Automated Preoperative Diagnosis and Postoperative Recovery Trajectory Prediction in Pediatric TOF

Article Impact Level: HIGH
Data Quality: STRONG
Summary of  eBioMedicine . https://doi.org/10.1016/j.ebiom.2026.106292
Dr. Qiang Gao et al.

Points

Congenital heart specialists developed an automated digital framework called DynaTOF to support clinical decision-making from preoperative diagnosis through postoperative risk prediction.
The deep learning system automatically recognizes standard apical and parasternal views from echocardiographic videos and measures key structural cardiac diameters.
A multimodal approach combining moving video features with quantitative measurements outperformed single-source data models to replicate how human cardiologists interpret clinical data.
Training protocols utilized multi-center data comprising healthy controls, confirmed tetralogy of Fallot cases, and highly challenging clinical mimicking conditions.
The framework successfully forecasted individual postoperative recovery trajectories and stratified long-term patient follow-up risks to help clinicians identify which children require closer monitoring.

Summary

Engineered to optimize the care pathway for tetralogy of Fallot (TOF), this study developed and validated an end-to-end artificial intelligence framework named DynaTOF. Tetralogy of Fallot stands as one of the most common cyanotic congenital heart defects, characterized by complex structural abnormalities that require highly precise preoperative evaluation, surgical corrective planning, and lifelong longitudinal tracking. While noninvasive echocardiography serves as the diagnostic cornerstone, manual interpretation introduces significant observer variability and heavy clinical workloads. The research sought to determine if a multimodal deep learning architecture could streamline clinical video interpretation and leverage preoperative data to predict post-surgical recovery trajectories.

Utilizing a multi-center data pipeline, investigators structured the diagnostic algorithm across several sequential steps to evaluate healthy controls, patients with confirmed TOF, and cohorts presenting with mimicking cardiac anomalies. The framework automatically recognizes standard apical and parasternal echocardiographic views before isolating and measuring key anatomical cardiac diameters to eliminate repetitive manual measurements. To mirror clinical decision-making, the system deploys a multimodal training architecture that integrates deep visual features from moving echocardiographic videos with quantitative structural metrics, demonstrating superior diagnostic accuracy compared to models analyzing video frames or physical measurements in isolation.

Evaluations focused on forecasting postoperative recovery patterns by synthesizing preoperative echocardiographic geometry, specific surgical intervention types, and follow-up timing sequences. The system demonstrated strong diagnostic and prognostic utility, reliably stratifying long-term follow-up risk and flagging patients predisposed to abnormal postoperative recovery score patterns. These predictive capabilities provide clinical teams with an early decision-support tool to individualize monitoring protocols for highly vulnerable pediatric patients. While prospective external validation across diverse ultrasound manufacturers is necessary to establish explicit long-term risk hazard ratios, this framework successfully connects structural imaging analysis directly to longitudinal congenital care.

Link to the article: https://www.thelancet.com/journals/ebiom/article/PIIS2352-3964(26)00174-X/fulltext

References

Gao, Q., Wang, A., Gao, Y., Xie, W., Yang, J., Yan, M., Chao, S., Zhao, L., Lu, H., Zhang, Y., Chen, L., & Yao, S. (2026). Echocardiography-based intelligent diagnosis and risk stratification management for tetralogy of Fallot. eBioMedicine, 129, 106292. https://doi.org/10.1016/j.ebiom.2026.106292