







































N. Zhou et al. /Future Technology                                                                                      February 2026| Volume 05 | Issue 01 | Pages 
254-262 

254 

 

 

 

Article 

Machine learning model for predicting symptom 

improvement rates in hospitalized deep vein 

thrombosis patients 
Nan Zhou1, Teck Han Ng1, Chai Nien Foo1, Lloyd Ling2, Yang Mooi Lim1*  

1M. Kandiah Faculty of Medicine and Health Science, Universiti Tunku Abdul Rahman, Kajang, Selangor 43000, Malaysia 
2Lee Kong Chian Faculty of Engineering and Science, Universiti Tunku Abdul Rahman, Kajang, Selangor 43000, Malaysia 

A R T I C L E   I N F O 
 

Article history: 
Received 30 August 2025  
Received in revised form 
06 October 2025 
Accepted 01 December 2025 
 
Keywords:  
Deep vein thrombosis, Machine learning, 
Treatment response prediction,  
Clinical decision support 
 
*Corresponding author 
Email address: 
ymlim@utar.edu.my 
 
DOI: 10.55670/fpll.futech.5.1.22 

A B S T R A C T 
 

Deep Vein Thrombosis (DVT) demonstrates considerable treatment response 
heterogeneity, with 40-60% of patients developing complications despite 
standard anticoagulation therapy. Accurate prediction of individual treatment 
outcomes remains an unmet clinical need. This study develops and validates a 
machine learning-based model to predict symptom Improvement Rate (IPR) 
using retrospective data from 403 hospitalized DVT patients (2018-2023). Six 
predictive features are identified using Random Forest-based Recursive 
Feature Elimination (RFE): age, white blood cell count, Activated Partial 
Thromboplastin Time (APTT), Thrombin Time (TT), surgical intervention 
status, and baseline symptom severity. The regularized eXtreme Gradient 
Boosting (XGBoost) algorithm achieves optimal performance with a test 
coefficient of determination (R²) of 0.60, Root Mean Square Error (RMSE) of 
12.36, and five-fold cross-validation R² of 0.58 ± 0.07. SHapley Additive 
exPlanations (SHAP) analysis reveals that APTT and surgical intervention are 
the strongest predictors of treatment response. The validated model is 
deployed as a publicly accessible web-based clinical decision support tool, 
enabling real-time outcome prediction at the point of care. This research 
establishes a practical framework bridging predictive analytics and clinical 
practice, facilitating evidence-based, personalized DVT management strategies. 

1. Introduction 

DVT = Deep Venous Thrombosis (DVT) refers to deep 
vein thrombosis, a type of venous system disease. Most often, 
it occurs in deep veins. It is mainly in the deep veins of the 
limbs, especially the lower extremities, and can extend to the 
pelvic veins and the lower half of the inferior vena cava [1]. 
DVT’s incidence among hospitalized patients rises 
considerably, reaching 100 - 200 for every 100,000 persons 
yearly all over the world, with much higher numbers in 
certain populations, post-surgical patients (2-3%), critically 
ill patients (5-10%), and those with malignancies (4-20 %) [2-
4]. The classical presentation includes unilateral limb 
edema/pain, erythema, and warmth; however, ~30% are 
clinically silent until complications [5,6]. Mature 
anticoagulation therapy protocols have been established, 
including low-molecular-weight heparin, direct oral 
anticoagulants, and vitamin K antagonists. However, there are 
still considerable differences in individual responses to the 
therapy, which are related to different genetic variations, 
comorbidities, and interactions with the medicine [7,8]. 
Clinical evidence shows that after standard treatment, 

between 40% and 60% of DVT patients still have a risk of 
complications like post-thrombotic syndrome. The symptoms 
of this syndrome include chronic leg pain, swelling, skin 
changes, and, in the worst cases, venous ulcers, all of which 
can seriously reduce a person's quality of life and use more 
medical resources [9]. This heterogeneity necessitates early 
identification of treatment-resistant patients to enable timely 
intervention with advanced therapies, including catheter-
directed thrombolysis, mechanical thrombectomy, or 
extended anticoagulation regimens [10]. In recent years, 
machine learning technology has made great progress in the 
field of health care, especially in predicting disease prognosis 
and evaluating symptom improvement [11]. The traditional 
treatment response assessment for DVT mainly relies on 
clinical experience and single-parameter judgment, without 
sufficient consideration of individual patients and 
multidimensional clinical data [12,13]. Machine learning 
algorithms include the demographic aspect of laboratory 
results, images, and therapy, and produce better forecast 
models. Studies demonstrate that machine learning 
approaches outperform conventional statistical methods in 

 

 Open Access Journal 

ISSN 2832-0379 

February 2026| Volume 05 | Issue 01 | Pages 254-262 

https://doi.org/10.55670/fpll.futech.5.1.22 

 

 

 

 

 

 

 

 

 

Journal homepage: https://fupubco.com/futech 

 

Future Technology 

mailto:ymlim@utar.edu.my
https://doi.org/10.55670/fpll.futech.5.1.22
https://fupubco.com/futech


N. Zhou et al. /Future Technology                                                                                      February 2026| Volume 05 | Issue 01 | Pages 254-262 

255 

 

predicting thrombotic events [14, 15]. The goal is to develop 
a machine learning-based prediction model to assess the 
response of DVT treatment with regard to symptom 
improvement rate among hospitalized DVT patients. We 
incorporate patients' clinical characteristics, laboratory 
parameters, and treatment parameters to construct a multi-
dimensional model. The model aims to accurately predict an 
individual patient's treatment improvement rate to assist 
clinicians in personalizing treatment plans. Furthermore, this 
study identifies key factors influencing symptom 
improvement, providing evidence-based guidance for 
therapeutic optimization. 

