Cardiac Echocardiographic Analysis with Multi-Scale Effective Fusion Module: A Novel Stroke Prediction Approach

Cardiac Echocardiographic Analysis with Multi-Scale Effective Fusion Module: A Novel Stroke Prediction Approach | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Cardiac Echocardiographic Analysis with Multi-Scale Effective Fusion Module: A Novel Stroke Prediction Approach Jiachun xie, Dianhuan Tan, Tingting Zheng, Liya Wei This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-5641383/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 4 You are reading this latest preprint version Abstract Purpose: To develop a stroke risk prediction model by integrating echocardiographic images (long-axis, short-axis, four-chamber views) and clinical indicators using a novel multi-scale effective fusion (MSEF) module. Methods: A total of 712 patients with 10,992 images and 27 clinical indicators were included. The MSEF module enhances multi-scale feature fusion by combining deep semantic and shallow high-resolution features. It consists of four components: Global Feature Fusion (GFF), Multi-Feature Reconstruction (MFR), Channel Attention, and Positional Attention, effectively improving small-target feature representation. The fused features and clinical indicators were used to train the stroke prediction model. Results: The proposed MSEF-based model achieved the highest performance, with an Accuracy of 76.8% and F1 Score of 64.7% on the test set. Ablation studies confirmed the importance of Channel Attention and Position Attention in enhancing feature representation. When integrating echocardiographic features with clinical indicators, the model achieved an Accuracy of 80.2% and F1 Score of 72.1% on the test set. Conclusion: The proposed MSEF-based approach effectively integrates imaging and clinical data, improving stroke risk prediction accuracy and offering a promising tool for clinical decision-making. Stroke prediction Echocardiographic images Feature fusion Multi-Scale Effective Fusion Attention mechanisms Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 1. Introduction Hypertension is a major chronic disease worldwide and a significant public health concern, with its prevalence increasing annually. It ranks among the top global health threats, as recognized by the World Health Organization 1 . The true danger of hypertension lies not in the condition itself but in its complications, which result in high rates of disability and mortality. Among these, cerebrovascular damage is particularly severe, with stroke being the most devastating consequence. Stroke is the second leading cause of death globally, and studies have shown that approximately 62% of stroke-related deaths are directly attributed to hypertension. A large-scale epidemiological study conducted in 2022 further highlighted this risk, reporting a cumulative stroke incidence of 78.9% among hypertensive patients 2 , 3 . Given the profound impact of stroke on individuals and healthcare systems, its prevention and early intervention are paramount. Developing accurate and reliable methods to predict and assess stroke risk in hypertensive patients is essential for mitigating its devastating consequences. Stroke remains one of the leading causes of morbidity and mortality worldwide, necessitating the development of accurate and reliable methods for early risk prediction and prevention. Traditional risk assessment models often rely on clinical features such as age, hypertension, diabetes, and previous cardiovascular events. While these models provide valuable insights, they may not fully capture the complexities of individual patient risk profiles 4 , 5 . Previous predictive models based on clinical features have played a significant role in advancing stroke risk assessment, providing valuable tools for identifying individuals at high risk and guiding preventive strategies 6 – 8 . These models have effectively utilized factors such as age, blood pressure, history of atrial fibrillation, and diabetes to stratify stroke risk in various populations, contributing to better clinical decision-making 9 – 11 . However, despite their usefulness, these models have limitations, particularly in their ability to capture the intricate structural and functional details of the heart. They often fail to incorporate the rich information available from echocardiographic imaging, such as subtle changes in cardiac morphology, valve abnormalities, and wall motion dynamics, which can be crucial indicators of stroke risk. As a result, the predictive accuracy of these models may be compromised, especially in cases where imaging features play a pivotal role in stroke pathogenesis. This underscores the importance of integrating echocardiographic data into stroke risk prediction models to achieve a more comprehensive and accurate assessment. Echocardiography, a widely used non-invasive imaging technique, provides comprehensive and detailed visualization of cardiac structures and functions, enabling clinicians to assess heart anatomy, chamber sizes, wall thickness, valve functionality, and blood flow dynamics 12 , 13 . It plays a crucial role in diagnosing and monitoring various cardiovascular conditions, offering real-time insights into cardiac performance and assisting in the evaluation of heart failure, valve diseases, congenital heart abnormalities, and stroke risk 14 . By accurately assessing factors such as left atrial size and function, left ventricular ejection fraction, valve abnormalities, intracardiac thrombus formation, aortic atherosclerotic plaques, and potential patent foramen ovale (PFO), echocardiography offers valuable information for predicting stroke risk. Its ability to deliver dynamic and high-resolution images makes echocardiography an indispensable tool in both routine clinical practice and advanced cardiac research, aiding in the prevention and management of stroke and other cardiovascular conditions 15 . One reason for the underutilization is the complexity and variability of echocardiographic data, which often require expert interpretation to extract meaningful insights. However, advancements in machine learning and deep learning techniques have opened up opportunities to harness the predictive power of these images more effectively. By integrating echocardiographic data from long-axis, short-axis, and four-chamber views with hypertension-related clinical features, such as patient history and risk factors, it is possible to develop more accurate models for stroke risk prediction. These models can analyze subtle changes in cardiac structure and function that may not be immediately apparent to the human eye, offering a more comprehensive assessment of stroke risk. This integration has the potential to improve early detection, guide preventive interventions, and enhance personalized treatment strategies, ultimately reducing the burden of stroke in at-risk populations 16 – 18 . Building on this success, deep learning models have the potential to revolutionize stroke risk prediction by analyzing echocardiographic images in ways that go beyond traditional clinical assessment 19 . These models can identify intricate patterns related to cardiac morphology, wall motion abnormalities, and hemodynamic changes that are indicative of stroke risk but might be overlooked in routine examinations 20 . By combining echocardiographic data with other clinical parameters, deep learning approaches can create a more comprehensive risk stratification model, allowing for earlier and more accurate identification of individuals at high risk for stroke 21 . This integration of deep learning into echocardiography could not only improve diagnostic accuracy but also facilitate personalized treatment plans, leading to better preventive strategies and outcomes for patients with cardiovascular disease. As the field continues to evolve, the application of deep learning in echocardiographic image analysis holds promise for advancing stroke prevention and enhancing our understanding of the complex relationship between cardiac function and cerebrovascular events. Integrating echocardiographic images, including long-axis, short-axis, and four-chamber views, poses unique challenges due to the heterogeneous nature and varying scales of the data. Each view captures distinct yet complementary cardiac information: long-axis views provide insights into left ventricular function and wall motion, short-axis views offer cross-sectional perspectives on chamber morphology and structural abnormalities, while four-chamber views visualize atria and ventricles, aiding in the detection of valve dysfunction, atrial enlargement, and thrombus formation. However, traditional feature fusion techniques, such as Feature Pyramid Networks (FPN), often fail to effectively integrate these multi-view features, particularly when addressing subtle structural abnormalities or small regions of interest critical to stroke risk prediction. To overcome these limitations, we introduce advanced fusion mechanisms, including the enhanced inter-layer feature correlation (EFC) and MSFC modules, which significantly enhance the representation and integration of multi-scale features. The EFC module focuses on spatial and channel-wise correlations through its GFF submodule, which generates spatial weights to emphasize key regions and enhance feature interactions, while the MFR submodule separates and reconstructs strong and weak features, preserving fine-grained details without interference. Meanwhile, the MSFC module improves the correlation between deep semantic features and shallow high-resolution features, enabling better multi-scale feature representation. By combining these modules, the proposed fusion strategy effectively captures subtle cardiac abnormalities and improves the overall predictive performance for stroke risk, particularly in cases where imaging features play a pivotal role in stroke pathogenesis. In this study, we propose a novel approach to stroke risk prediction by integrating deep learning-based analysis of echocardiographic images with traditional clinical risk factors. Our method consists of three main components: (1) developing a deep learning model to analyze multi-view echocardiographic images (long-axis, short-axis, and four-chamber views) to enhance feature extraction and representation; (2) introducing an advanced feature fusion strategy to integrate multi-scale echocardiographic features, improving the model's ability to capture subtle cardiac abnormalities related to stroke risk; and (3) combining the fused imaging features with hypertension-related clinical indicators to construct a hybrid predictive model, providing a comprehensive assessment of stroke risk. We hypothesize that the integration of echocardiographic imaging data and clinical features will enhance the accuracy and robustness of stroke risk prediction. By leveraging the strengths of both deep learning and traditional clinical assessment, our approach aims to provide a comprehensive tool for early detection and intervention, ultimately improving patient outcomes. To address the challenges associated with integrating multi-view echocardiographic data, including long-axis, short-axis, and four-chamber views, we improved traditional feature fusion strategies through the introduction of the Multi-Scale Feature Correlation (MSFC) module. While previous studies have applied the EFC module for multi-scale feature extraction, its limitations in fully capturing the spatial and semantic correlations across echocardiographic views prompted the need for further refinement 22 . The MSFC module improves upon EFC by enhancing the correlation between deep semantic features and shallow high-resolution features, allowing for better representation of multi-scale information. Specifically, the MSFC module refines spatial attention to emphasize subtle cardiac abnormalities, such as valve dysfunction, atrial enlargement, and wall motion irregularities, which are critical indicators of stroke risk. It further optimizes the integration of strong and weak features to preserve fine-grained details while mitigating feature interference, thereby improving the module’s robustness in detecting small or subtle targets within echocardiographic images. The following sections describe the methodology for developing and validating the deep learning and clinical models, the process of integrating these models into a hybrid framework, and the evaluation of their combined predictive performance. By incorporating the MSFC module and effectively integrating imaging and clinical data, our approach provides a more accurate and comprehensive assessment of stroke risk, paving the way for enhanced prevention and management strategies. 2. Materials and Methods This study was approved by the institutional review board, and the requirement for informed patient consent was waived due to its retrospective cohort design. 2.1 Data sources and preprocessing This retrospective study utilized echocardiographic examination data and relevant clinical information of confirmed hypertensive patients extracted from the electronic medical record database of Peking University Shenzhen Hospital, covering the period from September 2022 to December 2023. The study included patients with comprehensive clinical and imaging data, with stroke diagnoses confirmed through neuroimaging. A total of 712 patient records were analyzed, comprising 10,992 echocardiographic images from the left ventricular long-axis, short-axis, and apical four-chamber views, as well as 27 associated clinical features. Patients were classified into positive or negative outcome groups based on whether they experienced a stroke during the course of hypertension. The inclusion criteria were: (1) Age over 18 years; (2) Completion of echocardiographic examination with comprehensive imaging data, including long-axis, short-axis, and apical four-chamber views; (3) A confirmed diagnosis of hypertension documented in the medical history, current medical history, or disease course records; (4) Availability of complete clinical data (6) Patients admitted to the hospital between September 2022 and December 2023. The exclusion criteria were: (1) Incomplete clinical information or missing echocardiographic images; (2) Poor-quality echocardiographic images that fail to meet diagnostic standards (e.g., unclear visualization of cardiac structures); (3) Patients with severe comorbidities that could confound stroke risk assessment, such as late-stage cancer, end-stage renal disease, or severe heart failure; (4) Patients with no documented neuroimaging confirmation of stroke diagnosis. As shown in Fig. 1 , echocardiographic images from the long-axis, short-axis, and four-chamber views of 712 patients with confirmed diagnoses were processed to produce nine-channel TIFF files. Each file was created by stacking the three echocardiographic views, with each view contributing three channels. For each patient, the number of data units generated ranged from 3 to 9. If a patient had 10 or more images, the highest-quality images were selected to form the final data units. Conversely, patients with 3 or fewer images were considered to have poor-quality echocardiographic images and were excluded from the analysis. Ultimately, 3664 nine-channel TIFF files were integrated with the patients' corresponding clinical information, including 27 hypertension-related features, to form 3664 data units. Each data unit consists of a combined echocardiographic image and its associated clinical features, providing a robust basis for subsequent model development and analysis. Although the data units contain repeated clinical information, the data from the same patient were assigned to only one of the test, train, or validation sets to prevent data leakage and reduce the risk of overfitting. Regarding the 27 hypertension-related clinical features, a small portion of these features had missing values (< 10%), which were imputed using mean values to maintain data integrity. For clinical features with more than 10% missing values, they were excluded during the feature collection phase to ensure the robustness and reliability of the dataset. This preprocessing strategy ensures that the echocardiographic images and clinical features provide a clean and consistent input for subsequent model training and evaluation. Figure 2 presents representative examples of the selected ultrasound images, including the four-chamber, long-axis, and short-axis echocardiographic views. The image selection was performed manually by two experienced radiologists, each with at least two years of experience in echocardiographic imaging. 