495), along with additional positive effects from specific levels of GGT, urea and AST, while negative contributions came from the absence of CD and smoking history, a history of DM showed minimal positive effect, and CHD history demonstrated no predictive value. In the visual representations, orange indicates positive contributions and purple indicates negative contributions. LIME analysis provided additional interpretability at the individual level. For high-risk sample 50 (Figure 2F), the predicted probability of lower extremity arterial embolism and thrombosis reached 95%, with elevated Fib and AST, decreased MPV and GFR, advanced age, and CD history identified as key predictors.

3.5 MR and multivariable regression analyses

To further identify potential causal risk factors, MR analysis was performed for predictor variables from the model, excluding basic demographic and coagulation parameters, and was supplemented by multivariable logistic regression for validation. MR analyses were conducted for CD, DM, MPV, AST, urea, smoking, GFR, and GGT. The MR results (Supplementary Table S3) provided evidence of positive associations of CD (OR: 1.773, 95% CI: 1.043–3.015, P = 0.035) and GGT (OR: 1.327, 95% CI: 1.105–1.595, P = 0.003) with lower extremity arterial embolism and thrombosis, with no evidence of horizontal pleiotropy or heterogeneity (P > 0.05). MPV showed a negative association with lower extremity arterial embolism and thrombosis (OR: 0.879, 95% CI: 0.778–0.993, P = 0.038), with no evidence of horizontal pleiotropy but significant heterogeneity (P < 0.001). It should be noted that some degree of heterogeneity is acceptable in MR analyses, particularly when a larger number of SNP instruments are included. No evidence of a causal relationship was found for smoking, DM, urea, GFR, or AST (Supplementary Table S3). Figures 3A–L presents the MR results for the significant variables (CD, GGT, and MPV), including scatter plots, funnel plots, leave-one-out sensitivity analyses, and forest plots. Multivariable logistic regression (Table 3) showed that, among the variables with causal evidence from MR analysis, CD history (OR: 10.19, 95% CI: 4.36–23.82, P < 0.001) was positively associated with lower extremity arterial embolism and thrombosis, whereas MPV level (OR: 0.63, 95% CI: 0.52–0.76, P < 0.001) showed a negative association. GGT level demonstrated no significant association (OR: 1.00, 95% CI: 1.00–1.00, P = 0.843). Integrating the findings from both MR analysis and multivariable logistic regression, CD and MPV were identified as potential causal risk factors for lower extremity arterial embolism and thrombosis. Although GGT showed a significant association in the MR analysis, it did not remain independently associated with the outcome in the multivariable logistic regression model, suggesting that its clinical effect may be influenced by other correlated factors.

Figure 3

3.6 RCS analysis

To determine the critical threshold at which a continuous variable increases the risk of lower extremity arterial embolism and thrombosis, RCS analysis was performed for MPV. The RCS analysis (Figure 3M) revealed a significant overall association between MPV and thrombotic risk (overall P < 0.001), with evidence of nonlinearity (nonlinear P = 0.003). The spline curve showed a progressively increased risk with decreasing MPV levels, with a notably higher predicted probability of thrombotic events when MPV fell below 8.70 fl. This association was attenuated after multivariable adjustment, and the evidence for nonlinearity was no longer significant (Supplementary Figure S2).

3.7 Web calculator

The prediction model for lower extremity arterial embolism and thrombosis is available online at https://lixiaodong1999.shinyapps.io/thrombosis-risk-predictor/. This web-based platform allows researchers and clinicians to assess patient-specific risk by entering the predictor variables, and it returns both the estimated disease probability and the corresponding SHAP values for feature interpretation. However, it is crucial to emphasize that this tool is currently strictly for exploratory and research purposes. Given the lack of external validation, it should not be used as a standalone basis for direct clinical decision-making until its performance is rigorously validated in independent, multi-center cohorts. The layout of the user interface is illustrated in Supplementary material Figure S3.

4 Discussion

By integrating ML and MR analyses with multivariable logistic regression, this study identified a history of CD and lower MPV as independent predictors of lower extremity arterial embolism and thrombosis. RCS analysis further revealed a marked increase in risk when MPV decreased below 8.70 fl.