2. Methods 

2.1 Research design 
This study was conducted in accordance with the 

TRIPOD statement. The completed TRIPOD checklist is 
provided as Supplementary File 2. This retrospective 
observational study utilizes real-world data from the 
comprehensive hospital information system of [Masked for 
blind review] to evaluate treatment outcomes in patients 
diagnosed with DVT from 2018 to 2023. Specific inclusion and 
exclusion criteria were applied to ensure data completeness 
and relevance. Inclusion criteria: (1) DVT is the main 
diagnosis in medical records; (2) complete medical records 
required for the study are available. Complete medical 
records were defined as containing: (1) ultrasound-
confirmed DVT diagnosis; (2) baseline and discharge 
symptom scores; (3) laboratory data within 24 hours of 
admission; (4) therapeutic intervention records; and (5) 
documented outcomes. Planned diagnostic and treatment 
procedures referred to completion of the institutional 
protocol without premature discontinuation. Exclusion 
criteria: (1) Incomplete planned diagnostic and treatment 
procedures during hospitalization due to reasons such as 
transfer or treatment abandonment; (2) Unavailable data on 
confounding factors in medical records due to attending 
physician resignation or retirement. 

Between January 2018 and December 2023, 658 patients 
were identified. After excluding 215 patients (150 incomplete 
records, 65 transfers), 443 remained eligible. Subsequently, 
40 patients with missing confounding variables were 
excluded, leaving 403 patients in the final cohort (Figure 1). 
Cases meeting these criteria were systematically entered into 
an electronic data collection form designed specifically for 
this study. The data collection protocol encompasses a 
comprehensive set of variables (Supplementary File Table 1), 
including: patient hospitalization identification number 
(utilized solely for Source Data Verification purposes), 
demographic characteristics, hematological and coagulation 
function parameters assessed on the initial day of admission, 
duration of symptoms before admission, ultrasound-
confirmed thrombus localization, admission Wells score, 
history of DVT and associated comorbidities, therapeutic 
interventions for DVT and concomitant conditions 
(encompassing both pharmacological and physical 
modalities), surgical management strategies and a reference 
Diagnostic and Efficacy Criteria for Deep Venous Thrombosis 
of the Lower Extremities (Revised in 2015) symptom 
quantification assessment (Supplementary File Table 1) [16]. 
All patient identifiers were removed prior to analysis. Data 
extraction and de-identification were performed by 
personnel independent of the analytical team. This study was 
approved by the Ethics Committee of The Sun Simiao Hospital 
of Beijing University of Chinese Medicine (Approval number: 

SSMYY-KYPJ-2023-011). It conforms to the ethical standards 
of medical research. 

Total DVT patients screened

(January 2018 - December 2023)

n=658

After incomplete records/transfer 

exclusion= 443

Excluded (n=215)

• Incomplete medical records: n = 150

• Transfer before protocol completion: n = 65

After missing confounders exclusion

n=403

Excluded (n = 40)

• Missing confounding variables

Final analytical cohort

n=403

Eligible

Eligible

Included

Excluded

Excluded

 

Figure 1. Sequential patient selection flow diagram 

2.2 Data preprocessing and exploratory analysis 
Key variables included demographic data (age, sex), 

laboratory parameters (WBC, APTT, TT), therapeutic 
interventions, and symptom severity scores. Data cleaning 
processes attended to natural missingness via systematic 
imputation and deletion methods. Variables that had less than 
30% missing values were subjected to KNN imputation. KNN 
was implemented with k=5 neighbors using Euclidean 
distance for numerical variables and Hamming distance for 
categorical variables. Those above this figure were eliminated 
to avoid any form of analytical bias. The 30% threshold 
follows established guidelines for clinical prediction models. 
No variables exceeded this threshold. The final six features all 
demonstrated minimal missingness below the threshold. 
Disease classifications were standardized using ICD-10-CM 
coding, and drug names were unified systematically (specific 
formulations were converted to general names). The hospital 
system used ICD-10-CM natively; no manual mapping was 
required. Binary variables were coded as 0/1, ordinal 
variables were retained with their natural ordering, and one-
hot encoding was avoided due to sample size constraints. 
Continuous variables were standardized to avoid unit-based 
changes. Descriptive statistics of the dataset characteristics, 
with a summary of continuous variables including minimum 
value, maximum value, and median value; and a summary of 
categorical variables using frequency distributions. 
Skewness, kurtosis, and the disparity ratio are computed to 
evaluate the distribution of the data and identify potential 
limitations of the models, laying the groundwork for 
subsequent analytical activities. 

2.3 Variable definition 
This study used the Improvement Rate (IPR) to assess 

symptom improvement. IPR = [(Day 1 symptom score - 
discharge symptom score) / Day 1 symptom score)] × 100%, 
which indicates the percentage of symptom severity 
reduction from onset to discharge. Symptom scores (0-12 
points) were assessed using the 2015 DVT Diagnostic Criteria, 
evaluating swelling, pain, skin changes, and function. Patients 
with zero baseline scores were excluded; IPR was capped at 
100%. This standard measurement allows objective 
assessment of symptom resolution in the patient group. 



