Machine Learning–Based Prediction of Lower Extremity Deep Vein Thrombosis in Patients with Acute Exacerbation of Chronic Obstructive Pulmonary Disease

Abstract

Objective

Patients with acute exacerbation of chronic obstructive pulmonary disease (AECOPD) have an increased risk of lower-extremity deep vein thrombosis (DVT) due to inflammation, immobility, and coagulation abnormalities. This study aimed to develop a machine learning model for predicting DVT risk in AECOPD patients and identify important predictors.

Methods

This single-center retrospective study included 1,500 patients with AECOPD who were classified into DVT (n = 126) and non-DVT groups (n = 1,374) according to lower-limb venous ultrasonography findings. Clinical characteristics and laboratory parameters were collected. The Boruta algorithm was applied for feature selection, and three machine learning models, including decision tree (DT), multilayer perceptron (MLP), and extreme gradient boosting (XGBoost), were developed. Model performance was assessed using the receiver operating characteristic (ROC) curve, area under the curve (AUC), and multiple classification metrics. TOPSIS was used for comprehensive model evaluation, and SHAP analysis was performed to interpret the optimal model.

Results

Compared with the non-DVT group, patients with DVT had higher levels of C-reactive protein, D-dimer, and systemic inflammation–coagulation index (SCI) (all P < 0.05). The Boruta algorithm identified 14 key features. XGBoost showed the best performance in the testing set, with an AUC of 0.707, compared with MLP (0.684) and DT (0.667). TOPSIS analysis ranked XGBoost highest, with an overall score of 0.806. SHAP analysis identified D-dimer as the most influential predictor, followed by SCI, albumin, mean corpuscular volume, and C-reactive protein.

Conclusion

The XGBoost model showed potential for DVT risk assessment in patients with AECOPD. SCI, as an integrated marker of inflammation and coagulation, may provide complementary information for risk stratification. However, further validation in independent cohorts is required before its potential clinical application.

Background

Chronic obstructive pulmonary disease (COPD) is a common chronic respiratory disorder worldwide.¹ Acute exacerbations of COPD (AECOPD) are frequently accompanied by intensified systemic inflammation, reduced physical activity, and prolonged bed rest, all of which contribute to an increased risk of venous thromboembolism (VTE).^2–4 Deep vein thrombosis (DVT), a major component of VTE, can prolong hospital stay and may lead to severe complications such as pulmonary embolism, thereby increasing mortality risk.^5,6

In clinical practice, DVT in patients with AECOPD is often difficult to identify early because its symptoms are non-specific. Some patients present with subtle or even absent clinical signs, which can result in missed or delayed diagnosis.^7,8 Although D-dimer and other coagulation-related markers are commonly used for initial screening, their specificity is limited in the setting of infection,⁹ systemic inflammation,¹⁰ and chronic disease,¹¹ as they are frequently elevated under these conditions. Therefore, relying on a single biomarker is insufficient for accurate risk prediction, and early identification and risk stratification of DVT in AECOPD patients remain a clinical challenge.

In recent years, machine learning techniques have shown advantages in clinical risk prediction by integrating multiple variables and capturing complex, non-linear relationships among them, thereby improving predictive performance.^12,13 However, studies focusing on machine learning–based prediction of DVT in AECOPD patients are still limited. In particular, there remains a need for practical and interpretable models based on routine clinical and laboratory parameters that can be easily applied in clinical settings. Previous studies on DVT prediction have mainly relied on conventional clinical risk factors or single biomarkers, which may not fully capture the complex interactions between systemic inflammation and coagulation abnormalities in AECOPD patients. Therefore, an integrated approach incorporating multiple clinical variables and interpretable machine-learning algorithms may provide additional insights into individualized DVT risk assessment. Furthermore, considering the close interaction between inflammation and coagulation activation during thrombus formation, the systemic inflammation–coagulation index (SCI), which integrates platelet count, fibrinogen, and white blood cell count, may provide additional information regarding the inflammatory and hypercoagulable state associated with DVT development.

Based on this, the present study aimed to develop and compare several machine learning models using routine clinical and laboratory data to predict the risk of lower-extremity DVT in patients with AECOPD. The optimal model was selected through performance comparison and interpretability analysis, to explore the potential value of machine-learning approaches and SCI in DVT risk assessment among patients with AECOPD.