Beyond conventional coagulation markers, routine hematological parameters have emerged as valuable tools for risk stratification across a broad spectrum of thrombotic disorders. Blood cellular indices have demonstrated prognostic significance in acute pulmonary embolism (15), while baseline hemoglobin levels have been shown to predict clinical outcomes in venous thromboembolism (16). Collectively, these findings suggest that simple, cost-effective, and universally available hematological indices can provide valuable risk stratification across thrombotic disorders. MPV is generally regarded as an indicator of platelet activity, as larger platelets exhibit greater adhesion, aggregation, and procoagulant potential (17, 18). Consistent with this concept, elevated MPV has been associated with increased thrombotic or embolic risk in patients with coronary disease and CD (19–22). In peripheral arterial occlusive disease, however, studies have found an association between lower MPV levels and higher incidence of chronic limb-threatening ischemia (23), and these patients frequently present with thrombotic lumen occlusion (24, 25). Similar findings have also been reported in venous thrombosis. Studies in high-risk groups, including patients with cancer, have suggested that elevated MPV may inversely associate with venous thromboembolism (VTE) (26–28), and two large retrospective cohorts consistently observed lower MPV in patients with deep vein thrombosis (DVT) (29, 30).

Importantly, smaller platelets are not necessarily functionally impaired. Upon adenosine diphosphate (ADP) stimulation, smaller platelets have been shown to exhibit higher P-selectin expression (31), which may enhance leukocyte recruitment and stabilize platelet–leukocyte adhesion within thrombi. This effect can be further amplified by procoagulant mediators released from leukocytes, potentially accelerating thrombus propagation. Moreover, under comparable extracellular calcium concentrations, the reactivity of small and large platelets does not differ significantly (32), suggesting that small platelets retain full capacity to participate in thrombus formation. In addition, reduced platelet volume may be accompanied by increased platelet-derived microparticles with enhanced surface reactivity. Through high cross-reactivity with tissue factor (TF) and platelet aggregation receptors, smaller platelets ultimately promote arterial thrombus formation (33). The role of MPV in arterial embolism remains controversial. Although most studies report a positive association between MPV and thromboembolism, some evidence suggests that MPV is not associated with left atrial thrombus formation in patients with atrial fibrillation (34), implying a limited impact on embolic risk in this setting.

Clinically, our analysis identified a history of CD as an important risk factor for lower extremity arterial embolism or thrombosis. This association may reflect shared risk characteristics, including advanced age and metabolic abnormalities. In addition, CD can indicate systemic high-risk pathophysiology, such as advanced atherosclerosis or atrial fibrillation, which may also predispose to thrombotic events in the lower extremity arteries (35, 36). Existing evidence suggests that a history of CD, particularly stroke, is associated with an increased risk of peripheral arterial embolism (37). A higher prevalence of prior CD has also been reported among patients with ALI (38, 39). In patients with ALI, CD history has been identified as an independent risk factor for in-hospital complications (38). Mechanistically, elevated levels of neutrophil extracellular traps (NETs) have been observed in CD (40–43), and NETs may potentiate coagulation by interacting with coagulation factors and facilitating platelet–neutrophil aggregate formation (44–46).

Our ML model and SHAP analysis further identified advanced age as a prominent risk factor, with feature importance ranking second only to fibrinogen. Advanced age emerged as a strong predictor of lower extremity arterial embolism and thrombosis, likely due to its close association with the high prevalence of lower extremity atherosclerosis and cardiac arrhythmias (particularly atrial fibrillation). Epidemiological data indicate that the prevalence of peripheral artery disease increases continuously with advancing age (47). Aging itself is a key driver of atherosclerosis progression (48) and may further promote thrombosis through multiple pathways, including chronic inflammation (49), a hypercoagulable state (50), altered platelet function (51), and endothelial injury (52). Meanwhile, the prevalence of atrial fibrillation increases with age, significantly elevating the risk of cardioembolic events (53). Therefore, age not only drives in situ thrombosis by accelerating atherosclerosis but also exacerbates the burden of lower extremity arterial embolism by increasing the risk of cardioembolic events.

