Echocardiographic Evaluation in Children with Post-Acute Sequelae of SARS-CoV-2 infection Using Deep Learning

Echocardiographic Evaluation in Children with Post-Acute Sequelae of SARS-CoV-2 infection Using Deep Learning | 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 Echocardiographic Evaluation in Children with Post-Acute Sequelae of SARS-CoV-2 infection Using Deep Learning Yi-Chin Peng, Yi-Chen Huang, Xiao-Ling Liu, Jacky Chung-Hao Wu, and 9 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7134718/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Background: Post-acute sequelae of SARS-CoV-2 infection (PASC) is characterized by persistent symptoms following SARS-CoV-2 infection. Children with PASC are at risk of developing cardiac complications. Echocardiography has been instrumental in identifying cardiac abnormalities. This study applies deep learning to enhance the detection and understanding of echocardiographic changes in children with PASC. Methods: A case-control study was conducted at a pediatric tertiary center in central Taiwan. Children under 18 years who tested positive for SARS-CoV-2 and experienced symptoms for longer than four weeks were recruited between July 1, 2022, and July 31, 2023, during the Omicron variant surge. Echocardiographic data were also collected from a control group, consisting of children who presented with similar symptoms and received medical care in the same pediatric tertiary center in 2018. Children with congenital or structural heart disease, inflammatory conditions, or arrhythmias were excluded. Echocardiographic images were analyzed using a ResNet-50-based deep learning model to identify cardiac abnormalities. Results: A total of 270 children with PASC and 400 age-matched control children were included. No abnormalities were detected in the PASC group using conventional echocardiographic analysis. The deep learning model achieved an accuracy of 96.6%, sensitivity of 96.7%, specificity of 96.2%, and balanced accuracy of 96.4%. Conclusion: AI-assisted echocardiographic analysis demonstrated high performance in distinguishing cardiac function between PASC and controls. Deep learning models enhance the detection of subtle cardiac changes in children with PASC. Critical relevance statement: Deep learning enhances the detection of subtle cardiac abnormalities in children with post-COVID syndrome, improving diagnostic sensitivity beyond conventional echocardiographic interpretation. Pediatrics Artificial Intelligence and Machine Learning Post-acute sequelae of SARS-CoV-2 infection (PASC) children echocardiography Artificial intelligence Deep learning Pediatric cardiology Figures Figure 1 Figure 2 Figure 3 Figure 4 Article key points Deep learning improves echocardiographic analysis in pediatric PASC. ResNet-50 accurately detects subtle cardiac abnormalities in children with PASC, which are undetectable by conventional echocardiography. Despite normal echocardiographic parameters, children with PASC still show subtle but significant cardiac changes. AI-assisted imaging enhances early detection of post-COVID cardiovascular impact. This study bridges echocardiographic imaging and AI to improve pediatric cardiac screening. 1. Introduction After the pandemic of severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) infection, most patients recover from acute infection in a few days. However, there were certain proportion of patients suffering from new or persistent symptoms that develop during or after SARS-CoV-2 acute infection and not explainable by an alternative diagnosis were called “post-acute sequelae of SARS-CoV-2 infection” (PASC) [1, 2]. Heterogeneous symptoms affecting multiple organs are involved in PASC, and short- to mid-term, or long-term effect in personal health issues caused large impact on quality of life. The conditions was affected not only in adult patients, but also in pediatric patients. The prevalence of children with PASC estimates from 2% to 66% with wide variability [3, 4]. During the acute SARS-COV2 infection, host cells were infected through ACE2 receptors which are abundantly expressed in lung and heart. Clinically besides the respiratory involvement, cardiovascular manifestations, such as myocarditis and arrhythmia were common in pediatric patients [3]. The echocardiography has a major role to evaluate cardiac function in acute SARS-CoV-2 infection. But studies showed that conventional echocardiographic parameters were within normal limits during 1-year follow up [5]. More complex evaluation of ventricular functional parameters, such as longitudinal strain, showing transient biventricular alterations was reported in both adult [5] and pediatric patients 3 months after acute infection [6]. Although mild echocardiographic abnormalities are rarely observed in adults one weeks after recovering from coronavirus disease 2019 (COVID-19), they can still be detected in some cases. [7]. Patients with PASC may have long-term symptoms and effect in cardiac symptoms [3]. Up to now, echocardiographic studies used to evaluation cardiac function by brightness mode (B mode) and motion mode (M mode) have shown normal result in PASC patients [8, 9]. Previous echocardiographic studies focused on prognostic value in acute SARS-CoV-2 infection by myocardial longitudinal strain had confirmed its usefulness [8]. One study compared echocardiographic parameters, tissue Doppler, E/E' ratio, and left ventricular longitudinal strain six months after acute infection in adult PASC patients and showed no difference except for longitudinal strain [9,10]. However, complex parameters need complex off-line and time-consuming calculation. Currently, there is no practical way to detect echocardiographic abnormalities in children with PASC. Therefore, this study aims to explore whether echocardiography can be utilized in a novel way to evaluate previously unrecognized echocardiographic changes. 2. Materials and Methods 2.1 Patient selection and data collection This case-control study was conducted at a tertiary pediatric center and approved by the institutional review board. Children under 18 years with confirmed SARS-CoV-2 infection via RT-PCR or antigen test and persistent symptoms >4 weeks were recruited from an outpatient COVID-19 follow-up cohort [12] between July 1, 2022, and July 31, 2023, during the Omicron wave. Each child underwent clinical assessment including history, physical examination, lab tests, ECG, and echocardiography [12]. Pre-COVID control patients were identified from those who visited the hospital in 2018 with symptoms such as chest pain, palpitation, or dyspnea. Cases with congenital/structural heart disease, inflammatory conditions (e.g., myocarditis, Kawasaki disease, pericarditis), or arrhythmias were excluded. Echocardiographic images from these patients comprised the control group. 2.2 Echocardiography Echocardiographic examinations were performed using Philips IE33 or EPIQ 7G systems. B-mode ultrasound, widely used in diagnostic imaging, generates real-time two-dimensional images by processing echoes from tissue interfaces [13]. It offers advantages such as radiation-free imaging and cost-effectiveness. All echocardiograms were obtained with patients lying quietly without sedation; parents pacified younger children when necessary. Images were stored digitally in the hospital’s PACS and were available for later review. 2.3 Dataset and data splitting This study included 270 children with PASC and 400 pre-COVID controls without cardiac structural abnormalities (Table 1). A total of 12,457 B-mode ultrasound frames were collected from the PASC group and 1,910 frames from the control group. All images were reviewed by pediatric cardiologists. To maintain class distribution across the training, validation, and test sets, we used a stratified participant-based split. Each participant’s images were assigned entirely to one subset to avoid data leakage. The data were divided into training (64%), validation (16%), and test (20%) sets (Table 2) , reflecting the original class imbalance. The ratio of control to PASC images remained consistent across the three sets. 2.4 Data preprocessing In the preprocessing step, we masked the original B-mode ultrasound images ( Figure 1a ). First, we determined the region of interest (ROI, indicated by the blue area) in the image, then we used a full black mask to cover the areas outside the ROIs, preserving only the image content of the ROIs, as shown in Figure 1b . After this step, we obtained a masked image containing only the ROI but the background noise removed, as shown in Figure 1c . The main purpose of the masking technique is to filter out the noise and irrelevant regions in the image so that the model focuses only on the features in the ROI part, thus improving the accuracy of the subsequent analysis. 2.5 ResNet-50 based Model Architecture We applied a transfer learning strategy [14] using ResNet50 [15], a convolutional neural network pre-trained on ImageNet [16], for binary classification of echocardiographic images. ResNet50, with multiple residual blocks and a 2048-unit output layer, served as the backbone for extracting local and global features from B-mode ultrasound images. To adapt it to our task, we added a 512-unit fully connected layer with ReLU activation, followed by a dropout layer (rate = 0.2) to reduce overfitting. The final layer was a 2-unit fully connected layer mapping to the two output classes. As illustrated in Figure 2 , the overall architecture uses ResNet50 for feature extraction, followed by nonlinear transformation and classification layers to produce class probability predictions. Dropout regularization helps prevent overfitting and improves model generalization. 2.6 Experiment Setting This study presents experiments incorporating adjusted class weights during training and compares them to experiments without class weights to assess their impact. As shown in Table 1 , the current dataset suffers from a data imbalance issue. In binary classification tasks, a significant imbalance between the positive and negative sample sizes can cause the model to be biased toward learning the patterns of the majority class while neglecting the important information from the minority class. Specifically, we assigned each class a weight coefficient that is inversely proportional to the number of samples in that class. The weight for class was calculated as follows: During training, we select the model checkpoint corresponding to the highest balanced validation accuracy and evaluate its performance on the test set. The settings for the two experiments are the same: the Adam optimizer [17] is adopted with a learning rate of 10 -3 , the batch size is set to 32, and the maximum number of training epochs is 150. In one experiment, cross-entropy loss is guided by the adjusted class weights during model training, while in the other experiment, the experiment without class weights does not apply any weight adjustment and treats all samples equally during training. Both experiments are conducted with the PyTorch framework [18]. 