A Multitask Deep Learning Model for Pediatric Echocardiography Analysis Authors: Joseph Cho, Mrudang Mathur, Dhamanpreet Kaur, Matthew Duda, Adil Dahlan, Aravind Krishnan, Matthew Leipzig, Rohan Shad, Alexander K. Gonzalez, Joseph Logan, Christa Seidman, Robyn Fong, Abhinav Kumar, Cyril Zakka, Elizabeth Carter, Asif Padiyath, Andrea Jones, Michael D. Quartermain, Curtis P. Langlotz, Matthew A. Jolley, William Hiesinger Published: 2026 ABSTRACT BACKGROUND:Congenital heart defects afflict ≈1% of all births worldwide. Although deep learning has shown significant promise in automating and improving adult echocardiography analysis, existing pediatric-based models are often limited to single tasks and specific echocardiographic views. To address this, we introduce EchoAI-Peds, a multitask deep learning model for pediatric echocardiography. Our model was developed using the most comprehensive set of pediatric labels to date and is designed to integrate information from multiple views simultaneously.METHODS:We trained a video-based vision transformer to simultaneously detect 28 congenital heart defects, structural and functional abnormalities, repairs, and interventions directly from complete pediatric echocardiography studies with multiple videos. Our model was developed using >700 000 videos derived from >12 000 studies performed at Stanford Medicine between 2014 and 2021. Specifically, our model was trained on 10 815 studies (median age, 8 [IQR 2–14] years; 45% female) and validated on 1294 studies (median age, 9 [IQR 2–15] years; 46% female). Model efficacy was tested on an internal held-out data set of 1336 studies from that same period. In addition, model generalizability was tested on a spatially and temporally distinct patient cohort at the Children’s Hospital of Philadelphia using 2121 studies performed between 2024 and 2025.RESULTS:Our model achieved macroaveraged area under the receiver operating characteristic curve (AUROC) values of 0.91 (95% CI, 0.90–0.92) on the internal test set (median age, 8 [IQR 2–14] years; 47% female) and AUROC values of 0.89 (95% CI, 0.88–0.90) on the external test set (median age, 6 [IQR 0.92–13] years; 44.4% female). Moreover, EchoAI-Peds significantly outperformed adult-based echocardiography foundation models trained on substantially larger data sets, including EchoCLIP (internal AUROC=0.58, 95% CI, 0.56–0.60; external AUROC=0.61, 95% CI, 0.59–0.62) and EchoPrime (internal AUROC=0.58, 95% CI, 0.57–0.61; external AUROC=0.60, 95% CI, 0.59–0.62). Finally, our model demonstrated robust performance across patient age, patient sex, and studies with varying numbers of videos.CONCLUSIONS:Our findings demonstrate the remarkable potential for multitask deep learning models to aid the interpretation of pediatric echocardiograms. In addition, our results underscore the need for models that are specifically tailored to pediatric populations. Clinical Perspective What Is New? We present EchoAI-Peds, a multitask deep learning model that simultaneously detects 28 congenital heart defects, structural and functional abnormalities, repairs, and interventions directly from complete pediatric echocardiography studies with multiple videos. We present EchoAI-Peds, a multitask deep learning model that simultaneously detects 28 congenital heart defects, structural and functional abnormalities, repairs, and interventions directly from complete pediatric echocardiography studies with multiple videos. EchoAI-Peds significantly outperforms adult-based echocardiography foundation models trained on substantially larger data sets, underscoring the importance of pediatric-specific model development. EchoAI-Peds significantly outperforms adult-based echocardiography foundation models trained on substantially larger data sets, underscoring the importance of pediatric-specific model development. What Are the Clinical Implications? EchoAI-Peds could serve as an automated screening tool in settings with limited pediatric cardiology expertise. EchoAI-Peds could serve as an automated screening tool in settings with limited pediatric cardiology expertise. EchoAI-Peds could function as a secondary reader to assist sonographers and cardiologists in pediatric echocardiogram interpretation. EchoAI-Peds could function as a secondary reader to assist sonographers and cardiologists in pediatric echocardiogram interpretation. Editorial, see p 801 Congenital heart defects (CHDs) afflict ≈1% of all births worldwide and represent the most common category of birth anomaly.1 They comprise an exceptionally heterogeneous group of structural cardiac malformations