The XGBoost model developed in this study achieved an AUC of 0.951 in the test set. However, these high performance metrics must be interpreted with caution. Given the single-center, case-control design and the strong separation between the case and control groups, the reported AUC may reflect the optimism inherent in internal validation, potentially overestimating the model's real-world performance. The absence of external validation, temporal validation, or bootstrap optimism correction further underscores the need for prudence when applying these findings. Future multi-center prospective studies incorporating independent external validation cohorts are essential to rigorously evaluate the true predictive efficacy of this model in broader, more heterogeneous clinical populations. Similarly, the results of the RCS analysis should be interpreted with caution. Notably, this threshold (8.70 fl) falls below the standard reference interval (9.4–12.5 fl), suggesting that a markedly reduced MPV serves as a specific warning sign for high thrombotic risk. However, since MPV measurements vary across different hematology analyzers, this absolute cut-off may be assay- and center-specific. Therefore, it may not be universally applicable, and other institutions should recalibrate this threshold based on their local reference ranges before clinical implementation. Finally, although the proposed mechanisms provide a plausible explanation for the paradoxical association between low MPV and thrombotic risk, these biological interpretations remain largely speculative and require validation in large-scale prospective studies.

This study has several limitations. First, the retrospective design may introduce selection bias. Specifically, the exclusion of patients with missing key laboratory indicators not only reduced the sample size but also potentially introduced selection bias. Furthermore, arterial embolism and in situ arterial thrombosis were combined into a single endpoint due to the limited sample size, which precluded meaningful subgroup analyses despite their distinct pathophysiological mechanisms. Additionally, variable selection was primarily based on clinical prior knowledge and data availability, potentially omitting other influential factors. Second, the control group was not composed of completely healthy individuals, but rather patients with uncomplicated varicose veins. Although the majority of these controls did not have concurrent cardiovascular diseases, this choice inevitably introduced spectrum bias, meaning that the model's high performance may partly reflect these baseline disparities rather than its pure predictive efficacy in the general population. Third, the model demonstrated relatively low sensitivity, which limits its clinical utility as a standalone screening tool for this limb-threatening condition, and it lacked an independent external validation cohort, leaving its generalizability uncertain. Finally, the MR analysis has certain limitations. The evidence for MPV being protective was only marginally significant (P = 0.038) accompanied by significant heterogeneity (P < 0.001), which weakens the confidence in this specific causal interpretation. Furthermore, while the clinical data were derived from a Chinese population, the GWAS data used for causal inference were obtained from European populations, and future studies using Asian genomic data are needed to validate the cross-population applicability of the conclusions.

References

• 1.

  GillilandCShahJMartinJGMillerMJJr. Acute limb ischemia. Tech Vasc Interv Radiol. (2017) 20(4):274–80. 10.1053/j.tvir.2017.10.008

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 2.

  DobeshKNatourAKKabbaniLSRteilALeeANypaverTJet al. Patients with acute lower limb ischemia continue to have significant morbidity and mortality. Ann Vasc Surg. (2024) 108:127–40. 10.1016/j.avsg.2024.03.024

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 3.

  BjorckMEarnshawJJAcostaSBastos GoncalvesFCochennecFDebusESet al. Editor’s choice—European society for vascular surgery (ESVS) 2020 clinical practice guidelines on the management of acute limb ischaemia. Eur J Vasc Endovasc Surg. (2020) 59(2):173–218. 10.1016/j.ejvs.2019.09.006

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 4.

  FallujiNMukherjeeD. Critical and acute limb ischemia: an overview. Angiology. (2014) 65(2):137–46. 10.1177/0003319712470966

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 5.

  ChenJHAschSM. Machine learning and prediction in medicine—beyond the peak of inflated expectations. N Engl J Med. (2017) 376(26):2507–9. 10.1056/NEJMp1702071

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 6.

  TopolEJ. High-performance medicine: the convergence of human and artificial intelligence. Nat Med. (2019) 25(1):44–56. 10.1038/s41591-018-0300-7

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 7.

  RudinC. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nat Mach Intell. (2019) 1(5):206–15. 10.1038/s42256-019-0048-x

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 8.

  Ponce-BobadillaAVSchmittVMaierCSMensingSStodtmannS. Practical guide to SHAP analysis: explaining supervised machine learning model predictions in drug development. Clin Transl Sci. (2024) 17(11):e70056. 10.1111/cts.70056

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 9.