3. Results The basic demographic characteristics of the PASC and pre-COVID groups are listed in Table 1 . The PASC group consisted of 152 boys (56.3%) and 118 girls (43.7%), with a mean age of 9.08 ± 4.74 years. The pre-COVID group included 212 boys (53%) and 188 girls (47%), with a mean age of 8.91 ± 4.66 years. There was no statistically significant difference between these two groups. The various PASC symptoms are detailed in Supplemental Figure 1 . Among them, the most distinctive symptoms after fatigue were shortness of breath (35.2%), chest pain (31.9%), and palpitations (31.1%). Based on these findings, we selected the pre-COVID patients with the same chief complaint symptom complex for comparison. The results of blood tests in the PASC group were shown in S upplemental T able 1 . All the patients in PASC group showed no abnormality including complete blood cell count, inflammatory biomarkers (erythrocyte sedimentation rate (ESR) as well as high-sensitivity C-reactive protein (hsCRP), creatine phosphokinase (CPK), serum ferritin and D-dimer) and liver function. The conventional echocardiography parameters we compared were listed in Table 3 . These parameters including Left Ventricular Diastolic Internal Dimension (LVIDd), left ventricular systolic internal dimension (LVIDs), Left Atrium Dimension (LA), Aortic Root Dimension (Ao), Ejection Fraction (EF), Fractional Shortening (FS), Mitral valve E-point septal separation(EPSS) , LA/AO ratio to evaluation the cardiac systolic function and chambers size. Left ventricular dimensions showed no difference between the 2 groups of patients. Patient with PASC had significant lower percentages of EF and FS, and significant higher dimension of LA and longer length of MV EPSS than pre-COVID children, but these data of each parameter measured in both groups were all within normal range. The training and evaluation results of the ResNet-50 based model without adjusted class weights are showed in Fig. 3a and the confusion matrix of the test set is showed in Fig. 3b , and Fig 3c reveals the AUC of the ROC curve is 0.9937. The results of test performance, such as accuracy (97.2%), sensitivity (98.4%), specificity (89.9%), and the balanced accuracy (94.1%.) are shown in Table 4 . Fig. 4 shows the training curve ( Fig. 4a ), confusion matrix ( Fig. 4b ), ROC curve ( Fig. 4c ) experiment results with class weights. The results of test performance, such as accuracy (96.6%), sensitivity (96.7%), specificity (96.2%), and the balanced accuracy (96.4%.) are also shown in Table 4 . 4. Discussion The findings from this study revealed AI detectable differences in echocardiographic images between children with PASC and those from the pre-COVID control group. These results suggest that certain changes in pediatric PASC patients, which are undetectable by conventional echocardiographic imaging, can be identified using machine learning. Unlike traditional image evaluation based on human perception, machine learning enables the detection of subtle cardiac abnormalities. 4.1 Disease Characteristics in COVID-19 Patients Emerging literature has documented various cardiovascular impacts of PASC in adults. Several studies have highlighted that, in PASC patients, conventional echocardiographic measures such as EF and FS are often normal, whereas subtle abnormalities—most notably, a reduction in global longitudinal strain (GLS)—can serve as early indicators of myocardial involvement [6, 7, 10, 19, 20]. For instance, one multicenter study found that 8.2% of adult patients without prior cardiovascular comorbidities exhibited echocardiographic abnormalities, with reduced GLS being the predominant feature, especially in male patients [7]. However, despite these findings in adults, pediatric data remain scarce. A recently published report [9] assessing the cardiovascular effects of SARS-CoV-2 infection in adults found no significant differences in conventional echocardiographic parameters across disease severities, yet this study did not include a pediatric cohort. Given the gap in understanding PASC’s impact on children, our study contributes valuable insights by focusing on pediatric patients, not only reinforces the established findings in adults but also addresses the critical gap in pediatric research, thereby broadening our understanding of PASC’s cardiovascular impact. 4.2 Imaging-based Long-term Cardiac Monitoring in PASC Traditionally, speckle tracking using dynamic 2-dimensional grayscale echocardiographic images has been regarded as a reliable method for assessing both global and regional ventricular function [21]. Advanced echocardiographic techniques, such as speckle-tracking echocardiography (STE), have been increasingly recognized for their ability to detect subtle myocardial changes in both the acute and PASC phases [22]. However, these methods require high-quality dynamic imaging, involve complex imaging protocols, and necessitate extensive post-processing procedures, which may limit their routine clinical use. In contrast, our study uniquely employs deep learning models to analyze non-dynamic B-mode echocardiographic images—images that are routinely acquired in clinical practice. This innovative approach simplifies image acquisition while potentially enhancing the detection of subtle cardiac changes between PASC and pre-COVID patients. By leveraging advanced artificial intelligence techniques, our method may offer higher accuracy and sensitivity compared to conventional echocardiographic assessments. Furthermore, a review by Australian physicians [23] highlights the value of advanced echocardiographic techniques, such as 2D strain imaging, in revealing subtle myocardial damage in PASC patients. Although these advanced methods have proven effective, their complexity often restricts widespread clinical application. Our approach not only aligns with current trends in using artificial intelligence for diagnostic enhancement, but also extends the investigation to a pediatric population. By applying deep learning to conventional B-mode images, our study offers a more accessible and scalable solution for long-term cardiac monitoring in children with PASC. 4.3 Optimization of the ResNet-50 Model Using Class Weights The results from Figures 3 and 4 demonstrate the high diagnostic performance of our ResNet-50-based model, with AUC values consistently close to 1. To further improve the model's performance, we introduced class weights during training to effectively address the challenge of class imbalance. This modification significantly enhanced the model's specificity and balanced accuracy, ensuring a more reliable application in clinical settings where both sensitivity and specificity are critical. Our study tackled the inherent class imbalance between the PASC and pre-COVID groups by assigning higher weights to the minority class (PASC group) during model training. Without this adjustment, the model tended to be biased toward the majority class, potentially compromising its clinical utility. By applying class weight adjustments, the model achieved improved specificity and balanced accuracy, highlighting the importance of tailored machine learning strategies in optimizing AI-assisted diagnostic tools. Table 4 compares the performance metrics between the two experimental setups. The baseline, which predicts all samples as belonging to the majority class (the PASC group), serves as the reference. The experiment without class weights yielded higher sensitivity and overall accuracy but suffered from reduced specificity, indicating a bias toward the positive class. In contrast, incorporating class weights resulted in a more balanced performance across all metrics, with higher balanced accuracy and improved specificity. This demonstrates that the use of class weights not only mitigates the impact of class imbalance but also ensures a more stable and consistent diagnostic performance. 4.4 The potential of AI diagnosis In our study, the left atrial dimension was significantly larger in the PASC group compared to the pre-COVID group, although the LA/AO ratio remained within the normal range. This finding suggests that pediatric PASC patients may experience mild structural changes in the left atrium as a compensatory mechanism to maintain cardiac function under potential hemodynamic or inflammatory stress, while the preserved LA/AO ratio indicates that these changes have not yet led to overt cardiac dysfunction. Future longitudinal studies are warranted to determine whether these early alterations could serve as a precursor to long-term cardiovascular complications. Similarly, although differences in EF, FS, and MV EPSS reached statistical significance, all values remained within normal limits, implying the presence of subclinical cardiac changes or adaptive mechanisms in pediatric PASC patients. These subtle differences, though not immediately clinically evident, underscore the importance of continuous cardiovascular monitoring to detect potential long-term effects. While conventional echocardiographic parameters often appear normal in post-COVID conditions, advanced imaging techniques such as speckle tracking have revealed subtle abnormalities. In this context, our study demonstrates the potential of artificial intelligence to detect these nuanced yet statistically significant changes. The high sensitivity and specificity of our ResNet-50-based model indicate that AI could play a pivotal role in the early identification of PASC-related cardiac alterations, offering pediatricians a practical tool to monitor changes over time. Our findings suggest that AI-driven analysis can transform pediatric cardiovascular care by identifying subtle echocardiographic changes that traditional methods may overlook. Although our model