that differ widely in anatomical complexity, physiological impact, and clinical course.2 Accurate diagnosis of CHDs can be particularly challenging because of ongoing pediatric cardiac growth and complex postoperative anatomy after intervention.3 Echocardiography remains the principal modality for evaluation and follow-up because of its safety, accessibility, and ability to provide real-time assessment of cardiac structure and function.4,5 However, pediatric echocardiogram interpretation requires significant expertise and time, whereas shortages of trained specialists and persistent interobserver and intraobserver variability contribute to delays in patient evaluation and diagnostic inconsistency.5–7 Notably, deep learning has demonstrated significant potential in automating and improving echocardiographic analyses.8 For adult populations, models have progressed from those trained on relatively small data sets restricted to a single task and specific echocardiographic views9,10 to large-scale systems capable of performing numerous tasks using a wider array of views.11–13 However, similar progress has not been observed for pediatric populations. Pediatric echocardiography presents unique challenges because of a greater variation in heart rate, limited patient cooperation, and a wide spectrum of congenital and acquired heart diseases.14 Consequently, adult-based models have failed to generalize to pediatric populations.15 Despite this, existing pediatric-based models are often limited to small training data sets, narrow diagnostic scopes, and specific views, thus limiting their broader clinical utility.15–17 To address this gap, we developed EchoAI-Peds, a multitask deep learning model that simultaneously detects 28 CHDs, structural and functional abnormalities, repairs, and interventions directly from complete pediatric echocardiography studies with multiple videos. We sought to evaluate model performance on an internal held-out test set at Stanford Medicine (SM) and evaluate model generalizability on a spatially and temporally distinct patient cohort at the Children’s Hospital of Philadelphia (CHOP). Finally, we aimed to assess model robustness across patient age, patient sex, and the number of available videos per study. Methods This study adheres to the TRUE-AIM (Transparent Reporting and Evaluation of Artificial Intelligence Models) reporting recommendations,18 and a completed checklist is provided as Supplemental Material. Data Availability and Software Pediatric echocardiography studies from SM and the CHOP used to train, validate, and test our model are not publicly available becaue of patient privacy considerations. Data usage and ethical approval for this study was granted at both SM (institutional review board [IRB] No. 52440) and CHOP (IRB No. 25-023534) by their respective IRBs. Given the retrospective nature of this study, informed consent was waived at SM and CHOP by their respective IRBs. The inference code and model weights are publicly available at https://github.com/joseph-y-cho/echoai-peds. Data Sources We conducted a retrospective, diagnostic model development study with both internal and external validation using pediatric echocardiography studies collected from multiple institutions. Specifically, we collected 14 552 pediatric echocardiography studies that were conducted between 2014 and 2021 at SM (Figure 1A). We included studies performed at the Lucile Packard Children’s Hospital, for which no age restriction was applied, as well as those from Stanford Healthcare, in which patients were ≤18 years of age. At CHOP, we collected 2137 pediatric echocardiography studies that were conducted between 2024 and 2025 (Figure 1B). We did not exclude any studies based on patient age at CHOP. Overview of EchoAI-Peds framework. A, We collected pediatric echocardiography (echo.) studies from Stanford Medicine (SM) for model development. We excluded any study from patients with a previous history at the Children’s Hospital of Philadelphia (CHOP) and those with inadequate echocardiogram videos. The remaining studies were split into training, validation, and internal testing sets. B, Similarly, we collected pediatric echocardiography studies at CHOP. We excluded any study with a previous history at SM as well as those with inadequate echocardiogram videos or reports. The remaining studies served as our external test set. C, For training, one video was randomly selected from each study, from which one 16-frame clip was randomly selected. We trained the vision transformer (ViT) in EchoAI-Peds by minimizing a multilabel binary cross entropy (BCE) loss function. Loss values were computed between predictions derived from those selected clips and labels extracted from the corresponding clinical reports using a frozen, medically fine-tuned large language model (LLM). D, For inference, up to 4 nonoverlapping 16-frame clips were randomly sampled across all available videos within each study. Using the frozen, trained ViT, we first averaged predictions across clips within each video and subsequently averaged across all videos to simultaneously detect 28 clinical findings. ARD indicates aortic root dilation; ASD, atrial septal defect; BAV, bicuspid aortic valve; PDA, patent ductus arteriosus; PFO, patent foramen ovale; TOF, Tetralogy of Fallot; and VSD, ventricular septal defect. At both institutions, each study was accompanied by a text report containing clinical findings verified by a board-certified cardiologist. Patient demographics for both data sources are summarized in the Table. Further details on the data sources are found in Supplemental Section 1.1. Table.EchoAI-Peds Development and Evaluation Cohort Characteristics. EchoAI-Peds Development and Evaluation Cohort Characteristics. Definition of Outcomes Clinical labels for each pediatric echocardiography study performed at SM and CHOP were derived from their corresponding text reports. Outcomes of interest included 28 CHDs, structural and functional abnormalities, repairs, and interventions, each predicted individually. As a patient may present with multiple concurrent conditions, outcomes were not mutually exclusive. Further details on outcome definition are provided in Supplemental Section 1.2. Data Preprocessing Before model training, we preprocessed our internal and external data to remove patient overlaps between institutions and studies that described multiple patients. We then removed videos with insufficient frame counts and deidentified videos and text reports from all remaining studies. Next, we used MedGemma 27B19 to extract clinical class labels from each clinical report. We then used the identified labels for model development and evaluation. Finally, the SM data set was split at the patient level into a training set of 6549 patients (80%), a validation set of 819 patients (10%), and a held-out, internal testing set of 819 patients (10%). In contrast, we reserved studies from all 1979 patients at CHOP for external testing of our model. Further details on data preprocessing are found in Supplemental Section 1.2 and Tables S1 and S2. Model Training We trained a video-based, multiscale vision transformer20 to simultaneously predict 28 clinical findings directly from pediatric echocardiogram videos (Figure 1C). During training, we randomly selected a single video from each study, from which we then randomly sampled a 16-frame clip. This 16-frame clip was used as our model input to generate study-level predictions for all 28 clinical findings. The predictions of our model were compared against the large language model–extracted labels to compute the training loss. We initialized our model with pretrained Kinetics weights21 and minimized a multilabel binary cross-entropy loss. Please refer to Supplemental Section 1.3 for more details on model development and training. Model Evaluation We evaluated model performance on the SM and CHOP test sets using a multivideo, multiclip strategy (Figure 1D). Specifically, we randomly sampled up to 4 distinct 16-frame clips from each available video within a given study.13 We first averaged predictions across all clips from the same video and subsequently averaged the predictions across all videos to generate the study-level predictions. Because of class imbalance in the data set, we evaluated model performance using the area under the receiver operating characteristic curve (AUROC).22,23 Furthermore, we benchmarked our classification results against those of 2 state-of-the-art foundation models for adult echocardiography: EchoCLIP11 and EchoPrime.12 Please refer to Supplemental Sections 1.4–1.8 for more details on model evaluation, inference strategies, precision and recall, sensitivity and specificity, and explainability. Model performance across different inference strategies is presented in Tables S3–S6. Subgroup Analyses Age and sex are known to significantly influence cardiac structure and function.24,25 Furthermore, different echocardiographic views capture distinct anatomical features of the heart.26 Additionally, pediatric echocardiography studies significantly differ from those of adults.27 Specifically, they may include a larger number of views because of the varied