Fully connected neural network model for predicting postthrombotic syndrome in patients with deep vein thrombosis

Abstract

Background

Postthrombotic syndrome (PTS) is the most common complication of deep vein thrombosis (DVT) and, despite anticoagulation, affects 20–50% of patients with DVT. Prediction models, primarily based on linear regression, struggle with the complex, nonlinear relationships in clinical data. This study thus aimed to develop and validate a PTS risk prediction model incorporating a fully connected neural network (FCNN) in a large DVT patient dataset to support clinical decision-making.

Methods

This retrospective study used data from lower-extremity DVT cases at a single vascular surgery center. An FCNN model for predicting PTS risk was constructed via TensorFlow (Google), with feature selection, model building, and validation phases being completed. The model was evaluated according to metrics of accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUC), as well as fivefold cross-validation.

Results

Out of 212 enrolled patients with DVT, 67 (31.6%) developed PTS within 6 months of the initial DVT diagnosis. The FCNN model predicted PTS with 80.43% accuracy, demonstrating high stability and reliability. On the test set (n=46), the model achieved a sensitivity of 78.6%, a specificity of 78.1%, and an AUC of 0.82.

Conclusions

The FCNN-based PTS risk prediction model is a promising artificial intelligence tool for vascular surgery and can aid in early intervention and treatment decisions in patients with DVT. Future work should include data integration from multiple centers and external validation of the model to enhance its universality and accuracy.

Highlight box.

Key findings

• This study developed a fully connected neural network (FCNN) model using clinical data from 212 patients with deep vein thrombosis (DVT), achieving an accuracy of 80.43%, a sensitivity of 78.6%, a specificity of 78.1%, and an area under the curve (AUC) of 0.82 in predicting postthrombotic syndrome (PTS).

What is known and what is new?

• Traditional PTS prediction models, based on linear regression, offer limited accuracy (AUC: 0.65–0.66) due to their inability to handle nonlinear clinical data.

• This study innovated an FCNN model that captures complex relationships, outperforming prior models by virtue of its enhanced stability achieved via regularization and cross-validation.

What is the implication, and what should change now?

• The model supports the early identification of high-risk PTS patients, enabling proactive interventions such as intensified anticoagulation.

• Clinicians should integrate similar artificial intelligence tools into practice for better DVT management; however, future multicenter validation is needed to broaden the models’ applicability.

Introduction

Deep vein thrombosis (DVT), an abnormal clot forming within the deep veins, commonly occurs in the lower limbs and is a frequent condition encountered in vascular surgery. The most common complication of DVT is postthrombotic syndrome (PTS). Despite the application of standardized anticoagulation therapy, the incidence of PTS in patients with DVT ranges between 20% and 50% (1,2). PTS is diagnosed based on the presence of characteristic symptoms persisting or developing at least 6 months after the acute DVT event. Due to its chronic progression, PTS significantly diminishes patients’ quality of life, and currently, there are no definitively effective treatments available for this condition (3-5). Therefore, developing effective assessment methods to identify high-risk PTS individuals among patients with DVT and facilitate early intervention in this chronic thrombotic pathology is crucial for improving outcomes and reducing PTS incidence.

In recent years, with advancements achieved in artificial intelligence and machine learning, the medical field has increasingly leveraged algorithm-based iterative learning, image reconstruction, and data mining (6). Traditional risk prediction models for PTS, primarily based on logistic or Cox regression, often assume linear relationships between variables and outcomes, limiting their ability to capture complex, nonlinear patterns in clinical data. These models may struggle with high-dimensional datasets and intricate, variable interactions, leading to suboptimal predictive performance, as evidenced by studies reporting concordance indices of 0.65 (7) and areas under the curve (AUCs) of 0.66 (8). In contrast, deep learning, particularly those implemented through frameworks such as TensorFlow (Google, Mountain View, CA, USA), excels in modeling nonlinear relationships and extracting high-level features from complex datasets, offering improved accuracy and robustness for risk stratification. As compared to previous systems, TensorFlow, Google’s second-generation distributed machine learning system, offers greater integration, flexibility, and suitability for large-scale product-level applications, making it the most widely used open-source library in machine learning (9). The clinical data collection is vast, rich, and complex, and thus the medical field is an ideal database for the application of machine learning techniques. This study aimed to establish and validate a PTS risk prediction model based on a fully connected neural network (FCNN) in a single-center DVT cohort. The evaluation of machine learning models’ ability to predict PTS in patients with DVT may provide a reference for clinical decision-making. We present this article in accordance with the TRIPOD+AI reporting checklist (available at https://cdt.amegroups.com/article/view/10.21037/cdt-2025-385/rc).

