Multimodal Deep Learning for Predicting Deep Vein Thrombosis Risk After Total Knee Arthroplasty: A Clinical Study Authors: Kejia Zhu, Hang Li, Hui Zhang, Biao Wang, Bin Shen, Yong Nie Published: 2026 ABSTRACT ABSTRACTBackgroundDeep vein thrombosis (DVT) following total knee arthroplasty (TKA) remains a major postoperative complication. Conventional risk assessment tools such as the Caprini score show limited predictive performance (AUC 0.65–0.72). We developed a multimodal deep learning framework that integrates early postoperative radiographs with clinical data to improve DVT risk stratification.MethodsThis single‐center retrospective cohort study analyzed 1200 patients undergoing primary unilateral TKA (DVT incidence: 18.0%) from January 2018 to June 2023. A dual‐branch architecture was implemented: ViT‐B/16 or ResNet50 processed standardized radiographs (metal artifact suppression; 512 × 512 resolution), while Clinical‐BERT encoded structured electronic medical record (EMR) representations. A dynamic attention fusion module generated sample‐specific modality weights (α) for feature fusion (α·v_img + (1 − α)·v_txt). Performance was evaluated on a stratified test set (n = 180) against the Caprini score, unimodal models, and a nonattention concatenation fusion baseline. AUCs were compared using the DeLong test, and sensitivity/specificity differences were assessed using the McNemar test.ResultsThe multimodal model achieved an AUC of 0.89 (95% CI: 0.85–0.92; DeLong p < 0.001 vs. all comparators), sensitivity 83% (76%–89%), specificity 88% (82%–93%), and F1‐score 0.79. It identified 91.7% (11/12) of DVT cases labeled low‐risk by Caprini. The learned attention coefficient α suggested modality reliance patterns associated with specific findings (e.g., higher α in cases with prosthesis–bone lucency > 2 mm; higher text contribution in cases with D‐dimer > 1.0 mg/L), but these associations do not establish causality. Anatomical heatmaps consistently localized high‐risk regions, particularly the popliteal vein projection area and tibial interface.ConclusionThis attention‐based multimodal model integrates biomechanical (radiographic) and biochemical/clinical (EMR‐derived) signals to improve early postoperative DVT risk prediction while providing interpretable anatomical cues.Keywords: attention mechanism, deep vein thrombosis, multimodal deep learning, risk prediction, total knee arthroplasty Introduction Total knee arthroplasty (TKA) is an effective treatment for end‐stage knee osteoarthritis, with annual global surgical volume exceeding one million cases. However, deep vein thrombosis (DVT) remains a serious postoperative complication, with reported incidence of 5.8%–15.7% [1]. DVT can lead to pulmonary embolism (mortality 15%–30%) and postthrombotic syndrome, and it increases hospitalization costs by approximately 34% [2]. Current clinical practice often relies on the Caprini risk assessment model [3], yet its predictive accuracy is limited (AUC 0.65–0.72) with sensitivity frequently below 60% [4]. These limitations reflect both constrained variable coverage and incomplete representation of key risk dimensions. First, postoperative radiographs contain biomechanical information: prosthesis malalignment may alter popliteal venous hemodynamics (e.g., femoral component varus > 3° has been associated with increased venous compression risk) [5], and lucency at the bone–implant interface may correlate with local inflammatory activity. Second, laboratory indicators evolve dynamically (e.g., elevated early postoperative D‐dimer levels are associated with increased DVT risk) [6, 7]. Traditional scoring systems cannot effectively integrate heterogeneous imaging signals with longitudinal clinical information. Multimodal deep learning can jointly mine complementary information across heterogeneous medical data. In orthopedics, computer vision can quantify subtle imaging patterns that are difficult to capture via routine assessment [8, 9]: Vision Transformers (ViTs) can localize complex radiographic patterns via self‐attention, and convolutional neural networks (CNNs) can detect fine lucent lines at the implant interface with high sensitivity [10, 11]. In parallel, transformer‐based language models (e.g., Clinical‐BERT) can encode EMR semantics and structured clinical context. However, many prior fusion approaches exhibit key shortcomings: (1) naïve concatenation ignores nonlinear cross‐modal interactions; (2) modality heterogeneity is not explicitly handled; and (3) interpretability is limited, which can reduce clinical trust. To address these issues, we propose a dual Transformer + attention fusion architecture that (i) extracts radiographic features related to