nature of congenital heart disease and often contain sweeps around particular views to better visualize complex anatomy.27 Moreover, the number and quality of acquired views vary across patients and sonographers.28 As such, our model must also perform reliably across studies with differing numbers of views and videos. To assess this, we performed 3 distinct subgroup analyses based on patient age, patient sex, and the number of available echocardiogram videos per study for both the SM and CHOP test sets: Age-based analysis: We categorized studies into predefined groups per the patient’s age (in years) at the time of examination (A): A<1, 1≤A<3, 3≤A<8, 8≤A<12, 12≤A<18, and 18≤A.29 Age-based analysis: We categorized studies into predefined groups per the patient’s age (in years) at the time of examination (A): A<1, 1≤A<3, 3≤A<8, 8≤A<12, 12≤A<18, and 18≤A.29 Sex-based analysis: We categorized studies by the patient’s recorded sex and excluded those without a recorded sex. Sex-based analysis: We categorized studies by the patient’s recorded sex and excluded those without a recorded sex. View count–based analysis: We categorized studies into predefined groups per the number of available echocardiogram videos (N): N<25, 25≤N<50, 50≤N<100, and 100≤N. View count–based analysis: We categorized studies into predefined groups per the number of available echocardiogram videos (N): N<25, 25≤N<50, 50≤N<100, and 100≤N. In each of the above analyses, we included 5 clinical labels that had sufficient representation across all subgroups at both institutions, with a minimum of 2 positive cases per group. Statistical Analyses We used nonparametric bootstrapping with 1000 replicates to compute 95% CIs for the AUROC, area under the precision-recall curve, sensitivity, specificity, expected calibration error, and Brier score values on the SM and CHOP test sets. Next, the same bootstrapping procedure was used to obtain 2-sided P values assessing whether the AUROC differences between EchoAI-Peds and each competing model (EchoCLIP and EchoPrime) were statistically significant. In addition, we applied the Bonferroni correction to control the family-wise error rate in performing multiple comparisons. Finally, we repeated the bootstrapping procedure for AUROC calculations across subgroups stratified by patient age, patient sex, and the number of available echocardiogram videos per study. Normality of continuous variables, namely patient age, was assessed using quantile-quantile plots (Figure S1). Patients with missing race or ethnicity information were categorized as “unknown” in demographic reporting, and patients missing sex information were excluded from subgroup analyses stratified by patient sex. No imputation was performed, and missing demographic data did not affect model training or primary evaluation, as these variables were not used as model inputs. The statistical analyses were conducted in R (version 4.5) and Python (version 3.13). Results Internal Testing On the internal test set, EchoAI-Peds achieved a macroaveraged AUROC of 0.91 (95% CI, 0.90–0.92). Our model accurately detected CHDs, such as ventricular septal defect (VSD), with an AUROC of 0.89 (95% CI, 0.86–0.91), Tetralogy of Fallot, with an AUROC of 0.97 (95% CI, 0.95–0.98), and transposition of the great arteries, with an AUROC of 0.93 (95% CI, 0.90–0.95). It also performed strongly in detecting structural abnormalities, such as left ventricular dilation (LVDil.), with an AUROC of 0.93 (95% CI, 0.90–0.95), right ventricular dilation (RVDil.), with an AUROC of 0.92 (95% CI, 0.90–0.94), and hypertrophic cardiomyopathy, with an AUROC of 0.93 (95% CI, 0.87–0.98). In detecting functional abnormalities, our model also achieved high performance, including LVSysD., with an AUROC of 0.94 (95% CI, 0.91–0.96), left ventricular diastolic dysfunction, with an AUROC of 0.86 (95% CI, 0.82–0.91), and right ventricular systolic dysfunction, with an AUROC of 0.94 (95% CI, 0.91–0.96). Finally, our model performed remarkably in detecting repairs and interventions, such as VSD repair, with an AUROC of 0.95 (95% CI, 0.94–0.97), the Glenn procedure, with an AUROC of 0.99 (95% CI, 0.98–0.99), and heart transplant, with an AUROC of 0.98 (95% CI, 0.96–1.00). See Tables S6–S11 for a complete list of AUROC, area under the precision-recall curve, sensitivity and specificity values, alternative thresholding strategies, expected calibration error, and Brier score values. Please refer to Figures S2 and S3 for saliency maps and receiver operating characteristic curves of selected clinical labels, respectively. External Testing On the