N. Zhou et al. /Future Technology                                                                                      February 2026| Volume 05 | Issue 01 | Pages 254-262 

256 

 

2.4 Model development and evaluation 
The dataset was split 80 - 20 on the train-test sets. Five-

fold cross-validation was performed on the training set to 
ensure robust performance estimation. A MinMax scaler was 
applied to normalise the data to the range [0,1]. Feature 
selection used the RFE from the random forest model, 
checking model performance with 3-12 kept features to 
determine the best feature set size. RFE employed Random 
Forest with 100 estimators, using R² as the selection metric. 
No significant multicollinearity was detected among retained 
features (VIF < 5).  

Four models were implemented: linear regression, SVR, 
random forest and xgboost. Hyperparameter optimization 
was conducted using the Optuna framework to optimize the 
model. For XGBoost, the search space included: learning_rate 
(0.01-0.3), max_depth (3-8), n_estimators (50-300), 
subsample (0.6-1.0), colsample_bytree (0.6-1.0), reg_alpha 
(0-10), and reg_lambda (1-10). Optimized parameters were: 
learning_rate=0.01, max_depth=3, n_estimators=150, 
subsample=0.8, colsample_bytree=0.8, reg_alpha=5.0, 
reg_lambda=10.0. Early stopping with 30 rounds was applied 
to prevent overfitting.   

Model evaluation metrics included the coefficient of 
determination (R²), Mean Squared Error (MSE), Mean 
Absolute Error (MAE), and Root Mean Squared Error (RMSE) 
to evaluate the model’s performance across diverse modeling 
strategies. Bootstrap resampling (1000 iterations) was used 
to estimate 95% confidence intervals for all performance 
metrics. Model calibration was assessed using calibration 
plots (Figure 2). The calibration slope was 0.84, calibration-
in-the-large was 8.43, and the Brier score was 0.006, 
indicating acceptable calibration with minor systematic bias. 

 

 
Figure 2. Calibration plot for XGBoost model 

2.5 Model interpretation 
SHAP (SHapley Additive exPlanations) analysis was 

performed on the optimal performing model to explore 
features. TreeExplainer was used for SHAP value 
computation, which is optimized for tree-based ensemble 
methods including XGBoost. Values are calculated using 
cooperative game theory and measure a feature's marginal 

contribution to model predictions by averaging the effects of 
all possible feature combinations. It delivers both global 
feature importance ratings and local understandability for 
separate predictions. The SHAP framework assigns each 
feature a value that reflects its importance for the model's 
outputs. It shows how features interact and influence model 
output without losing accuracy. It helps people see how 
models make decisions. In this analysis of explainable results, 
we gain essential knowledge of the patterns behind the 
model's decisions, enabling us to transform those 
mathematical results into useful information for clinics. 

2.6 Model deployment 
To make it easier for doctors to use, we made a simple 

computer program using a tool called Gradio. This program is 
like a window where doctors can see the results of our model 
in a way that’s easy for them to understand. The interface was 
deployed to the Hugging Face platform and is publicly 
available to any healthcare practitioners. The model is 
accessible at 
https://huggingface.co/spaces/Curvature/DVT_Managemen
tSource code, trained model, and environment specifications 
(Python 3.9, XGBoost 1.7.0, random seed 42) are available on 
GitHub. Users must confirm healthcare professional status 
before accessing.  This deployment strategy overcomes 
technical obstacles, allowing for immediate use of the model 
without specialized programming skills. The whole project, 
together with its source code and documentation, was 
published under the MIT license, promoting open 
collaboration and enabling everyone to freely modify and 
distribute it. It ensures that the know-how is widely known, 
so it can be easily used when caring for sick people, 
connecting what researchers discover with how doctors 
actually care for patients. 

3. Results 

3.1 Descriptive statistics 
The study population comprised 403 patients with a 

mean age of 61.32 ± 14.84 years, and anthropometric 
measurements were done for the same, with an average 
height of 161.37 ± 6.89 cm and weight of 61.43 ± 10.77 kg.  
Males accounted for 54.09% and females for 45.91%. 
Comprehensive laboratory parameters revealed mean values 
of WBC 9.90±2.15×10⁹/L, RBC 5.04±0.37×10¹²/L, Hgb 
15.03±1.07g/dL, PLT 165.20±64.74×10⁹/L, and HCT 
45.10±1.42%, suggesting mild leukocytosis with otherwise 
normal hematological profiles. Coagulation profiles showed 
PT 13.29±1.93s, INR 1.09±0.17, APTT 32.94±5.08s, TT 
25.69±5.17s, Fibrinogen 3.53±0.87g/L, D-dimer 
12.30±4.91mg/L, and FDP 19.49±10.45mg/L, indicating 
hypercoagulability with markedly elevated D-dimer and FDP 
levels characteristic of venous thromboembolism.  