biomechanical and interface changes using ViT (or ResNet50), (ii) encodes clinical context from structured EMR‐derived sequences using Clinical‐BERT, and (iii) dynamically weights modality contributions via a gated attention fusion mechanism. Accordingly, this study addressed the following objectives: To determine whether integrating postoperative radiographs with structured EMR data can improve early DVT risk prediction after TKA. To determine whether integrating postoperative radiographs with structured EMR data can improve early DVT risk prediction after TKA. To evaluate whether attention‐based multimodal fusion outperforms traditional risk assessment and conventional fusion strategies. To evaluate whether attention‐based multimodal fusion outperforms traditional risk assessment and conventional fusion strategies. To explore whether the model captures clinically meaningful imaging and clinical patterns relevant to postoperative DVT. To explore whether the model captures clinically meaningful imaging and clinical patterns relevant to postoperative DVT. Materials and Methods Study Population This study was approved by the Institutional Ethics Committee (approval number: 1975 of 2023) and adopted a retrospective cohort design. Consecutive patients who underwent primary unilateral TKA at the Orthopedic Department of West China Hospital from January 2018 to June 2023 were included. Inclusion criteria: primary osteoarthritis; primary unilateral TKA with posterior‐stabilized prostheses (Zimmer Biomet Persona or Depuy Attune); standard anteroposterior and lateral digital radiography within 24 h postoperatively (DICOM); complete EMR including demographics, laboratory indicators, medication use, and perioperative parameters; preoperative ultrasound excluding DVT; ASA I–III; eGFR > 60 mL/min/1.73m2; and systematic postoperative follow‐up. Exclusion criteria: revision surgery; bilateral TKA; intraoperative conversion to semi‐constrained prostheses; operative time > 180 min; active infection; malignancy history; contraindications to anticoagulation; radiographs with metal artifacts obscuring key anatomical structures (Metal Artifact Reduction Sequence [MARS] visibility score < 3); missing key variables > 20%; or loss to follow‐up. A total of 1200 patients were included. DVT was diagnosed by lower extremity venous ultrasound within 14 days postoperatively following International Society on Thrombosis and Hemostasis criteria (venous noncompressibility or filling defect). Radiographs were obtained using the Siemens AXIOM Aristos FX DR system (60–70 kV, 5–8 mAs, focal‐film distance 100 cm). Rationale for using an early single postoperative radiograph: Radiographs acquired within 24 h were selected to reflect a clinically feasible early risk stratification window aligned with routine postoperative imaging workflows. We acknowledge that DVT risk is time‐dependent and that longitudinal radiographic changes may contain additional predictive information; this is addressed as a limitation and a future direction. Image Preprocessing (Metal Artifact Suppression and Standardization) Metal implant artifacts can degrade radiographic interpretability near the prosthesis–bone interface. We applied a MARS to suppress high‐intensity streak artifacts while preserving overall anatomical geometry. Because radiographic artifact suppression can potentially alter local intensities, we minimized the risk of distorting anatomical landmarks through (i) exclusion of low‐visibility images using the predefined MARS visibility threshold (< 3) and (ii) standardized postprocessing across all cases. After artifact suppression, we performed windowing (bone window: width 2000 HU, level 500 HU) to enhance the bone–implant interface, resampled images to 512 × 512 pixels via bicubic interpolation, and normalized pixel values to [0,1]. Model Architecture The imaging branch used ViT‐B/16 or ResNet50 as the backbone network. The ViT was configured with 12 layers of Transformer encoders and 12 attention heads, embedding 16 × 16 pixel patches into 768‐dimensional vectors. Finally, the global features corresponding to the [CLS] token were extracted and dimensionally increased to 1024 via a fully connected layer. The text branch adopted the pretrained Clinical‐BERT model in the biomedical field. Input sequences (maximum length 128 tokens) were processed using the WordPiece tokenizer, outputting a 768‐dimensional semantic vector of the [CLS] token. The multimodal fusion module innovatively introduced a dynamic attention mechanism: first, image features were projected to 768 dimensions via a linear layer to match text feature dimensionality. Then, the projected image features and text features were