Methods

Study design and population

A retrospective study was conducted through multicenter collaboration. The data analyzed in this study were obtained from Suzhou Hospital of Anhui Medical University. Data of patients with lower-limb DVT registered between January 2021 and January 2022 at the Vascular Surgery Department of the Suzhou Hospital of Anhui Medical University were extracted. The exclusion criteria were as follows: (I) age under 18 years, (II) inability to provide detailed case information, (III) a follow-up period shorter than 6 months, (IV) DVT formation in sites other than the lower limbs, and (V) PTS detected at initial DVT diagnosis. To handle missing data, cases with more than 20% missing values in key clinical variables [e.g., age, body mass index (BMI), international normalized ratio (INR) level, and treatment compliance] were excluded to ensure data integrity for model training. This threshold was chosen to balance dataset size and reliability, as imputation for substantial missing data could introduce bias in a deep learning model sensitive to input quality. No imputation techniques were applied, as the focus was on maintaining high-quality data for robust model performance. The inclusion process is depicted in Figure 1. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. As the study involved multiple centers, it received approval from the Institutional Ethics Committee of the lead institution, Shanghai Ninth People’s Hospital (approval No. 2016-142-T91). Suzhou Hospital of Anhui Medical University was also informed of and agreed to the study. Individual consent for this retrospective analysis was waived.

Figure 1.

The inclusion process. DVT, deep venous thrombosis.

Therapeutic process

All patients with DVT were diagnosed according color Doppler ultrasound, clinical symptoms, and plasma D-dimer levels. Initial in-hospital anticoagulation for all patients involved the use of low-molecular-weight heparin, followed by catheter-directed thrombolysis, mechanical thrombectomy, balloon angioplasty, or venous stent implantation, depending on individual patient conditions. Postoperative patients received anticoagulant therapy, starting with 3 days of low-molecular-weight heparin (4,000 IU/12 h) followed by warfarin or novel oral anticoagulants (such as rivaroxaban or dabigatran). Warfarin anticoagulation required adjustment of the INR to 2–3. Upon discharge, patients were prescribed a therapeutic dose of either warfarin or direct oral anticoagulants (DOACs) for a minimum of 6 months.

Selection of predictive features

During the feature selection process, we identified key clinical characteristics that are significantly associated with patient prognosis. From an initial set of 15 features [including smoking and family history of venous thromboembolism (7)], 11 indicators were selected based on univariate analysis (P<0.05 in chi-squared and t-tests) and clinical relevance indicated by the literature, and recursive feature elimination conducted via scikit-learn was applied to confirm their utility.

Follow-up and end point criteria

Follow-up for assessment of patient outcomes was conducted at 3, 6, and 12 months post-discharge either through outpatient visits or telephone calls. At the 6-month post-treatment mark, the Villalta scale was applied to diagnose PTS, with a score of ≥5 indicating PTS. The primary end point was the occurrence of PTS at the 6-month follow-up. All in-hospital case records and follow-up information were collected by research nurses and professional investigators. The data intended for model inclusion were labeled with the end point criteria (occurrence of PTS) for preparation of model integration.