2.2 Clinical variables A total of 27 stroke-related risk factors, listed in Table 1 , were included in the study. These factors are clinically relevant indicators that can be easily assessed in practice. All risk factors were transformed into variables for model development. Table 1 The risk factors with a definition in this study Risk factors Definition Age Age of the patients, ranging from 40 years to the maximum observed age. Gender Female/Male Systolic pressure Systolic blood pressure of the patient, measured in mmHg Diastolic pressure Diastolic blood pressure of the patient, measured in mmHg Medical history Encoded information about the patient's medical history (categorical variable) Height Height of the patient, measured in centimeters (cm) Weight Weight of the patient, measured in kilograms (kg) BMI Body Mass Index of the patient, calculated as weight (kg) divided by height (m) squared Drink Drinking status of the patient, encoded as 0 (non-drinker) or 1 (drinker) Smoke Smoking status of the patient, encoded as 0 (non-smoker) or 1 (smoker) Family history Presence of family history of stroke, encoded as 0 (no) or 1 (yes) Blood fat Blood fat level, encoded (specific encoding not provided) Blood sugar Blood sugar level, encoded (specific encoding not provided). Use drugs Use of antihypertensive drugs, encoded as 0 (no) or 1 (yes). COVID19 COVID−19 infection status, encoded as 0 (negative) or 1 (positive). Function change Change in function, encoded (specific encoding not provided). Structural change Structural changes in the heart, encoded (specific encoding not provided) RWT Relative wall thickness, a measure used in echocardiography. LVMI Left ventricular mass index, a measure used in echocardiography AAO Ascending aorta diameter, a measure used in echocardiography LA Left atrium diameter, a measure used in echocardiography LV Left ventricle diameter, a measure used in echocardiography IVSD Interventricular septum diameter, a measure used in echocardiography LVPW Left ventricular posterior wall thickness, a measure used in echocardiography EF Ejection fraction, a measure of the heart's pumping efficiency, expressed as a percentage Septal_E Early diastolic velocity of the septal mitral annulus, a measure used in echocardiography Lateral wall Early diastolic velocity of the lateral mitral annulus, a measure used in echocardiography Risk factors with a P-value < 0.05 in the Chi-square test were considered statistically significant, listed in Table 2 . A total of 13 variables were identified as significant. To determine the optimal variables for constructing the prediction model, a multivariable logistic regression analysis was performed, and the results were expressed in terms of P-values. The area under the curve (AUC) was utilized to evaluate the performance and predictive accuracy of the model. Seven variables (Systolic Pressure, Medical History, Age, Height, Weight, RWT, AAO) showed a statistically significant difference (P < 0.05) in the multivariable logistic regression analysis. The results of the multivariable logistic regression analysis are displayed as forest plots in Fig. 3 . Table 2 Baseline characteristics of the total cohort Risk Factor Stroke Mean Non-Stroke Mean P-Value Systolic pressure 174.48 164.71 0.01 Diastolic pressure 97.41 97.53 0.938 Medical history 12.03 7.63 0.01 Age 66.78 56.89 0.001 Height 163.24 164.84 0.104 Weight 67.1 69.44 0.128 BMI 25.16 25.45 0.535 Gender 0.52 0.55 0.585 Drink 0.15 0.08 0.051 Smoke 0.3 0.24 0.191 Family history 0.49 0.59 0.065 Blood fat 0.7 0.59 0.045 Blood sugar 0.42 0.3 0.023 Use antihypertensive drugs 0.79 0.68 0.025 COVID19 positive 0.48 0.52 0.469 Function change 0.01 0.01 0.795 Structural change 0.64 0.4 0.001 RWT 0.11 0.04 0.004 LVMI 0.3 0.19 0.013 AAO 0.51 0.28 0.001 LA 0.22 0.33 0.495 LV 0.07 0.04 0.391 IVSD 0.26 0.15 0.01 LVPW 0.03 0.01 0.331 EF 0.01 0.01 0.734 Septal E 0.46 0.31 0.005 Lateral wall 0.58 0.42 0.005 2.3 FPN and EFC modules Feature Pyramid Networks (FPN) and the EFC module play a critical role in improving feature extraction and fusion for ultrasound imaging. The FPN is designed to effectively integrate multi-scale features, ensuring that both low-level spatial details and high-level semantic information are preserved. However, traditional feature fusion strategies within FPNs often suffer from weak correlations between layers and redundant features, which can hinder the accurate detection of small and complex lesions. To address these limitations, the EFC module was introduced, as illustrated in Fig. 4 . The EFC module consists of two key components: the Grouped Feature Focus (GFF) and Feature Reconstruction (FR). The GFF selectively emphasizes critical feature groups, enhancing the representation of important structures. The FR further refines the fused features, recovering lost information and improving small object detection. Compared to the traditional fusion approach, the EFC module demonstrates superior performance by reducing redundancy and strengthening feature correlations, leading to more accurate lesion detection and classification. 2.4 MSEF module To further improve upon the EFC module and address the limitations of traditional multi-scale feature fusion strategies, we propose the MSEF module. While EFC effectively enhances feature correlation and reduces redundancy, its performance in handling small targets under complex and dense backgrounds can still be optimized. MSEF builds on EFC by introducing additional mechanisms to strengthen the coordination between deep semantic features and shallow high-resolution features, resulting in more effective multi-scale fusion and improved small target representation. As shown in Fig. 5 , MSEF integrates advanced feature processing techniques to enhance feature correlation, minimize redundancy, and maintain computational efficiency. It incorporates four key submodules: 1. Grouped Feature Focus (GFF): This submodule captures contextual information to strengthen the representation of small targets while generating spatial weights to focus on critical regions, ensuring small targets are not overlooked in complex backgrounds. By dividing feature maps into groups based on channels, it enables localized interaction and enhances channel-wise feature correlation, improving feature alignment and representation. 2. Multi-Layer Feature Reconstruction (MFR): MFR separates strong and weak features to avoid interference, ensuring that weak features are independently optimized while preserving the integrity of strong features. It applies fine-grained operations like 1×1 convolutions to refine strong features, while using lightweight depth-wise separable convolutions to extract richer information from weak features. The reconstructed features are fused layer by layer, preserving small target details and semantic information for better representation. 3. Channel Attention Submodule: This submodule applies global average pooling to capture global context and uses a 1D convolution to model channel interactions, enhancing the relevance of feature channels. Unlike traditional methods, it avoids dimension reduction, ensuring that critical channel features are preserved, thereby improving overall feature correlation. The mechanisms involved are as follows: $$\:s=W\ast\:z$$ 1 The above equation represents the operation where the weight \(\:W\) acts on the input vector \(\:z\) through a convolution, resulting in a new channel attention representation \(\:s\) . $$\:{z}_{c}=\frac{1}{H\cdot\:W}{\sum\:}_{i=1}^{H}{\sum\:}_{j=1}^{W}{X}_{c}\left(i,j\right),c=\text{1,2},,C$$ 2 The above equation describes the global average pooling operation applied to each channel \(\:\text{c}\) of the input feature map \(\:\text{X}\) . This computes the global contextual information \(\:{\text{z}}_{\text{c}}\) for each channel by averaging all spatial locations. 4. Position Attention Submodule: By applying attention along the horizontal and vertical axes of the feature maps, this submodule extracts spatial structure information through pooling operations. It enables precise focus on key spatial regions, enhancing the model's ability to locate and represent small targets accurately. The mechanisms involved are as follows: $$\:{P}_{ℎ}\left(i,c\right)=\frac{1}{W}{\sum\:}_{j=1}^{W}{X}_{c}\left(i,j\right)$$ 3 This formula represents the global average pooling operation along the horizontal axis (width \(\:\text{W}\) ) for the \(\:\text{c}\) -th channel of the input feature map \(\:\text{X}\) . It aggregates spatial information across the horizontal dimension at height position \(\:\text{i}\) . $$\:{\widehat{X}}_{c}\left(i,j\right)={\alpha\:}_{ℎ}\left(i,c\right)\cdot\:{X}_{c}\left(i,j\right)+{\alpha\:}_{v}\left(j,c\right)\cdot\:{X}_{c}\left(i,j\right)$$ 4 This formula represents the position attention mechanism for enhancing feature maps. Specifically, it combines the horizontal attention weight \(\:{\alpha\:}_{ℎ}\left(i,c\right)\) and the vertical attention weight \(\:{{\alpha\:}}_{\text{v}}\left(\text{j},\text{c}\right)\) , which are applied to the \(\:\text{c}\) -th channel of the input feature map \(\:{\text{X}}_{\text{c}}\left(\text{i},\text{j}\right)\) . The resulting weighted feature map \(\:{\widehat{\text{X}}}_{\text{c}}\left(\text{i},\text{j}\right)\) is obtained by aggregating spatial information along both the horizontal and vertical axes, ensuring that important spatial relationships are captured across the input feature space. The MSEF module offers several advantages, making it a robust and efficient solution for multi-scale feature fusion. It enhances feature correlation by improving the spatial and semantic alignment of multi-scale features, which significantly strengthens the representation of small targets. The MFR submodule further reduces redundancy through precise feature reconstruction, ensuring that valuable information is retained while minimizing unnecessary fusion. Additionally, its lightweight design optimizes computational efficiency, effectively reducing both parameters and FLOPs without compromising detection accuracy. Moreover, the MSEF module's plug-and-play compatibility allows seamless integration into various backbone networks, highlighting its versatility and applicability across a wide range of computer vision tasks. In summary, the MSEF module effectively overcomes the limitations of traditional feature fusion strategies by providing a robust and efficient approach for multi-scale feature integration. Through grouped feature focus, feature reconstruction, and attention mechanisms, it enhances small target representation and delivers superior performance in complex ultrasound images. This makes the MSEF module particularly well-suited for tasks requiring high precision, such as small target detection, segmentation, and classification in cardiac ultrasound imaging. As shown in Fig. 6 , the MSEF module plays a crucial role in integrating multi-view echocardiographic features for the prediction of stroke risk. By fusing features extracted from long-axis, four-chamber, and short-axis views, MSEF ensures that spatial and semantic information from multiple perspectives is effectively combined. To further enhance the quality of feature integration, the long-axis and short-axis features are first fused individually with the four-chamber view features, which serves as a baseline due to its comprehensive representation of the heart's overall structure. This stepwise fusion approach effectively leverages the complementary nature of different echocardiographic views, balancing global structural information with detailed local features while reducing feature redundancy and conflicts. Such a strategy also ensures that nuanced patterns, particularly small targets and subtle abnormalities, are accurately captured. The multi-scale feature integration enabled by MSEF allows the model to better capture subtle variations critical for stroke risk prediction, as these variations are often linked to cardiac structure and function abnormalities. The attention mechanisms within MSEF further emphasize key regions of the images, enabling the model to focus on clinically relevant features while minimizing noise. Overall, the application of MSEF significantly improves the robustness and accuracy of echocardiographic-based stroke risk prediction, addressing the challenges posed by complex and heterogeneous ultrasound data. 2.5 Experimental Setup The code for both EFC and MSEF modules is provided in Supplementary_code.py, and all required Python package versions are listed in Supplementary_version_list.txt. The experiments were conducted using Python 3.11 on an Ubuntu 18.04 operating system. For network implementation, either PyTorch or the MMPretrain framework can be used. It is necessary to modify the transform functions to accommodate TIFF image inputs, as it have specific channel characteristics. The mean and standard deviation (std) of ultrasound image channels typically hover around 40, but it is recommended to adjust these values based on the specific dataset being used. The training process was conducted with a batch size of 32, a learning rate of 0.0001, and the SGD (Stochastic Gradient Descent) optimizer. Manual tuning was performed for the learning rate and batch size, initially starting with a learning rate of 0.001 and progressively reducing it to 0.0001 based on validation performance. The model was trained on a single NVIDIA A100 GPU with 80GB memory, with each epoch taking approximately 1.2 minutes, resulting in a total training time of 2 hours over 100 epochs. For devices with limited GPU memory, we recommend proportionally reducing the batch size and learning rate while maintaining their ratio (e.g., halving the batch size to 16 and lowering the learning rate to 0.00005). Alternatively, simplifying the network architecture can help alleviate memory constraints during training. 3. Results 3.1 Comparison of Fusion Module Performance To objectively evaluate the effectiveness of our proposed method, we utilized a set of quantitative metrics, including Accuracy, Precision, Recall, and F1 Score, to ensure a comprehensive performance assessment. Table 3 presents a summary of the results for three fusion methods—FPN, EFC and MSEF —evaluated on both the validation set and the test set. These metrics provide an in-depth comparison of each method's performance. The results clearly demonstrate that the MSEF method is the most robust across both datasets. It consistently achieves the highest Accuracy and F1 Score, showcasing its ability to effectively balance Precision and Recall. While the EFC method achieves strong Precision, its lower Recall reduces its overall performance. In contrast, FPN exhibits moderate performance but struggles to remain competitive, particularly on the test set. These findings indicate that the multi-scale fusion approach employed by MSEF significantly enhances the model's generalization capability, resulting in superior performance on unseen data. Table 3. The performance metrics of three methods evaluated on the validation set and the test set. Method Data Split Accuracy Precision Recall F1 Score FPN Validation Set 74.4% 72.4% 66.7% 67.8% EFC Validation Set 73.7% 70.5% 68.2% 69.0% MSEF Validation Set 74.6% 71.9% 68.1% 69.1% FPN Test Set 70.5% 66.6% 63.0% 63.6% EFC Test Set 72.2% 69.7% 63.3% 63.9% MSEF Test Set 76.8% 66.2% 63.3% 64.7% 3.2 Ablation Study of the MSEF Module To gain a deeper understanding of the contribution of each component within the MSEF module, we conducted a systematic ablation study. Specifically, we removed the Channel Attention and Position Attention mechanisms independently to assess their respective roles in the module's overall performance. The results of these experiments, summarized in Table 4, compare three model variants: MSEF without Channel Attention, MSEF without Position Attention, and the Full MSEF Module, evaluated on both the validation set and the test set. Ablative experiments are crucial for evaluating the effectiveness of individual components within a model, as they provide insights into how specific mechanisms contribute to the overall functionality. The comparative results clearly demonstrate that both Channel Attention and Position Attention play significant roles in enhancing the performance of the MSEF module. The removal of Channel Attention results in a noticeable drop in Recall on both datasets, with the validation set Recall decreasing to 63.2% and the test set Recall dropping to 69.2%. This reduction indicates that Channel Attention is particularly effective in capturing inter-channel relationships and enhancing feature representations by focusing on the most informative channels. By selectively prioritizing critical channels, this mechanism ensures that the network retains relevant information and suppresses noise, which is particularly important for tasks requiring high sensitivity, as reflected in the Recall metric. On the other hand, removing Position Attention produces slightly higher Accuracy (77.5% on the validation set) but lowers Precision and F1 Score. The Precision drops to 67.6% on the validation set, highlighting that Position Attention is essential for balancing overall performance. Position Attention focuses on the spatial relationships within the feature maps, ensuring the model