concatenated to form a 1536‐dimensional joint vector; a learnable weight matrix was used to calculate the modal contribution coefficient α (formula: α = sigmoid(W·[v_img_proj⊕v_txt] + b)); finally, weighted fusion features α*v_img_proj + (1‐α)*v_txt were generated. After being enhanced by two fully connected networks (1024,512,256 nodes, ReLU activation, Dropout = 0.5), the fusion features were input into a Sigmoid classifier to output the DVT risk probability. The feature concatenation fusion baseline simply concatenated image and text features into a 1792D vector, which was then fed into a fully connected network (1024,512,256 nodes) for classification, without attention‐based weighting. Training and Statistical Analysis The dataset was divided using stratified random sampling (training set: 840 cases; validation set: 180 cases; test set: 180 cases), maintaining a balanced DVT positive rate across subsets (18.1% ± 0.3%). The AdamW optimizer (learning rate 1e‐5, weight decay 1e‐4) was used to minimize the Focal Loss function (γ = 2, α = 0.75), which effectively alleviated data imbalance by suppressing the weight of easily classified samples. The training early stopping strategy was set to terminate when the validation set AUC did not improve for five consecutive epochs. Baseline models included three control groups: the traditional Caprini scoring system, unimodal models (imaging‐only or clinical data‐only), and a feature concatenation fusion model without an attention mechanism. Statistical Analysis Methods The AREA under the receiver operating characteristic curve (AUC) of the test set and its 95% confidence interval were calculated using the Delong test, and differences were compared with baseline models. Classification performance indicators included sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV), with statistical significance evaluated using the McNemar test (α = 0.05). All analyses were performed in Python 3.9, using the PyTorch 1.12 deep learning framework and the HuggingFace Transformers 4.24 library. Results Clinical Data A total of 1200 patients who underwent primary unilateral TKA were finally included in this study, among whom 216 cases (18.0% incidence) were diagnosed with DVT via ultrasound within 14 days after surgery. As shown in Table 1, patients in the DVT‐positive group exhibited significant high‐risk characteristics: the average age in this group reached 71.6 ± 6.8 years, which was 4.2 years higher than that in the negative group (67.4 ± 7.1 years) (p < 0.001), and the BMI was also significantly higher (30.1 ± 3.9 vs. 28.0 ± 3.4 kg/m2, p < 0.001). In terms of comorbidities, the prevalence rates of hypertension (63.9% vs. 51.8%, p = 0.001), diabetes (36.1% vs. 21.3%, p < 0.001), and cardiovascular diseases (25.0% vs. 14.0%, p < 0.001) in the DVT‐positive group were significantly higher than those in the negative group. It is particularly noteworthy that the proportion of patients with a history of venous thromboembolism in the positive group reached 16.7%, which was 4.5 times higher than that in the negative group (3.7%) (p < 0.001). TABLE 1Baseline characteristics and group comparison of patients.VariableTotal (n = 1200)DVT‐positive group (n = 216)DVT‐negative group (n = 984)Statistic value p Demographic characteristicsAge (years)68.3 ± 7.271.6 ± 6.867.4 ± 7.1 t = 8.16< 0.001Female (n, %)732 (61.0%)142 (65.7%)590 (60.0%) χ 2 = 2.410.12BMI (kg/m2)28.4 ± 3.630.1 ± 3.928.0 ± 3.4 t = 7.38< 0.001Comorbidities (n, %)Hypertension648 (54.0%)138 (63.9%)510 (51.8%) χ 2 = 10.440.001Diabetes mellitus288 (24.0%)78 (36.1%)210 (21.3%) χ 2 = 21.18< 0.001Cardiovascular diseases192 (16.0%)54 (25.0%)138 (14.0%) χ 2 = 15.83< 0.001History of venous thromboembolism72 (6.0%)36 (16.7%)36 (3.7%) χ 2 = 49.67< 0.001Laboratory indicatorsD‐dimer (mg/L)0.62 ± 0.381.18 ± 0.420.52 ± 0.31 t = 21.95< 0.001> 0.5 mg/L (n, %)672 (56.0%)186 (86.1%)486 (49.4%) χ 2 = 96.54< 0.001Platelet count (×109/L)228 ± 64265 ± 71221 ± 60 t = 8.47< 0.001Surgical parametersOperation duration (min)112 ± 23128 ± 26108 ± 21 t = 10.45< 0.001Tourniquet time (min)78 ± 1692 ± 1875 ± 15 t = 12.92< 0.001Intraoperative blood loss (mL)235 ± 85298 ± 92220 ± 78 t = 11.58< 0.001Anticoagulant regimens (n, %)Rivaroxaban576 (48.0%)84 (38.9%)492 (50.0%) χ 2 = 8.750.003Enoxaparin432 (36.0%)90 (41.7%)342 (34.8%) χ 2 = 3.720.05Aspirin192 (16.0%)42 (19.4%)150 (15.2%) χ 2 = 2.300.11Traditional scoresCaprini score7.2 ± 2.19.1 ± 1.86.8 ± 2.0 t = 16.60< 0.001High risk (≥ 5 points) (n, %)864 (72.0%)204 (94.4%)660 (67.1%) χ 2 = 65.71< 0.001 Note: Data are presented as mean ± standard deviation or n (%). Continuous variables were compared using the independent‐samples t test, and categorical variables