Methods

Study Population

This single-center retrospective observational study included patients with chronic obstructive pulmonary disease (COPD) admitted to the Department of Respiratory Medicine at the Central Hospital of Enshi Tujia and Miao Autonomous Prefecture between January 2023 and December 2025. All patients met the diagnostic criteria for COPD according to the Global Initiative for Chronic Obstructive Lung Disease (GOLD 2017) guidelines.¹⁴ Patients hospitalized for acute exacerbation of COPD (AECOPD) who underwent lower-extremity venous ultrasonography during admission were included for evaluation of deep vein thrombosis (DVT). Laboratory measurements included in the prediction models were obtained at admission before the diagnosis of DVT by ultrasonography to ensure that the model represented risk prediction rather than discrimination of established thrombosis. Exclusion criteria were as follows: (1) missing more than 20% of key clinical or laboratory data; (2) absence of lower-limb venous ultrasound examination during hospitalization; (3) conditions known to significantly affect coagulation, such as active malignancy, severe hepatic or renal failure, or recent major surgery; and (4) repeated admissions, in which only the first hospitalization record was retained for analysis. The selection process is shown in Figure 1. The study was approved by the Ethics Committee of the Central Hospital of Enshi Tujia and Miao Autonomous Prefecture (Approval No. 202607001) and conducted in accordance with the Declaration of Helsinki. Given the retrospective design, the requirement for written informed consent was waived. All patient data were anonymized and kept confidential, and patient privacy was strictly protected throughout the study.

Figure 1.

Flowchart of Patient Enrollment and Inclusion Criteria.

Data Collection

Clinical information and laboratory data were extracted from medical records. Baseline characteristics included age, sex, and comorbidities such as hypertension, diabetes mellitus, heart failure, prior stroke, arrhythmia, and smoking history. Laboratory parameters included serum biochemistry: albumin (ALB), alanine aminotransferase (ALT), aspartate aminotransferase (AST), creatinine (CREA), urea (UREA), uric acid (UA), total bilirubin (TBIL), direct bilirubin (DBIL), C-reactive protein (CRP), D-dimer, glucose (GLU), and fibrinogen. Complete blood count–related variables included hemoglobin (HGB), red blood cell count (RBC), hematocrit (HCT), mean corpuscular volume (MCV), mean corpuscular hemoglobin (MCH), mean corpuscular hemoglobin concentration (MCHC), white blood cell count (WBC), lymphocyte count (LY), monocyte count (MONO), platelet count (PLT), mean platelet volume (MPV), plateletcrit (PCT), and platelet distribution width (PDW). Arterial blood gas analysis included pH. In addition, the systemic inflammation–coagulation index (SCI) was calculated as: SCI = platelet count × fibrinogen / white blood cell count.¹⁵ Missing values were imputed using multiple imputation by chained equations (MICE) before model development.

Feature Selection

To reduce redundancy and minimize overfitting, the dataset was first randomly divided into a training set and a testing set at a ratio of 7:3. The Boruta algorithm was subsequently applied exclusively within the training set for feature selection. Boruta is a wrapper method based on random forest importance ranking, which introduces shuffled shadow features and compares them with original variables to identify features significantly associated with the outcome.

Model Development

After feature selection. Three machine learning models were developed using the training cohort: extreme gradient boosting (XGBoost), decision tree (DT), and multilayer perceptron (MLP). All model training and hyperparameter optimization procedures were conducted exclusively within the training set. Hyperparameters were optimized using five-fold cross-validation to identify the best-performing parameter combinations. Model training and parameter optimization were performed on the training set, and performance was evaluated on the independent testing set to assess generalizability for predicting DVT risk in patients with AECOPD. Predictive performance was assessed using the receiver operating characteristic (ROC) curve and area under the curve (AUC). Additional metrics included accuracy, sensitivity, specificity, precision, recall, and average precision (AP), providing a more comprehensive evaluation of classification performance. To integrate multiple evaluation indicators, the technique for order preference by similarity to ideal solution (TOPSIS) was applied to rank model performance. This method determines the relative closeness of each model to an ideal solution by calculating its distance from both the ideal and anti-ideal reference points. Model calibration was assessed using calibration curves to evaluate the agreement between predicted probabilities and observed outcomes.