Thrombus locations predominantly involved the 
popliteal vein (55.58%), common femoral vein (52.36%), iliac 
vein (45.16%), posterior tibial vein (43.92%), and inter-
muscular veins of the calf (37.97%). Percentages exceed 
100% as patients frequently presented with multi-site 
involvement, demonstrating extensive thrombosis of both 
proximal and distal lower extremity vasculature. Therapeutic 
Management was comprehensive. Enoxaparin 62.78% was 
the most common medication, followed by Diosmin 44.67%, 
and Rivaroxaban 37.22%. Treatment protocols followed 
institutional guidelines; therapeutic variations were 
accounted for in model development through inclusion of 
intervention status as a predictor variable. Non-
pharmacological therapies were limb elevation at 30 degrees 



N. Zhou et al. /Future Technology                                                                                      February 2026| Volume 05 | Issue 01 | Pages 254-262 

257 

 

(73.70%), Bingxiao external application (46.65%), and 
compression stockings (35.73%). Most procedures 
performed are Venography (70.97%), Therapeutic Treatment 
(Thrombolysis, Thrombectomy) (70.47%), Filter 
Placement/Retrieval ( 47.89%), Angioplasty / Stent (33.00). 
The comorbidity analysis showed that essential hypertension 
(28.54%), type 2 diabetes mellitus (14.64%), and other 
venous disorders (11.91%) were the most common, with 
13.65% of patients having no comorbidities. The occupational 
distribution showed farmers (26.55%), unemployed 
individuals (23.08%), and retired persons (11.41%) as the 
major parts of the cohort, possibly indicating socio-economic 
elements related to DVT creation and management. 

3.2 Feature selection 
RFE analysis was performed to improve feature 

selection, and R² was evaluated across different feature 
subsets. The R2 curve showed optimal performance between 
6 and 8 features, with a maximum R2 of 0.58. 6 Key predictors 
were selected. The final feature set comprised one 
demographic parameter (age), three laboratory parameters 
(white blood cell count, activated partial thromboplastin 
time, thrombin time), one therapeutic parameter 
(thrombolysis/thrombectomy), and one clinical parameter 
(Day 1 symptom score). Feature definitions were as follows: 
age (years), WBC (×10⁹/L), APTT (seconds), TT (seconds), 
thrombolysis/thrombectomy (binary, 1=performed), and 
Day 1 symptom score (0-12 points). These features were the 
most influential predictors that kept the model efficient 
(Figure 3). 

 
Figure 3. Feature selection optimization: R² performance analysis 
across varying feature dimensions 

3.3 Model evaluation 
A comparison of the four machine learning models shows 

different performances on both training and testing datasets. 
The overall assessment of prediction accuracy via true-
versus-predicted scatter plots shows that XGBoost reached 
the best alignment with the ideal diagonal line, meaning that 
it has the strongest predictive ability and the least systematic 
bias. After applying regularization to address overfitting, 
XGBoost achieved a test R² of 0.60, RMSE of 12.36, and five-
fold cross-validation R² of 0.58 ± 0.07. The second-best  model 
was Random Forest, and it had an R² of 0.573 and an RMSE of 
12.77. The residuals were fairly spread out across all of the 
predictions, but they were slightly more scattered than the 
XGBoost model. On the other hand, both the Linear 
Regression and Support Vector Machine models yielded much 
worse results, with test R² values of 0.333 and 0.242, 
respectively, indicating a poor ability to model the underlying 

pattern in the data. Their corresponding residual plots 
showed obvious heteroscedasticity (Breusch-Pagan test, 
p<0.05), especially in areas with high predictions, where large 
deviations from the actual values were observed, suggesting 
a systematic error in prediction that increases with the size of 
the target. Detailed residual analysis again showed that 
XGBoost has better prediction stability, with the most 
concentrated and symmetrical residual distribution centered 
on 0, few outliers, and consistent error variation across the 
entire prediction range. This stability is quantified by 
XGBoost’s test-set MAE of 7.82 and MSE of 152.80, the 
smallest errors among all models. After regularization, the 
gap between training R² (0.75) and test R² (0.60) was reduced 
to 0.15, compared to 0.35 before regularization, 
demonstrating effective mitigation of overfitting. This 
improved generalization makes the model suitable for clinical 
deployment. Random Forest showed a larger train-test gap 
(training R² of 0.9396 versus test R² of 0.5729). Both the 
quantitative performance metrics and the graphical 
diagnostics from the scatter and residual plots support 
selecting XGBoost as the best model to deploy in a clinical 
setting for providing the most accurate predictions of 
symptom improvement rate in DVT patients (Figure 4-7; 
Table 1). Temporal validation was performed by training on 
2018-2021 data and testing on 2022-2023 data, yielding 
comparable performance (R²=0.61), supporting model 
generalizability. Decision curve analysis demonstrated 
positive net benefit across threshold probabilities of 30-70%, 
indicating clinical utility compared to default strategies. 

Table 1. Performance metrics comparison of machine learning 
models for prediction across training and test sets 

 

3.4 SHAP analysis 
SHAP analysis showed different patterns of feature 

importance and its impact on the model-predicted IPR. APTT 
had the biggest impact on the model, with SHAP values going 
between -15 and +15, high APTT values were associated with 
lower IPR predictions, which may reflect clinical 
considerations regarding bleeding risk. Surgical intervention 
(T&T) had a clear bimodal distribution. Positive T&T was 
associated with greater IPR prediction (SHAP values +10 to 
+15) as it indicated the benefit of early thrombus removal, 
and negative T&T was associated with lesser IPR prediction 
(SHAP values -10 to -15). WBC count had a moderate bilateral 
impact: low counts (SHAP values -5 to 0) predicted good 
outcomes by lowering inflammation, and high counts (SHAP 
values 0 to +5) predicted poor outcomes.  