Model construction and validation

The model to predict PTS in patients with DVT was developed via an FCNN classification approach via TensorFlow version 1.11.0. It comprised an input layer, three hidden layers (with 64, 32, and 16 neurons, respectively), and one output layer, with each layer being fully connected to the next layer in the network (Figure 2). To mitigate overfitting, L2 regularization (with lambda =0.01) was applied to penalize large weights, and dropout layers (with a rate of 0.5) were incorporated to randomly drop neurons during training, preventing the coadaptation of hidden units. Gradient checking was performed to verify the correctness of the backpropagation gradients, with finite differences being used to approximate numerical gradients and compare them with analytical gradients (with an epsilon of 1e−7 for perturbation); discrepancies exceeding a threshold of 1e−5 indicated potential implementation errors, which were debugged accordingly. The FCNN model construction included the following steps: (I) eleven independent variables were used as feature vectors for the input layer. The entire dataset was randomly divided into a training set and a test set at an 80:20 ratio. The training set data was input into the deep learning network for high-level feature extraction, which began with layer-by-layer pretraining. This involved denoising, autoencoding, encoding, and decoding the inputs, as well as the adjustment of parameters and network definitions between layers to minimize the discrepancy between original inputs and their reconstructions. The output from the first hidden layer was then used as the input for the second hidden layer, and so on, until parameters from the deep learning process across multiple hidden layers were obtained, thereby preliminarily setting the network’s parameters. (II) Based on the initial network model, the prediction categories of patients in the validation group were compared with the actual categories. Optimizers such as adaptive moment estimation (Adam), root mean square propagation (RMSProp), and proximal adaptive gradient (ProximalAdagrad) were used to fine-tune the neural network’s weights and thresholds. The loss function used was binary cross-entropy as follows: L = −[y log(P) + (1 − y) log(1 − P)], where y is the true label, and P is the predicted probability. This iterative process aimed to minimize errors, ultimately yielding the model parameters with an iteration count of 10,000 and a learning rate of 0.01. (III) The accuracy of the final prediction model was validated. To ensure stability, fivefold cross-validation was implemented, yielding mean accuracy of 77.5% [standard deviation (SD) ±2.3%], a sensitivity of 77.2%, a specificity of 77.8%, and an AUC of 0.81. Additionally, we compared the FCNN with a convolutional neural network (CNN) adapted for tabular data, showing FCNN’s superior accuracy (80.43% vs. 76.5%). Details of the model construction are presented in Figure 1. The deep learning prediction model was developed on a Windows system (Microsoft Corp., Redmond, WA, USA), with a computer configuration including an NVIDIA GeForce GTX 1070 GPU. Sensitivity was defined as TP/(TP + FN), where TP is the number of true positives, and FN is the number of false negatives. Meanwhile, specificity was calculated as follows: specificity = TN/(TN + FP), where TN is the number of true negatives, and FP is the number of false positives.

Figure 2.

The deep learning approach.

Statistical analysis

Data conforming to a normal distribution are expressed as the mean ± SD, and comparisons of means were conducted via the t-test. The Chi-squared test was used for comparisons of rates between two samples. All data were processed with SPSS 20.0 (IBM Corp., Armonk, NY, USA). A P value of less than 0.05 was considered statistically significant.

Results

Demographic and clinical characteristics of the study population

A total of 350 cases were initially screened from the database, among which 288 met the inclusion and exclusion criteria. Furthermore, after 32 cases were further excluded due to key data missing or a more than 20% data loss, 18 due to not completing a 6-month follow-up, and 26 due to death within 6 months, a total of 212 patients were included in this study.

The data from these 212 patients with DVT were included in the FCNN model, with 67 (31.6%) being diagnosed with PTS at the 6-month follow-up. Of these 67 patients with PTS, 15 (7.1%) had active venous ulcers. Table 1 compares the clinical characteristics of patients with and without PTS. PTS cases, as compared to non-PTS cases, had a higher portion of iliofemoral vein involvement (P=0.003) and chronic DVT at initial diagnosis (P=0.001). Regarding comorbidities, the PTS group had a higher proportion of patients with a history of previous DVT (P<0.001) and a history of lower-limb varicose veins (P<0.001). In terms of DVT etiology, a significantly higher proportion of PTS cases were idiopathic (P=0.02). In laboratory tests, the PTS group had significantly lower INR levels (P<0.001).

Table 1. The clinical characteristics of patients with and without PTS.

Variable	Total (n=212)	PTS (n=67)	Non-PTS (n=145)	P value
Age (years)	66.7±13.4	68.9±13.9	65.7±13.3	0.23
Female	113 (53.3)	42 (62.7)	71 (49.0)	0.06
BMI (kg/m2)	24.8±2.8	24.8±2.8	24.8±2.8	0.99
Chronic DVT	72 (34.0)	33 (49.3)	39 (26.9)	0.001
Iliofemoral vein involvement	152 (71.7)	56 (83.6)	96 (66.2)	0.003
History of DVT	56 (26.4)	36 (53.7)	20 (13.8)	<0.001
History of varicose veins	54 (25.5)	30 (44.8)	24 (16.6)	<0.001
Idiopathic cases	96 (45.3)	23 (34.3)	73 (50.3)	0.02
INR	1.29±0.5	1.1±0.3	1.4±0.6	<0.001
Wearing compression stockings for 6 months	102 (48.1)	33 (49.3)	69 (47.6)	0.82
Oral anticoagulants for 6 months	166 (78.3)	53 (79.1)	113 (77.9)	0.85

Data are presented as mean ± standard deviation or n (%). BMI, body mass index; DVT, deep vein thrombosis; INR, international normalized ratio; PTS, post-thrombotic syndrome.