were compared using the chi‐square test. Statistic values are reported as t or χ 2, respectively. Baseline characteristics and group comparison of patients. Note: Data are presented as mean ± standard deviation or n (%). Continuous variables were compared using the independent‐samples t test, and categorical variables were compared using the chi‐square test. Statistic values are reported as t or χ 2, respectively. Model Construction and Evaluation The model framework is shown in Figure 1. In the performance evaluation on the test set, the multimodal fusion model constructed in this study demonstrated significant predictive advantages. As presented in Table 2, the AUC of the model reached 0.89 (95% CI: 0.85–0.92), which was significantly superior to the traditional Caprini score (AUC = 0.71, Delong test p < 0.001) and unimodal models—including the imaging branch model (ViT backbone, AUC = 0.78) and the clinical branch model (BERT backbone, AUC = 0.74). In terms of key classification indicators, the multimodal model achieved a sensitivity of 83% (95% CI: 76%–89%) and a specificity of 88% (95% CI: 82%–93%), with an F1‐score of 0.79. This represented an 11.3% improvement compared to the feature concatenation fusion model without an attention mechanism (ΔAUC = 0.06, p = 0.002). Notably, the model exhibited excellent ability to identify high‐risk patients missed by the traditional scoring system: among 12 DVT patients with a Caprini score < 5, 11 were successfully Warning (only 1 case of intramuscular venous thrombosis with prosthesis varus < 2° was missed), and the NPV reached 97.2%. Architecture of the multimodal deep learning framework for predicting postoperative deep vein thrombosis after total knee arthroplasty. The framework includes two parallel feature‐extraction branches. The imaging branch processes postoperative knee radiographs using a convolutional neural network backbone (ResNet50 shown in the schematic; ViT‐B/16 was alternatively used in model experiments) to obtain image embeddings, which are projected into a 768‐dimensional latent space by a fully connected layer. The clinical branch encodes structured perioperative electronic medical record data using Clinical‐BERT to generate a corresponding 768‐dimensional feature vector. These two modality‐specific representations are subsequently integrated within a multimodal fusion module and classified through fully connected layers followed by a sigmoid activation function, yielding the probability of postoperative DVT. The figure illustrates the overall workflow of feature extraction, multimodal integration, and binary risk prediction. TABLE 2Model evaluation.ModelAUC (95% CI)SensitivitySpecificityF1‐scoreTraditional methodCaprini score0.71 (0.66–0.76)0.630.750.58Unimodal modelsImaging‐only (ViT‐B/16)0.78 (0.73–0.83)0.720.810.67Imaging‐only (ResNet50)0.75 (0.70–0.80)0.680.790.63Clinical‐only (BERT)0.74 (0.69–0.79)0.660.790.61Multimodal fusion modelsFeature concatenation fusion0.83 (0.78–0.87)0.760.850.71Attention fusion (ours)0.89 (0.85–0.92)0.830.880.79 Model evaluation. The average weight coefficient α of the imaging modality was 0.63 ± 0.12, and that of the clinical text modality was 0.37 ± 0.09. Their distribution showed clear clinical regularity: when imaging detected a prosthesis‐bone interface lucent zone > 2 mm, the α value significantly increased to 0.82 ± 0.07 (OR = 3.2, 95% CI: 1.8–5.7, p = 0.01); when “D‐dimer > 1.0 mg/L” appeared in the text, the text weight increased to 0.58 ± 0.10 (OR = 4.1, 95% CI: 2.5–6.7). Analysis of typical cases further confirmed this mechanism: a 71‐year‐old female patient was correctly identified as high‐risk due to a 5°varus of the prosthesis (imaging weight α = 0.86) accompanied by a mild increase in D‐dimer (0.4 mg/L, text weight 0.14), and popliteal vein thrombosis was confirmed by ultrasound on the 5th day after surgery; another 68‐year‐old male, although with normal laboratory indicators (D‐dimer 0.3 mg/L), was also accurately warned because the lucent zone around the tibial prosthesis reached 3.2 mm (α = 0.89). These findings indicate data‐driven modality reliance patterns, but they should be interpreted as correlational rather than causal. Clinical Value Application Key imaging features were visualized for anatomical localization using Gradient‐Weighted Class Activation Mapping (Grad‐CAM) (Figure 2). The heatmaps showed that the model primarily focused on two anatomical regions: the posterior part of the femoral intercondylar fossa (popliteal vein projection area) and the tibial prosthesis–bone cement interface. In DVT‐positive cases, the average heat value in the popliteal vein area was 0.87 ± 0.05, which was significantly higher than that