 

Model Dataset R² 
5-Fold 
CV R² 

MSE RMSE MAE 

XGBoost(re
gularized) 

Train 
0.7500

- 
0.58 ± 
0.07 

104.05 10.20 6.85 

 Test 0.6000 - 152.80 12.36 7.82 

Random 
Forest 

Train 0.9396 
0.52 ± 
0.08 

25.12 5.01 2.85 

 Test 0.5729 - 163.18 12.77 8.11 

Linear 
Regression 

Train 0.4874 
0.31 ± 
0.05 

213.31 14.61 9.62 

 Test 0.3333 - 254.72 15.96 
10.5

9 

SVM Train 0.3802 
0.22 ± 
0.06 

257.88 16.06 
10.9

5 

 Test 0.2417 - 289.75 17.02 
11.0

1 



N. Zhou et al. /Future Technology                                                                                      February 2026| Volume 05 | Issue 01 | Pages 254-262 

258 

 

(a) 

(b) 
Figure 4. Linear regression: predictive accuracy assessment through 
(a) true-predicted correlation, (b) residual distribution analysis 

 

(a) 

(b) 
Figure 5. Random forest: predictive accuracy assessment through (a) 
true-predicted correlation, (b) residual distribution analysis 

(a) 

(b) 

Figure 6. Support vector machine: predictive accuracy assessment 
through (a) true-predicted correlation (b) residual distribution 
analysis 

(a) 

(b) 

Figure 7. XGBoost: Predictive accuracy assessment through (a) 
true-predicted correlation (b) residual distribution analysis 

 



N. Zhou et al. /Future Technology                                                                                      February 2026| Volume 05 | Issue 01 | Pages 254-262 

259 

 

Age showed asymmetric SHAP values between -5 and 
+10, with older age having a negative effect on IPR prediction, 
as older patients have less tolerance for treatments and more 
complications. The initial symptom scores displayed balanced 
bidirectional impacts, with zero as the center, and severe 
baseline symptoms indicating a larger possible margin of 
improvement. TT has the most concentrated distribution, 
indicating stabilization of the predictions. Parallel coordinate 
visualization validated these associations; optimal IPR 
predictions (80 - 100 %) showed low APTT, positive surgical 
status, and moderate WBCs, poor outcomes (20 - 40 %) linked 
to high APTT, negative surgical status, and poor inflammation 
markers（Figures 8 and Figure 9). 

 
Figure 8. Feature impact distribution analysis through SHAP value 
visualization 

 

 
Figure 9. Model output response analysis with feature value 
distribution 

3.5 User interface development and deployment 
Deployed DVT Symptom Improvement Management 

System with Predictive Analytics into Clinical Workflow via a 
Web-based Platform. This implementation takes 
multidimensional clinical parameters, labs, and standardized 
symptom evaluation and creates a quantitative measure of 
the rate of improvement. The system uses SHAP-based model 
interpretation, and the clinicians can see which parameters 
contribute the most to the model’s predictions. This evidence-
based decision-support framework aids risk stratification, 
therapeutic strategy enhancement, and outcome prediction in 
the management of DVT. Implementing the platform in 
clinical assessment would lead to more consistent assessment 
of the patient's current state, generating data-driven 
information to build a tailored treatment plan. This 
implementation bridges the translation gap between high-
end machine learning algorithms and clinical practice and 
establishes a standard method for predicting evidence-based 
DVT symptom improvement in routine medical care 
(Supplementary File Figure 2). 

 

4. Discussion  

This study developed and validated a machine learning-
based prediction model of DVT symptom improvement rates, 
achieving good predictive performance in the test set (R² = 
0.60). Among the four machine learning methods, the 
XGBoost algorithm had the best performance with an RMSE 
of 12.36 and an MAE of 7.82. By SHAP analysis, 6 key 
predictive features were identified, with APTT and whether 
surgery is performed being the most influential ones. It is 
successfully applied as a web-based clinical decision support 
system that can predict the improvement rate of symptoms in 
real time. These findings provide a novel approach to 
predicting individual treatment response in DVT patients by 
integrating multiple clinical parameters into an accessible 
decision support framework. 

Deep vein thrombosis management is complicated by 
disease complexity, variable treatment responses, and 
potential serious complications. Treatment modalities 
include anticoagulant therapy, thrombolysis, and mechanical 
interventions, each producing different effects depending on 
patient characteristics and treatment compliance. Low-
molecular-weight heparin (LMWH) is effective for both 
proximal and isolated distal DVT [17]. However, 
anticoagulation response varies considerably among 
patients, influenced by age, weight, renal function, and 
pharmacogenetic factors [18]. Elderly patients exhibit altered 
pharmacokinetics, requiring careful monitoring and dose 
adjustment to maintain therapeutic levels while minimizing 
bleeding risk [19].  

Surgical strategy significantly influences treatment 
outcomes. Catheter-directed thrombolysis (CDT) is 
recommended as first-line treatment for acute lower 
extremity DVT in patients with high thrombus burden [20]. 
However, procedural complications, including bleeding and 
embolization, may occur, potentially leading to post-
thrombotic syndrome (PTS) [21]. Clinical studies 
demonstrate that multimodal intervention protocols achieve 
significantly lower PTS rates compared to anticoagulation 
alone [22, 23]. Treatment adherence is a critical prognostic 
determinant; sustained anticoagulation compliance 
maintains therapeutic efficacy and reduces recurrent 
thromboembolic risk [24]. However, treatment complexity, 
adverse effects, and limited patient awareness often 
compromise adherence [25]. These challenges underscore 
the clinical significance of standardized decision support 
tools to enhance physician-patient communication and 
improve treatment compliance. 

