Abstract
Congenital heart disease (CHD) is the most prevalent birth defect worldwide, yet its detection during the first trimester (11–14 weeks) remains a significant clinical challenge. This study, published in npj Digital Medicine, introduces an interpretable deep learning (DL) framework designed to mimic the diagnostic logic of expert clinicians. By focusing on diastolic blood flow patterns in the four-chamber view, the model achieves expert-level accuracy while providing transparent, "human-readable" explanations. The framework significantly enhances clinical sensitivity, offering a vital opportunity for early intervention.
Introduction
Early detection of CHD is crucial for reproductive planning and prenatal care. However, first-trimester screening is notoriously difficult due to the small size of the fetal heart and the rapid heart rate, leading to a clinical sensitivity of only 14–17%. While artificial intelligence has shown potential in medical imaging, the "black box" nature of many models limits their clinical utility. This research addresses these hurdles by developing a model that prioritizes interpretability and aligns with established clinical reasoning.
Methods
The researchers developed an end-to-end trainable architecture that processes Doppler flow four-chamber view (4CV) images through a structured four-step pipeline:
- Structure-Guided Feature Learning: Utilizing a Vision Transformer (ViT) encoder, the model employs an Intersection-over-Union (IoU) loss against clinician-drawn masks to ensure focus on internal cardiac structures.
- Flow Pattern Analysis: The model predicts intermediate clinical features, such as the number of filled chambers and flow symmetry, rather than jumping directly to a diagnosis.
- Flow-Based Anomaly Prediction: A hierarchical strategy filters "normal" cases based on these predicted patterns, mirroring the deductive process of a cardiologist.
- Cross-Attention Subtype Classification: For abnormal cases, a cross-attention mechanism relates specific flow features to disease subtypes like Atrioventricular Septal Defect (AVSD).
Dataset and Validation
The study leveraged a massive multicenter cohort from seven tertiary hospitals in China, involving 108,521 individual cases. A curated dataset of 8,062 high-quality Doppler flow 4CV images (5,830 normal and 2,232 abnormal) was used for training and validation. The model's generalizability was confirmed through testing on internal and three distinct external validation sets, including prospective data.
Key Results
The model demonstrated near-perfect performance in distinguishing normal from abnormal hearts:
- Binary Detection: Achieved an AUROC of 1.000 on internal data and 0.925–1.000 across external sites.
- Subtype Accuracy: For complex defects like Functional Single Ventricle (FSV), the model achieved an AUROC of 0.962.
- Expert Comparison: In head-to-head trials, the AI matched or surpassed the performance of clinicians with 5–8 years of experience.
- Human-AI Synergy: When clinicians were provided with the AI’s predictions and explanations, their diagnostic AUROC improved by up to 8.8%, and sensitivity for complex defects increased by over 20%.
Discussion and Limitations
The shift from "black box" classification to "clinical reasoning emulation" is the core innovation of this work. By forcing the model to predict intermediate flow patterns, the researchers made the AI’s decision-making process legible to medical professionals. Furthermore, the AI proved more robust than humans at identifying subtle flow abnormalities often dismissed as technical artifacts.
However, limitations remain. The current system requires manual selection of the best diastolic frame and relies solely on the four-chamber view, which may miss certain conotruncal anomalies. Future work aims to automate frame selection and integrate multi-modality imaging (B-mode and Doppler) to create a more comprehensive screening tool.
Summary updated