captures spatial context effectively. This mechanism is particularly important for preserving structural information, enabling the network to achieve a better balance between Precision and Recall, as evidenced by the higher F1 Score in the Full MSEF Module. The Full MSEF Module, which integrates both Channel Attention and Position Attention, achieves the best overall performance across all metrics and datasets. On the validation set, it achieves a Recall of 68.1% and an F1 Score of 69.1%, while on the test set, it attains an Accuracy of 76.8% and an F1 Score of 64.7%. This demonstrates that the synergy between the two attention mechanisms enables the model to extract both channel-wise and spatial features effectively, enhancing its ability to generalize to unseen data. The ablation study underscores the critical importance of the MSEF module's structure, where Channel Attention and Position Attention work collaboratively to address complementary aspects of feature representation. Channel Attention ensures that the model focuses on the most relevant feature channels, improving sensitivity to key patterns, while Position Attention captures spatial dependencies, ensuring that the spatial structure of the data is preserved. By combining these two mechanisms, the Full MSEF Module achieves a superior trade-off between Precision and Recall, leading to robust and reliable performance across diverse datasets. In sum, the ablation study validates the design choices within the MSEF module, demonstrating that the integration of Channel Attention and Position Attention is essential for maximizing the model's overall effectiveness and generalization capability. Table 4. The performance metrics of three model variants evaluated on the validation set and the test set. Model Variant Data Split Accuracy Precision Recall F1 Score MSEF without Channel Attention Validation Set 72.8% 71.2% 63.2% 63.8% MSEF without Position Attention Validation Set 77.5% 67.6% 64.7% 66.1% Full MSEF Module Validation Set 74.6% 71.9% 68.1% 69.1% MSEF without Channel Attention Test Set 71.0% 68.1% 69.2% 68.5% MSEF without Position Attention Test Set 73.5% 70.1% 67.8% 68.6% Full MSEF Module Test Set 76.8% 66.2% 63.3% 64.7% 3.3 Stroke Risk Prediction with Multimodal Integration of Clinical and Imaging Data Accurate stroke risk prediction relies on effectively integrating clinical data and imaging information, two critical sources of patient information. While imaging data provides detailed anatomical and functional insights, traditional clinical features such as age, blood pressure, cholesterol levels, and comorbidities remain cornerstone predictors for stroke risk, as supported by extensive clinical studies. Leveraging these well-established clinical indicators alongside advanced imaging features ensures a robust and clinically interpretable prediction model. Table 5 presents the performance metrics of three multimodal prediction models—Logistic Regression, Random Forest, and XGBoost—evaluated on both the validation set and the test set. These models integrate both imaging features extracted from the MSEF module and traditional clinical data to predict stroke risk. Table 5. The performance metrics of three multimodal prediction model. Method Data Split Accuracy Precision Recall F1 Score Logistic Regression Validation Set 80.5% 74.0% 71.4% 72.7% Random Forest Validation Set 80.7% 74.4% 71.8% 73.0% XGBoost Validation Set 81.1% 74.8% 72.3% 73.6% Logistic Regression Test Set 79.9% 72.8% 70.1% 71.4% Random Forest Test Set 80.1% 73.1% 70.5% 71.8% XGBoost Test Set 80.2% 73.4% 70.8% 72.1% The integration of traditional clinical features with imaging-derived data is rooted in established clinical practice and research. Clinical features such as age, hypertension, diabetes, and lifestyle factors have been extensively validated as independent stroke predictors in prior studies. Imaging features, particularly those extracted through advanced fusion methods like MSEF, capture subtle and spatially distributed changes that are often invisible to manual inspection. By combining these two modalities, the models benefit from: (1) Enhanced Predictive Power: Clinical features provide a solid baseline prediction, while imaging features refine the model's ability to capture complex patterns indicative of stroke risk. (2) Clinical Interpretability: Traditional clinical predictors offer transparency and trust, enabling clinicians to validate and interpret the model's output more easily. (3) Generalizability: The inclusion of clinical data stabilizes the model's performance across datasets, reducing overfitting to imaging-specific noise. In sum, the results demonstrate that integrating clinical features with imaging-derived data significantly enhances stroke risk prediction performance. Among the evaluated models, XGBoost consistently achieves the best results, underscoring its ability to handle multimodal data effectively. This multimodal integration approach aligns with prior research and clinical practice, highlighting the importance of leveraging well-established clinical predictors alongside advanced imaging features to achieve reliable, interpretable, and clinically applicable stroke prediction models. 4. Discussion In this study, we proposed a novel stroke risk prediction framework that integrates clinical indicators with echocardiographic imaging features using a Multi-Scale Effective Fusion module. By combining traditional clinical factors and advanced imaging analysis, we achieved enhanced performance and demonstrated the importance of multimodal data integration for stroke risk assessment. 4.1 Significance of Integrating Clinical and Imaging Data Traditional stroke risk models primarily rely on clinical indicators, such as blood pressure, age, and medical history, which have long been validated as strong predictors in clinical practice 23 – 25 . However, these models often fall short in accounting for subtle cardiac morphological and functional changes detectable through imaging. Echocardiographic data, including long-axis, short-axis, and four-chamber views, offer critical insights into cardiac structure and function, such as valve abnormalities, wall motion dynamics, and atrial enlargement, which are pivotal for stroke pathogenesis. Integrating these imaging-derived features with clinical data bridges the gap between visual assessments and quantitative predictions, leading to a more comprehensive and interpretable risk prediction model. As shown in Section 3.3 , the multimodal integration approach demonstrated superior performance compared to single-modality methods. This highlights the advantages of advanced machine learning algorithms in handling complex, heterogeneous data, where clinical and imaging features complement one another to refine predictions. 4.2 The Role of the MSEF Module The MSEF module, introduced in this study, played a critical role in improving feature extraction from echocardiographic images. Previous studies have demonstrated that multi-scale feature fusion enhances representation learning, especially for subtle or small abnormalities. Our ablation study in Section 3.2 revealed that both Channel Attention and Position Attention mechanisms significantly contribute to feature enhancement. Channel Attention prioritizes critical channels, improving the detection of informative patterns, while Position Attention captures spatial relationships, enabling the model to preserve structural details. By combining these mechanisms, the MSEF module effectively addresses the challenges posed by heterogeneous echocardiographic data, ensuring robust and accurate feature integration. The superior performance of the Full MSEF Module compared to its ablated variants underscores the necessity of a multi-dimensional attention strategy. Removing either Channel or Position Attention led to performance degradation, confirming that the combination of both mechanisms achieves optimal results for stroke risk prediction. 4.3 Clinical Implications Our findings align with existing clinical knowledge, emphasizing the continued relevance of traditional clinical features while leveraging advancements in imaging and machine learning. Clinical predictors such as age, blood pressure, and medical history remain highly interpretable and trusted by clinicians. Integrating these indicators with echocardiographic features enhances predictive power while ensuring the model’s clinical applicability and transparency. This multimodal framework allows for the early identification of stroke risk, enabling timely preventive interventions and personalized treatment strategies for at-risk populations. Furthermore, the success of the MSEF-based approach highlights the potential of deep learning models in clinical settings. The ability to analyze echocardiographic data at a granular level enables the detection of subtle cardiac abnormalities, such as early atrial enlargement or wall motion irregularities, which may be overlooked during manual interpretation. Combined with clinical factors, these insights provide a more holistic understanding of stroke risk. 4.4 Limitations and Future Work Despite its promising results, this study has some limitations. First, the retrospective nature of the dataset may introduce selection bias, and future studies should validate the proposed framework using prospective data. Second, while the MSEF module effectively enhances feature extraction, further improvements in model efficiency could enable real-time clinical applications. Additionally, incorporating other imaging modalities, such as cardiac MRI, may provide additional insights and further improve predictive accuracy. Future work will focus on expanding the dataset to include diverse patient populations and exploring other fusion strategies for multimodal data integration. 4.5 Conclusion In conclusion, this study demonstrates the importance of integrating clinical indicators with advanced imaging-derived features for stroke risk prediction. The proposed MSEF module effectively captures multi-scale echocardiographic information, improving predictive accuracy and robustness. By leveraging the strengths of both clinical data and machine learning-based imaging analysis, our approach provides a clinically interpretable and highly accurate tool for early stroke risk assessment, paving the way for improved prevention and personalized interventions. Declarations Data availability The data used in this study contain sensitive patient information and are therefore not publicly available to protect patient privacy. Funding This work was supported by the Science and Technology Project of Shenzhen (JCYJ20230807095209018, JCYJ20210324110211031, JCYJ20210324131402008, KXCFZ202002011010487), National Key Research and Development Program of China (2023YFC3402605), the Natural Science Foundation of Guangdong Province (2022A1515010986, 2022A1515010296), the Shenzhen Key Medical Discipline Construction Fund (SZXK051), and the Sanming Project of Medicine in Shenzhen (SZSM202111011), Peking University Shenzhen Hospital(LCYJZD2021010). Human Ethics and Consent to Participate 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. Ethics Approval The study was approved by the Medical Ethics Committee of Peking University Shenzhen Hospital (Approval Number: Peking University Shenzhen Hospital Ethics Committee (Research) [2024] No. 156), and the requirement of individual consent for this retrospective analysis was waived. Consent to Participate declaration This study was approved by the institutional review board, and the requirement for informed patient consent was waived due to its retrospective cohort design. Clinical trial number not applicable Author Contribution J.X and D.T contributed equally to this work and were primarily responsible for data collection and statistical analysis. T.Z and L.W jointly conceived the study, designed the methodology, and secured the necessary funding to support this research. All authors contributed to the preparation and revision of the manuscript and approved the final version for submission. References Pei Z, Liu J, Liu M, et al. Risk-Predicting Model for Incident of Essential Hypertension Based on Environmental and Genetic Factors with Support Vector Machine. Interdiscip Sci . 2018;10(1):126-130. doi:10.1007/s12539-017-0271-2 Kim H, Hwang S, Lee S, Kim Y. Classification and Prediction on Hypertension with Blood Pressure Determinants in a Deep Learning Algorithm. Int J Environ Res Public Health . 2022;19(22)doi:10.3390/ijerph192215301 Li A-l, Ji Y, Zhu S, et al. Risk probability and influencing factors of stroke in followed-up hypertension patients. BMC Cardiovasc Disord . 2022;22(1):328. doi:10.1186/s12872-022-02780-w Arafa A, Kokubo Y, Sheerah HA, et al. Developing a Stroke Risk Prediction Model Using Cardiovascular Risk Factors: The Suita Study. Cerebrovasc Dis . 2022;51(3):323-330. doi:10.1159/000520100 Vu T, Kokubo Y, Inoue M, et al. Machine Learning Approaches for Stroke Risk Prediction: Findings from the Suita Study. J Cardiovasc Dev Dis . 2024;11(7)doi:10.3390/jcdd11070207 Heo J, Yoon JG, Park H, Kim YD, Nam HS, Heo JH. Machine Learning-Based Model for Prediction of Outcomes in Acute Stroke. Stroke . 2019;50(5):1263-1265. doi:10.1161/STROKEAHA.118.024293 Nishi H, Oishi N, Ishii A, et al. Predicting Clinical Outcomes of Large Vessel Occlusion Before Mechanical Thrombectomy Using Machine Learning. Stroke . 2019;50(9):2379-2388. doi:10.1161/STROKEAHA.119.025411 Hendrix P, Sofoluke N, Adams MD, et al. Risk Factors for Acute Ischemic Stroke Caused by Anterior Large Vessel Occlusion. Stroke . 2019;50(5):1074-1080. doi:10.1161/STROKEAHA.118.023917 Miyazaki Y, Kawakami M, Kondo K, et al. Logistic regression analysis and machine learning for predicting post-stroke gait independence: a retrospective study. Sci Rep . 2024;14(1):21273. doi:10.1038/s41598-024-72206-4 van der Vliet R, Selles RW, Andrinopoulou E-R, et al. Predicting Upper Limb Motor Impairment Recovery after Stroke: A Mixture Model. Ann Neurol . 2020;87(3):383-393. doi:10.1002/ana.25679 Islam MS, Hussain I, Rahman MM, Park SJ, Hossain MA. Explainable Artificial Intelligence Model for Stroke Prediction Using EEG Signal. Sensors (Basel) . 2022;22(24)doi:10.3390/s22249859 Saha SK, Kiotsekoglou A. Value of speckle tracking echocardiography for prediction of stroke risk in atrial fibrillation: Time to spare a stare outside the box? Echocardiography . 2018;35(5):589-591. doi:10.1111/echo.14005 Tufano A, Galderisi M. Can echocardiography improve the prediction of thromboembolic risk in atrial fibrillation? Evidences and perspectives. Intern Emerg Med . 2020;15(6):935-943. doi:10.1007/s11739-020-02303-5 Chamsi-Pasha MA, Sengupta PP, Zoghbi WA. Handheld Echocardiography: Current State and Future Perspectives. Circulation . 2017;136(22):2178-2188. doi:10.1161/CIRCULATIONAHA.117.026622 Gillam LD, Marcoff L. Echocardiography: Past, Present, and Future. Circ Cardiovasc Imaging . 2024;17(4):e016517. doi:10.1161/CIRCIMAGING.124.016517 Belfilali H, Bousefsaf F, Messadi M. Left ventricle analysis in echocardiographic images using transfer learning. Phys Eng Sci Med . 2022;45(4):1123-1138. doi:10.1007/s13246-022-01179-3 Moradi S, Oghli MG, Alizadehasl A, et al. MFP-Unet: A novel deep learning based approach for left ventricle segmentation in echocardiography. Phys Med . 2019;67:58-69. doi:10.1016/j.ejmp.2019.10.001 Zeng Y, Tsui P-H, Pang K, et al. MAEF-Net: Multi-attention efficient feature fusion network for left ventricular segmentation and quantitative analysis in two-dimensional echocardiography. Ultrasonics . 2023;127:106855. doi:10.1016/j.ultras.2022.106855 Fu H, Hou D, Xu R, et al. Risk prediction models for deep venous thrombosis in patients with acute stroke: A systematic review and meta-analysis. Int J Nurs Stud . 2024;149:104623. doi:10.1016/j.ijnurstu.2023.104623 LeCun Y, Bengio Y, Hinton G. Deep learning. Nature . 2015;521(7553):436-444. doi:10.1038/nature14539 Rui Chen M, Fangqi Guo, MM, Jia Guo, MD, Jiaqi Zhao, MD. Application and Prospect of AI and ABVS-based in Breast Ultrasound Diagnosis. Advanced Ultrasound in Diagnosis and Therapy . 2023-06-30 2023;7(2):130-135. doi:10.37015/audt.2023.230017 Xiao Y, Xu T, Yu X, Fang Y, Li J. A Lightweight Fusion Strategy with Enhanced Inter-layer Feature Correlation for Small Object Detection. IEEE Transactions on Geoscience and Remote Sensing . 2024; Liu Y, Kong Y, Yan Y, Hui P. Explore the value of carotid ultrasound radiomics nomogram in predicting ischemic stroke risk in patients with type 2 diabetes mellitus. Front Endocrinol (Lausanne) . 2024;15:1357580. doi:10.3389/fendo.2024.1357580 Yang Z, Ye L, Yang L, Lu Q, Yu A, Bai D. Early screening of post-stroke fall risk: A simultaneous multimodal fNIRs-EMG study. CNS Neurosci Ther . 2024;30(9):e70041. doi:10.1111/cns.70041 Xue B, Li D, Lu C, et al. Use of Machine Learning to Develop and Evaluate Models Using Preoperative and Intraoperative Data to Identify Risks of Postoperative Complications. JAMA Netw Open . 