  VimbiVShaffiNMahmudM. Interpreting artificial intelligence models: a systematic review on the application of LIME and SHAP in Alzheimer’s disease detection. Brain Inform. (2024) 11(1):10. 10.1186/s40708-024-00222-1

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 10.

  AlabiROElmusratiMLeivoIAlmangushAMäkitieAA. Machine learning explainability in nasopharyngeal cancer survival using LIME and SHAP. Sci Rep. (2023) 13(1):8984. 10.1038/s41598-023-35795-0

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 11.

  ElbeltagiASrivastavaACaoXBilaliAERazaAKhadkeLet al. An interpretable machine learning approach based on SHAP, Sobol and LIME values for precise estimation of daily soybean crop coefficients. Sci Rep. (2025) 15(1):36594. 10.1038/s41598-025-20386-y

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 12.

  DaviesNMHolmesMVDavey SmithG. Reading Mendelian randomisation studies: a guide, glossary, and checklist for clinicians. Br Med J. (2018) 362:k601. 10.1136/bmj.k601

  • CrossRef
  • Google Scholar
  • View reference in article

• 13.

  RasoolyDPelosoGM. Two-sample multivariable Mendelian randomization analysis using R. Curr Protoc. (2021) 1(12):e335. 10.1002/cpz1.335

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 14.

  PalmerTMLawlorDAHarbordRMSheehanNATobiasJHTimpsonNJet al. Using multiple genetic variants as instrumental variables for modifiable risk factors. Stat Methods Med Res. (2012) 21(3):223–42. 10.1177/0962280210394459

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 15.

  SiddiquiFTafurAHussainMGarcia-OrtegaADarkiAFareedJet al. The prognostic value of blood cellular indices in pulmonary embolism. Am J Hematol. (2024) 99(9):1704–11. 10.1002/ajh.27379

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 16.

  SiniscalchiCDi MiccoPTufanoAPerisMLLópez-MiguelPAlda-LozanoAet al. Baseline hemoglobin values and clinical outcomes in acute venous thromboembolism: insights from the RIETE registry. Am J Hematol. (2025) 100(8):1444–7. 10.1002/ajh.27707

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 17.

  ProvidênciaRFaustinoAPaivaLFernandesABarraSPimentaJet al. Mean platelet volume is associated with the presence of left atrial stasis in patients with non-valvular atrial fibrillation. BMC Cardiovasc Disord. (2013) 13:40. 10.1186/1471-2261-13-40

  • CrossRef
  • Google Scholar
  • View reference in article

• 18.

  ProvidenciaRFerreiraMJGoncalvesLFaustinoAPaivaLFernandesAet al. Mean corpuscular volume and red cell distribution width as predictors of left atrial stasis in patients with non-valvular atrial fibrillation. Am J Cardiovasc Dis. (2013) 3(2):91–102.

  • Pubmed Abstract
  • Google Scholar
  • View reference in article

• 19.

  GreiseneggerSEndlerGHsiehKTentschertSMannhalterCLalouschekW. Is elevated mean platelet volume associated with a worse outcome in patients with acute ischemic cerebrovascular events?Stroke. (2004) 35(7):1688–91. 10.1161/01.STR.0000130512.81212.a2

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 20.

  BayarNArslanSCagirciGUreyenCMCaySYukselIOet al. Usefulness of mean platelet volume for predicting stroke risk in paroxysmal atrial fibrillation patients. Blood Coagul Fibrinolysis. (2015) 26(6):669–72. 10.1097/MBC.0000000000000334

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 21.

  TurfanMErdoganEErtasGDuranMMuratSNCelikEet al. Usefulness of mean platelet volume for predicting stroke risk in atrial fibrillation patients. Blood Coagul Fibrinolysis. (2013) 24(1):55–8. 10.1097/MBC.0b013e32835a0850

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 22.

  XuXFJiangFLOuMJZhangZH. The association between mean platelet volume and chronic atrial fibrillation and the presence of thrombotic events. Biomed Rep. (2015) 3(3):388–94. 10.3892/br.2015.418

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 23.