shows promising performance based on high-quality single-center data, further validation in larger, multicenter cohorts is necessary to establish its generalizability and clinical utility across diverse populations. Artificial intelligence has been widely applied in medical image analysis, including echocardiography [24, 25]. Most applications of AI in echocardiography have focused on detecting structural or anatomical lesions. Our previous work using AI to detect congenital ventricular septal defects [26] and acquired coronary aneurysms in Kawasaki Disease [27] has yielded promising results, as has AI-assisted assessment of valvular heart diseases [28]. To the best of our knowledge, this is the first study to utilize AI to detect subtle changes in echocardiographic images. Additionally, these findings may shed new light on the natural course of PASC/COVID syndrome. The clinical implications of our study are significant. The higher incidence of cardiac abnormalities in children with PASC underscores the need for ongoing cardiovascular monitoring in this population. Early detection and intervention can mitigate potential long-term cardiovascular sequelae, ultimately improving overall health outcomes. Pediatricians and cardiologists should be aware of these risks and consider incorporating routine echocardiographic assessments for children presenting with PASC. 4.5 Limitations and Future Perspectives This study has several strengths, including its study design, the use of a well-defined control group, and the application of advanced deep learning techniques to analyze echocardiographic data. These factors contribute to the robustness and reliability of our findings. However, there are also limitations to consider. First, the study was conducted at a single medical center, which may limit the generalizability of the results to other populations and settings. Second, the retrospective nature of the control group data collection might introduce selection bias, despite our efforts to match the groups based on symptom presentation and exclusion criteria. Finally, while the deep learning model provided high accuracy, but the black-box nature of deep learning makes the realization of its working process difficult! It is essential to validate these findings with larger, multicenter studies to confirm their applicability across diverse patient populations. Further research should explore the long-term cardiovascular outcomes of children with PASC, with larger and more diverse populations. Investigating the underlying mechanisms of cardiac involvement, including potential inflammatory and autoimmune responses, could inform targeted therapeutic interventions. Continued integration of deep learning in medical imaging will enhance diagnostic accuracy and predictive capabilities. 5. Conclusion In conclusion, our findings highlight the value of deep learning in pediatric echocardiographic evaluation, providing a novel and efficient approach to assess cardiac health in children with PASC. By integrating AI-driven image analysis with conventional clinical assessments, healthcare providers can enhance diagnostic accuracy and ultimately improve outcomes for affected children. Future studies should validate these findings in larger, more diverse cohorts and further explore the underlying biological mechanisms driving these echocardiographic changes. Abbreviations ACE2 Angiotensin-Converting Enzyme 2 AI Artificial Intelligence AUC Area Under the Curve B-mode Brightness Mode COVID-19 Coronavirus Disease 2019 CPK Creatine Phosphokinase DL Deep Learning ECG Electrocardiogram EF Ejection Fraction EPSS E-point Septal Separation ESR Erythrocyte Sedimentation Rate FS Fractional Shortening GLS Global Longitudinal Strain hsCRP High-Sensitivity C-Reactive Protein LA Left Atrium LV Left Ventricle LVIDd Left Ventricular Internal Diameter at End-Diastole LVIDs Left Ventricular Internal Diameter at End-Systole M-mode Motion Mode MV Mitral Valve PACS Picture Archiving and Communication System PASC Post-Acute Sequelae of SARS-CoV-2 Infection ReLU Rectified Linear Unit ROC Receiver Operating Characteristic ROI Region of Interest RT-PCR Reverse Transcriptase Polymerase Chain Reaction STE Speckle-Tracking Echocardiography VSD Ventricular Septal Defect Declarations Ethics approval and consent to participate This study was conducted in accordance with the Declaration of Helsinki and was approved by the Ethics Committee of China Medical University Children’s Hospital (Approval numbers: CMUH111-REC2-113 and CMUH111-REC2-122). Written informed consent was obtained from the legal guardians of all pediatric participants in the post-COVID group. For the retrospective control group, the requirement for informed consent was waived by the IRB due to the use of de-identified imaging data. Consent for publication Not applicable. This manuscript does not include identifiable individual data, images, or videos. Availability of data and material The datasets generated or analyzed during the current study are not publicly available due to institutional and ethical restrictions regarding the sharing of patient medical data, but are available from the corresponding author on reasonable request. Competing interests The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Funding This research was supported by the following grants: NSTC 113-2314-B-039-057 from the National Science and Technology Council, Taiwan; a research grant (1JA8) from the Center for Allergy, Immunology, and Microbiome (A.I.M.), China Medical University Hospital, Taichung, Taiwan; and DMR-113-114, DMR-114-024, and DMR-114-109 from China Medical University Hospital, Taichung, Taiwan. Authors' contributions YCP collected and interpreted the clinical data. YJH and JCW developed and implemented the deep learning algorithm, and assisted in visualization and model validation. XLL contributed to data curation and manuscript writing. PYL, YLH, and PCC coordinated clinical data collection and project administration. LSHW provided methodological guidance. HJT and WWC participated in manuscript review and editing. KSH, HSL, and JYW supervised the project and contributed to study design, funding acquisition, and final manuscript revision. All authors read and approved the final manuscript. Acknowledgements I sincerely thank all members of this research project for their dedication, expertise, and commitment, which were instrumental in achieving our research objectives. Their valuable insights, collaborative spirit, and unwavering support greatly contributed to this study. Authors' information (optional) Not applicable. References Thaweethai T, Jolley SE, Karlson EW, Levitan EB, Levy B, McComsey GA et al (2023) Development of a Definition of Postacute Sequelae of SARS-CoV-2 Infection. JAMA 329:1934–1946 Sherif ZA, Gomez CR, Connors TJ, Henrich TJ, Reeves WB, Force RMPT (2023) Pathogenic mechanisms of post-acute sequelae of SARS-CoV-2 infection (PASC). eLife 12:e86002 Rao S, Gross RS, Mohandas S, Stein CR, Case A, Dreyer B et al (2024) Postacute Sequelae of SARS-CoV-2 in Children. Pediatrics 153:e2023062570 Rao S, Lee GM, Razzaghi H, Lorman V, Mejias A, Pajor NM et al (2022) Clinical Features and Burden of Postacute Sequelae of SARS-CoV-2 Infection in Children and Adolescents. JAMA Pediatr 176:1000–1009 Olszanecka A, Wojciechowska W, Bednarek A, Kusak P, Wizner B, Terlecki M et al (2023) Serial echocardiographic evaluation of COVID-19 patients without prior history of structural heart disease: a 1-year follow-up CRACoV-HHS study. Front Cardiovasc Med 10:1230669 Sirico D, Di Chiara C, Costenaro P, Bonfante F, Cozzani S, Plebani M et al (2022) Left ventricular longitudinal strain alterations in asymptomatic or mildly symptomatic paediatric patients with SARS-CoV-2 infection. Eur Heart J Cardiovasc Imaging 23:1083–1089 Garcia-Zamora S, Picco JM, Lepori AJ, Galello MI, Saad AK, Ayón M et al (2023) Abnormal echocardiographic findings after COVID-19 infection: a multicenter registry. Int J Cardiovasc Imaging 39:77–85 Li Y, Li H, Zhu S, Xie Y, Wang B, He L et al (2020) Prognostic Value of Right Ventricular Longitudinal Strain in Patients With COVID-19. JACC Cardiovasc Imaging 13:2287–2299 Flores R, Pires O, Alves J, Pereira VH (2023) An Echocardiographic Insight Into Post-COVID-19 Symptoms. Cureus 15:e38039 Vidal-Perez R, Grapsa J, Bouzas-Mosquera A, Fontes-Carvalho R, Vazquez-Rodriguez JM (2023) Current role and future perspectives of artificial intelligence in echocardiography. World J Cardiol 15:284–292 Anon (2013) Diagnostic Ultrasound Imaging: Inside Out He K, Zhang X, Ren S, Sun J (2016) Deep Residual Learning for Image Recognition. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)., :770–778 Deng J, Dong W, Socher R, Li L-J, Li K, Fei-Fei L (2009) ImageNet: A large-scale hierarchical image database. In: 2009 IEEE Conference on Computer Vision and Pattern Recognition., :248–255 Kingma DP, Ba J, Adam (2017) A Method for Stochastic Optimization Paszke A, Gross S, Massa F, Lerer A, Bradbury J, Chanan G et al (2019) PyTorch: An Imperative Style, High-Performance Deep Learning Library Kirkpatrick JN, Swaminathan M, Adedipe A, Garcia-Sayan E, Hung J, Kelly N et al (2023) American Society of Echocardiography COVID-19 Statement Update: Lessons Learned and Preparation for Future Pandemics. J Am Soc Echocardiogr 36:1127–1139 Ds JW, Y K, G T-A Y, Sf N Preserved left ventricular twist and circumferential deformation, but depressed longitudinal and radial deformation in patients with diastolic heart failure. Eur Heart J 2008;29 Kersten J, Schellenberg J, Jerg A, Kirsten J, Persch H, Liu Y et al (2023) Strain Echocardiography in Acute COVID-19 and Post-COVID Syndrome: More than Just a Snapshot. Biomedicines 11:1236 Altersberger M, Goliasch G, Khafaga M, Schneider M, Cho Y, Winkler R et al (2023) Echocardiography and Lung Ultrasound in Long COVID and Post-COVID Syndrome, a Review Document of the Austrian Society of Pneumology and the Austrian Society of Ultrasound in Medicine. J Ultrasound Med 42:269–277 Vidal-Perez R, Grapsa J, Bouzas-Mosquera A, Fontes-Carvalho R, Vazquez-Rodriguez JM (2023) Current role and future perspectives of artificial intelligence in echocardiography. World J Cardiol 15:284 Chen S-H, Wang C-W, Tai I-H, Weng K-P, Chen Y-H, Hsieh K-S (2021) Modified YOLOv4-DenseNet Algorithm for Detection of Ventricular Septal Defects in Ultrasound Images. Int J Interact Multimedia Artif Intell 