Model-Driven Analysis of ECG Using Reinforcement Learning.

Area of Science:

• Computational Biology
• Biomedical Engineering
• Pediatric Cardiology

Background:

• Understanding complex physiological signals like the electrocardiogram (ECG) requires robust modeling techniques.
• Analyzing ECGs in infants is crucial for early detection of cardiac abnormalities and monitoring development.

Purpose of the Study:

• To develop and validate a model-driven framework for analyzing infant ECG signals.
• To decompose ECG signals into interpretable lognormal components using deep neural networks.
• To investigate the relationship between extracted ECG parameters and infant age.

Main Methods:

• A systematic framework was developed to decompose ECG signals into overlapping lognormal components.
• Reinforcement learning was employed to train a deep neural network for parameter estimation.
• The model was applied to ECG data from 751,510 PQRST complexes in infants aged 1-24 months.

Main Results:

• 82.7% of modeled PQRST complexes yielded a signal-to-noise ratio suitable for analysis (>5 dB).
• 10 out of 24 modeling parameters showed statistical significance (p<0.01) with age, demonstrating sensitivity.
• Kendall rank correlation coefficients ranged from 0.27 to 0.51, indicating moderate associations.

Conclusions:

• The model-driven approach effectively captures sensitive ECG parameters in infants.
• The extracted parameters exhibit age-dependent variations, providing physiological interpretability.
• This framework offers a window into latent variables influencing heart function and autonomic nervous system control in infants.