  GuetlKRaggamRBMusterVGressenbergerPVujicJAvianAet al. The white blood cell count to mean platelet volume ratio for the prediction of chronic limb-threatening ischemia in lower extremity artery disease. J Clin Med. (2019) 8(10):1593. 10.3390/jcm8101593

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 24.

  NarulaNDannenbergAJOlinJWBhattDLJohnsonKWNadkarniGet al. Pathology of peripheral artery disease in patients with critical limb ischemia. J Am Coll Cardiol. (2018) 72(18):2152–63. 10.1016/j.jacc.2018.08.002

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 25.

  JinnouchiHSakakuraKHatoriMTsukuiTWatanabeYSeguchiMet al. Association between medial calcification and chronic thrombus: an optical frequency domain imaging study. Catheter Cardiovasc Interv. (2025) 106(1):109–19. 10.1002/ccd.31524

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 26.

  Rupa-MatysekJGilLKroll-BalcerzakRBaranskaMKomarnickiM. Mean platelet volume as a predictive marker for venous thromboembolism and mortality in patients treated for diffuse large B-cell lymphoma. Hematol Oncol. (2017) 35(4):456–64. 10.1002/hon.2321

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 27.

  RiedlJKaiderAReitterEMMarosiCJägerUSchwarzingerIet al. Association of mean platelet volume with risk of venous thromboembolism and mortality in patients with cancer. Results from the Vienna cancer and thrombosis study (CATS). Thromb Haemost. (2014) 111(4):670–8. 10.1160/TH13-07-0603

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 28.

  WangZChenXWuJZhouQLiuHWuYet al. Low mean platelet volume is associated with deep vein thrombosis in older patients with hip fracture. Clin Appl Thromb Hemost. (2022) 28:10760296221078837. 10.1177/10760296221078837

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 29.

  LiJLiangY. Associations between mean platelet volume and risk of deep vein thrombosis: a Mendelian randomization study and a retrospective study. Int J Gen Med. (2023) 16:515–24. 10.2147/IJGM.S401059

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 30.

  LippiGBuonocoreRCervellinG. The mean platelet volume is decreased in patients diagnosed with venous thromboembolism in the emergency department. Semin Thromb Hemost. (2016) 42(6):632–5. 10.1055/s-0036-1571335

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 31.

  HandtkeSSteilLPalankarRConradJCauhanSKrausLet al. Role of platelet size revisited-function and protein composition of large and small platelets. Thromb Haemost. (2019) 119(3):407–20. 10.1055/s-0039-1677875

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 32.

  HandtkeSWescheJPalankarRGreinacherAThieleT. Function of large and small platelets differs, depending on extracellular calcium availability and type of inductor. Thromb Haemost. (2020) 120(7):1075–86. 10.1055/s-0040-1712447

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 33.

  RiefPRaggamRBAvianAHafnerFSarebanMWehrschuetzMet al. Low mean platelet volume is associated with critical limb ischemia in peripheral arterial occlusive disease. Sci Rep. (2018) 8(1):6718. 10.1038/s41598-018-25058-8

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 34.

  YuceMCakiciMDavutogluVOzerOSariIErcanSet al. Relationship between mean platelet volume and atrial thrombus in patients with atrial fibrillation. Blood Coagul Fibrinolysis. (2010) 21(8):722–5. 10.1097/MBC.0b013e32833d6eda

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 35.

  WasilewskaMGosk-BierskaI. Thromboembolism associated with atrial fibrillation as a cause of limb and organ ischemia. Adv Clin Exp Med. (2013) 22(6):865–73.

  • Pubmed Abstract
  • Google Scholar

• 36.

  MenkeJLüthjeLKastrupALarsenJ. Thromboembolism in atrial fibrillation. Am J Cardiol. (2010) 105(4):502–10. 10.1016/j.amjcard.2009.10.018

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 37.

  FrostLEngholmGJohnsenSMøllerHHennebergEWHustedS. Incident thromboembolism in the aorta and the renal, mesenteric, pelvic, and extremity arteries after discharge from the hospital with a diagnosis of atrial fibrillation. Arch Intern Med. (2001) 161(2):272–6. 10.1001/archinte.161.2.272

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 38.