This study established a machine learning-based 
symptom recovery prediction model within a web application 
framework, creating a clinical decision support system. Using 
the SHAP analysis methodology, the model displayed the 
contributions of the clinical indicators and the mechanisms 
that affected the results. From the SHAP value analysis, there 
is a significant negative correlation between APTT and IPR, 
suggesting that elevated APTT may be associated with 
reduced treatment efficacy. This relationship may reflect 
considerations in clinical decision-making, as elevated APTT 
indicates increased bleeding risk, potentially leading 
clinicians to adopt more conservative therapeutic approaches 
[7, 26]. Surgical intervention (thrombolysis/thrombectomy) 
emerged as a strong positive predictor, likely attributable to 
the benefits of early thrombus removal. Existing studies have 
shown that patients who underwent surgical interventions 
had better symptom improvement than those who did not, 
especially at earlier stages of the disease. This finding 



N. Zhou et al. /Future Technology                                                                                      February 2026| Volume 05 | Issue 01 | Pages 254-262 

260 

 

supports the clinical value of early, personalized surgical 
decision-making [27]. 

 WBC is an important inflammatory marker that 
indicates the level of inflammation, which is an important 
factor in the development of DVT clinical limb symptoms. This 
study found that lower initial WBCs were associated with 
better treatment response, which may be explained by 
thrombus-inflammation interactions. After a thrombus is 
formed, injured vascular endothelial cells and platelets will 
produce pro-inflammatory cytokines to attract leucocytes 
that invade the vessel wall. Activated leukocytes go on to 
release more cytokines and proteases, which can make the 
endothelial damage worse and lead to more reactions that 
cause blood to clot. Moreover, the released NETs directly take 
part in thrombosis and stimulate platelets to be activated, 
thus forming a vicious circle of inflammation-thrombosis and 
causing aggravated local tissue injury and continued clinical 
manifestations. Thus, lower WBC levels may indicate milder 
inflammatory responses and lesser thrombosis–
inflammation cycle, which explains its link to better treatment 
response [28, 29]. 

Implementing a graphical user interface represents a 
significant advancement in translating machine learning 
algorithms into practical clinical tools [30]. The web-based 
platform reduces implementation barriers by eliminating 
requirements for specialized programming skills or 
computational infrastructure [31]. Interactive SHAP 
visualization transforms mathematical predictions into 
clinically interpretable information, enabling clinicians to 
understand both predictions and underlying reasoning [32]. 
This addresses the "black box" perception that undermines 
clinician confidence by providing a clear visualization of each 
parameter's influence through intuitive force plots [33]. The 
dynamic interface allows variable adjustment with 
immediate prediction updates, creating a simulation-like 
experience for treatment planning [34]. The system 
distinguishes modifiable from non-modifiable factors, 
guiding clinical attention toward areas with the greatest 
potential impact [35]. This transforms the model from a 
passive assessment tool into an active decision aid supporting 
personalized treatment optimization. This study has some 
limitations. As a single-center retrospective study at The Sun 
Simiao Hospital of Beijing University of Chinese Medicine, it is 
uncertain if our findings are generalizable to other regions or 
healthcare settings. Retrospective data collection may have 
selection bias. It may impact the robustness of our results 
even with thorough inclusion/exclusion criteria. In addition, 
we examined only in-hospital outcomes at discharge. We did 
not use long-term follow-up data to assess the model’s 
forecasting performance for longer-term treatments and 
complications. 

5. Conclusion 

This study develops a new machine learning approach to 
predict DVT treatment response and demonstrates that this 
method has a better predictive performance than other 
methods through the XGBoost algorithm with an R2 of 0.60 
and an RMSE of 12.36. Comprehensive SHAP analysis 
revealed that APTT level and surgical intervention are major 
factors determining the outcome of treatment, providing 
essential information on the mechanism of DVT treatment 
response. We have successfully deployed our model as a web-
based clinical decision support tool. This is a major advance 
in taking complex algorithmic predictions and putting them 
to practical use in a clinical setting. This implementation 
connects theory with real-world medicine, providing doctors 

with a common, proven method to figure out how well their 
patients will respond to treatment. Our findings help 
understand how DVT is treated and set a basis for personal 
ways to treat DVT.  

Ethical issue 
The authors are aware of and comply with best practices in 
publication ethics, specifically regarding authorship 
(avoidance of guest authorship), dual submission, 
manipulation of figures, competing interests, and compliance 
with research ethics policies. The authors adhere to 
publication requirements that the submitted work is original 
and has not been published elsewhere. 

Data availability statement 
The manuscript contains all the data. However, more data will 

be available upon request from the authors. 

Conflict of interest 

The authors declare no potential conflict of interest. 

References 

[1] S. Wolf et al., "Epidemiology of deep vein thrombosis," 

Vasa, 2024.doi: 10.1024/0301-1526/a001145 

[2] G. Wagner et al., "Prevalence and incidence of venous 

thromboembolism in geriatric patients admitted to 

long-term care hospitals," Scientific reports, vol. 14, 

no. 1, p. 17737, 2024. doi:10.1038/s41598-024-

67480-1 

[3] S. Cheng et al., "Analysis of risk factors of 

postoperative lower extremity deep venous 

thrombosis in patients with cervical cancer," Clinical 

and Applied Thrombosis/Hemostasis, vol. 30, p. 