Model Interpretability

The best-performing model was further interpreted using Shapley Additive exPlanations (SHAP). Based on cooperative game theory, SHAP assigns each feature a contribution value for individual predictions, allowing both the direction and magnitude of variable effects to be quantified. Feature importance plots and SHAP beeswarm plots were used to visualize global and local interpretability.

Statistical Analysis

Statistical analyses were performed using R software (version 4.3.0) and Python (version 3.9). Continuous variables were tested for normality. Normally distributed data were expressed as mean ± standard deviation (SD) and compared using independent-samples t-tests. Non-normally distributed variables were presented as median (interquartile range, IQR) and compared using the Mann–Whitney U-test. Categorical variables were expressed as numbers and percentages (n, %) and analyzed using the chi-square test or Fisher’s exact test where appropriate. All tests were two-sided, and a P value < 0.05 was considered statistically significant.

Results

Baseline Characteristics

As shown in Table 1, a total of 1,500 patients with AECOPD were included in this study, including 126 patients in the DVT group and 1,374 in the non-DVT group. No significant differences were observed between the two groups in terms of comorbidities, including hypertension, diabetes mellitus, heart failure, prior stroke, and arrhythmia (all P > 0.05). Regarding laboratory findings, patients in the DVT group showed significantly lower albumin levels (P < 0.001), higher C-reactive protein (P = 0.001), higher D-dimer (P < 0.001), and elevated direct bilirubin (P = 0.007). In addition, red blood cell–related parameters, including hematocrit, hemoglobin, and RBC count, were all significantly reduced in the DVT group (P = 0.013, 0.007, and 0.004, respectively). Lymphocyte count was also lower in the DVT group (P < 0.001), while monocyte count showed a modest but significant difference (P = 0.036). No significant differences were found in platelet-related parameters (MPV, PCT, PDW) or oxygenation index between the two groups (all P > 0.05). Among other biochemical and blood gas variables, mean corpuscular volume (MCV) was slightly higher in the DVT group (P = 0.039), and arterial pH also differed significantly (P = 0.007). However, ALT, AST, creatinine, urea, uric acid, total bilirubin, and glucose showed no significant differences between groups (all P > 0.05). The systemic inflammation–coagulation index was significantly higher in the DVT group than in the non-DVT group (106.63 vs 90.52, P = 0.001).

Table 1.

Baseline Characteristics

Variable	Non-VTE (n=1374)	VTE (n=126)	P-value
HBP=yes (%)	319 (23.2)	28 (22.2)	0.8
DM=yes (%)	95 (6.9)	7 (5.6)	0.562
HF=yes (%)	93 (6.8)	7 (5.6)	0.601
Stroke=yes (%)	99 (7.2)	6 (4.8)	0.304
Arrhythmia=yes (%)	84 (6.1)	10 (7.9)	0.419
Smoking=yes (%)	865 (63.0)	62 (49.2)	0.002*
ALB (median [IQR])	35.92 [33.21, 38.48]	33.86 [30.82, 36.59]	<0.001*
ALT (median [IQR])	16.00 [11.00, 23.00]	15.00 [11.00, 21.75]	0.261
AST (median [IQR])	21.00 [17.00, 26.00]	21.00 [17.00, 28.00]	0.752
CREA (median [IQR])	69.15 [57.70, 83.60]	69.45 [55.68, 84.77]	0.486
CRP (median [IQR])	7.30 [2.24, 24.31]	11.80 [3.51, 47.67]	0.001*
Dimer (median [IQR])	0.15 [0.09, 0.28]	0.37 [0.19, 0.78]	<0.001*
DBIL (median [IQR])	2.90 [2.10, 3.90]	3.20 [2.42, 4.20]	0.007*
GLU (median [IQR])	6.17 [5.22, 7.74]	6.15 [5.22, 7.68]	0.882
HCT (median [IQR])	44.10 [39.05, 48.40]	41.80 [37.76, 46.79]	0.013*
HGB (median [IQR])	130.00 [118.00, 142.00]	125.50 [115.00, 137.50]	0.007*
LY (median [IQR])	1.19 [0.79, 1.63]	0.86 [0.59, 1.25]	<0.001*
MCH (median [IQR])	30.70 [29.60, 31.90]	30.60 [29.42, 32.48]	0.685
MCHC (median [IQR])	332.00 [325.00, 338.00]	330.50 [321.00, 336.00]	0.061
MCV (median [IQR])	92.50 [89.60, 95.68]	93.45 [89.32, 97.85]	0.039*
MONO (median [IQR])	0.46 [0.36, 0.58]	0.42 [0.33, 0.55]	0.036*
MPV (median [IQR])	10.10 [9.40, 11.00]	10.00 [9.33, 11.10]	0.545
PCT (median [IQR])	0.19 [0.16, 0.22]	0.18 [0.14, 0.23]	0.213
PDW (median [IQR])	16.20 [15.93, 16.50]	16.20 [16.00, 16.50]	0.896
PF (median [IQR])	331.88 [288.00, 374.33]	332.50 [283.45, 372.68]	0.952
PH (median [IQR])	7.41 [7.39, 7.44]	7.42 [7.40, 7.45]	0.007*
RBC (median [IQR])	4.25 [3.87, 4.62]	4.04 [3.71, 4.47]	0.004*
TBIL (median [IQR])	10.70 [8.10, 14.78]	11.55 [8.60, 15.45]	0.153
UA (median [IQR])	307.46 [248.17, 387.67]	303.25 [227.80, 382.00]	0.382
UREA (median [IQR])	6.43 [5.10, 7.94]	6.52 [4.85, 8.26]	0.808
SCI (median [IQR])	90.52 [65.00, 118.08]	106.63 [72.35, 146.75]	0.001*