external test set, EchoAI-Peds achieved a macroaveraged AUROC of 0.89 (95% CI, 0.88–0.90). Our model accurately detected CHDs, such as VSD, with an AUROC of 0.83 (95% CI, 0.80–0.85), Tetralogy of Fallot, with an AUROC of 0.92 (95% CI, 0.90–0.94), and transposition of the great arteries, with an AUROC of 0.88 (95% CI, 0.86–0.91). It also performed strongly in detecting structural abnormalities such as LVDil., with an AUROC of 0.91 (95% CI, 0.88–0.94), RVDil., with an AUROC of 0.91 (95% CI, 0.89–0.92), and hypertrophic cardiomyopathy, with an AUROC of 0.95 (95% CI, 0.87–1.00). In detecting functional abnormalities, our model also achieved high performance, including LVSysD., with an AUROC of 0.95 (95% CI, 0.93–0.97), left ventricular diastolic dysfunction, with an AUROC of 0.81 (95% CI, 0.69–0.91), and right ventricular systolic dysfunction, with an AUROC of 0.93 (95% CI, 0.91–0.95). Finally, our model performed remarkably in detecting repairs and interventions, such as VSD repair, with an AUROC of 0.92 (95% CI, 0.90–0.95), the Glenn procedure, with an AUROC of 0.99 (95% CI, 0.99–0.99), and heart transplant, with an AUROC of 0.99 (95% CI, 0.98–1.00). See Tables S12 and S13 for a complete list of AUROC and area under the precision-recall curve values, respectively. Notably, model performance was comparable to that on the internal test set (Figure 2). However, model performance in detecting major aortopulmonary collateral arteries was substantially lower on the external test set, with an AUROC of 0.75 (95% CI, 0.67–0.82) versus 0.92 (95% CI, 0.88–0.96) on the internal test set. Comparison of EchoAI-Peds performance across institutions. EchoAI-Peds demonstrates comparable classification performance on the Stanford Medicine (SM) and the Children’s Hospital of Philadelphia (CHOP) test sets, as measured by area under the receiver operating characteristic curve (AUROC) values across 28 clinical findings. ARD indicates aortic root dilation; ASD, atrial septal defect; BAV, bicuspid aortic valve; DORV, double outlet right ventricle; HCM, hypertrophic cardiomyopathy; LVDiaD., left ventricular diastolic dysfunction; LVDil., left ventricular dilation; LVH, left ventricular hypertrophy; MAPCAs, major aortopulmonary collateral arteries; PDA, patent ductus arteriosus; PE, pleural effusion; PFO, patent foramen ovale; PulmAtr., pulmonary atresia; RVDil., right ventricular dilation; RVH, right ventricular hypertrophy; TGA, transposition of the great arteries; TOF, Tetralogy of Fallot; RVSysD., right ventricular systolic dysfunction; and VSD, ventricular septal defect. Comparison Against Adult Echocardiography Models We compared the performance of our model with that of state-of-the-art foundation models for adult echocardiography: EchoCLIP11 and EchoPrime.12 On the internal test set, EchoCLIP and EchoPrime obtained significantly lower macroaveraged AUROC values of 0.58 (95% CI, 0.56–0.60) and 0.58 (95% CI, 0.57–0.61), respectively (P<0.001 for both). Notably, our model achieved higher performance against both models on all 28 clinical classes (Figure 3A). Additionally, our model significantly outperformed EchoCLIP on all 28 clinical classes (P<0.001 for all) and EchoPrime on 27 of the 28 clinical classes (P<0.001 for 26 classes; P<0.01 for 1). Comparison of EchoAI-Peds against adult-based models across institutions. EchoAI-Peds substantially outperforms adult-based echocardiography foundation models (EchoCLIP and EchoPrime) in classification tasks on the Stanford Medicine (SM) test set (A) and the Children’s Hospital of Philadelphia (CHOP) test set (B), as measured by area under the receiver operating characteristic curve (AUROC) values across 28 clinical findings. ARD indicates aortic root dilation; ASD, atrial septal defect; BAV, bicuspid aortic valve; DORV, double outlet right ventricle; HCM, hypertrophic cardiomyopathy; LVDiaD., left ventricular diastolic dysfunction; LVDil., left ventricular dilation; LVH, left ventricular hypertrophy; MAPCAs, major aortopulmonary collateral arteries; PDA, patent ductus arteriosus; PE, pleural effusion; PFO, patent foramen ovale; PulmAtr., pulmonary atresia; RVDil., right ventricular dilation; RVH, right ventricular hypertrophy; TGA, transposition of the great arteries; TOF, Tetralogy of Fallot; RVSysD., right ventricular systolic dysfunction; and VSD, ventricular septal defect. On the external test set, EchoCLIP and EchoPrime also obtained significantly lower macroaveraged AUROC values of 0.61 (95% CI, 0.59–0.62) and 0.60 (95% CI, 0.59–0.62), respectively. Notably, our model achieved higher performance against both models