Extending Multiscale Characterization of Heart Rate Variability via Deep Learning for Mortality Risk Prediction

Area of Science:

• Cardiology
• Biomedical Engineering
• Data Science

Background:

• Traditional linear analyses of heart rate variability (HRV) often miss nonlinear scaling patterns crucial for mortality risk prediction.
• Autonomic dysfunction, reflected in HRV, is a key indicator of mortality risk.

Purpose of the Study:

• To enhance mortality risk prediction from HRV signals by incorporating nonlinear scaling patterns using detrended moving average (DMA) analysis.
• To combine DMA with convolutional neural networks (CNNs) for improved feature extraction from HRV data.

Main Methods:

• Detrended moving average (DMA) curves were calculated from 2-hour windows of 24-hour Holter ECG recordings.
• A convolutional neural network (CNN) was trained on DMA curves to predict mortality risk in 916 survivors and 70 nonsurvivors.
• The CNN model performance was benchmarked against traditional HRV and clinical feature-based models.

Main Results:

• The CNN model achieved an ROC-AUC of 0.72, outperforming standard models in mortality risk prediction.
• Distinct DMA scaling patterns were identified, with reduced short-term slopes indicating impaired autonomic adaptability in nonsurvivors.
• Reduced long-term scaling slopes in specific patient groups were strongly associated with increased mortality risk.

Conclusions:

• The combination of DMA analysis and CNNs significantly improves HRV-based mortality risk stratification.
• This approach reveals novel physiological scaling patterns linked to survival outcomes, offering insights into HRV dynamics.
• The findings suggest potential for personalized health monitoring and improved clinical decision-making.