6:101–108 Kuo H-C, Chen S-H, Chen Y-H, Lin Y-C, Chang C-Y, Wu Y-C et al (2022) Detection of coronary lesions in Kawasaki disease by Scaled-YOLOv4 with HarDNet backbone. Front Cardiovasc Med 9:1000374 Nedadur R, Wang B, Tsang W (2022) Artificial intelligence for the echocardiographic assessment of valvular heart disease. Heart 108:1592–1599 Akkus Z, Aly YH, Attia IZ, Lopez-Jimenez F, Arruda-Olson AM, Pellikka PA et al (2021) Artificial Intelligence (AI)-Empowered Echocardiography Interpretation: A State-of-the-Art Review. J Clin Med 10:1391 Karlsen S, Dahlslett T, Grenne B, Sjøli B, Smiseth O, Edvardsen T et al (2019) Global longitudinal strain is a more reproducible measure of left ventricular function than ejection fraction regardless of echocardiographic training. Cardiovasc Ultrasound 17:18 Barry T, Farina JM, Chao C-J, Ayoub C, Jeong J, Patel BN et al (2023) The Role of Artificial Intelligence in Echocardiography. J Imaging 9:50 Hsu Y-L, Chen P-C, Tsai Y-F, Wei C-H, Wu LS-H, Hsieh K-S et al (2024) Clinical Features and Vaccination Effects among Children with Post-Acute Sequelae of COVID-19 in Taiwan. Vaccines (Basel) 12:910 Caruana R (1997) Multitask Learning. Mach Learn 28:41–75 Additional Declarations The authors declare no competing interests. Cite Share Download PDF Status: Posted Version 1 posted 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-7134718","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":486032305,"identity":"53f11669-55bd-4e7d-8169-ed94d8002ccf","order_by":0,"name":"Yi-Chin Peng","email":"","orcid":"","institution":"Department of Pediatric Cardiology, China Medical University Children’s Hospital, China Medical University, Taichung 404327, Taiwan.","correspondingAuthor":false,"prefix":"","firstName":"Yi-Chin","middleName":"","lastName":"Peng","suffix":""},{"id":486032306,"identity":"f7f2dbd1-2c24-4f72-8c48-9c5948525315","order_by":1,"name":"Yi-Chen Huang","email":"","orcid":"","institution":"Institute of Statistics, National Yang Ming Chiao Tung University, Hsinchu 300093, Taiwan.","correspondingAuthor":false,"prefix":"","firstName":"Yi-Chen","middleName":"","lastName":"Huang","suffix":""},{"id":486032307,"identity":"12de3921-1381-41e9-b1bc-fcec5d163f14","order_by":2,"name":"Xiao-Ling Liu","email":"","orcid":"https://orcid.org/0009-0000-3860-975X","institution":"Research Center of Allergy, Immunology and Microbiome (A.I.M.), China Medical University Hospital, China Medical University, Taichung 404327, Taiwan.","correspondingAuthor":false,"prefix":"","firstName":"Xiao-Ling","middleName":"","lastName":"Liu","suffix":""},{"id":486033171,"identity":"e6a6aee4-4fca-44f9-9d93-b613ee3c6ab1","order_by":3,"name":"Jacky Chung-Hao Wu","email":"","orcid":"","institution":"Biomedical Artificial Intelligence Academy, Kaohsiung Medical University, Kaohsiung 807378, Taiwan.","correspondingAuthor":false,"prefix":"","firstName":"Jacky","middleName":"Chung-Hao","lastName":"Wu","suffix":""},{"id":486033172,"identity":"c9a02987-f96d-46ab-bb08-7182c813a6ed","order_by":4,"name":"Pang-Yan Liu","email":"","orcid":"","institution":"Department of Teaching and Research, China Medical University Children’s Hospital, China Medical University, Taichung 404327, Taiwan","correspondingAuthor":false,"prefix":"","firstName":"Pang-Yan","middleName":"","lastName":"Liu","suffix":""},{"id":486033173,"identity":"6f7647ce-804f-4c2a-aa30-06ea877a05c7","order_by":5,"name":"Yu-Lung Hsu","email":"","orcid":"","institution":"Department of Pediatric Infectious Diseases, China Medical University Children’s Hospital, China Medical University, Taichung 404327, Taiwan.","correspondingAuthor":false,"prefix":"","firstName":"Yu-Lung","middleName":"","lastName":"Hsu","suffix":""},{"id":486033174,"identity":"5ab8f028-f5a9-40a4-a189-753fd3815436","order_by":6,"name":"Pei-Chi Chen","email":"","orcid":"","institution":"Department of Microbiology \u0026 Immunology, College of Medicine, National Cheng Kung University, Tainan, Taiwan.","correspondingAuthor":false,"prefix":"","firstName":"Pei-Chi","middleName":"","lastName":"Chen","suffix":""},{"id":486033175,"identity":"a8a61a23-13a7-4228-8f8b-7e0585b453ec","order_by":7,"name":"Lawrence Shih-Hsin Wu","email":"","orcid":"","institution":"Institute of Biomedical Sciences, College of Medicine, China Medical University, Taichung, Taiwan","correspondingAuthor":false,"prefix":"","firstName":"Lawrence","middleName":"Shih-Hsin","lastName":"Wu","suffix":""},{"id":486033176,"identity":"7b724615-f36b-4f1d-a298-f6aae704b6ac","order_by":8,"name":"Hui-Ju Tsai","email":"","orcid":"","institution":"Institute of Population Health Sciences, National Health Research Institutes, Zhunan, Taiwan.","correspondingAuthor":false,"prefix":"","firstName":"Hui-Ju","middleName":"","lastName":"Tsai","suffix":""},{"id":486033177,"identity":"b8d91bc1-2bad-4c5f-8965-11a0063efe50","order_by":9,"name":"Wei-Wen Chen","email":"","orcid":"","institution":"Institute of Statistics, National Yang Ming Chiao Tung University, Hsinchu 300093, Taiwan.","correspondingAuthor":false,"prefix":"","firstName":"Wei-Wen","middleName":"","lastName":"Chen","suffix":""},{"id":486033178,"identity":"8d307bd7-7f3d-4be7-b12d-1e4ac9e9817f","order_by":10,"name":"Kai-Sheng Hsieh","email":"","orcid":"","institution":"Department of Pediatric Cardiology, China Medical University Children’s Hospital, China Medical University, Taichung 404327, Taiwan","correspondingAuthor":false,"prefix":"","firstName":"Kai-Sheng","middleName":"","lastName":"Hsieh","suffix":""},{"id":486033179,"identity":"24bc1d44-b0f5-46b4-aca7-b055694e1c7f","order_by":11,"name":"Henry Horng-Shing Lu","email":"","orcid":"","institution":"11.School of Post-Baccalaureate Medicine, Kaohsiung Medical University, Kaohsiung 807378, Taiwan.","correspondingAuthor":false,"prefix":"","firstName":"Henry","middleName":"Horng-Shing","lastName":"Lu","suffix":""},{"id":486033180,"identity":"1060c50f-8dfb-4a14-b1d9-ef313eb9d049","order_by":12,"name":"Jiu-Yao Wang","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA7ElEQVRIie3RPQrCMBiA4S8E0iW2a4uiV2gpCII/V6lk1dXFQaXQSfeKl9AlcyWDi+jqqHgBO6ngYFQQFxNHh7xDvhB4ICEAJtM/hvFyfwKgttyj4eMk0xGLsCCVhHyQ1/yaQ6tFKufvxInBLzY6zRKBwjKfJlC2d5F1pAriCojCLmfyYjbzFgmE3i5CsYqAgIx1OZaEVtEhgfb8QToKURFoKGp88CYDLfEFRjHi4kXkxSJfRwJBMJrwFSWYhl66cYPp+jCa3RSkvN2eL1febznWOMjHvXrFXrEsT1XPf4efqwug+0mTyWQy6bsD6WpF7VtldWAAAAAASUVORK5CYII=","orcid":"https://orcid.org/0000-0003-4540-9822","institution":"Department of Allergy and Immunology, China Medical University Children’s Hospital, China Medical University, Taichung 404327, Taiwan","correspondingAuthor":true,"prefix":"","firstName":"Jiu-Yao","middleName":"","lastName":"Wang","suffix":""}],"badges":[],"createdAt":"2025-07-16 01:43:10","currentVersionCode":1,"declarations":{"humanSubjects":true,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":true,"humanSubjectConsent":true,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-7134718/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7134718/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":87051670,"identity":"e99e29f2-76af-4a96-afe8-8040008cb765","added_by":"auto","created_at":"2025-07-18 15:05:11","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":147562,"visible":true,"origin":"","legend":"\u003cp\u003ePre-processing steps applied to B-mode ultrasound images. (a) The original B-mode ultrasound image. (b) A binary mask delineating the region of interest (ROI), where the blue area represents the ROI, and the black area indicates masked-out regions. (c) The final masked image, preserving only the ROI while removing background noise.\u003c/p\u003e","description":"","filename":"FIGURE01.png","url":"https://assets-eu.researchsquare.com/files/rs-7134718/v1/8daa6ae759c948a7fbf11ca2.png"},{"id":87054031,"identity":"48f735cc-9843-4a86-bd50-ca9df1e48b31","added_by":"auto","created_at":"2025-07-18 15:21:11","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":51734,"visible":true,"origin":"","legend":"\u003cp\u003eProposed ResNet-50 based model architecture.\u003c/p\u003e","description":"","filename":"FIGURE02.png","url":"https://assets-eu.researchsquare.com/files/rs-7134718/v1/95cbea5099c13f994f1cefbd.png"},{"id":87051673,"identity":"aa726153-3a4d-46c3-baf5-25d6f21e7a4d","added_by":"auto","created_at":"2025-07-18 15:05:11","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":311865,"visible":true,"origin":"","legend":"\u003cp\u003eTraining and evaluation results of the ResNet-50-based model without adjusted class weights. (a) Training curves showing model convergence. (b) Confusion matrix of the test set. (c) Receiver operating characteristic (ROC) curve of the test set.\u003c/p\u003e","description":"","filename":"FIGURE03.png","url":"https://assets-eu.researchsquare.com/files/rs-7134718/v1/4c97e8ac2d3c4faf8d1547aa.png"},{"id":87054805,"identity":"ce804d10-f068-4364-a5de-3f45f36bf935","added_by":"auto","created_at":"2025-07-18 15:29:11","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":298702,"visible":true,"origin":"","legend":"\u003cp\u003eTraining and evaluation results of the ResNet-50-based model without adjusted class weights. (a) Training curves showing model convergence. (b) Confusion matrix of the test set. (c) Receiver operating characteristic (ROC) curve of the test set.\u003c/p\u003e","description":"","filename":"FIGURE04.png","url":"https://assets-eu.researchsquare.com/files/rs-7134718/v1/7b8a87058b36f7e6a7ee0dfd.png"},{"id":87054994,"identity":"a360ef7c-3f46-437e-a5f5-6218b99e1906","added_by":"auto","created_at":"2025-07-18 15:37:12","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1413568,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7134718/v1/cbcd6ac7-1cb4-4acf-98fa-d1cf0b4ee9d4.pdf"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003eEchocardiographic Evaluation in Children with Post-Acute Sequelae of SARS-CoV-2 infection Using Deep Learning\u003c/p\u003e","fulltext":[{"header":"Article key points","content":"\u003col\u003e\n \u003cli\u003eDeep learning improves echocardiographic analysis in pediatric PASC.\u003c/li\u003e\n \u003cli\u003eResNet-50 accurately detects subtle cardiac abnormalities in children with PASC, which are undetectable by conventional echocardiography.\u003c/li\u003e\n \u003cli\u003eDespite normal echocardiographic parameters, children with PASC still show subtle but significant cardiac changes.\u003c/li\u003e\n \u003cli\u003eAI-assisted imaging enhances early detection of post-COVID cardiovascular impact.