Automated HFrEF Diagnosis Using an Optimized TimeSformer Model in Echocardiography.

Area of Science:

• Artificial Intelligence in Medicine
• Cardiovascular Imaging Analysis
• Deep Learning for Medical Diagnosis

Background:

• Diagnosing heart failure with reduced ejection fraction (HFrEF) is challenging, particularly in advanced stages.
• Deep learning (DL) models show promise for automated HFrEF detection but struggle with small, imbalanced clinical datasets.
• Current methods require improvement for reliable HFrEF diagnosis in diverse clinical settings.

Purpose of the Study:

• To develop and evaluate a novel deep learning approach for enhanced HFrEF detection using echocardiographic videos.
• To adapt and apply the TimeSformer architecture for spatiotemporal feature extraction in echocardiography.
• To improve model performance by incorporating domain-informed left ventricle (LV) masking for focused analysis.

Main Methods:

• Utilized the TimeSformer architecture, a Transformer-based model, for analyzing echocardiographic video data.
• Implemented a novel domain-informed left ventricle (LV) masking technique using image segmentation.
• Evaluated the methodology on a large-scale benchmark dataset and a specialized, smaller clinical dataset after fine-tuning.

Main Results:

• The proposed framework achieved a 3% improvement in accuracy and AUC on the benchmark dataset.
• On the specialized clinical dataset, improvements reached 7% in accuracy and 30% in AUC values.
• TimeSformer with LV masking consistently outperformed conventional methods, demonstrating significant performance gains.

Conclusions:

• The novel deep learning framework offers a practical and generalizable strategy for improving automated HFrEF diagnosis.
• The approach enhances diagnostic performance, particularly in data-scarce healthcare environments.
• Findings support the potential of this method for clinical decision support in cardiovascular medicine.