Optimized multi-stage network with multi-dimensional spatiotemporal interactions for septal and apical hypertrophic...

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Hypertrophic cardiomyopathy (HCM) is a genetic heart condition and a primary cause of sudden cardiac death in adolescents.
• Septal hypertrophy (SH) and apical hypertrophy (AH) are key subtypes of HCM, increasing risks of heart failure and arrhythmias.
• Accurate identification of SH and AH via 12-lead electrocardiography (ECG) is vital for early diagnosis and patient management.

Purpose of the Study:

• To develop an advanced deep learning model for precise detection of SH and AH in HCM patients using ECG data.
• To overcome limitations of existing 1D or 2D ECG analysis methods that may lose critical spatiotemporal information.
• To enhance diagnostic accuracy for HCM subtypes through a novel multi-dimensional approach.

Main Methods:

• An optimized multi-stage network with multi-dimensional spatiotemporal interactions (Ms-MdST) was designed.
• The Ms-MdST model integrates 1D convolutions for temporal features and 2D convolutions for spatial features across ECG leads.
• A global-local interactive attention mechanism (GLIA) and multi-loss optimization were used for feature fusion.

Main Results:

• The Ms-MdST model achieved high F1-scores: 0.9672 (CONTROL), 0.7250 (SH), and 0.8009 (AH).
• The model demonstrated superior performance compared to existing ECG classification techniques.
• The proposed method showed interpretability, suggesting potential for clinical application.

Conclusions:

• The Ms-MdST model offers a significant advancement in detecting SH and AH in HCM patients using ECG.
• This spatiotemporal deep learning approach improves diagnostic accuracy by preserving essential data dimensions.
• The model's interpretability and performance indicate its promise for clinical integration in cardiovascular diagnostics.