\u003c/li\u003e\n \u003cli\u003eThis study bridges echocardiographic imaging and AI to improve pediatric cardiac screening.\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"1. Introduction","content":"\u003cp\u003eAfter the pandemic of severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) infection, most patients recover from acute infection in a few days. However, there were certain proportion of patients suffering from new or persistent symptoms that develop during or after SARS-CoV-2 acute infection and not explainable by an alternative diagnosis were called “post-acute sequelae of SARS-CoV-2 infection” (PASC) [1, 2].\u003c/p\u003e\n\u003cp\u003eHeterogeneous symptoms affecting multiple organs are involved in PASC, and short- to mid-term, or long-term effect in personal health issues caused large impact on quality of life. The conditions was affected not only in adult patients, but also in pediatric patients. The prevalence of children with PASC estimates from 2% to 66% with wide variability [3, 4]. During the acute SARS-COV2 infection, host cells were infected through ACE2 receptors which are abundantly expressed in lung and heart. Clinically besides the respiratory involvement, cardiovascular manifestations, such as myocarditis and arrhythmia were common in pediatric patients [3]. The echocardiography has a major role to evaluate cardiac function in acute SARS-CoV-2 infection. But studies showed that conventional echocardiographic parameters were within normal limits during 1-year follow up [5]. More complex evaluation of ventricular functional parameters, such as longitudinal strain, showing transient biventricular alterations was reported in both adult [5] and pediatric patients 3 months after acute infection [6]. Although mild echocardiographic abnormalities are rarely observed in adults one weeks after recovering from coronavirus disease 2019 (COVID-19), they can still be detected in some cases. [7].\u003c/p\u003e\n\u003cp\u003ePatients with PASC may have long-term symptoms and effect in cardiac symptoms [3]. Up to now, echocardiographic studies used to evaluation cardiac function by brightness mode (B mode) and motion mode (M mode) have shown normal result in PASC patients [8, 9]. Previous echocardiographic studies focused on prognostic value in acute SARS-CoV-2 infection by myocardial longitudinal strain had confirmed its usefulness [8]. One study compared echocardiographic parameters, tissue Doppler, E/E' ratio, and left ventricular longitudinal strain six months after acute infection in adult PASC patients and showed no difference except for longitudinal strain [9,10]. However, complex parameters need complex off-line and time-consuming calculation.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eCurrently, there is no practical way to detect echocardiographic abnormalities in children with PASC. Therefore, this study aims to explore whether echocardiography can be utilized in a novel way to evaluate previously unrecognized echocardiographic changes.\u003c/p\u003e"},{"header":"2. Materials and Methods","content":"\u003cp\u003e\u003cem\u003e2.1 Patient selection and data collection\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eThis case-control study was conducted at a tertiary pediatric center and approved by the institutional review board. Children under 18 years with confirmed SARS-CoV-2 infection via RT-PCR or antigen test and persistent symptoms \u0026gt;4 weeks were recruited from an outpatient COVID-19 follow-up cohort [12] between July 1, 2022, and July 31, 2023, during the Omicron wave. Each child underwent clinical assessment including history, physical examination, lab tests, ECG, and echocardiography [12].\u003c/p\u003e\n\u003cp\u003ePre-COVID control patients were identified from those who visited the hospital in 2018 with symptoms such as chest pain, palpitation, or dyspnea. Cases with congenital/structural heart disease, inflammatory conditions (e.g., myocarditis, Kawasaki disease, pericarditis), or arrhythmias were excluded. Echocardiographic images from these patients comprised the control group.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e2.2 Echocardiography\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eEchocardiographic examinations were performed using Philips IE33 or EPIQ 7G systems. B-mode ultrasound, widely used in diagnostic imaging, generates real-time two-dimensional images by processing echoes from tissue interfaces [13]. It offers advantages such as radiation-free imaging and cost-effectiveness. All echocardiograms were obtained with patients lying quietly without sedation; parents pacified younger children when necessary. Images were stored digitally in the hospital’s PACS and were available for later review.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e2.3 Dataset and data splitting\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eThis study included 270 children with PASC and 400 pre-COVID controls without cardiac structural abnormalities\u003cstrong\u003e\u0026nbsp;(Table 1).\u003c/strong\u003e A total of 12,457 B-mode ultrasound frames were collected from the PASC group and 1,910 frames from the control group. All images were reviewed by pediatric cardiologists. To maintain class distribution across the training, validation, and test sets, we used a stratified participant-based split. Each participant’s images were assigned entirely to one subset to avoid data leakage. The data were divided into training (64%), validation (16%), and test (20%) sets\u0026nbsp;\u003cstrong\u003e(Table 2)\u003c/strong\u003e, reflecting the original class imbalance. The ratio of control to PASC images remained consistent across the three sets.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e2.4 Data preprocessing\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eIn the preprocessing step, we masked the original B-mode ultrasound images (\u003cstrong\u003eFigure 1a\u003c/strong\u003e). First, we determined the region of interest (ROI, indicated by the blue area) in the image, then we used a full black mask to cover the areas outside the ROIs, preserving only the image content of the ROIs, as shown in \u003cstrong\u003eFigure 1b\u003c/strong\u003e. After this step, we obtained a masked image containing only the ROI but the background noise removed, as shown in \u003cstrong\u003eFigure 1c\u003c/strong\u003e. The main purpose of the masking technique is to filter out the noise and irrelevant regions in the image so that the model focuses only on the features in the ROI part, thus improving the accuracy of the subsequent analysis.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e2.5 ResNet-50 based Model Architecture\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eWe applied a transfer learning strategy [14] using ResNet50 [15], a convolutional neural network pre-trained on ImageNet [16], for binary classification of echocardiographic images. ResNet50, with multiple residual blocks and a 2048-unit output layer, served as the backbone for extracting local and global features from B-mode ultrasound images. To adapt it to our task, we added a 512-unit fully connected layer with ReLU activation, followed by a dropout layer (rate = 0.2) to reduce overfitting. The final layer was a 2-unit fully connected layer mapping to the two output classes. As illustrated in\u0026nbsp;\u003cstrong\u003eFigure 2\u003c/strong\u003e, the overall architecture uses ResNet50 for feature extraction, followed by nonlinear transformation and classification layers to produce class probability predictions. Dropout regularization helps prevent overfitting and improves model generalization.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e2.6 Experiment Setting\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eThis study presents experiments incorporating adjusted class weights during training and compares them to experiments without class weights to assess their impact. As shown in \u003cstrong\u003eTable 1\u003c/strong\u003e, the current dataset suffers from a data imbalance issue. In binary classification tasks, a significant imbalance between the positive and negative sample sizes can cause the model to be biased toward learning the patterns of the majority class while neglecting the important information from the minority class. Specifically, we assigned each class a weight coefficient that is inversely proportional to the number of samples in that class. The weight for class was calculated as follows:\u003c/p\u003e\n\u003cp\u003e\u003cimg width=\"47\" height=\"34\" src=\"data:image/png;base64,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\" alt=\"image\"\u003e\u003c/p\u003e\n\u003cp\u003eDuring training, we select the model checkpoint corresponding to the highest balanced validation accuracy and evaluate its performance on the test set. The settings for the two experiments are the same: the Adam optimizer [17] is adopted with a learning rate of 10\u003csup\u003e-3\u003c/sup\u003e, the batch size is set to 32, and the maximum number of training epochs is 150. In one experiment, cross-entropy loss is guided by the adjusted class weights during model training, while in the other experiment, the experiment without class weights does not apply any weight adjustment and treats all samples equally during training. Both experiments are conducted with the PyTorch framework [18].\u003c/p\u003e"},{"header":"3. Results","content":"\u003cp\u003eThe basic demographic characteristics of the PASC and pre-COVID groups are listed in \u003cstrong\u003eTable\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003e1\u003c/strong\u003e. The PASC group consisted of 152 boys (56.3%) and 118 girls (43.7%), with a mean age of 9.08 ± 4.74 years. The pre-COVID group included 212 boys (53%) and 188 girls (47%), with a mean age of 8.91 ± 4.66 years. There was no statistically significant difference between these two groups. The various PASC symptoms are detailed in\u003cstrong\u003e\u0026nbsp;Supplemental Figure 1\u003c/strong\u003e. Among them, the most distinctive symptoms after fatigue were shortness of breath (35.2%), chest pain (31.9%), and palpitations (31.1%). Based on these findings, we selected the pre-COVID patients with the same chief complaint symptom complex for comparison. The results of blood tests in the PASC group were shown in\u003cstrong\u003eS\u003c/strong\u003e\u003cstrong\u003eupplemental\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003eT\u003c/strong\u003e\u003cstrong\u003eable 1\u003c/strong\u003e. All the patients in PASC group showed no abnormality including complete blood cell count, inflammatory biomarkers (erythrocyte sedimentation rate\u0026nbsp;(ESR)\u0026nbsp;as well as\u0026nbsp;high-sensitivity\u0026nbsp;C-reactive protein\u0026nbsp;(hsCRP), creatine phosphokinase\u0026nbsp;(CPK),\u0026nbsp;serum ferritin and D-dimer) and liver function.