in the negative group (0.32 ± 0.11, p < 0.001). Figure 2A illustrates a typical high‐risk image: the anteroposterior radiograph shows a 4.2° varus of the femoral prosthesis, and the heatmap (red area) is concentrated in the popliteal vein projection area (within 10 mm posterior to the intercondylar fossa); the patient's D‐dimer level increased to 1.8 mg/L after surgery. Figure 2B presents the lucent zone around the tibial prosthesis (indicated by the arrow) with a heat value of 0.92, corresponding to the area with insufficient bone cement filling during the operation. These visualization results were highly consistent with the venous compression locations identified by ultrasound (κ = 0.81, p < 0.001), confirming that the model has learned clinically meaningful anatomical features. Representative Grad‐CAM heatmaps highlighting radiographic regions associated with model‐predicted postoperative DVT risk. Heatmaps are superimposed on postoperative radiographs to visualize the image regions that contributed most strongly to the model output. The color bar represents normalized Grad‐CAM intensity, with red indicating the highest contribution and blue indicating the lowest contribution. (A) Heatmap of a representative high‐risk case showing focal model attention in the posterior femoral intercondylar region, anatomically corresponding to the projection area of the popliteal vein. (B) Heatmap of a representative case showing concentrated activation around the tibial implant–bone interface, where radiolucent changes are present. These findings suggest that the model captures clinically plausible imaging patterns related to venous compression risk and prosthesis‐interface abnormalities. Discussion Development of a Multimodal Model for Early Postoperative DVT Prediction After TKA The multimodal deep learning model constructed in this study has increased the AUC for predicting DVT risk after TKA to 0.89 by dynamically fusing postoperative X‐ray images and clinical text data, significantly outperforming the traditional Caprini scoring system. The core of this breakthrough lies in the model's successful capture of the synergistic pathological mechanism between imaging features and coagulation indicators: when prosthesis varus > 3° coexists with D‐dimer > 0.5 mg/L, the DVT risk surges by 7.2 times (95% CI: 4.1–12.6), a phenomenon severely overlooked in previous unimodal studies. Notably, Grad‐CAM visualization technology revealed two key anatomical risk areas—the popliteal vein projection area (posterior to the femoral intercondylar fossa, with a heat value of 0.87 ± 0.05) and the tibial prosthesis‐bone cement interface (heat value of 0.85 ± 0.07). This not only explains the model's high sensitivity (83%) but also provides precise targets for intraoperative technical optimization: for example, reducing uneven bone cement filling can lower the incidence of interface lucent zones (OR = 3.2), and precise control of femoral prosthesis rotation angle can alleviate popliteal vein compression (Chen et al. confirmed that prosthesis varus > 3° reduces venous cross‐sectional area by 42%). Comparative Performance and Incremental Value Over Existing Approaches Compared with existing studies, the innovative value of this model is reflected in three aspects. First, the traditional view holds that X‐rays can only evaluate prosthesis position, but this study found that micromotion at the bone‐implant interface may activate the coagulation cascade. This provides imaging evidence for the mechano‐inflammation‐thrombosis hypothesis proposed [12, 13]. Second, unlike Gupta et al., who only used CNN to detect structural abnormalities, the gated attention mechanism introduced in this study realizes sample‐level modal weight allocation: in patients with D‐dimer > 1.0 mg/L, the text weight automatically increases to 0.58 ± 0.10; while for those with prosthesis position abnormalities, the imaging weight reaches 0.82 ± 0.07. This dynamic adjustment simulates the individualized decision‐making logic of senior physicians. Finally, the model successfully identified 91.7% (11/12 cases) of actual DVT patients classified as low‐risk by the Caprini score, all of whom had high‐risk imaging features (e.g., prosthesis varus > 3°), confirming that traditional scales underestimate the risk of patients with [14, 15, 16]. Clinical Interpretability, Limitations, and Future Directions Future research should focus on three directions: launching multicenter prospective trials to verify the model's actual impact on DVT incidence; accurately quantifying the spatial correlation between popliteal vein compression degree and prosthesis position using intraoperative OCT; and integrating postoperative early