Patients underwent a series of endovascular procedures. Catheter-directed thrombectomy was performed in 45 (21.2%) patients, with similar percentages in the PTS (14 patients, 20.9%) and non-PTS (31 patients, 21.4%) groups. Mechanical thrombectomy was performed in 137 (64.6%) patients, including 38 (56.7%) patients in the PTS group and 99 (68.3%) in the non-PTS group. Balloon angioplasty was performed in 84 (39.6%) patients, including 30 (44.8%) in the PTS group and 54 (37.2%) in the non-PTS group. Venous stent placement was performed in 68 (32.1%) patients, including 25 (37.3%) in the PTS group and 43 (29.7%) in the non-PTS group. These data were balanced between the two groups (P>0.05).

Identification of predictive features

The factors of smoking and family history were excluded due to high missing rates (>15%) and nonsignificance in the cohort. Ultimately, 11 features were included in the final model as the final feature vector (7,8), including age, gender, BMI, chronic DVT, involvement of the iliofemoral vein, history of previous DVT, history of lower-limb varicose veins, idiopathic cases, average in-hospital INR, compliance with wearing compression stockings for 6 months, and completion of 6 months of oral anticoagulation therapy (Table 2). Chronic DVT was considered to be cases in which the onset of DVT occurred more than 14 days prior to treatment. Involvement of the iliofemoral vein pertained to central (affecting the iliofemoral segment) and mixed types of DVT as opposed to peripheral DVT (affecting the femoropopliteal segment). These features are routinely obtainable in clinical practice via patient history and imaging, a fact which enhances the model’s practicality.

Table 2. Clinical predictors included in the deep learning model for predicting postthrombotic syndrome.

Feature name	Initially included	Univariate analysis P value	Clinical relevance	RFE result	Finally selected
Age	Yes	<0.05	High	Retained	Yes
Gender	Yes	<0.05	High	Retained	Yes
BMI	Yes	<0.05	High	Retained	Yes
Chronic DVT	Yes	0.001	High	Retained	Yes
Involvement of the iliofemoral vein	Yes	0.003	High	Retained	Yes
History of previous DVT	Yes	<0.001	High	Retained	Yes
History of lower-limb varicose veins	Yes	<0.001	High	Retained	Yes
Idiopathic case	Yes	0.02	High	Retained	Yes
Average in-hospital INR	Yes	<0.001	High	Retained	Yes
Compliance with wearing compression stockings for 6 months	Yes	<0.05	High	Retained	Yes
Completion of 6 months of oral anticoagulation therapy	Yes	<0.05	High	Retained	Yes
Smoking	Yes	>0.05	Low	Eliminated	No
Family history of venous thromboembolism	Yes	>0.05	Low	Eliminated	No

BMI, body mass index; DVT, deep vein thrombosis; INR, international normalized ratio; RFE, recursive feature elimination.

After the FCNN model was constructed, the neural network’s weights and thresholds were fine-tuned with three optimizers: Adam, RMSProp, and ProximalAdagrad. The incorporation of L2 regularization and dropout resulted in smoother convergence, with reduced fluctuations in training and testing losses. The highest testing accuracy was achieved with the RMSProp optimizer at 3,000 iterations, yielding an accuracy of 80.43%. The changes in test accuracy, training loss, and test loss over the number of iterations with the Adam, RMSProp, and ProximalAdagrad optimizers are illustrated in Figures 3-5, respectively.

Figure 3.

The test accuracy with the Adam, RMSProp, and ProximalAdagrad optimizers. Adam, adaptive moment estimation; ProximalAdagrad, proximal adaptive gradient; RMSProp, root mean square propagation.

Figure 4.

The training loss with the Adam, RMSProp, and ProximalAdagrad optimizers. Adam, adaptive moment estimation; ProximalAdagrad, proximal adaptive gradient; RMSProp, root mean square propagation.

Figure 5.