\u003c/p\u003e\n\u003cp\u003eThe conventional echocardiography parameters we compared were listed in \u003cstrong\u003eTable\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003e3\u003c/strong\u003e. These parameters including Left Ventricular Diastolic Internal Dimension (LVIDd), left ventricular systolic internal dimension (LVIDs), Left Atrium Dimension (LA), Aortic Root Dimension (Ao), Ejection Fraction (EF), Fractional Shortening (FS), Mitral valve E-point septal separation(EPSS) , LA/AO ratio to evaluation the cardiac systolic function and chambers size. Left ventricular dimensions showed no difference between the 2 groups of patients. Patient with PASC had significant lower percentages of EF and FS, and significant higher dimension of LA and longer length of MV EPSS than pre-COVID children, but these data of each parameter measured in both groups were all within normal range.\u003c/p\u003e\n\u003cp\u003eThe training and evaluation results of the ResNet-50 based model without adjusted class weights are showed in \u003cstrong\u003eFig. 3a\u003c/strong\u003e and the confusion matrix of the test set is showed in \u003cstrong\u003eFig. 3b\u003c/strong\u003e, and\u003cstrong\u003e\u0026nbsp;Fig 3c\u003c/strong\u003e reveals the AUC of the ROC curve is 0.9937. The results of test performance, such as accuracy (97.2%), sensitivity (98.4%), specificity (89.9%), and the balanced accuracy (94.1%.) are shown in \u003cstrong\u003eTable 4\u003c/strong\u003e. \u003cstrong\u003eFig. 4\u003c/strong\u003e shows the training curve\u0026nbsp;(\u003cstrong\u003eFig. 4a\u003c/strong\u003e), confusion matrix (\u003cstrong\u003eFig. 4b\u003c/strong\u003e), ROC curve (\u003cstrong\u003eFig. 4c\u003c/strong\u003e) experiment results with class weights. The results of test performance, such as accuracy (96.6%), sensitivity (96.7%), specificity (96.2%), and the balanced accuracy (96.4%.) are also shown in \u003cstrong\u003eTable 4\u003c/strong\u003e.\u003c/p\u003e"},{"header":"4. Discussion","content":"\u003cp\u003eThe findings from this study revealed AI detectable differences in echocardiographic images between children with PASC and those from the pre-COVID control group. These results suggest that certain changes in pediatric PASC patients, which are undetectable by conventional echocardiographic imaging, can be identified using machine learning. Unlike traditional image evaluation based on human perception, machine learning enables the detection of subtle cardiac abnormalities.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e4.1 Disease Characteristics in COVID-19 Patients\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eEmerging literature has documented various cardiovascular impacts of PASC in adults. Several studies have highlighted that, in PASC patients, conventional echocardiographic measures such as EF and FS are often normal, whereas subtle abnormalities—most notably, a reduction in global longitudinal strain (GLS)—can serve as early indicators of myocardial involvement [6, 7, 10, 19, 20]. For instance, one multicenter study found that 8.2% of adult patients without prior cardiovascular comorbidities exhibited echocardiographic abnormalities, with reduced GLS being the predominant feature, especially in male patients [7]. However, despite these findings in adults, pediatric data remain scarce.\u003c/p\u003e\n\u003cp\u003eA recently published report [9] assessing the cardiovascular effects of SARS-CoV-2 infection in adults found no significant differences in conventional echocardiographic parameters across disease severities, yet this study did not include a pediatric cohort. Given the gap in understanding PASC’s impact on children, our study contributes valuable insights by focusing on pediatric patients, not only reinforces the established findings in adults but also addresses the critical gap in pediatric research, thereby broadening our understanding of PASC’s cardiovascular impact.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e4.2 Imaging-based Long-term Cardiac Monitoring in PASC\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eTraditionally, speckle tracking using dynamic 2-dimensional grayscale echocardiographic images has been regarded as a reliable method for assessing both global and regional ventricular function [21]. Advanced echocardiographic techniques, such as speckle-tracking echocardiography (STE), have been increasingly recognized for their ability to detect subtle myocardial changes in both the acute and PASC phases [22]. However, these methods require high-quality dynamic imaging, involve complex imaging protocols, and necessitate extensive post-processing procedures, which may limit their routine clinical use.\u003c/p\u003e\n\u003cp\u003eIn contrast, our study uniquely employs deep learning models to analyze non-dynamic B-mode echocardiographic images—images that are routinely acquired in clinical practice. This innovative approach simplifies image acquisition while potentially enhancing the detection of subtle cardiac changes between PASC and pre-COVID patients. By leveraging advanced artificial intelligence techniques, our method may offer higher accuracy and sensitivity compared to conventional echocardiographic assessments.\u003c/p\u003e\n\u003cp\u003eFurthermore, a review by Australian physicians [23] highlights the value of advanced echocardiographic techniques, such as 2D strain imaging, in revealing subtle myocardial damage in PASC patients. Although these advanced methods have proven effective, their complexity often restricts widespread clinical application. Our approach not only aligns with current trends in using artificial intelligence for diagnostic enhancement, but also extends the investigation to a pediatric population. By applying deep learning to conventional B-mode images, our study offers a more accessible and scalable solution for long-term cardiac monitoring in children with PASC.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e4.3 Optimization of the ResNet-50 Model Using Class Weights\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eThe results from\u0026nbsp;\u003cstrong\u003eFigures 3 and 4\u003c/strong\u003e demonstrate the high diagnostic performance of our ResNet-50-based model, with AUC values consistently close to 1.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTo further improve the model's performance, we introduced class weights during training to effectively address the challenge of class imbalance. This modification significantly enhanced the model's specificity and balanced accuracy, ensuring a more reliable application in clinical settings where both sensitivity and specificity are critical.\u003c/p\u003e\n\u003cp\u003eOur study tackled the inherent class imbalance between the PASC and pre-COVID groups by assigning higher weights to the minority class (PASC group) during model training. Without this adjustment, the model tended to be biased toward the majority class, potentially compromising its clinical utility. By applying class weight adjustments, the model achieved improved specificity and balanced accuracy, highlighting the importance of tailored machine learning strategies in optimizing AI-assisted diagnostic tools.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003e4\u003c/strong\u003e compares the performance metrics between the two experimental setups. The baseline, which predicts all samples as belonging to the majority class (the PASC group), serves as the reference. The experiment without class weights yielded higher sensitivity and overall accuracy but suffered from reduced specificity, indicating a bias toward the positive class. In contrast, incorporating class weights resulted in a more balanced performance across all metrics, with higher balanced accuracy and improved specificity. This demonstrates that the use of class weights not only mitigates the impact of class imbalance but also ensures a more stable and consistent diagnostic performance.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e4.4 The potential of AI diagnosis\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eIn our study, the left atrial dimension was significantly larger in the PASC group compared to the pre-COVID group, although the LA/AO ratio remained within the normal range. This finding suggests that pediatric PASC patients may experience mild structural changes in the left atrium as a compensatory mechanism to maintain cardiac function under potential hemodynamic or inflammatory stress, while the preserved LA/AO ratio indicates that these changes have not yet led to overt cardiac dysfunction. Future longitudinal studies are warranted to determine whether these early alterations could serve as a precursor to long-term cardiovascular complications.\u003c/p\u003e\n\u003cp\u003eSimilarly, although differences in EF, FS, and MV EPSS reached statistical significance, all values remained within normal limits, implying the presence of subclinical cardiac changes or adaptive mechanisms in pediatric PASC patients. These subtle differences, though not immediately clinically evident, underscore the importance of continuous cardiovascular monitoring to detect potential long-term effects.\u003c/p\u003e\n\u003cp\u003eWhile conventional echocardiographic parameters often appear normal in post-COVID conditions, advanced imaging techniques such as speckle tracking have revealed subtle abnormalities. In this context, our study demonstrates the potential of artificial intelligence to detect these nuanced yet statistically significant changes. The high sensitivity and specificity of our ResNet-50-based model indicate that AI could play a pivotal role in the early identification of PASC-related cardiac alterations, offering pediatricians a practical tool to monitor changes over time.