2021;4(3):e212240. doi:10.1001/jamanetworkopen.2021.2240 Additional Declarations No competing interests reported. Supplementary Files Supplementaryversionlist.txt Supplementarycode.py Cite Share Download PDF Status: Under Review Version 1 posted Editorial decision: Revision requested 27 Dec, 2024 Editor assigned by journal 19 Dec, 2024 Submission checks completed at journal 19 Dec, 2024 First submitted to journal 13 Dec, 2024 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-5641383","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":392437070,"identity":"40675946-9824-4412-ab6b-b10f0f95b908","order_by":0,"name":"Jiachun xie","email":"","orcid":"","institution":"Peking University Shenzhen Hospital, Shenzhen Peking University-The Hong Kong University of Science and Technology Medical Center","correspondingAuthor":false,"prefix":"","firstName":"Jiachun","middleName":"","lastName":"xie","suffix":""},{"id":392437071,"identity":"3ae44c25-bae4-4a27-ab2d-f64fd07eaff5","order_by":1,"name":"Dianhuan Tan","email":"","orcid":"","institution":"Peking University Shenzhen Hospital, Shenzhen Peking University-The Hong Kong University of Science and Technology Medical Center","correspondingAuthor":false,"prefix":"","firstName":"Dianhuan","middleName":"","lastName":"Tan","suffix":""},{"id":392437073,"identity":"3e494de7-6954-48af-8e51-536e52031b60","order_by":2,"name":"Tingting Zheng","email":"","orcid":"","institution":"Peking University Shenzhen Hospital, Shenzhen Peking University-The Hong Kong University of Science and Technology Medical Center","correspondingAuthor":false,"prefix":"","firstName":"Tingting","middleName":"","lastName":"Zheng","suffix":""},{"id":392437075,"identity":"e87aa48d-1001-4cfd-8322-1303c9cc5333","order_by":3,"name":"Liya Wei","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAwElEQVRIiWNgGAWjYBACfvbG9p8fKv7x8BOtRbLncIO0xJkDcpINxGoxmJHeIMHbcsDY4ADRWngONhhINtxJ3Hw8eQPDj4pthLWYszc2JBTueJa47cyzAsaeM7cJa7HsOdhwQPIMc+K2GzkGzIxtRGgxuJHY2MDbxpy4eQYJWpoZeNsOGxtIEKtFsudgG7PEmTQ5CaBfDhLlF3729meMHypsePjbkzc++FFBhBYkkEB81CC0kKpjFIyCUTAKRggAALIqRPrE3e5KAAAAAElFTkSuQmCC","orcid":"","institution":"Peking University Shenzhen Hospital, Shenzhen Peking University-The Hong Kong University of Science and Technology Medical Center","correspondingAuthor":true,"prefix":"","firstName":"Liya","middleName":"","lastName":"Wei","suffix":""}],"badges":[],"createdAt":"2024-12-14 04:08:33","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-5641383/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-5641383/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":72376679,"identity":"62380175-5cca-47ee-a802-9d586b09ac12","added_by":"auto","created_at":"2024-12-26 08:25:06","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":57238,"visible":true,"origin":"","legend":"\u003cp\u003eFlowchart of patient selection and dataset creation.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-5641383/v1/bc0df359b82e8d3e9f458a1c.png"},{"id":72374864,"identity":"445120ce-2f4f-4864-9391-dd364e08cd2d","added_by":"auto","created_at":"2024-12-26 08:17:06","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":129951,"visible":true,"origin":"","legend":"\u003cp\u003e(A) The four-chamber view shows all four heart chambers (left/right atria and ventricles), essential for evaluating chamber size, function, and relationships. (B) The long-axis view highlights the left ventricle, aorta, and valves (mitral and aortic), aiding in the assessment of ventricular function and valve abnormalities. (C) The short-axis view provides cross-sectional images of the heart at various levels, showing structures like papillary muscles and left ventricular wall thickness.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-5641383/v1/d7fd95ec66dcca0b62a83caf.png"},{"id":72374867,"identity":"6aa330d9-fc16-4256-bc4b-e1f831d32cec","added_by":"auto","created_at":"2024-12-26 08:17:06","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":82583,"visible":true,"origin":"","legend":"\u003cp\u003eThe risk factors in multivariable logistic regression analysis.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-5641383/v1/b56fead6c71c1fe22d357e2f.png"},{"id":72374872,"identity":"c943d498-eff7-4165-8303-d7cf1d9721a3","added_by":"auto","created_at":"2024-12-26 08:17:07","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":113515,"visible":true,"origin":"","legend":"\u003cp\u003eWhile the traditional method suffers from weak feature correlations and redundancy, the EFC module incorporates Grouped Feature Focus and Feature Reconstruction to enhance feature representation and recover small object details, resulting in improved detection performance.\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-5641383/v1/a7766c2e4e5e81db3d70955e.png"},{"id":72374877,"identity":"a57e96f9-06b1-4e38-a56d-34d548944ce8","added_by":"auto","created_at":"2024-12-26 08:17:07","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":28982,"visible":true,"origin":"","legend":"\u003cp\u003eThe architecture of the MSEF module. Input feature maps are processed through 1×1 convolutions (Conv1 and Conv2) to extract initial representations. Channel Attention selectively enhances important features across channels, while the Grouped Feature Focus refines spatial and grouped information. Simultaneously, Feature Reconstruction recovers lost details and complements the grouped focus. The refined features are further processed by Depth-wise Convolution to extract fine-grained details efficiently, followed by Dynamic Reweighting, which adjusts feature importance dynamically. The final optimized output features are then produced for subsequent network operations.\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-5641383/v1/6b23ed070ba139d8f772c4e2.png"},{"id":72374891,"identity":"54dd75ac-41ca-455e-99cd-e4c769944695","added_by":"auto","created_at":"2024-12-26 08:17:07","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":104208,"visible":true,"origin":"","legend":"\u003cp\u003eMulti-view echocardiographic image classification pipeline. In the Multi-View Feature Extraction stage, features are extracted independently from three echocardiographic views. These features are then fused to integrate spatial and semantic information. The fused features are passed through a ResNet-based Backbone to further refine the representations. Finally, the Classification Head is applied to reduce the feature dimensions and produce the final classification probabilities.\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-5641383/v1/845e6d1a9a7431885a801c72.png"},{"id":72377935,"identity":"fdb77d5c-9fec-4a79-be8c-ff0c99abdb14","added_by":"auto","created_at":"2024-12-26 08:41:08","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1329753,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-5641383/v1/db4e0895-b29c-4b12-8206-8c6c5b357ee1.pdf"},{"id":72377328,"identity":"08ce6eb5-6684-4fe9-aab9-d32f1b06e818","added_by":"auto","created_at":"2024-12-26 08:33:06","extension":"txt","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":7130,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementaryversionlist.txt","url":"https://assets-eu.researchsquare.com/files/rs-5641383/v1/93c2de6f94e25c2f0c605e0e.txt"},{"id":72374869,"identity":"30d3387a-3749-478f-9255-3878acb09dda","added_by":"auto","created_at":"2024-12-26 08:17:07","extension":"py","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":9294,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementarycode.py","url":"https://assets-eu.researchsquare.com/files/rs-5641383/v1/f5b56ddfd8507a3163896ba3.py"}],"financialInterests":"No competing interests reported.","formattedTitle":"Cardiac Echocardiographic Analysis with Multi-Scale Effective Fusion Module: A Novel Stroke Prediction Approach","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eHypertension is a major chronic disease worldwide and a significant public health concern, with its prevalence increasing annually. It ranks among the top global health threats, as recognized by the World Health Organization\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e. The true danger of hypertension lies not in the condition itself but in its complications, which result in high rates of disability and mortality. Among these, cerebrovascular damage is particularly severe, with stroke being the most devastating consequence. Stroke is the second leading cause of death globally, and studies have shown that approximately 62% of stroke-related deaths are directly attributed to hypertension. A large-scale epidemiological study conducted in 2022 further highlighted this risk, reporting a cumulative stroke incidence of 78.9% among hypertensive patients\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e,\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e. Given the profound impact of stroke on individuals and healthcare systems, its prevention and early intervention are paramount. Developing accurate and reliable methods to predict and assess stroke risk in hypertensive patients is essential for mitigating its devastating consequences.\u003c/p\u003e \u003cp\u003eStroke remains one of the leading causes of morbidity and mortality worldwide, necessitating the development of accurate and reliable methods for early risk prediction and prevention. Traditional risk assessment models often rely on clinical features such as age, hypertension, diabetes, and previous cardiovascular events. While these models provide valuable insights, they may not fully capture the complexities of individual patient risk profiles\u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e,\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003ePrevious predictive models based on clinical features have played a significant role in advancing stroke risk assessment, providing valuable tools for identifying individuals at high risk and guiding preventive strategies\u003csup\u003e\u003cspan additionalcitationids=\"CR7\" citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e. These models have effectively utilized factors such as age, blood pressure, history of atrial fibrillation, and diabetes to stratify stroke risk in various populations, contributing to better clinical decision-making\u003csup\u003e\u003cspan additionalcitationids=\"CR10\" citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e. However, despite their usefulness, these models have limitations, particularly in their ability to capture the intricate structural and functional details of the heart. They often fail to incorporate the rich information available from echocardiographic imaging, such as subtle changes in cardiac morphology, valve abnormalities, and wall motion dynamics, which can be crucial indicators of stroke risk. As a result, the predictive accuracy of these models may be compromised, especially in cases where imaging features play a pivotal role in stroke pathogenesis. This underscores the importance of integrating echocardiographic data into stroke risk prediction models to achieve a more comprehensive and accurate assessment.\u003c/p\u003e \u003cp\u003eEchocardiography, a widely used non-invasive imaging technique, provides comprehensive and detailed visualization of cardiac structures and functions, enabling clinicians to assess heart anatomy, chamber sizes, wall thickness, valve functionality, and blood flow dynamics\u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e,\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e. It plays a crucial role in diagnosing and monitoring various cardiovascular conditions, offering real-time insights into cardiac performance and assisting in the evaluation of heart failure, valve diseases, congenital heart abnormalities, and stroke risk\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e. By accurately assessing factors such as left atrial size and function, left ventricular ejection fraction, valve abnormalities, intracardiac thrombus formation, aortic atherosclerotic plaques, and potential patent foramen ovale (PFO), echocardiography offers valuable information for predicting stroke risk. Its ability to deliver dynamic and high-resolution images makes echocardiography an indispensable tool in both routine clinical practice and advanced cardiac research, aiding in the prevention and management of stroke and other cardiovascular conditions\u003csup\u003e\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e. One reason for the underutilization is the complexity and variability of echocardiographic data, which often require expert interpretation to extract meaningful insights. However, advancements in machine learning and deep learning techniques have opened up opportunities to harness the predictive power of these images more effectively. By integrating echocardiographic data from long-axis, short-axis, and four-chamber views with hypertension-related clinical features, such as patient history and risk factors, it is possible to develop more accurate models for stroke risk prediction. These models can analyze subtle changes in cardiac structure and function that may not be immediately apparent to the human eye, offering a more comprehensive assessment of stroke risk. This integration has the potential to improve early detection, guide preventive interventions, and enhance personalized treatment strategies, ultimately reducing the burden of stroke in at-risk populations\u003csup\u003e\u003cspan additionalcitationids=\"CR17\" citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eBuilding on this success, deep learning models have the potential to revolutionize stroke risk prediction by analyzing echocardiographic images in ways that go beyond traditional clinical assessment\u003csup\u003e\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e. These models can identify intricate patterns related to cardiac morphology, wall motion abnormalities, and hemodynamic changes that are indicative of stroke risk but might be overlooked in routine examinations\u003csup\u003e\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e. By combining echocardiographic data with other clinical parameters, deep learning approaches can create a more comprehensive risk stratification model, allowing for earlier and more accurate identification of individuals at high risk for stroke\u003csup\u003e\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u003c/sup\u003e. This integration of deep learning into echocardiography could not only improve diagnostic accuracy but also facilitate personalized treatment plans, leading to better preventive strategies and outcomes for patients with cardiovascular disease. As the field continues to evolve, the application of deep learning in echocardiographic image analysis holds promise for advancing stroke prevention and enhancing our understanding of the complex relationship between cardiac function and cerebrovascular events.\u003c/p\u003e \u003cp\u003eIntegrating echocardiographic images, including long-axis, short-axis, and four-chamber views, poses unique challenges due to the heterogeneous nature and varying scales of the data. Each view captures distinct yet complementary cardiac information: long-axis views provide insights into left ventricular function and wall motion, short-axis views offer cross-sectional perspectives on chamber morphology and structural abnormalities, while four-chamber views visualize atria and ventricles, aiding in the detection of valve dysfunction, atrial enlargement, and thrombus formation. However, traditional feature fusion techniques, such as Feature Pyramid Networks (FPN), often fail to effectively integrate these multi-view features, particularly when addressing subtle structural abnormalities or small regions of interest critical to stroke risk prediction. To overcome these limitations, we introduce advanced fusion mechanisms, including the enhanced inter-layer feature correlation (EFC) and MSFC modules, which significantly enhance the representation and integration of multi-scale features. The EFC module focuses on spatial and channel-wise correlations through its GFF submodule, which generates spatial weights to emphasize key regions and enhance feature interactions, while the MFR submodule separates and reconstructs strong and weak features, preserving fine-grained details without interference. Meanwhile, the MSFC module improves the correlation between deep semantic features and shallow high-resolution features, enabling better multi-scale feature representation. By combining these modules, the proposed fusion strategy effectively captures subtle cardiac abnormalities and improves the overall predictive performance for stroke risk, particularly in cases where imaging features play a pivotal role in stroke pathogenesis.