10760296241240747, 2024. 

doi:10.1177/10760296241240747 

[4] J. Björklund et al., "Risk of venous thromboembolic 

events after surgery for cancer," JAMA network open, 

vol. 7, no. 2, pp. e2354352-e2354352, 2024. 

doi:10.1001/jamanetworkopen.2023.54352 

[5] M.-E. Mathieu et al., "Management and outcomes of 

superficial vein thrombosis: a single-center 

retrospective study," Research and Practice in 

Thrombosis and Haemostasis, vol. 8, no. 1, p. 102263, 

2024. doi:10.1016/j.rpth.2023.102263 

[6] M. Betensky et al., "Recommendations for 

standardized definitions, clinical assessment, and 

future research in pediatric clinically unsuspected 

venous thromboembolism: communication from the 

ISTH SSC subcommittee on pediatric and neonatal 

thrombosis and hemostasis," Journal of Thrombosis 

and Haemostasis, vol. 20, no. 7, pp. 1729-1734, 2022. 

doi:10.1111/jth.15731 

[7] B. Cross, R. M. Turner, J. E. Zhang, and M. 

Pirmohamed, "Being precise with anticoagulation to 

reduce adverse drug reactions: are we there yet?," 

The pharmacogenomics journal, vol. 24, no. 2, p. 7, 

2024. doi:10.1038/s41397-024-00329-y 

[8] A. Kholmukhamedov, D. Subbotin, A. Gorin, and R. 

Ilyassov, "Anticoagulation management: current 

landscape and future trends," Journal of Clinical 

Medicine, vol. 14, no. 5, p. 1647, 2025. 

doi:10.3390/jcm14051647 

[9] L. Khider et al., "Acute phase determinant of post-

thrombotic syndrome: A review of the literature," 



N. Zhou et al. /Future Technology                                                                                      February 2026| Volume 05 | Issue 01 | Pages 254-262 

261 

 

Thrombosis Research, vol. 238, pp. 11-18, 2024. 

doi:10.1016/j.thromres.2024.04.004 

[10] M. R. Gil, J. Pantanowitz, and H. H. Rashidi, "Venous 

thromboembolism in the era of machine learning and 

artificial intelligence in medicine," Thrombosis 

Research, vol. 242, p. 109121, 2024. 

doi:10.1016/j.thromres.2024.109121 

[11] S. A. Alowais et al., "Revolutionizing healthcare: the 

role of artificial intelligence in clinical practice," BMC 

medical education, vol. 23, no. 1, p. 689, 2023. 

doi:10.1186/s12909-023-04698-z 

[12] A.-D. Anghele, V. Marina, L. Dragomir, C. A. Moscu, M. 

Anghele, and C. Anghel, "Predicting Deep Venous 

Thrombosis Using Artificial Intelligence: A Clinical 

Data Approach," Bioengineering, vol. 11, no. 11, p. 

1067, 2024. doi:10.3390/bioengineering11111067 

[13] K. Nothnagel and M. F. Aslam, "Evaluating the benefits 

of machine learning for diagnosing deep vein 

thrombosis compared with gold standard ultrasound: 

a feasibility study," BJGP open, vol. 8, no. 4, 2024. 

doi:10.3399/BJGPO.2024.0057 

[14] W. Sheng, X. Wang, W. Xu, Z. Hao, H. Ma, and S. Zhang, 

"Development and validation of machine learning 

models for venous thromboembolism risk 

assessment at admission: a retrospective study," 

Frontiers in Cardiovascular Medicine, vol. 10, p. 

1198526, 2023. doi:10.3389/fcvm.2023.1198526 

[15] N. Cahan et al., "Multimodal fusion models for 

pulmonary embolism mortality prediction," Scientific 

Reports, vol. 13, no. 1, p. 7544, 2023. 

doi:10.1038/s41598-023-34303-8 

[16] L. Z. Hou Yufen, "Diagnostic and Efficacy Criteria for 

Deep Venous Thrombosis of the Lower Extremities 

(Revised in 2015)," Chin J Integr Tradit West Med 

SurgeryZhong Guo Zhong Xi Yi Jie He Wai Ke Za Zhi, 

vol. 22, no. 5, pp. 520-520, 2016. 

[17] F. Dentali et al., "Clinical course of isolated distal deep 

vein thrombosis in patients with active cancer: a 

multicenter cohort study," Journal of thrombosis and 

haemostasis, vol. 15, no. 9, pp. 1757-1763, 2017. 

doi:10.1111/jth.13761 

[18] P. C. Kruger, J. W. Eikelboom, J. D. Douketis, and G. J. 

Hankey, "Deep vein thrombosis: update on diagnosis 

and management," Medical Journal of Australia, vol. 