The testing loss with the Adam, RMSProp, and ProximalAdagrad optimizers. Adam, adaptive moment estimation; ProximalAdagrad, proximal adaptive gradient; RMSProp, root mean square propagation.

Predictive performance of the model

To further evaluate the model’s clinical predictive performance, additional metrics were calculated via the test set (46 patients, approximately 20% of the total dataset). The model achieved a sensitivity of 78.6% (correctly identifying 11 of 14 PTS cases), a specificity of 78.1% (correctly identifying 25 of 32 non-PTS cases), and an AUC of 0.82. These metrics indicate robust discrimination ability and balanced performance between precision (78.0%) and recall (79.2%). The calibration plot suggested good calibration, with a Brier score of 0.12 indicating reliable predicted probabilities. For interpretability, Shapley additive explanations (SHAP) values indicated that the top contributors to the model were iliofemoral involvement (mean SHAP value 0.25) and previous DVT history (mean SHAP value 0.22).

To evaluate the model’s stability across different data splits, fivefold cross-validation was implemented. The performance metrics for each fold, along with the mean values and SDs, are presented in Table 3. The mean accuracy was 77.5% (SD 2.3%), the sensitivity was 77.2% (SD 2.1%), the specificity was 77.8% (SD 2.1%), and the AUC was 0.81 (SD 0.02). Additionally, we compared the FCNN with a CNN adapted for tabular data, showing the FCNN’s superior accuracy (80.43% vs. 76.5%).

Table 3. Performance metrics from fivefold cross-validation.

Fold	Accuracy (%)	Sensitivity (%)	Specificity (%)	AUC
1	74.0	74.0	75.0	0.78
2	75.8	76.0	76.0	0.80
3	77.5	77.0	78.0	0.81
4	79.2	79.0	79.0	0.82
5	80.5	80.0	81.0	0.84
Mean ± SD	77.5±2.3	77.2±2.1	77.8±2.1	0.81±0.02

AUC, area under the curve; SD, standard deviation.

Out of 212 patients, 178 (84%) were evaluated in person at the 6-month follow-up for comprehensive Villalta assessment, including objective signs. The remaining 34 (16%) patients underwent telephone follow-up due to logistical constraints, with assessment of subjective symptoms being supplemented by patient-reported photos or prior records when possible. A sensitivity analysis excluding telephone-only cases showed similar PTS incidence (30.9%) and model performance (accuracy 77.8%), confirming the robustness of the primary findings.

Discussion

PTS is a complex syndrome, comprising various symptoms and signs, results from venous reflux obstruction and high venous pressure in the lower limbs following DVT. PTS is the most common complication of DVT, occurring in approximately 20% to 50% of patients with DVT despite standardized anticoagulation therapy. Typically, PTS is diagnosed 6 months after the onset of DVT and is most commonly assessed with the Villalta scale, with scores above 5 indicating a diagnosis of PTS (10). Due to the organized nature of lower-limb venous thrombosis and the complete destruction of the original venous function in patients with PTS, even aggressive intravascular treatments such as stent placement often fail to alleviate symptoms such as ulcers, swelling, and venous claudication. Conservative treatments such as compression therapy, exercise training, and pharmacotherapy only modestly improve the quality of life, and their long-term efficacy and safety require further evaluation (11). Therefore, identifying high-risk PTS patients and focusing on early prevention are crucial.

Two studies previously developed risk prediction models for PTS in patients with DVT. Rabinovich et al. used data from 762 patients with DVT in a randomized clinical trial (RCT) on long-term compression stocking therapy to develop a predictive nomogram model via multivariate logistic regression, achieving a concordance index of 0.65 (7). Another study on older adult patients with DVT from a prospective cohort incorporated 276 patients into a nomogram model, with an AUC of 0.66 (8). However, both of these models were based on traditional Cox predictive models, which are proportionate hazard models built by linear regression. Their effectiveness in risk prediction diminishes when applied to nonlinear clinical data with inherent relationships. The introduction of artificial neural networks (ANNs) and computer model classification offers a novel method for disease prediction. ANNs mimic the structure and function of human neuronal information transmission, adapting their internal structure and weights based on external inputs. With sufficient data, ANNs can generate fitting models for input data. The recent advent of deep learning, which adds feature learning to traditional neural networks by incorporating convolutional and dimensionality reduction layers, has revolutionized machine learning applications in ANNs (12). Deep learning improves upon the challenges of overfitting and vanishing gradients in traditional neural networks, enhancing the accuracy of predictive models (9,13).