\u003c/p\u003e \u003cp\u003eIn this study, we propose a novel approach to stroke risk prediction by integrating deep learning-based analysis of echocardiographic images with traditional clinical risk factors. Our method consists of three main components: (1) developing a deep learning model to analyze multi-view echocardiographic images (long-axis, short-axis, and four-chamber views) to enhance feature extraction and representation; (2) introducing an advanced feature fusion strategy to integrate multi-scale echocardiographic features, improving the model's ability to capture subtle cardiac abnormalities related to stroke risk; and (3) combining the fused imaging features with hypertension-related clinical indicators to construct a hybrid predictive model, providing a comprehensive assessment of stroke risk.\u003c/p\u003e \u003cp\u003eWe hypothesize that the integration of echocardiographic imaging data and clinical features will enhance the accuracy and robustness of stroke risk prediction. By leveraging the strengths of both deep learning and traditional clinical assessment, our approach aims to provide a comprehensive tool for early detection and intervention, ultimately improving patient outcomes.\u003c/p\u003e \u003cp\u003eTo address the challenges associated with integrating multi-view echocardiographic data, including long-axis, short-axis, and four-chamber views, we improved traditional feature fusion strategies through the introduction of the Multi-Scale Feature Correlation (MSFC) module. While previous studies have applied the EFC module for multi-scale feature extraction, its limitations in fully capturing the spatial and semantic correlations across echocardiographic views prompted the need for further refinement\u003csup\u003e\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e. The MSFC module improves upon EFC by enhancing the correlation between deep semantic features and shallow high-resolution features, allowing for better representation of multi-scale information. Specifically, the MSFC module refines spatial attention to emphasize subtle cardiac abnormalities, such as valve dysfunction, atrial enlargement, and wall motion irregularities, which are critical indicators of stroke risk. It further optimizes the integration of strong and weak features to preserve fine-grained details while mitigating feature interference, thereby improving the module\u0026rsquo;s robustness in detecting small or subtle targets within echocardiographic images.\u003c/p\u003e \u003cp\u003eThe following sections describe the methodology for developing and validating the deep learning and clinical models, the process of integrating these models into a hybrid framework, and the evaluation of their combined predictive performance. By incorporating the MSFC module and effectively integrating imaging and clinical data, our approach provides a more accurate and comprehensive assessment of stroke risk, paving the way for enhanced prevention and management strategies.\u003c/p\u003e"},{"header":"2. Materials and Methods","content":"\u003cp\u003eThis study was approved by the institutional review board, and the requirement for informed patient consent was waived due to its retrospective cohort design.\u003c/p\u003e\n\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\n \u003ch2\u003e2.1 Data sources and preprocessing\u003c/h2\u003e\n \u003cp\u003eThis retrospective study utilized echocardiographic examination data and relevant clinical information of confirmed hypertensive patients extracted from the electronic medical record database of Peking University Shenzhen Hospital, covering the period from September 2022 to December 2023. The study included patients with comprehensive clinical and imaging data, with stroke diagnoses confirmed through neuroimaging. A total of 712 patient records were analyzed, comprising 10,992 echocardiographic images from the left ventricular long-axis, short-axis, and apical four-chamber views, as well as 27 associated clinical features. Patients were classified into positive or negative outcome groups based on whether they experienced a stroke during the course of hypertension.\u003c/p\u003e\n \u003cp\u003eThe inclusion criteria were: (1) Age over 18 years; (2) Completion of echocardiographic examination with comprehensive imaging data, including long-axis, short-axis, and apical four-chamber views; (3) A confirmed diagnosis of hypertension documented in the medical history, current medical history, or disease course records; (4) Availability of complete clinical data (6) Patients admitted to the hospital between September 2022 and December 2023.\u003c/p\u003e\n \u003cp\u003eThe exclusion criteria were: (1) Incomplete clinical information or missing echocardiographic images; (2) Poor-quality echocardiographic images that fail to meet diagnostic standards (e.g., unclear visualization of cardiac structures); (3) Patients with severe comorbidities that could confound stroke risk assessment, such as late-stage cancer, end-stage renal disease, or severe heart failure; (4) Patients with no documented neuroimaging confirmation of stroke diagnosis.\u003c/p\u003e\n \u003cp\u003eAs shown in Fig. \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e, echocardiographic images from the long-axis, short-axis, and four-chamber views of 712 patients with confirmed diagnoses were processed to produce nine-channel TIFF files. Each file was created by stacking the three echocardiographic views, with each view contributing three channels. For each patient, the number of data units generated ranged from 3 to 9. If a patient had 10 or more images, the highest-quality images were selected to form the final data units. Conversely, patients with 3 or fewer images were considered to have poor-quality echocardiographic images and were excluded from the analysis. Ultimately, 3664 nine-channel TIFF files were integrated with the patients\u0026apos; corresponding clinical information, including 27 hypertension-related features, to form 3664 data units. Each data unit consists of a combined echocardiographic image and its associated clinical features, providing a robust basis for subsequent model development and analysis.\u003c/p\u003e\n \u003cp\u003eAlthough the data units contain repeated clinical information, the data from the same patient were assigned to only one of the test, train, or validation sets to prevent data leakage and reduce the risk of overfitting. Regarding the 27 hypertension-related clinical features, a small portion of these features had missing values (\u0026lt;\u0026thinsp;10%), which were imputed using mean values to maintain data integrity. For clinical features with more than 10% missing values, they were excluded during the feature collection phase to ensure the robustness and reliability of the dataset. This preprocessing strategy ensures that the echocardiographic images and clinical features provide a clean and consistent input for subsequent model training and evaluation.\u003c/p\u003e\n \u003cp\u003eFigure 2 presents representative examples of the selected ultrasound images, including the four-chamber, long-axis, and short-axis echocardiographic views. The image selection was performed manually by two experienced radiologists, each with at least two years of experience in echocardiographic imaging.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e\n \u003ch2\u003e2.2 Clinical variables\u003c/h2\u003e\n \u003cp\u003eA total of 27 stroke-related risk factors, listed in Table\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e, were included in the study. These factors are clinically relevant indicators that can be easily assessed in practice. All risk factors were transformed into variables for model development.\u003c/p\u003e\n \u003cdiv class=\"gridtable\"\u003e\u0026nbsp;\u003ctable id=\"Tab1\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eThe risk factors with a definition in this study\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003ccolgroup cols=\"3\"\u003e\u003c/colgroup\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eRisk factors\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\u0026nbsp;\u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eDefinition\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAge\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAge of the patients, ranging from 40 years to the maximum observed age.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eGender\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eFemale/Male\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSystolic pressure\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSystolic blood pressure of the patient, measured in mmHg\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eDiastolic pressure\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eDiastolic blood pressure of the patient, measured in mmHg\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMedical history\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eEncoded information about the patient\u0026apos;s medical history (categorical variable)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eHeight\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eHeight of the patient, measured in centimeters (cm)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eWeight\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eWeight of the patient, measured in kilograms (kg)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eBMI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eBody Mass Index of the patient, calculated as weight (kg) divided by height (m) squared\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eDrink\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eDrinking status of the patient, encoded as 0 (non-drinker) or 1 (drinker)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSmoke\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSmoking status of the patient, encoded as 0 (non-smoker) or 1 (smoker)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eFamily history\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ePresence of family history of stroke, encoded as 0 (no) or 1 (yes)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eBlood fat\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eBlood fat level, encoded (specific encoding not provided)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eBlood sugar\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eBlood sugar level, encoded (specific encoding not provided).\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eUse drugs\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eUse of antihypertensive drugs, encoded as 0 (no) or 1 (yes).\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCOVID19\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCOVID\u0026minus;19 infection status, encoded as 0 (negative) or 1 (positive).\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eFunction change\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eChange in function, encoded (specific encoding not provided).\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eStructural change\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eStructural changes in the heart, encoded (specific encoding not provided)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eRWT\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eRelative wall thickness, a measure used in echocardiography.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLVMI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLeft ventricular mass index, a measure used in echocardiography\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAAO\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAscending aorta diameter, a measure used in echocardiography\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLA\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLeft atrium diameter, a measure used in echocardiography\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLV\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLeft ventricle diameter, a measure used in echocardiography\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eIVSD\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eInterventricular septum diameter, a measure used in echocardiography\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLVPW\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLeft ventricular posterior wall thickness, a measure used in echocardiography\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eEF\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eEjection fraction, a measure of the heart\u0026apos;s pumping efficiency, expressed as a percentage\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSeptal_E\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eEarly diastolic velocity of the septal mitral annulus, a measure used in echocardiography\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLateral wall\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eEarly diastolic velocity of the lateral mitral annulus, a measure used in echocardiography\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003cp\u003eRisk factors with a P-value\u0026thinsp;\u0026lt;\u0026thinsp;0.05 in the Chi-square test were considered statistically significant, listed in Table \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e. A total of 13 variables were identified as significant. To determine the optimal variables for constructing the prediction model, a multivariable logistic regression analysis was performed, and the results were expressed in terms of P-values. The area under the curve (AUC) was utilized to evaluate the performance and predictive accuracy of the model. Seven variables (Systolic Pressure, Medical History, Age, Height, Weight, RWT, AAO) showed a statistically significant difference (P\u0026thinsp;\u0026lt;\u0026thinsp;0.05) in the multivariable logistic regression analysis. The results of the multivariable logistic regression analysis are displayed as forest plots in Fig. \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e.\u003c/p\u003e\n \u003cdiv class=\"gridtable\"\u003e\u0026nbsp;\u003ctable id=\"Tab2\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eBaseline characteristics of the total cohort\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003ccolgroup cols=\"4\"\u003e\u003c/colgroup\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eRisk Factor\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eStroke Mean\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eNon-Stroke Mean\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eP-Value\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSystolic pressure\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e174.48\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e164.71\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eDiastolic pressure\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e97.41\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e97.53\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.938\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMedical history\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e12.03\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e7.63\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAge\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e66.78\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e56.89\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eHeight\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e163.24\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e164.84\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.104\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eWeight\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e67.