210, no. 11, pp. 516-524, 2019. 

doi:10.5694/mja2.50201 

[19] F. Xing, L. Li, Y. Long, and Z. Xiang, "Admission 

prevalence of deep vein thrombosis in elderly 

Chinese patients with hip fracture and a new 

predictor based on risk factors for thrombosis 

screening," BMC musculoskeletal disorders, vol. 19, 

no. 1, p. 444, 2018. doi:10.1186/s12891-018-2371-5 

[20] W. Li, Z. Chuanlin, M. Shaoyu, C. H. Yeh, C. Liqun, and 

Z. Zeju, "Catheter-directed thrombolysis for patients 

with acute lower extremity deep vein thrombosis: a 

meta-analysis," Revista latino-americana de 

enfermagem, vol. 26, p. e2990, 2018. 

doi:10.1590/1518-8345.2309.2990 

[21] S. Z. Goldhaber, E. A. Magnuson, K. M. 

Chinnakondepalli, D. J. Cohen, and S. Vedantham, 

"Catheter-directed thrombolysis for deep vein 

thrombosis: 2021 update," Vascular Medicine, vol. 26, 

no. 6, pp. 662-669, 2021. 

doi:10.1177/1358863X211042930 

[22] X. Du et al., "Long-term outcome of catheter-directed 

thrombolysis in pregnancy-related venous 

thrombosis," Medical Science Monitor: International 

Medical Journal of Experimental and Clinical 

Research, vol. 25, p. 3771, 2019. 

doi:10.12659/msm.914592 

[23] M. J. Garcia et al., "Endovascular management of deep 

vein thrombosis with rheolytic thrombectomy: final 

report of the prospective multicenter PEARL 

(Peripheral Use of AngioJet Rheolytic Thrombectomy 

with a Variety of Catheter Lengths) registry," Journal 

of Vascular and Interventional Radiology, vol. 26, no. 

6, pp. 777-785, 2015. doi:10.1016/j.jvir.2015.01.036 

[24] H. İner, "The treatment indication affects the time in 

therapeutic range," Cardiovascular Surgery and 

Interventions, vol. 9, no. 3, pp. 147-151, 2022. 

doi:10.5606/e-cvsi.2022.1392 

[25] J. A. Kline, Z. P. Kahler, and D. M. Beam, "Outpatient 

treatment of low-risk venous thromboembolism with 

monotherapy oral anticoagulation: patient quality of 

life outcomes and clinician acceptance," Patient 

preference and adherence, pp. 561-569, 2016. 

doi:10.2147/ppa.s104446 

[26] A. Abdel-Hafez et al., "Predicting therapeutic 

response to unfractionated heparin therapy: machine 

learning approach," Interactive Journal of Medical 

Research, vol. 11, no. 2, p. e34533, 2022. 

doi:10.2196/34533 

[27] A. Shaikh et al., "Six-month outcomes of mechanical 

thrombectomy for treating deep vein thrombosis: 

analysis from the 500-patient CLOUT registry," 

Cardiovascular and interventional radiology, vol. 46, 

no. 11, pp. 1571-1580, 2023. doi:10.1007/s00270-

023-03509-8 

[28] M. Yao et al., "Neutrophil extracellular traps mediate 

deep vein thrombosis: from mechanism to therapy," 

Frontiers in immunology, vol. 14, p. 1198952, 2023. 

doi:10.3389/fimmu.2023.1198952 

[29] J. Ding, X. Yue, X. Tian, Z. Liao, R. Meng, and M. Zou, 

"Association between inflammatory biomarkers and 

venous thromboembolism: a systematic review and 

meta-analysis," Thrombosis Journal, vol. 21, no. 1, p. 

82, 2023. doi:10.1186/s12959-023-00526-y 

[30] N. Marlow, M. Eckert, G. Sharplin, I. Gwilt, and K. 

Carson-Chahhoud, "Graphical User Interface 

Development for a Hospital-Based Predictive Risk 

Tool: Protocol for a Co-Design Study," JMIR Research 

Protocols, vol. 12, no. 1, p. e47717, 2023. 

doi:10.2196/47717 

[31] S. Sachdeva et al., "Unraveling the role of cloud 

computing in health care system and biomedical 

sciences," Heliyon, vol. 10, no. 7, 2024. 

doi:10.1016/j.heliyon.2024.e29044 

[32] A. V. Ponce‐Bobadilla, V. Schmitt, C. S. Maier, S. 

Mensing, and S. Stodtmann, "Practical guide to SHAP 

analysis: Explaining supervised machine learning 

model predictions in drug development," Clinical and 



N. Zhou et al. /Future Technology                                                                                      February 2026| Volume 05 | Issue 01 | Pages 254-262 

262 

 

translational science, vol. 17, no. 11, p. e70056, 2024. 

doi:10.1111/cts.70056 

[33] I. D. Mienye et al., "A survey of explainable artificial 

intelligence in healthcare: Concepts, applications, and 

challenges," Informatics in Medicine Unlocked, vol. 

51, p. 101587, 2024. doi:10.1016/j.imu.2024.101587 

[34] J. S. Chang et al., "Continuous multimodal data supply 

chain and expandable clinical decision support for 

oncology," npj Digital Medicine, vol. 8, no. 1, p. 128, 

2025. doi:10.1038/s41746-025-01508-2 

[35] B. Vasey et al., "Reporting guideline for the early stage 

clinical evaluation of decision support systems driven 

by artificial intelligence: DECIDE-AI," bmj, vol. 377, 

2022. doi:10.1038/s41591-022-01772-9 

 

  

This article is an open-access article distributed under the 

terms and conditions of the Creative Commons Attribution 

(CC BY) license 

(https://creativecommons.org/licenses/by/4.0/). 

https://creativecommons.org/licenses/by/4.0/