on all 28 clinical classes (Figure 3B). Additionally, our model significantly outperformed EchoCLIP on 25 of the 28 clinical classes (P<0.001 for 24 classes; P<0.01 for 1) and EchoPrime on 27 of the 28 clinical classes (P<0.001 for 26; P<0.05 for 1). Complete lists of AUROC values for model comparison on internal and external test sets are provided in Tables S6 and S12, respectively. Complete lists of area under the precision-recall curve values for model comparison on internal and external test sets are provided in Tables S7 and S13, respectively. Detailed analyses comparing EchoCLIP and EchoPrime performance at SM and CHOP are included in Supplemental Section 1.9 and Tables S14–S16. Performance Across Subgroups To test the sensitivity of our model to patient age, patient sex, and the number of available echocardiogram videos per study, we performed 3 distinct subgroup analyses for the internal and external test sets. Age-Based Results We stratified studies from both our test sets into 6 groups based on patient age (A): group A1 (A<1), group A2 (1≤A<3), group A3 (3≤A<8), group A4 (8≤A<12), group A5 (12≤A<18), and group A6 (18≤A). On the internal test set, our model achieved consistent and strong results for VSD repair and cardiac surgery across all patient age groups. For VSD repair, AUROC values ranged from 0.93 (95% CI, 0.85–0.99) for group A5 to 0.96 for groups A1 (95% CI, 0.93–0.99), A4 (95% CI, 0.93–0.99), and A6 (95% CI, 0.91–0.99; Figure 4A). In addition, AUROC values for cardiac surgery ranged from 0.92 (95% CI, 0.87–0.96) for group A1 to 0.99 (95% CI, 0.98–1.00) for group A6. Although our model performed strongly for LVDil., RVDil., and LVSysD. across all age groups, its performance varied considerably by age. Specifically, the AUROC value for LVDil. was 0.84 (95% CI, 0.75–0.92) for group A1 but 1.00 (95% CI, 0.98–1.00) for group A6. Next, for RVDil., the AUROC value was 0.84 (95% CI, 0.78–0.89) for group A1 but 0.96 (95% CI, 0.92–0.99) for group A4. Finally, for LVSysD., the AUROC value was 0.84 (95% CI, 0.69–0.95) for group A6 but 0.98 for groups A2 (95% CI, 0.96–1.00), A3 (95% CI, 0.95–1.00), and A4 (95% CI, 0.96–0.99). Comparison of EchoAI-Peds performance across subgroups. EchoAI-Peds exhibits comparable classification performance for 5 clinical findings across subgroups defined by patient age (A; A), patient sex (B), and number of videos per study (N; C). This performance is observed in both the Stanford Medicine (SM) and the Children’s Hospital of Philadelphia (CHOP) test sets, as measured by area under the receiver operating characteristic curve (AUROC) values. LVDil. indicates left ventricular dilation; RVDil., right ventricular dilation; and VSD, ventricular septal defect. On the external test set, our model achieved consistently strong results for RVDil., LVSysD., and cardiac surgery across all age groups. For RVDil., the AUROC values ranged from 0.87 (95% CI, 0.84–0.90) for group A1 to 0.94 (95% CI, 0.91–0.96) for group A3. In addition, AUROC values for LVSysD. ranged from 0.91 for groups A1 (95% CI, 0.87–0.95) and A6 (95% CI, 0.84–0.96) to 0.98 (95% CI, 0.96–0.99) for group A5. Furthermore, the AUROC values for cardiac surgery ranged from 0.91 (95% CI, 0.85–0.96) for group A6 to 0.97 (95% CI, 0.94–0.98) for group A3. Although our model performed strongly for LVDil. and VSD repair across all age groups, its performance varied considerably. Specifically, the AUROC value for LVDil. was 0.83 (95% CI, 0.73–0.93) for group A3 but 0.97 (95% CI, 0.95–0.99) for group A4. Moreover, the AUROC value for VSD repair was 0.87 (95% CI, 0.70–0.98) for group A6 but 0.97 (95% CI, 0.95–0.98) for group A1. See Supplemental Table S17 for a complete list of AUROC values. Overall, our model demonstrated strong performance across patient age groups, suggesting that its performance was not dependent on age. Sex-Based Results Despite being trained on fewer female patients, our model achieved slightly higher performance for female patients in 3 of the 5 clinical classes: LVDil., RVDil., and LVSysD. (Figure 4B). Notably, this result was observed across both the internal and external test sets. For example, on the internal test set, our model achieved an AUROC of 0.94 (95% CI, 0.91–0.97) for females versus 0.91 (95% CI, 0.87–0.94) for boys in detecting LVDil. Similarly, on the external test set, our model achieved an AUROC value of 0.92 (95% CI, 0.88–0.95) for females versus 0.90 (95% CI, 0.85–0.94) for boys in detecting LVDil. See Table S18 for a complete list of AUROC values. Overall, the performance of our model remained consistently strong