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e69.44\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.128\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eBMI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e25.16\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e25.45\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.535\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eGender\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.52\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.55\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.585\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eDrink\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.15\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.08\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.051\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSmoke\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.24\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.191\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eFamily history\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.49\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.59\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.065\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eBlood fat\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.59\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.045\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eBlood sugar\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.42\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.023\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eUse antihypertensive drugs\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.79\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.68\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.025\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCOVID19 positive\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.48\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.52\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.469\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eFunction change\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.795\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eStructural change\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.64\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eRWT\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.11\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.04\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.004\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLVMI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.19\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.013\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAAO\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.51\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.28\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLA\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.22\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.33\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.495\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLV\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.07\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.04\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.391\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eIVSD\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.26\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.15\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLVPW\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.03\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.331\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eEF\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.734\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSeptal E\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.46\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.31\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.005\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLateral wall\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.58\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.42\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.005\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec5\" class=\"Section2\"\u003e\n \u003ch2\u003e2.3 FPN and EFC modules\u003c/h2\u003e\n \u003cp\u003eFeature Pyramid Networks (FPN) and the EFC module play a critical role in improving feature extraction and fusion for ultrasound imaging. The FPN is designed to effectively integrate multi-scale features, ensuring that both low-level spatial details and high-level semantic information are preserved. However, traditional feature fusion strategies within FPNs often suffer from weak correlations between layers and redundant features, which can hinder the accurate detection of small and complex lesions.\u003c/p\u003e\n \u003cp\u003eTo address these limitations, the EFC module was introduced, as illustrated in Fig. \u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e. The EFC module consists of two key components: the Grouped Feature Focus (GFF) and Feature Reconstruction (FR). The GFF selectively emphasizes critical feature groups, enhancing the representation of important structures. The FR further refines the fused features, recovering lost information and improving small object detection. Compared to the traditional fusion approach, the EFC module demonstrates superior performance by reducing redundancy and strengthening feature correlations, leading to more accurate lesion detection and classification.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec6\" class=\"Section2\"\u003e\n \u003ch2\u003e2.4 MSEF module\u003c/h2\u003e\n \u003cp\u003eTo further improve upon the EFC module and address the limitations of traditional multi-scale feature fusion strategies, we propose the MSEF module. While EFC effectively enhances feature correlation and reduces redundancy, its performance in handling small targets under complex and dense backgrounds can still be optimized. MSEF builds on EFC by introducing additional mechanisms to strengthen the coordination between deep semantic features and shallow high-resolution features, resulting in more effective multi-scale fusion and improved small target representation.\u003c/p\u003e\n \u003cp\u003eAs shown in Fig. \u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003e, MSEF integrates advanced feature processing techniques to enhance feature correlation, minimize redundancy, and maintain computational efficiency. It incorporates four key submodules:\u003c/p\u003e\u003cspan\u003e\n \u003cp\u003e1. Grouped Feature Focus (GFF): This submodule captures contextual information to strengthen the representation of small targets while generating spatial weights to focus on critical regions, ensuring small targets are not overlooked in complex backgrounds. By dividing feature maps into groups based on channels, it enables localized interaction and enhances channel-wise feature correlation, improving feature alignment and representation.\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e2. Multi-Layer Feature Reconstruction (MFR): MFR separates strong and weak features to avoid interference, ensuring that weak features are independently optimized while preserving the integrity of strong features. It applies fine-grained operations like 1\u0026times;1 convolutions to refine strong features, while using lightweight depth-wise separable convolutions to extract richer information from weak features. The reconstructed features are fused layer by layer, preserving small target details and semantic information for better representation.\u003c/p\u003e\n \u003c/span\u003e \u003cspan\u003e\n \u003cp\u003e3. Channel Attention Submodule: This submodule applies global average pooling to capture global context and uses a 1D convolution to model channel interactions, enhancing the relevance of feature channels. Unlike traditional methods, it avoids dimension reduction, ensuring that critical channel features are preserved, thereby improving overall feature correlation. The mechanisms involved are as follows:\u003c/p\u003e\n \u003c/span\u003e\n \u003cdiv id=\"Equ1\" class=\"Equation\"\u003e\n \u003cdiv class=\"mathdisplay\" id=\"FileID_Equ1\" name=\"EquationSource\"\u003e$$\\:s=W\\ast\\:z$$\u003c/div\u003e\n \u003cdiv class=\"EquationNumber\"\u003e1\u003c/div\u003e\n \u003c/div\u003e\n \u003cp\u003eThe above equation represents the operation where the weight \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:W\\)\u003c/span\u003e\u003c/span\u003e acts on the input vector \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:z\\)\u003c/span\u003e\u003c/span\u003e through a convolution, resulting in a new channel attention representation \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:s\\)\u003c/span\u003e\u003c/span\u003e.\u003c/p\u003e\n \u003cdiv id=\"Equ2\" class=\"Equation\"\u003e\n \u003cdiv class=\"mathdisplay\" id=\"FileID_Equ2\" name=\"EquationSource\"\u003e$$\\:{z}_{c}=\\frac{1}{H\\cdot\\:W}{\\sum\\:}_{i=1}^{H}{\\sum\\:}_{j=1}^{W}{X}_{c}\\left(i,j\\right),c=\\text{1,2},,C$$\u003c/div\u003e\n \u003cdiv class=\"EquationNumber\"\u003e2\u003c/div\u003e\n \u003c/div\u003e\n \u003cp\u003eThe above equation describes the global average pooling operation applied to each channel \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\text{c}\\)\u003c/span\u003e\u003c/span\u003e of the input feature map \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\text{X}\\)\u003c/span\u003e\u003c/span\u003e. This computes the global contextual information \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{z}}_{\\text{c}}\\)\u003c/span\u003e\u003c/span\u003e for each channel by averaging all spatial locations.\u003c/p\u003e\n \u003cp\u003e4. Position Attention Submodule: By applying attention along the horizontal and vertical axes of the feature maps, this submodule extracts spatial structure information through pooling operations. It enables precise focus on key spatial regions, enhancing the model\u0026apos;s ability to locate and represent small targets accurately. The mechanisms involved are as follows:\u003c/p\u003e\n \u003cdiv id=\"Equ3\" class=\"Equation\"\u003e\n \u003cdiv class=\"mathdisplay\" id=\"FileID_Equ3\" name=\"EquationSource\"\u003e$$\\:{P}_{ℎ}\\left(i,c\\right)=\\frac{1}{W}{\\sum\\:}_{j=1}^{W}{X}_{c}\\left(i,j\\right)$$\u003c/div\u003e\n \u003cdiv class=\"EquationNumber\"\u003e3\u003c/div\u003e\n \u003c/div\u003e\n \u003cp\u003eThis formula represents the global average pooling operation along the horizontal axis (width \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\text{W}\\)\u003c/span\u003e\u003c/span\u003e) for the \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\text{c}\\)\u003c/span\u003e\u003c/span\u003e-th channel of the input feature map \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\text{X}\\)\u003c/span\u003e\u003c/span\u003e. It aggregates spatial information across the horizontal dimension at height position \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\text{i}\\)\u003c/span\u003e\u003c/span\u003e.\u003c/p\u003e\n \u003cdiv id=\"Equ4\" class=\"Equation\"\u003e\n \u003cdiv class=\"mathdisplay\" id=\"FileID_Equ4\" name=\"EquationSource\"\u003e$$\\:{\\widehat{X}}_{c}\\left(i,j\\right)={\\alpha\\:}_{ℎ}\\left(i,c\\right)\\cdot\\:{X}_{c}\\left(i,j\\right)+{\\alpha\\:}_{v}\\left(j,c\\right)\\cdot\\:{X}_{c}\\left(i,j\\right)$$\u003c/div\u003e\n \u003cdiv class=\"EquationNumber\"\u003e4\u003c/div\u003e\n \u003c/div\u003e\n \u003cp\u003eThis formula represents the position attention mechanism for enhancing feature maps. Specifically, it combines the horizontal attention weight \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\alpha\\:}_{ℎ}\\left(i,c\\right)\\)\u003c/span\u003e\u003c/span\u003e and the vertical attention weight \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{{\\alpha\\:}}_{\\text{v}}\\left(\\text{j},\\text{c}\\right)\\)\u003c/span\u003e\u003c/span\u003e, which are applied to the \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\text{c}\\)\u003c/span\u003e\u003c/span\u003e-th channel of the input feature map \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{X}}_{\\text{c}}\\left(\\text{i},\\text{j}\\right)\\)\u003c/span\u003e\u003c/span\u003e. The resulting weighted feature map \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\widehat{\\text{X}}}_{\\text{c}}\\left(\\text{i},\\text{j}\\right)\\)\u003c/span\u003e\u003c/span\u003e is obtained by aggregating spatial information along both the horizontal and vertical axes, ensuring that important spatial relationships are captured across the input feature space.\u003c/p\u003e\n \u003cp\u003eThe MSEF module offers several advantages, making it a robust and efficient solution for multi-scale feature fusion. It enhances feature correlation by improving the spatial and semantic alignment of multi-scale features, which significantly strengthens the representation of small targets. The MFR submodule further reduces redundancy through precise feature reconstruction, ensuring that valuable information is retained while minimizing unnecessary fusion. Additionally, its lightweight design optimizes computational efficiency, effectively reducing both parameters and FLOPs without compromising detection accuracy. Moreover, the MSEF module\u0026apos;s plug-and-play compatibility allows seamless integration into various backbone networks, highlighting its versatility and applicability across a wide range of computer vision tasks.\u003c/p\u003e\n \u003cp\u003eIn summary, the MSEF module effectively overcomes the limitations of traditional feature fusion strategies by providing a robust and efficient approach for multi-scale feature integration. Through grouped feature focus, feature reconstruction, and attention mechanisms, it enhances small target representation and delivers superior performance in complex ultrasound images. This makes the MSEF module particularly well-suited for tasks requiring high precision, such as small target detection, segmentation, and classification in cardiac ultrasound imaging.\u003c/p\u003e\n \u003cp\u003eAs shown in Fig. \u003cspan class=\"InternalRef\"\u003e6\u003c/span\u003e, the MSEF module plays a crucial role in integrating multi-view echocardiographic features for the prediction of stroke risk. By fusing features extracted from long-axis, four-chamber, and short-axis views, MSEF ensures that spatial and semantic information from multiple perspectives is effectively combined. To further enhance the quality of feature integration, the long-axis and short-axis features are first fused individually with the four-chamber view features, which serves as a baseline due to its comprehensive representation of the heart\u0026apos;s overall structure. This stepwise fusion approach effectively leverages the complementary nature of different echocardiographic views, balancing global structural information with detailed local features while reducing feature redundancy and conflicts. Such a strategy also ensures that nuanced patterns, particularly small targets and subtle abnormalities, are accurately captured. The multi-scale feature integration enabled by MSEF allows the model to better capture subtle variations critical for stroke risk prediction, as these variations are often linked to cardiac structure and function abnormalities. The attention mechanisms within MSEF further emphasize key regions of the images, enabling the model to focus on clinically relevant features while minimizing noise. Overall, the application of MSEF significantly improves the robustness and accuracy of echocardiographic-based stroke risk prediction, addressing the challenges posed by complex and heterogeneous ultrasound data.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e\n \u003ch2\u003e2.5 Experimental Setup\u003c/h2\u003e\n \u003cp\u003eThe code for both EFC and MSEF modules is provided in Supplementary_code.py, and all required Python package versions are listed in Supplementary_version_list.txt. The experiments were conducted using Python 3.11 on an Ubuntu 18.04 operating system.\u003c/p\u003e\n \u003cp\u003eFor network implementation, either PyTorch or the MMPretrain framework can be used. It is necessary to modify the transform functions to accommodate TIFF image inputs, as it have specific channel characteristics. The mean and standard deviation (std) of ultrasound image channels typically hover around 40, but it is recommended to adjust these values based on the specific dataset being used.\u003c/p\u003e\n \u003cp\u003eThe training process was conducted with a batch size of 32, a learning rate of 0.0001, and the SGD (Stochastic Gradient Descent) optimizer. Manual tuning was performed for the learning rate and batch size, initially starting with a learning rate of 0.001 and progressively reducing it to 0.0001 based on validation performance. The model was trained on a single NVIDIA A100 GPU with 80GB memory, with each epoch taking approximately 1.2 minutes, resulting in a total training time of 2 hours over 100 epochs. For devices with limited GPU memory, we recommend proportionally reducing the batch size and learning rate while maintaining their ratio (e.g., halving the batch size to 16 and lowering the learning rate to 0.00005). Alternatively, simplifying the network architecture can help alleviate memory constraints during training.\u003c/p\u003e\n\u003c/div\u003e"},{"header":"3. Results","content":"\u003cp\u003e\u003cstrong\u003e\u003cem\u003e3.1 Comparison of Fusion Module Performance\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo objectively evaluate the effectiveness of our proposed method, we utilized a set of quantitative metrics, including Accuracy, Precision, Recall, and F1 Score, to ensure a comprehensive performance assessment. \u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTable 3 presents a summary of the results for three fusion methods\u0026mdash;FPN, EFC and MSEF\u0026nbsp;\u0026mdash;evaluated on both the validation set and the test set. These metrics provide an in-depth comparison of each method\u0026apos;s performance. \u0026nbsp;\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe results clearly demonstrate that the MSEF method is the most robust across both datasets. It consistently achieves the highest Accuracy and F1 Score, showcasing its ability to effectively balance Precision and Recall. While the EFC method achieves strong Precision, its lower Recall reduces its overall performance. In contrast, FPN exhibits moderate performance but struggles to remain competitive, particularly on the test set. \u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThese findings indicate that the multi-scale fusion approach employed by MSEF significantly enhances the model\u0026apos;s generalization capability, resulting in superior performance on unseen data.