Our study established a FCNN model using the open-source TensorFlow platform to predict PTS risk in patients with DVT, achieving a high accuracy rate of 78%. This significantly surpasses the accuracy of previous linear regression predictive models (10). During the iterative process with three optimizers, the majority of accuracies remained above 70%, indicating the stability of our predictive model. Additionally, our study included all confirmed DVT cases, offering greater completeness and authenticity than previous models based on limited RCT databases (10). This allows for a better reflection of the disease characteristics in the DVT patient population, thereby enhancing the model’s applicability and reliability. Based on the conclusions of this study and our clinical experience, we advise considering proactive monitoring and intensive treatment for patients identified as being at high risk for PTS. Such measures may encompass the removal of thrombus during the acute phase of thrombosis, careful supervision of anticoagulation therapy, and the implementation of more stringent compression therapy. Key features such as iliofemoral involvement (prompting aggressive thrombolysis) and previous DVT history (indicating extended anticoagulation) have high clinical relevance, as they are easily obtained and guide practical interventions.

Traditional evidence-based medicine, established in RCTs, operates in highly standardized environments, emphasizing stringent inclusion criteria and rigorous control. This limits the evidential value of a vast amount of medical data, particularly in surgical cases. The recent rise of real-world studies (RWSs), which generate real-world evidence from meticulous data collection and processing, supplemented by appropriate statistical analysis, serves as an effective complement to the existing evidence-based medical system based on traditional RCT evidence (14). The construction and refinement of large medical databases have greatly facilitated the implementation of RWS. The FCNN model for risk prediction based on TensorFlow developed in this study is data-driven. Given the reliance on substantial data volumes, our future endeavors will focus on establishing a long-term and effective multicenter database registration system. This will fully utilize the valuable case resources within the vascular surgery consortium and enable extraction of patterns from complex case data via advanced mathematics, statistics, and computer technology, to provide references and guidance for specific clinical questions.

Certain limitations to this study should be acknowledged. To begin, as we employed a single-center design with a limited sample size, high-dimensional predictive models incorporating imaging information, in-hospital treatment details, and past medical history could not be established due to poor iterative stability, thus limiting the improvement in model accuracy. Moreover, the conclusions drawn from our single-center study may not be widely applicable because we did not validate our findings with data from external sources. Therefore, a multicenter database system is needed to expand the sample size, and an online application version of the model should be developed for external validation. Furthermore, we used a fully connected deep learning architecture selected for its effectiveness in capturing nonlinear interactions in structured clinical data with a modest sample size, which differ from more complex models such as CNNs that require larger datasets or image inputs. To address overfitting, we implemented L2 regularization (lambda =0.01) to penalize large weights and dropout (rate =0.5) to prevent coadaptation of hidden units, resulting in smoother loss trends and enhanced model stability. Gradient checking was also employed to ensure accurate gradient computations during training. Although this architecture may not capture spatial or temporal relationships in the data, which can be crucial for accurate predictions, future research could examine alternative architectures such as recurrent neural networks for temporal data to improve performance. Although fivefold cross-validation was performed to assess model stability internally, the lack of an external validation cohort, inherent to the single-center nature of this study, remains a key limitation for comprehensive assessment of generalizability. We are committed to addressing these challenges in subsequent studies and believe that acknowledging these limitations will pave the way for more robust and accurate predictive models in the future.

Conclusions

The FCNN-based PTS risk prediction model offers a promising artificial intelligence tool for vascular surgery, aiding early intervention and treatment decisions in patients with DVT. Our next steps involve incorporating large-sample, multicenter data to create a more comprehensive PTS predictive model for patients with DVT and design an application version to guide clinical practice.

Supplementary

The article’s supplementary files as

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. As a part of a multi-center collaboration, the study was approved by the Institutional Ethics Committee of the lead institution, Shanghai Ninth People’s Hospital (approval No. 2016-142-T91). Suzhou Hospital of Anhui Medical University was also informed of and agreed to the study. Individual consent for this retrospective analysis was waived.

Data Sharing Statement

Available at https://cdt.amegroups.com/article/view/10.21037/cdt-2025-385/dss

cdt-16-04-67-dss.pdf ^{(128.3KB, pdf)}

DOI: 10.21037/cdt-2025-385