  TsujimuraTTakaharaMIidaOKohsakaSSogaYFujiharaMet al. In-hospital outcomes after endovascular therapy for acute limb ischemia: a report from a Japanese nationwide registry [J-EVT registry]. J Atheroscler Thromb. (2021) 28(11):1145–52. 10.5551/jat.60053

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 39.

  JohnerFThalhammerCJacomellaVHusmannMAmann-VestiB. Differences in cardiovascular risk factors between patients with acute limb ischemia and intermittent claudication. Angiology. (2014) 65(6):497–500. 10.1177/0003319713487428

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 40.

  GenchiASemeranoAGullottaGSStramboDSchwarzGBergamaschiAet al. Cerebral thrombi of cardioembolic etiology have an increased content of neutrophil extracellular traps. J Neurol Sci. (2021) 423:117355. 10.1016/j.jns.2021.117355

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 41.

  DenormeFPortierIRustadJLCodyMJde AraujoCVHokiCet al. Neutrophil extracellular traps regulate ischemic stroke brain injury. J Clin Invest. (2022) 132(10):e154225. 10.1172/JCI154225

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 42.

  SchneiderHMünchDSpalartVStroobantsMBerlisAMaurerCJet al. Neutrophil extracellular traps and delayed cerebral ischemia in patients with subarachnoid hemorrhage. J Am Heart Assoc. (2025) 14(19):e041325. 10.1161/JAHA.124.041325

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 43.

  JinJZhaoXLiWWangFTianJWangNet al. Neutrophil extracellular traps: a novel therapeutic target for intracranial hemorrhage. Thromb Res. (2022) 219:1–13. 10.1016/j.thromres.2022.08.024

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 44.

  GaoXZhaoXLiJLiuCLiWZhaoJet al. Neutrophil extracellular traps mediated by platelet microvesicles promote thrombosis and brain injury in acute ischemic stroke. Cell Commun Signal. (2024) 22(1):50. 10.1186/s12964-023-01379-8

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 45.

  DenormeFRustadJLCampbellRA. Brothers in arms: platelets and neutrophils in ischemic stroke. Curr Opin Hematol. (2021) 28(5):301–7. 10.1097/MOH.0000000000000665

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 46.

  ZhouYXuZLiuZ. Impact of neutrophil extracellular traps on thrombosis formation: new findings and future perspective. Front Cell Infect Microbiol. (2022) 12:910908. 10.3389/fcimb.2022.910908

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 47.

  ZhouJShanSShenGTanBYingJRudanIet al. National, regional and provincial prevalence of peripheral artery disease in Chinese adults in 2023: an updated systematic review and modelling study. J Glob Health. (2026) 16:04155. 10.7189/jogh.16.04155

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 48.

  WangJCBennettM. Aging and atherosclerosis: mechanisms, functional consequences, and potential therapeutics for cellular senescence. Circ Res. (2012) 111(2):245–59. 10.1161/CIRCRESAHA.111.261388

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 49.

  SchlaudeckerJBeckerR. Inflammatory response and thrombosis in older individuals. Semin Thromb Hemost. (2014) 40(6):669–74. 10.1055/s-0034-1387882

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 50.

  WilkersonWRSaneDC. Aging and thrombosis. Semin Thromb Hemost. (2002) 28(6):555–68. 10.1055/s-2002-36700

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 51.

  MohebaliDKaplanDCarlisleMSupianoMARondinaMT. Alterations in platelet function during aging: clinical correlations with thromboinflammatory disease in older adults. J Am Geriatr Soc. (2014) 62(3):529–35. 10.1111/jgs.12700

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 52.

  SepúlvedaCPalomoIFuentesE. Mechanisms of endothelial dysfunction during aging: predisposition to thrombosis. Mech Ageing Dev. (2017) 164:91–9. 10.1016/j.mad.2017.04.011

  • CrossRef
  • Google Scholar

• 53.

  ShiSTangYZhaoQYanHYuBZhengQet al. Prevalence and risk of atrial fibrillation in China: a national cross-sectional epidemiological study. Lancet Reg Health West Pac. (2022) 23:100439. 10.1016/j.lanwpc.2022.100439

  • Pubmed Abstract
  • CrossRef
  • Google Scholar