\u003c/p\u003e\n\u003cp\u003eOur findings suggest that AI-driven analysis can transform pediatric cardiovascular care by identifying subtle echocardiographic changes that traditional methods may overlook. Although our model shows promising performance based on high-quality single-center data, further validation in larger, multicenter cohorts is necessary to establish its generalizability and clinical utility across diverse populations.\u003c/p\u003e\n\u003cp\u003eArtificial intelligence has been widely applied in medical image analysis, including echocardiography [24, 25]. Most applications of AI in echocardiography have focused on detecting structural or anatomical lesions. Our previous work using AI to detect congenital ventricular septal defects [26] and acquired coronary aneurysms in Kawasaki Disease [27] has yielded promising results, as has AI-assisted assessment of valvular heart diseases [28]. To the best of our knowledge, this is the first study to utilize AI to detect subtle changes in echocardiographic images. Additionally, these findings may shed new light on the natural course of PASC/COVID syndrome.\u003c/p\u003e\n\u003cp\u003eThe clinical implications of our study are significant. The higher incidence of cardiac abnormalities in children with PASC underscores the need for ongoing cardiovascular monitoring in this population. Early detection and intervention can mitigate potential long-term cardiovascular sequelae, ultimately improving overall health outcomes. Pediatricians and cardiologists should be aware of these risks and consider incorporating routine echocardiographic assessments for children presenting with PASC.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e4.5 Limitations and Future Perspectives\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eThis study has several strengths, including its study design, the use of a well-defined control group, and the application of advanced deep learning techniques to analyze echocardiographic data. These factors contribute to the robustness and reliability of our findings. However, there are also limitations to consider. First, the study was conducted at a single medical center, which may limit the generalizability of the results to other populations and settings. Second, the retrospective nature of the control group data collection might introduce selection bias, despite our efforts to match the groups based on symptom presentation and exclusion criteria. Finally, while the deep learning model provided high accuracy, but the black-box nature of deep learning makes the realization of its working process difficult! It is essential to validate these findings with larger, multicenter studies to confirm their applicability across diverse patient populations.\u003c/p\u003e\n\u003cp\u003eFurther research should explore the long-term cardiovascular outcomes of children with PASC, with larger and more diverse populations. Investigating the underlying mechanisms of cardiac involvement, including potential inflammatory and autoimmune responses, could inform targeted therapeutic interventions. Continued integration of deep learning in medical imaging will enhance diagnostic accuracy and predictive capabilities.\u003c/p\u003e"},{"header":"5. Conclusion","content":"\u003cp\u003eIn conclusion, our findings highlight the value of deep learning in pediatric echocardiographic evaluation, providing a novel and efficient approach to assess cardiac health in children with PASC. By integrating AI-driven image analysis with conventional clinical assessments, healthcare providers can enhance diagnostic accuracy and ultimately improve outcomes for affected children. Future studies should validate these findings in larger, more diverse cohorts and further explore the underlying biological mechanisms driving these echocardiographic changes.\u003c/p\u003e\n\u003cp\u003e\u003cbr\u003e\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cp\u003e\u003cstrong\u003eACE2\u0026nbsp;\u0026nbsp;\u0026nbsp;\u003c/strong\u003eAngiotensin-Converting Enzyme 2\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAI\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;\u0026nbsp;\u003c/strong\u003eArtificial Intelligence\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAUC\u0026nbsp; \u0026nbsp;\u0026nbsp;\u0026nbsp;\u003c/strong\u003eArea Under the Curve\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eB-mode\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;\u003c/strong\u003eBrightness Mode\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCOVID-19\u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;\u003c/strong\u003eCoronavirus Disease 2019\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCPK\u0026nbsp; \u0026nbsp;\u0026nbsp;\u0026nbsp;\u003c/strong\u003eCreatine Phosphokinase\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDL\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;\u003c/strong\u003eDeep Learning\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eECG\u0026nbsp; \u0026nbsp;\u0026nbsp;\u003c/strong\u003eElectrocardiogram\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEF\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;\u003c/strong\u003eEjection Fraction\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEPSS \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u003c/strong\u003eE-point Septal Separation\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eESR\u0026nbsp; \u0026nbsp;\u0026nbsp;\u0026nbsp;\u003c/strong\u003eErythrocyte Sedimentation Rate\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFS\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;\u003c/strong\u003eFractional Shortening\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eGLS\u0026nbsp; \u0026nbsp;\u0026nbsp;\u0026nbsp;\u003c/strong\u003eGlobal Longitudinal Strain\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ehsCRP\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;\u0026nbsp;\u003c/strong\u003eHigh-Sensitivity C-Reactive Protein\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eLA\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;\u003c/strong\u003eLeft Atrium\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eLV\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;\u003c/strong\u003eLeft Ventricle\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eLVIDd\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;\u0026nbsp;\u003c/strong\u003eLeft Ventricular Internal Diameter at End-Diastole\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eLVIDs\u0026nbsp;\u0026nbsp;\u003c/strong\u003eLeft Ventricular Internal Diameter at End-Systole\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eM-mode\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;\u0026nbsp;\u003c/strong\u003eMotion Mode\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMV\u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;\u003c/strong\u003eMitral Valve\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ePACS\u0026nbsp;\u0026nbsp;\u0026nbsp;\u003c/strong\u003ePicture Archiving and Communication System\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ePASC\u0026nbsp;\u0026nbsp;\u0026nbsp;\u003c/strong\u003ePost-Acute Sequelae of SARS-CoV-2 Infection\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eReLU\u0026nbsp;\u0026nbsp;\u0026nbsp;\u003c/strong\u003eRectified Linear Unit\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eROC\u0026nbsp; \u0026nbsp;\u0026nbsp;\u003c/strong\u003eReceiver Operating Characteristic\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eROI\u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;\u003c/strong\u003eRegion of Interest\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eRT-PCR\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;\u003c/strong\u003eReverse Transcriptase Polymerase Chain Reaction\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eSTE\u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;\u003c/strong\u003eSpeckle-Tracking Echocardiography\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eVSD \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;\u003c/strong\u003eVentricular Septal Defect\u003cstrong\u003e\u003cbr\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study was conducted in accordance with the Declaration of Helsinki and was approved by the Ethics Committee of China Medical University Children’s Hospital (Approval numbers: CMUH111-REC2-113 and CMUH111-REC2-122). Written informed consent was obtained from the legal guardians of all pediatric participants in the post-COVID group. For the retrospective control group, the requirement for informed consent was waived by the IRB due to the use of de-identified imaging data.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for publication\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable. This manuscript does not include identifiable individual data, images, or videos.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and material\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe datasets generated or analyzed during the current study are not publicly available due to institutional and ethical restrictions regarding the sharing of patient medical data, but are available from the corresponding author on reasonable request.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis research was supported by the following grants: NSTC 113-2314-B-039-057 from the National Science and Technology Council, Taiwan; a research grant (1JA8) from the Center for Allergy, Immunology, and Microbiome (A.I.M.), China Medical University Hospital, Taichung, Taiwan; and DMR-113-114, DMR-114-024, and DMR-114-109 from China Medical University Hospital, Taichung, Taiwan.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthors' contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eYCP collected and interpreted the clinical data. YJH and JCW developed and implemented the deep learning algorithm, and assisted in visualization and model validation. \u0026nbsp;XLL contributed to data curation and manuscript writing. PYL, YLH, and PCC coordinated clinical data collection and project administration. LSHW provided methodological guidance. HJT and WWC participated in manuscript review and editing. KSH, HSL, and JYW supervised the project and contributed to study design, funding acquisition, and final manuscript revision. \u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAll authors read and approved the final manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eI sincerely thank all members of this research project for their dedication, expertise, and commitment, which were instrumental in achieving our research objectives. Their valuable insights, collaborative spirit, and unwavering support greatly contributed to this study.