Note: * p<0.05.

Abbreviations: HBP, hypertension; DM, diabetes mellitus; HF, heart failure; ALB, albumin; ALT, alanine aminotransferase; AST, aspartate aminotransferase; CREA, creatinine; CRP, C-reactive protein; Dimer, D-dimer; DBIL, direct bilirubin; GLU, glucose; HCT, hematocrit; HGB, hemoglobin; LY, lymphocyte count; MCH, mean corpuscular hemoglobin; MCHC, mean corpuscular hemoglobin concentration; MCV, mean corpuscular volume; MONO, monocyte count; MPV, mean platelet volume; PCT, plateletcrit; PDW, platelet distribution width; PF, oxygenation index; RBC, red blood cell count; TBIL, total bilirubin; UA, uric acid; SCI, systemic inflammation-coagulation index.

Boruta Feature Selection

As shown in Figure 2, the Boruta algorithm was used to identify relevant variables for model development. The left panel presents the distribution of variable importance scores, while the right panel shows the stability of importance rankings across iterations. With increasing iterations, feature importance gradually stabilized. After filtering, 14 key variables were selected for model construction: ALB, ALT, AST, CRP, D-dimer, DBIL, HGB, MCH, MCV, PCT, PDW, RBC, TBIL, and SCI. These variables were used as the final input features for subsequent modeling.

Figure 2.

Boruta and LASSO screening results for important predictor variables.

Model Performance

As shown in Figure 3, three machine learning models (DT, MLP, and XGBoost) were evaluated in both the training and testing sets using multiple performance metrics. Overall, XGBoost showed the most consistent and balanced performance. In the testing cohort, it achieved an AUC of 0.707, outperforming MLP (0.684) and DT (0.667). It also demonstrated stable performance in accuracy (0.676), sensitivity (0.619), specificity (0.681), and F1-score (0.263). The MLP model performed relatively well in the training set, with an accuracy of 0.774 and an AUC of 0.852, but its sensitivity dropped notably in the testing set (0.500), suggesting reduced generalizability. The decision tree model showed the weakest overall performance, particularly in precision and F1-score compared with the other models. As shown in Figure 4, TOPSIS-based multi-criteria evaluation ranked XGBoost as the best-performing model, with the highest composite score (0.806), followed by DT (0.490) and MLP (0.480). This supported the selection of XGBoost as the final optimal model. Calibration curves demonstrated acceptable agreement between predicted and observed probabilities for the three models, with XGBoost showing relatively better calibration performance (Figure 5).

Figure 3.

Heatmap for test set and training set evaluation.

Figure 4.