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 3.\u003c/strong\u003e The performance metrics of three methods evaluated on the validation set and the test set.\u0026nbsp;\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eMethod\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eData Split\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAccuracy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePrecision\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eRecall\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eF1 Score\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eFPN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eValidation Set\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e74.4%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e72.4%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e66.7%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e67.8%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eEFC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eValidation Set\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e73.7%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e70.5%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e68.2%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e69.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eMSEF\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eValidation Set\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e74.6%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e71.9%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e68.1%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e69.1%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eFPN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTest Set\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e70.5%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e66.6%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e63.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e63.6%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eEFC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTest Set\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e72.2%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e69.7%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;63.3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;63.9%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eMSEF\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTest Set\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003e76.8%\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003e66.2%\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003e63.3%\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003e64.7%\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e3.2 Ablation Study of the MSEF Module\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo gain a deeper understanding of the contribution of each component within the MSEF module, we conducted a systematic ablation study. Specifically, we removed the Channel Attention and Position Attention mechanisms independently to assess their respective roles in the module\u0026apos;s overall performance. The results of these experiments, summarized in Table 4, compare three model variants: MSEF without Channel Attention, MSEF without Position Attention, and the Full MSEF Module, evaluated on both the validation set and the test set.\u003c/p\u003e\n\u003cp\u003eAblative experiments are crucial for evaluating the effectiveness of individual components within a model, as they provide insights into how specific mechanisms contribute to the overall functionality. The comparative results clearly demonstrate that both Channel Attention and Position Attention play significant roles in enhancing the performance of the MSEF module.\u003c/p\u003e\n\u003cp\u003eThe removal of Channel Attention results in a noticeable drop in Recall on both datasets, with the validation set Recall decreasing to 63.2% and the test set Recall dropping to 69.2%. This reduction indicates that Channel Attention is particularly effective in capturing inter-channel relationships and enhancing feature representations by focusing on the most informative channels. By selectively prioritizing critical channels, this mechanism ensures that the network retains relevant information and suppresses noise, which is particularly important for tasks requiring high sensitivity, as reflected in the Recall metric.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eOn the other hand, removing Position Attention produces slightly higher Accuracy (77.5% on the validation set) but lowers Precision and F1 Score. The Precision drops to 67.6% on the validation set, highlighting that Position Attention is essential for balancing overall performance. Position Attention focuses on the spatial relationships within the feature maps, ensuring the model captures spatial context effectively. This mechanism is particularly important for preserving structural information, enabling the network to achieve a better balance between Precision and Recall, as evidenced by the higher F1 Score in the Full MSEF Module.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe Full MSEF Module, which integrates both Channel Attention and Position Attention, achieves the best overall performance across all metrics and datasets. On the validation set, it achieves a Recall of 68.1% and an F1 Score of 69.1%, while on the test set, it attains an Accuracy of 76.8% and an F1 Score of 64.7%. This demonstrates that the synergy between the two attention mechanisms enables the model to extract both channel-wise and spatial features effectively, enhancing its ability to generalize to unseen data. \u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe ablation study underscores the critical importance of the MSEF module\u0026apos;s structure, where Channel Attention and Position Attention work collaboratively to address complementary aspects of feature representation. Channel Attention ensures that the model focuses on the most relevant feature channels, improving sensitivity to key patterns, while Position Attention captures spatial dependencies, ensuring that the spatial structure of the data is preserved. By combining these two mechanisms, the Full MSEF Module achieves a superior trade-off between Precision and Recall, leading to robust and reliable performance across diverse datasets. In sum, the ablation study validates the design choices within the MSEF module, demonstrating that the integration of Channel Attention and Position Attention is essential for maximizing the model\u0026apos;s overall effectiveness and generalization capability.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 4.\u003c/strong\u003e The performance metrics of three model variants evaluated on the validation set and the test set.\u0026nbsp;\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eModel Variant\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eData Split\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAccuracy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePrecision\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eRecall\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eF1 Score\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eMSEF without Channel Attention\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eValidation Set\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e72.8%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e71.2%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e63.2%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e63.8%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eMSEF without Position Attention\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eValidation Set\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e77.5%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e67.6%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e64.7%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e66.1%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eFull MSEF Module\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eValidation Set\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e74.6%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e71.9%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e68.1%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e69.1%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eMSEF without Channel Attention\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTest Set\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e71.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e68.1%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e69.2%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e68.5%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eMSEF without Position Attention\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTest Set\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e73.5%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e70.1%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e67.8%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e68.6%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eFull MSEF Module\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTest Set\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003e76.8%\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003e66.2%\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003e63.3%\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003e64.7%\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e3.3 Stroke Risk Prediction with Multimodal Integration of Clinical and Imaging Data\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAccurate stroke risk prediction relies on effectively integrating clinical data and imaging information, two critical sources of patient information. While imaging data provides detailed anatomical and functional insights, traditional clinical features such as age, blood pressure, cholesterol levels, and comorbidities remain cornerstone predictors for stroke risk, as supported by extensive clinical studies. Leveraging these well-established clinical indicators alongside advanced imaging features ensures a robust and clinically interpretable prediction model.\u003c/p\u003e\n\u003cp\u003eTable 5 presents the performance metrics of three multimodal prediction models\u0026mdash;Logistic Regression, Random Forest, and XGBoost\u0026mdash;evaluated on both the validation set and the test set. These models integrate both imaging features extracted from the MSEF module and traditional clinical data to predict stroke risk.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 5.\u003c/strong\u003e The performance metrics of three multimodal prediction model.\u0026nbsp;\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eMethod\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eData Split\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eAccuracy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003ePrecision\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eRecall\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eF1 Score\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eLogistic Regression\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eValidation Set\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e80.5%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e74.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e71.4%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e72.7%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eRandom Forest\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eValidation Set\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e80.7%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e74.4%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e71.8%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e73.0%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eXGBoost\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eValidation Set\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e81.1%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e74.8%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e72.3%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e73.6%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eLogistic Regression\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTest Set\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e79.9%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e72.8%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e70.1%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e71.4%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eRandom Forest\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTest Set\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e80.1%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e73.1%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e70.5%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e71.8%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eXGBoost\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003eTest Set\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e80.2%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e73.4%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e70.8%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\"\u003e\n \u003cp\u003e72.1%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eThe integration of traditional clinical features with imaging-derived data is rooted in established clinical practice and research. Clinical features such as age, hypertension, diabetes, and lifestyle factors have been extensively validated as independent stroke predictors in prior studies. Imaging features, particularly those extracted through advanced fusion methods like MSEF, capture subtle and spatially distributed changes that are often invisible to manual inspection.\u003c/p\u003e\n\u003cp\u003eBy combining these two modalities, the models benefit from:\u003c/p\u003e\n\u003cp\u003e(1) Enhanced Predictive Power: Clinical features provide a solid baseline prediction, while imaging features refine the model\u0026apos;s ability to capture complex patterns indicative of stroke risk.\u003c/p\u003e\n\u003cp\u003e(2) Clinical Interpretability: Traditional clinical predictors offer transparency and trust, enabling clinicians to validate and interpret the model\u0026apos;s output more easily.\u003c/p\u003e\n\u003cp\u003e(3) Generalizability: The inclusion of clinical data stabilizes the model\u0026apos;s performance across datasets, reducing overfitting to imaging-specific noise.\u003c/p\u003e\n\u003cp\u003eIn sum, the results demonstrate that integrating clinical features with imaging-derived data significantly enhances stroke risk prediction performance. Among the evaluated models, XGBoost consistently achieves the best results, underscoring its ability to handle multimodal data effectively. This multimodal integration approach aligns with prior research and clinical practice, highlighting the importance of leveraging well-established clinical predictors alongside advanced imaging features to achieve reliable, interpretable, and clinically applicable stroke prediction models.\u003c/p\u003e"},{"header":"4. Discussion","content":"\u003cp\u003eIn this study, we proposed a novel stroke risk prediction framework that integrates clinical indicators with echocardiographic imaging features using a Multi-Scale Effective Fusion module. By combining traditional clinical factors and advanced imaging analysis, we achieved enhanced performance and demonstrated the importance of multimodal data integration for stroke risk assessment.