across sex, with no evidence of decline in either group. Video Count–Based Results We stratified studies from both our test sets into 4 groups based on video count (N): group N1 (N<25), group N2 (25≤N≤50), group N3 (50≤N<100), and group N4 (100≤N). On the internal test set, our model achieved consistent and strong results for LVDil., VSD repair, and cardiac surgery (Figure 4C). For LVDil., AUROC values ranged from 0.89 (95% CI, 0.80–0.97) for group N1 to 0.97 (95% CI, 0.94–0.99) for group N4. In addition, AUROC values for VSD repair ranged from 0.91 (95% CI, 0.85–0.97) for group N4 to 0.97 for groups N1 (95% CI, 0.93–0.99) and N3 (95% CI, 0.96–0.98). Furthermore, AUROC values for cardiac surgery ranged from 0.95 (95% CI, 0.88–1.00) for group N4 to 0.97 for groups N1 (95% CI, 0.94–0.99), N2 (95% CI, 0.94–0.98), and N3 (95% CI, 0.95–0.99). Although our model performed strongly for RVDil. and LVSysD. across most video-count groups, its performance varied considerably by group. First, the AUROC value for RVDil. was 0.84 (95% CI, 0.76–0.91) for group N4 but 0.94 (95% CI, 0.90–0.96) for group N2. In addition, the AUROC value for LVSysD. was 0.79 (95% CI, 0.40–1.00) for group N4 but 0.96 (95% CI, 0.94–0.98) for group N2. On the external test set, our model achieved consistent and strong results for LVDil., RVDil., LVSysD., and cardiac surgery. For LVDil., AUROC values ranged from 0.89 (95% CI, 0.84–0.93) for group N2 to 0.94 (95% CI, 0.90–0.97) for group N3. For RVDil., AUROC values ranged from 0.87 (95% CI, 0.82–0.92) for group N1 to 0.93 for groups N3 (95% CI, 0.91–0.95) and N4 (95% CI, 0.87–0.97). For LVSysD., AUROC values ranged from 0.94 for groups N1 (95% CI, 0.89–0.98) and N2 (95% CI, 0.92–0.96) to 0.96 (95% CI, 0.93–0.99) for group N3. For cardiac surgery, AUROC values ranged from 0.93 (95% CI, 0.91–0.95) for group N2 to 1.00 (95% CI, 1.00–1.00) for group N4. However, our model demonstrated strong but inconsistent performance for VSD repair. Specifically, the AUROC value was 0.88 (95% CI, 0.79–0.94) for group N1 but 1.00 (95% CI, 1.00–1.00) for group N4. See Table S19 for a complete list of AUROC values. In addition, please refer to Supplemental Section 1.10 and Table S20 for a detailed analysis of model performance in detecting LVSysD. for group N4 on the internal test set. Overall, our model achieved strong performance across video-count groups, suggesting that its performance was not dependent on the number of available videos within a given study. Discussion In this work, we introduce EchoAI-Peds, a multitask deep learning model for pediatric echocardiography. EchoAI-Peds uses a video-based vision transformer to simultaneously predict 28 clinical conditions directly from complete pediatric echocardiography studies with multiple videos. Much like expert clinicians, our model leverages information from multiple echocardiographic views simultaneously to produce unified, study-level predictions.27,30 To our knowledge, we curated the most comprehensive set of pediatric labels to date for deep learning model development and evaluation, spanning CHDs, structural and functional abnormalities, repairs, and interventions. EchoAI-Peds was developed using >12 000 studies at SM, and model performance was evaluated on an internal held-out test set. In addition, we assessed the generalizability of EchoAI-Peds on a spatially and temporally distinct patient cohort at CHOP. To emphasize the importance of pediatric-focused model development, we demonstrate that EchoAI-Peds outperforms existing foundation models for adult echocardiography that were trained on substantially larger data sets. Finally, we demonstrate that the performance of EchoAI-Peds is largely insensitive to patient age, patient sex, and the number of videos within a given study. Deep learning has made significant progress in medical image analysis for adult populations, ranging from zero-shot disease diagnosis31 to accurate prediction of long-term time-to-event patient outcomes.32 However, advances in pediatric populations remain severely limited. For example, Chen et al developed a model capable of detecting CHDs from pediatric ECGs, but it could not differentiate between specific CHD types.33 More recently, Mayourian et al used deep learning to identify CHDs from ECGs but omitted pediatric repairs or interventions.34 In pediatric echocardiography, Reddy et al introduced a model trained on ≈6000 videos to estimate left ventricular ejection fraction from apical 4-chamber and parasternal short-axis views.15 Similarly, Jiang et al developed a model trained on ≈13 000 frames to detect CHDs from