activity data monitored by wearable devices to construct a dynamic risk prediction model. It is worth emphasizing that the value of this model lies not only in the improvement of statistical indicators but also in revealing the pathological correlation between imaging markers and coagulation mechanisms—for example, each 1 mm increase in the tibial prosthesis lucent zone is associated with a 0.15 mg/L increase in the daily average D‐dimer level (R 2 = 0.63, p < 0.001), which provides an interpretable decision basis for personalized anticoagulation. With the continuous growth of joint replacement surgeries, such multimodal fusion technologies will promote the paradigm shift of postoperative risk management from experience‐driven to data‐driven. Importantly, the dynamic attention coefficient α should not be interpreted as evidence of causal mechanisms. While α varied systematically with certain imaging or clinical patterns, such associations may reflect correlations and dataset‐specific dependencies. Residual confounding remains possible, particularly from perioperative factors that influence both radiographic appearance and DVT risk, including operative duration, tourniquet time, intraoperative blood loss, and anticoagulant protocols. Limitations and Future Directions Several limitations should be acknowledged. First, this was a single‐center retrospective study, which may limit generalizability due to institutional practice patterns, imaging protocols, ultrasound thresholds, and anticoagulation strategies. However, the single‐center setting also ensured consistency in radiographic acquisition, ultrasound diagnosis, and EMR structure, which is advantageous for initial model development. Multicenter external validation and prospective evaluation are required before clinical deployment. Second, we relied on a single early postoperative radiograph (within 24 h). DVT risk is dynamic, and longitudinal radiographs (e.g., postoperative Day 1 vs. Day 3) and serial laboratory measurements may provide additional predictive information. Future work will explore longitudinal multimodal modeling (e.g., temporal transformers or recurrent fusion) that integrates multitimepoint imaging with time‐stamped laboratory trajectories and postoperative mobility/rehabilitation signals. Third, although metal artifact suppression was applied to improve visibility, any artifact reduction procedure can alter local image intensity distributions. We reduced this concern through standardized preprocessing and exclusion of low‐visibility images, but future studies should evaluate robustness across different artifact reduction pipelines and imaging devices. Conclusions This attention‐based model synergistically analyzes biomechanical (e.g., implant position) and biochemical (e.g., D‐dimer level) risk factors. This integrated approach significantly improves the accuracy of DVT prediction while simultaneously providing interpretable, anatomically relevant insights to guide personalized anticoagulation therapy strategies. Author Contributions Kejia Zhu: conceptualization, data curation, formal analysis, software, validation. Hang Li: formal analysis, investigation, methodology. Hui Zhang: data curation, investigation, methodology, project administration. Biao Wang: formal analysis, methodology, software. Bin Shen: writing – original draft, resources, supervision, writing – review and editing. Yong Nie: funding acquisition, project administration, resources, supervision, writing – review and editing. All the authors participated in the analysis of the results and provided critical input. Funding The study was funded by the National Key Research and Development Program (2023YFB4606700); the Sichuan Provincial International Science and Technology Innovation Cooperation (2025YFHZ0110); 1·3·5 Projects for Artificial Intelligence West China Hospital (ZYAI24038); 1·3·5 Project for Disciplines of Excellence–Clinical Research Fund, West China Hospital (2023HXFH024). Ethics Statement The investigation must conform to the principles outlined in the Declaration of Helsinki. Detailed information please see the original data. Consent They have approved the version to be published and have reached an agreement on the journal to which the article has been submitted. Additionally, the authors accept full responsibility for all aspects of the work and are willing to be held accountable. Conflicts of Interest The authors declare no conflicts of interest. Acknowledgments The authors have nothing to report. Data Availability Statement The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions. References Associated Data Data Availability Statement The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.