\u003c/p\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e4.1 Significance of Integrating Clinical and Imaging Data\u003c/h2\u003e \u003cp\u003eTraditional stroke risk models primarily rely on clinical indicators, such as blood pressure, age, and medical history, which have long been validated as strong predictors in clinical practice\u003csup\u003e\u003cspan additionalcitationids=\"CR24\" citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e. However, these models often fall short in accounting for subtle cardiac morphological and functional changes detectable through imaging. Echocardiographic data, including long-axis, short-axis, and four-chamber views, offer critical insights into cardiac structure and function, such as valve abnormalities, wall motion dynamics, and atrial enlargement, which are pivotal for stroke pathogenesis. Integrating these imaging-derived features with clinical data bridges the gap between visual assessments and quantitative predictions, leading to a more comprehensive and interpretable risk prediction model. As shown in Section \u003cspan refid=\"Sec11\" class=\"InternalRef\"\u003e3.3\u003c/span\u003e, the multimodal integration approach demonstrated superior performance compared to single-modality methods. This highlights the advantages of advanced machine learning algorithms in handling complex, heterogeneous data, where clinical and imaging features complement one another to refine predictions.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003e4.2 The Role of the MSEF Module\u003c/h2\u003e \u003cp\u003eThe MSEF module, introduced in this study, played a critical role in improving feature extraction from echocardiographic images. Previous studies have demonstrated that multi-scale feature fusion enhances representation learning, especially for subtle or small abnormalities. Our ablation study in Section \u003cspan refid=\"Sec10\" class=\"InternalRef\"\u003e3.2\u003c/span\u003e revealed that both Channel Attention and Position Attention mechanisms significantly contribute to feature enhancement. Channel Attention prioritizes critical channels, improving the detection of informative patterns, while Position Attention captures spatial relationships, enabling the model to preserve structural details. By combining these mechanisms, the MSEF module effectively addresses the challenges posed by heterogeneous echocardiographic data, ensuring robust and accurate feature integration. The superior performance of the Full MSEF Module compared to its ablated variants underscores the necessity of a multi-dimensional attention strategy. Removing either Channel or Position Attention led to performance degradation, confirming that the combination of both mechanisms achieves optimal results for stroke risk prediction.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003e4.3 Clinical Implications\u003c/h2\u003e \u003cp\u003eOur findings align with existing clinical knowledge, emphasizing the continued relevance of traditional clinical features while leveraging advancements in imaging and machine learning. Clinical predictors such as age, blood pressure, and medical history remain highly interpretable and trusted by clinicians. Integrating these indicators with echocardiographic features enhances predictive power while ensuring the model\u0026rsquo;s clinical applicability and transparency. This multimodal framework allows for the early identification of stroke risk, enabling timely preventive interventions and personalized treatment strategies for at-risk populations.\u003c/p\u003e \u003cp\u003eFurthermore, the success of the MSEF-based approach highlights the potential of deep learning models in clinical settings. The ability to analyze echocardiographic data at a granular level enables the detection of subtle cardiac abnormalities, such as early atrial enlargement or wall motion irregularities, which may be overlooked during manual interpretation. Combined with clinical factors, these insights provide a more holistic understanding of stroke risk.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003e4.4 Limitations and Future Work\u003c/h2\u003e \u003cp\u003eDespite its promising results, this study has some limitations. First, the retrospective nature of the dataset may introduce selection bias, and future studies should validate the proposed framework using prospective data. Second, while the MSEF module effectively enhances feature extraction, further improvements in model efficiency could enable real-time clinical applications. Additionally, incorporating other imaging modalities, such as cardiac MRI, may provide additional insights and further improve predictive accuracy. Future work will focus on expanding the dataset to include diverse patient populations and exploring other fusion strategies for multimodal data integration.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003e4.5 Conclusion\u003c/h2\u003e \u003cp\u003eIn conclusion, this study demonstrates the importance of integrating clinical indicators with advanced imaging-derived features for stroke risk prediction. The proposed MSEF module effectively captures multi-scale echocardiographic information, improving predictive accuracy and robustness. By leveraging the strengths of both clinical data and machine learning-based imaging analysis, our approach provides a clinically interpretable and highly accurate tool for early stroke risk assessment, paving the way for improved prevention and personalized interventions.\u003c/p\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eData availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe data used in this study contain sensitive patient information and are therefore not publicly available to protect patient privacy.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis work was supported by the Science and Technology Project of Shenzhen (JCYJ20230807095209018, JCYJ20210324110211031, JCYJ20210324131402008, KXCFZ202002011010487), National Key Research and Development Program of China (2023YFC3402605), the Natural Science Foundation of Guangdong Province (2022A1515010986, 2022A1515010296), \u0026nbsp;the Shenzhen Key Medical Discipline Construction Fund (SZXK051), and the Sanming Project of Medicine in Shenzhen (SZSM202111011), \u0026nbsp;Peking University Shenzhen Hospital(LCYJZD2021010).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eHuman Ethics and Consent to Participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe 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.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics Approval\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe study was approved by the Medical Ethics Committee of Peking University Shenzhen Hospital (Approval Number: Peking University Shenzhen Hospital Ethics Committee (Research) [2024] No. 156), and the requirement of individual consent for this retrospective analysis was waived.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent to Participate declaration\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study was approved by the institutional review board, and the requirement for informed patient consent was waived due to its retrospective cohort design.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eClinical trial number\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003enot applicable\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eJ.X and D.T contributed equally to this work and were primarily responsible for data collection and statistical analysis. T.Z and L.W jointly conceived the study, designed the methodology, and secured the necessary funding to support this research. All authors contributed to the preparation and revision of the manuscript and approved the final version for submission.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003ePei Z, Liu J, Liu M, et al. Risk-Predicting Model for Incident of Essential Hypertension Based on Environmental and Genetic Factors with Support Vector Machine. \u003cem\u003eInterdiscip Sci\u003c/em\u003e. 2018;10(1):126-130. doi:10.1007/s12539-017-0271-2\u003c/li\u003e\n\u003cli\u003eKim H, Hwang S, Lee S, Kim Y. Classification and Prediction on Hypertension with Blood Pressure Determinants in a Deep Learning Algorithm. \u003cem\u003eInt J Environ Res Public Health\u003c/em\u003e. 2022;19(22)doi:10.3390/ijerph192215301\u003c/li\u003e\n\u003cli\u003eLi A-l, Ji Y, Zhu S, et al. Risk probability and influencing factors of stroke in followed-up hypertension patients. \u003cem\u003eBMC Cardiovasc Disord\u003c/em\u003e. 2022;22(1):328. doi:10.1186/s12872-022-02780-w\u003c/li\u003e\n\u003cli\u003eArafa A, Kokubo Y, Sheerah HA, et al. Developing a Stroke Risk Prediction Model Using Cardiovascular Risk Factors: The Suita Study. \u003cem\u003eCerebrovasc Dis\u003c/em\u003e. 2022;51(3):323-330. doi:10.1159/000520100\u003c/li\u003e\n\u003cli\u003eVu T, Kokubo Y, Inoue M, et al. Machine Learning Approaches for Stroke Risk Prediction: Findings from the Suita Study. \u003cem\u003eJ Cardiovasc Dev Dis\u003c/em\u003e. 2024;11(7)doi:10.3390/jcdd11070207\u003c/li\u003e\n\u003cli\u003eHeo J, Yoon JG, Park H, Kim YD, Nam HS, Heo JH. Machine Learning-Based Model for Prediction of Outcomes in Acute Stroke. \u003cem\u003eStroke\u003c/em\u003e. 2019;50(5):1263-1265. doi:10.1161/STROKEAHA.118.024293\u003c/li\u003e\n\u003cli\u003eNishi H, Oishi N, Ishii A, et al. Predicting Clinical Outcomes of Large Vessel Occlusion Before Mechanical Thrombectomy Using Machine Learning. \u003cem\u003eStroke\u003c/em\u003e. 2019;50(9):2379-2388. doi:10.1161/STROKEAHA.119.025411\u003c/li\u003e\n\u003cli\u003eHendrix P, Sofoluke N, Adams MD, et al. Risk Factors for Acute Ischemic Stroke Caused by Anterior Large Vessel Occlusion. \u003cem\u003eStroke\u003c/em\u003e. 2019;50(5):1074-1080. doi:10.1161/STROKEAHA.118.023917\u003c/li\u003e\n\u003cli\u003eMiyazaki Y, Kawakami M, Kondo K, et al. Logistic regression analysis and machine learning for predicting post-stroke gait independence: a retrospective study. \u003cem\u003eSci Rep\u003c/em\u003e. 2024;14(1):21273. doi:10.1038/s41598-024-72206-4\u003c/li\u003e\n\u003cli\u003evan der Vliet R, Selles RW, Andrinopoulou E-R, et al. Predicting Upper Limb Motor Impairment Recovery after Stroke: A Mixture Model. \u003cem\u003eAnn Neurol\u003c/em\u003e. 2020;87(3):383-393. doi:10.1002/ana.25679\u003c/li\u003e\n\u003cli\u003eIslam MS, Hussain I, Rahman MM, Park SJ, Hossain MA. Explainable Artificial Intelligence Model for Stroke Prediction Using EEG Signal. \u003cem\u003eSensors (Basel)\u003c/em\u003e. 2022;22(24)doi:10.3390/s22249859\u003c/li\u003e\n\u003cli\u003eSaha SK, Kiotsekoglou A. Value of speckle tracking echocardiography for prediction of stroke risk in atrial fibrillation: Time to spare a stare outside the box? \u003cem\u003eEchocardiography\u003c/em\u003e. 2018;35(5):589-591. doi:10.1111/echo.14005\u003c/li\u003e\n\u003cli\u003eTufano A, Galderisi M. Can echocardiography improve the prediction of thromboembolic risk in atrial fibrillation? Evidences and perspectives. \u003cem\u003eIntern Emerg Med\u003c/em\u003e. 2020;15(6):935-943. doi:10.1007/s11739-020-02303-5\u003c/li\u003e\n\u003cli\u003eChamsi-Pasha MA, Sengupta PP, Zoghbi WA. Handheld Echocardiography: Current State and Future Perspectives. \u003cem\u003eCirculation\u003c/em\u003e. 2017;136(22):2178-2188. doi:10.1161/CIRCULATIONAHA.117.026622\u003c/li\u003e\n\u003cli\u003eGillam LD, Marcoff L. Echocardiography: Past, Present, and Future. \u003cem\u003eCirc Cardiovasc Imaging\u003c/em\u003e. 2024;17(4):e016517. doi:10.1161/CIRCIMAGING.124.016517\u003c/li\u003e\n\u003cli\u003eBelfilali H, Bousefsaf F, Messadi M. Left ventricle analysis in echocardiographic images using transfer learning. \u003cem\u003ePhys Eng Sci Med\u003c/em\u003e. 2022;45(4):1123-1138. doi:10.1007/s13246-022-01179-3\u003c/li\u003e\n\u003cli\u003eMoradi S, Oghli MG, Alizadehasl A, et al. MFP-Unet: A novel deep learning based approach for left ventricle segmentation in echocardiography. \u003cem\u003ePhys Med\u003c/em\u003e. 2019;67:58-69. doi:10.1016/j.ejmp.2019.10.001\u003c/li\u003e\n\u003cli\u003eZeng Y, Tsui P-H, Pang K, et al. MAEF-Net: Multi-attention efficient feature fusion network for left ventricular segmentation and quantitative analysis in two-dimensional echocardiography. \u003cem\u003eUltrasonics\u003c/em\u003e. 2023;127:106855. doi:10.1016/j.ultras.2022.106855\u003c/li\u003e\n\u003cli\u003eFu H, Hou D, Xu R, et al. Risk prediction models for deep venous thrombosis in patients with acute stroke: A systematic review and meta-analysis. \u003cem\u003eInt J Nurs Stud\u003c/em\u003e. 2024;149:104623. doi:10.1016/j.ijnurstu.2023.104623\u003c/li\u003e\n\u003cli\u003eLeCun Y, Bengio Y, Hinton G. Deep learning. \u003cem\u003eNature\u003c/em\u003e. 2015;521(7553):436-444. doi:10.1038/nature14539\u003c/li\u003e\n\u003cli\u003eRui Chen M, Fangqi Guo, MM, Jia Guo, MD, Jiaqi Zhao, MD. Application and Prospect of AI and ABVS-based in Breast Ultrasound Diagnosis. \u003cem\u003eAdvanced Ultrasound in Diagnosis and Therapy\u003c/em\u003e. 2023-06-30 2023;7(2):130-135. doi:10.37015/audt.2023.230017\u003c/li\u003e\n\u003cli\u003eXiao Y, Xu T, Yu X, Fang Y, Li J. A Lightweight Fusion Strategy with Enhanced Inter-layer Feature Correlation for Small Object Detection. \u003cem\u003eIEEE Transactions on Geoscience and Remote Sensing\u003c/em\u003e. 2024;\u003c/li\u003e\n\u003cli\u003eLiu Y, Kong Y, Yan Y, Hui P. Explore the value of carotid ultrasound radiomics nomogram in predicting ischemic stroke risk in patients with type 2 diabetes mellitus. \u003cem\u003eFront Endocrinol (Lausanne)\u003c/em\u003e. 2024;15:1357580. doi:10.3389/fendo.2024.1357580\u003c/li\u003e\n\u003cli\u003eYang Z, Ye L, Yang L, Lu Q, Yu A, Bai D. Early screening of post-stroke fall risk: A simultaneous multimodal fNIRs-EMG study. \u003cem\u003eCNS Neurosci Ther\u003c/em\u003e. 2024;30(9):e70041. doi:10.1111/cns.70041\u003c/li\u003e\n\u003cli\u003eXue B, Li D, Lu C, et al. Use of Machine Learning to Develop and Evaluate Models Using Preoperative and Intraoperative Data to Identify Risks of Postoperative Complications. \u003cem\u003eJAMA Netw Open\u003c/em\u003e. 2021;4(3):e212240. doi:10.1001/jamanetworkopen.2021.2240\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-imaging","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmim","sideBox":"Learn more about [BMC Medical Imaging](http://bmcmedimaging.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bmim/default.aspx","title":"BMC Medical Imaging","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Stroke prediction, Echocardiographic images, Feature fusion, Multi-Scale Effective Fusion, Attention mechanisms","lastPublishedDoi":"10.21203/rs.3.rs-5641383/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-5641383/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003ePurpose: \u003c/strong\u003eTo develop a stroke risk prediction model by integrating echocardiographic images (long-axis, short-axis, four-chamber views) and clinical indicators using a novel multi-scale effective fusion (MSEF) module.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods: \u003c/strong\u003eA total of 712 patients with 10,992 images and 27 clinical indicators were included. The MSEF module enhances multi-scale feature fusion by combining deep semantic and shallow high-resolution features. It consists of four components: Global Feature Fusion (GFF), Multi-Feature Reconstruction (MFR), Channel Attention, and Positional Attention, effectively improving small-target feature representation. The fused features and clinical indicators were used to train the stroke prediction model.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults: \u003c/strong\u003eThe proposed MSEF-based model achieved the highest performance, with an Accuracy of 76.8% and F1 Score of 64.7% on the test set. Ablation studies confirmed the importance of Channel Attention and Position Attention in enhancing feature representation. When integrating echocardiographic features with clinical indicators, the model achieved an Accuracy of 80.2% and F1 Score of 72.1% on the test set.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusion: \u003c/strong\u003eThe proposed MSEF-based approach effectively integrates imaging and clinical data, improving stroke risk prediction accuracy and offering a promising tool for clinical decision-making.\u003c/p\u003e","manuscriptTitle":"Cardiac Echocardiographic Analysis with Multi-Scale Effective Fusion Module: A Novel Stroke Prediction Approach","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-12-26 08:17:02","doi":"10.21203/rs.3.rs-5641383/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2024-12-27T17:15:19+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2024-12-19T14:52:00+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2024-12-19T14:49:51+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Medical Imaging","date":"2024-12-14T04:00:05+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-imaging","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmim","sideBox":"Learn more about [BMC Medical Imaging](http://bmcmedimaging.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bmim/default.aspx","title":"BMC Medical Imaging","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"7e4f25f5-dec2-4463-9e69-9f5a3e3e0360","owner":[],"postedDate":"December 26th, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-05-20T01:39:42+00:00","versionOfRecord":[],"versionCreatedAt":"2024-12-26 08:17:02","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-5641383","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-5641383","identity":"rs-5641383","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}