one of 7 standard views.35 Notably, these models were limited to single tasks and specific views. Moreover, they were restricted to smaller data sets, in part, because of their reliance on manual physician annotations. In contrast, our work leverages recent advancements in open-source large language models19,36 to accurately extract numerous clinical labels at scale. This approach, a first for pediatric echocardiography, enables us to address a diverse range of clinically relevant tasks. For instance, detecting major aortopulmonary collateral arteries and identifying Fontan procedures were unexplored in previous deep learning work and underscore the clinical novelty of our model. Similar to our work, Holste et al developed a novel, multitask model for adult echocardiography.13 However, their training labels were derived from structured echocardiography reporting fields that explicitly indicated the presence of specific conditions, an approach also adopted in electrocardiography.37 This limits model training and evaluation to institutions that maintain echocardiographic data in structured databases.38 In contrast, our large language model–based framework accommodates any text-based clinical report and broadens the applicability of our methods to institutions globally that may lack structured echocardiographic databases. Beyond multitask modeling, Christensen et al and Vukadinovic et al developed vision-language foundation models for adult echocardiography, EchoCLIP and EchoPrime, respectively. Despite their large-scale training on >200 000 studies, both models performed poorly on our pediatric data sets. Reddy et al observed a similar outcome for pediatric left ventricular ejection fraction regression,15 albeit for a smaller adult-based model. Together, these results demonstrate that adult-based models cannot be directly applied to pediatric populations because of fundamental differences in cardiovascular features and clinical conditions. We envision 2 primary clinical use cases for EchoAI-Peds. First, in hospitals or regional centers with limited pediatric cardiology expertise, the model could serve as an automated screening tool to flag studies likely to contain clinically important abnormalities, enabling earlier detection and more timely referral.39 Second, in settings in which pediatric cardiology expertise is available, EchoAI-Peds could function as a secondary reader to assist sonographers and cardiologists during interpretation.40 More broadly, EchoAI-Peds represents a foundational step toward fully automated pediatric echocardiography interpretation that could enable more equitable and continuous care of pediatric patients. Finally, our study is subject to several limitations. First, our model produces binary predictions even for conditions that are typically graded by severity, such as LVDil. Second, our model was trained on data from a single institution, which may limit its robustness across other pediatric populations. Third, although our model predicts numerous clinical findings, it currently omits several congenital and acquired heart diseases, repairs, and interventions. Fourth, our model cannot predict quantitative measurements, such as fractional shortening. In this work, we focused primarily on clinical conditions specific to pediatric patients and therefore excluded measurement-based labels. Fifth, model performance is lower for detecting a few rare but clinically important conditions, specifically patent ductus arteriosus and bicuspid aortic valve. Finally, EchoAI-Peds has not been prospectively evaluated in either screening or physician-assist workflows, so its integration into clinical practice remains to be studied. Future updates to EchoAI-Peds will include training on larger, multi-institutional data sets to improve generalizability, enhance performance on existing clinical classes, and expand coverage across a wider spectrum of disease states. In addition, future work will focus on enabling our model to generate both categorical and quantitative predictions, such as disease severity and functional measurements, to further improve its clinical interpretability and utility. In conclusion, our findings reveal that multitask deep learning models hold significant potential for streamlining pediatric echocardiography analyses. Moreover, our results highlight the importance of developing echocardiography artificial intelligence systems specifically tailored to pediatric populations. We hope this work will accelerate the development of pediatric-focused deep learning models with broad clinical utility. Article Information