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthors' information (optional)\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eThaweethai T, Jolley SE, Karlson EW, Levitan EB, Levy B, McComsey GA et al (2023) Development of a Definition of Postacute Sequelae of SARS-CoV-2 Infection. JAMA 329:1934\u0026ndash;1946\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSherif ZA, Gomez CR, Connors TJ, Henrich TJ, Reeves WB, Force RMPT (2023) Pathogenic mechanisms of post-acute sequelae of SARS-CoV-2 infection (PASC). eLife 12:e86002\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRao S, Gross RS, Mohandas S, Stein CR, Case A, Dreyer B et al (2024) Postacute Sequelae of SARS-CoV-2 in Children. Pediatrics 153:e2023062570\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRao S, Lee GM, Razzaghi H, Lorman V, Mejias A, Pajor NM et al (2022) Clinical Features and Burden of Postacute Sequelae of SARS-CoV-2 Infection in Children and Adolescents. JAMA Pediatr 176:1000\u0026ndash;1009\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eOlszanecka A, Wojciechowska W, Bednarek A, Kusak P, Wizner B, Terlecki M et al (2023) Serial echocardiographic evaluation of COVID-19 patients without prior history of structural heart disease: a 1-year follow-up CRACoV-HHS study. Front Cardiovasc Med 10:1230669\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSirico D, Di Chiara C, Costenaro P, Bonfante F, Cozzani S, Plebani M et al (2022) Left ventricular longitudinal strain alterations in asymptomatic or mildly symptomatic paediatric patients with SARS-CoV-2 infection. Eur Heart J Cardiovasc Imaging 23:1083\u0026ndash;1089\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGarcia-Zamora S, Picco JM, Lepori AJ, Galello MI, Saad AK, Ay\u0026oacute;n M et al (2023) Abnormal echocardiographic findings after COVID-19 infection: a multicenter registry. Int J Cardiovasc Imaging 39:77\u0026ndash;85\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLi Y, Li H, Zhu S, Xie Y, Wang B, He L et al (2020) Prognostic Value of Right Ventricular Longitudinal Strain in Patients With COVID-19. JACC Cardiovasc Imaging 13:2287\u0026ndash;2299\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eFlores R, Pires O, Alves J, Pereira VH (2023) An Echocardiographic Insight Into Post-COVID-19 Symptoms. Cureus 15:e38039\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eVidal-Perez R, Grapsa J, Bouzas-Mosquera A, Fontes-Carvalho R, Vazquez-Rodriguez JM (2023) Current role and future perspectives of artificial intelligence in echocardiography. World J Cardiol 15:284\u0026ndash;292\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAnon (2013) Diagnostic Ultrasound Imaging: Inside Out\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHe K, Zhang X, Ren S, Sun J (2016) Deep Residual Learning for Image Recognition. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)., :770\u0026ndash;778\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDeng J, Dong W, Socher R, Li L-J, Li K, Fei-Fei L (2009) ImageNet: A large-scale hierarchical image database. In: 2009 IEEE Conference on Computer Vision and Pattern Recognition., :248\u0026ndash;255\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKingma DP, Ba J, Adam (2017) A Method for Stochastic Optimization\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePaszke A, Gross S, Massa F, Lerer A, Bradbury J, Chanan G et al (2019) PyTorch: An Imperative Style, High-Performance Deep Learning Library\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKirkpatrick JN, Swaminathan M, Adedipe A, Garcia-Sayan E, Hung J, Kelly N et al (2023) American Society of Echocardiography COVID-19 Statement Update: Lessons Learned and Preparation for Future Pandemics. J Am Soc Echocardiogr 36:1127\u0026ndash;1139\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDs JW, Y K, G T-A Y, Sf N Preserved left ventricular twist and circumferential deformation, but depressed longitudinal and radial deformation in patients with diastolic heart failure. Eur Heart J 2008;29\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKersten J, Schellenberg J, Jerg A, Kirsten J, Persch H, Liu Y et al (2023) Strain Echocardiography in Acute COVID-19 and Post-COVID Syndrome: More than Just a Snapshot. Biomedicines 11:1236\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAltersberger M, Goliasch G, Khafaga M, Schneider M, Cho Y, Winkler R et al (2023) Echocardiography and Lung Ultrasound in Long COVID and Post-COVID Syndrome, a Review Document of the Austrian Society of Pneumology and the Austrian Society of Ultrasound in Medicine. J Ultrasound Med 42:269\u0026ndash;277\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eVidal-Perez R, Grapsa J, Bouzas-Mosquera A, Fontes-Carvalho R, Vazquez-Rodriguez JM (2023) Current role and future perspectives of artificial intelligence in echocardiography. World J Cardiol 15:284\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChen S-H, Wang C-W, Tai I-H, Weng K-P, Chen Y-H, Hsieh K-S (2021) Modified YOLOv4-DenseNet Algorithm for Detection of Ventricular Septal Defects in Ultrasound Images. Int J Interact Multimedia Artif Intell 6:101\u0026ndash;108\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKuo H-C, Chen S-H, Chen Y-H, Lin Y-C, Chang C-Y, Wu Y-C et al (2022) Detection of coronary lesions in Kawasaki disease by Scaled-YOLOv4 with HarDNet backbone. Front Cardiovasc Med 9:1000374\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eNedadur R, Wang B, Tsang W (2022) Artificial intelligence for the echocardiographic assessment of valvular heart disease. Heart 108:1592\u0026ndash;1599\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAkkus Z, Aly YH, Attia IZ, Lopez-Jimenez F, Arruda-Olson AM, Pellikka PA et al (2021) Artificial Intelligence (AI)-Empowered Echocardiography Interpretation: A State-of-the-Art Review. J Clin Med 10:1391\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKarlsen S, Dahlslett T, Grenne B, Sj\u0026oslash;li B, Smiseth O, Edvardsen T et al (2019) Global longitudinal strain is a more reproducible measure of left ventricular function than ejection fraction regardless of echocardiographic training. Cardiovasc Ultrasound 17:18\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBarry T, Farina JM, Chao C-J, Ayoub C, Jeong J, Patel BN et al (2023) The Role of Artificial Intelligence in Echocardiography. J Imaging 9:50\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHsu Y-L, Chen P-C, Tsai Y-F, Wei C-H, Wu LS-H, Hsieh K-S et al (2024) Clinical Features and Vaccination Effects among Children with Post-Acute Sequelae of COVID-19 in Taiwan. Vaccines (Basel) 12:910\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eCaruana R (1997) Multitask Learning. Mach Learn 28:41\u0026ndash;75\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"China Medical University Children's Hospital","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Post-acute sequelae of SARS-CoV-2 infection (PASC), children, echocardiography, Artificial intelligence, Deep learning, Pediatric cardiology","lastPublishedDoi":"10.21203/rs.3.rs-7134718/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7134718/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eBackground:\u003c/strong\u003e Post-acute sequelae of SARS-CoV-2 infection (PASC) is characterized by persistent symptoms following SARS-CoV-2 infection. Children with PASC are at risk of developing cardiac complications. Echocardiography has been instrumental in identifying cardiac abnormalities. This study applies deep learning to enhance the detection and understanding of echocardiographic changes in children with PASC.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods: \u003c/strong\u003eA case-control study was conducted at a pediatric tertiary center in central Taiwan. Children under 18 years who tested positive for SARS-CoV-2 and experienced symptoms for longer than four weeks were recruited between July 1, 2022, and July 31, 2023, during the Omicron variant surge. Echocardiographic data were also collected from a control group, consisting of children who presented with similar symptoms and received medical care in the same pediatric tertiary center in 2018. Children with congenital or structural heart disease, inflammatory conditions, or arrhythmias were excluded. Echocardiographic images were analyzed using a ResNet-50-based deep learning model to identify cardiac abnormalities.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults: \u003c/strong\u003eA total of 270 children with PASC and 400 age-matched control children were included. No abnormalities were detected in the PASC group using conventional echocardiographic analysis. The deep learning model achieved an accuracy of 96.6%, sensitivity of 96.7%, specificity of 96.2%, and balanced accuracy of 96.4%.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusion:\u003c/strong\u003e AI-assisted echocardiographic analysis demonstrated high performance in distinguishing cardiac function between PASC and controls. Deep learning models enhance the detection of subtle cardiac changes in children with PASC.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCritical relevance statement: \u003c/strong\u003eDeep learning enhances the detection of subtle cardiac abnormalities in children with post-COVID syndrome, improving diagnostic sensitivity beyond conventional echocardiographic interpretation.\u003c/p\u003e","manuscriptTitle":"Echocardiographic Evaluation in Children with Post-Acute Sequelae of SARS-CoV-2 infection Using Deep Learning","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-07-18 15:05:06","doi":"10.21203/rs.3.rs-7134718/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"56180e89-e9ba-4b40-9db1-75884f3490b0","owner":[],"postedDate":"July 18th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":51595460,"name":"Pediatrics"},{"id":51595461,"name":"Artificial Intelligence and Machine Learning"}],"tags":[],"updatedAt":"2025-07-18T15:05:07+00:00","versionOfRecord":[],"versionCreatedAt":"2025-07-18 15:05:06","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7134718","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7134718","identity":"rs-7134718","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}