TOPSIS Closeness Coefficients for Different Models.

Figure 5.

Decision curve analysis of the prediction models.

Model Interpretability

As shown in Figure 6, SHAP analysis was applied to the best-performing model to explore the contribution of individual variables to predictions of DVT in patients with AECOPD. The SHAP beeswarm plot indicated that D-dimer had the strongest impact on model output, followed by the SCI. Albumin, MCV, MCH, and CRP also contributed to model predictions to varying degrees. The SHAP feature importance plot further confirmed this ranking based on mean absolute SHAP values, with D-dimer, SCI, ALB, MCV, MCH, and CRP among the most influential predictors.

Figure 6.

A, Feature attributes in SHAP; b, SHAP-based feature importance ranking.

Discussion

In this study, we developed and evaluated a DVT prediction model in 1,500 patients with AECOPD. Patients with DVT showed significantly higher levels of CRP and D-dimer, along with lower albumin levels. The SCI was also markedly increased in the DVT group. After Boruta-based feature selection, 14 variables were retained for model construction, providing a stable input set for subsequent analysis. Among the three models, XGBoost achieved the best overall performance in the testing cohort (ROC-AUC = 0.707) and received the highest TOPSIS score (0.806), suggesting relatively better predictive performance compared with MLP and decision tree models. SHAP analysis further indicated that D-dimer and SCI were the most influential predictors, followed by albumin, MCV, MCH, and CRP. The increased risk of lower-extremity DVT in patients with AECOPD is generally explained by Virchow’s triad, which includes venous stasis, endothelial injury, and a hypercoagulable state. During acute exacerbations, patients often experience reduced mobility, prolonged bed rest, and intensified systemic inflammation, all of which contribute to slowed venous return. At the same time, inflammatory mediators may impair endothelial function and facilitate activation of the coagulation cascade. Together, these processes create a biological environment that favors thrombosis, highlighting the close interaction between inflammation and coagulation in the development of DVT.

Fibrinogen, as a key substrate in the coagulation system, contributes not only to fibrin formation but also increases blood viscosity, thereby promoting thrombus development.^16–18 Platelets play a central role in the early stages of thrombosis through adhesion, aggregation, and microthrombus formation.^19,20 Leukocytes, on the other hand, reflect systemic inflammatory activity and can participate in thrombosis through the release of inflammatory mediators and the formation of neutrophil extracellular traps (NETs).^21,22 In AECOPD, acute exacerbations are often accompanied by a systemic inflammatory surge and oxidative stress,^23,24 which may further elevate fibrinogen levels, enhance platelet activation, and disturb leukocyte homeostasis, ultimately contributing to a prothrombotic state.

In the SHAP analysis, SCI ranked second in importance, just after D-dimer. D-dimer is widely used in clinical practice as a biomarker for thrombotic events and has established diagnostic value. However, it mainly reflects fibrin degradation and is therefore more representative of downstream thrombotic activity rather than early pathophysiological changes. In addition, its levels can be influenced by infection, systemic inflammation, malignancy, and aging. In the context of AECOPD, where systemic inflammation is prominent, its specificity for early risk stratification is limited. In contrast, SCI is a composite index integrating fibrinogen, platelet count, and white blood cell count. It reflects coagulation substrate availability, platelet involvement, and inflammatory activation simultaneously, providing a more comprehensive description of the inflammatory–coagulatory state.^25,26 More importantly, SCI may reflect biological conditions associated with a prothrombotic state rather than merely representing a consequence of clot formation. This may explain why SCI ranked highly in the SHAP analysis and suggests that it offers complementary information to D-dimer in risk assessment.

Overall, our findings suggest that DVT risk in AECOPD is not driven by a single coagulation marker but rather results from a complex interaction between inflammation and coagulation pathways. The model integrates downstream markers such as D-dimer and upstream indicators such as SCI, providing additional information for thrombotic risk assessment. The relatively high importance of SCI supports its potential value in reflecting inflammation-driven hypercoagulability and suggests that it may provide complementary information alongside traditional coagulation markers for DVT risk assessment. However, considering the moderate predictive performance of the model and the lack of external validation, further studies using independent cohorts are required to confirm its clinical applicability.

Disclosure

The authors report no conflicts of interest in this work.