Deep learning-based dynamic ventilatory threshold estimation from electrocardiograms.

Area of Science:

• Cardiorespiratory physiology
• Biomedical engineering
• Machine learning in healthcare

Background:

• Ventilatory threshold (VT) indicates aerobic to anaerobic metabolism transition, crucial for cardiorespiratory endurance assessment.
• Conventional VT assessment via cardiopulmonary exercise testing requires gas analysis.
• Heart rate variability (HRV) from ECGs offers an alternative but has limitations like minimum recording duration and preprocessing dependency.

Purpose of the Study:

• To develop and validate a deep learning model for estimating VT using only ECG signals and physical characteristics.
• To overcome the limitations of traditional HRV-based VT assessment methods.

Main Methods:

• A deep learning model combining Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN) was developed for lead II ECG.
• Physical characteristics and ECG signals were used as inputs.
• Joint optimization of CNN layers and LSTM integration were applied for VT estimation.
• The model was trained and evaluated on datasets from the Bruce and multiple tasks (MT) protocols.

Main Results:

• The model demonstrated acceptable performance on both Bruce (-0.28[-1.91,1.34] ml/min/kg) and MT (0.07[-3.14,3.28] ml/min/kg) protocols.
• High coefficients of determination (0.84 Bruce, 0.73 MT) and Pearson correlation coefficients (0.93 Bruce, 0.97 MT) were achieved.
• Root mean square errors were 0.868 for Bruce and 3.373 for MT protocols.

Conclusions:

• The proposed deep learning model can effectively estimate VT from successive ECGs.
• Incorporating physical variables and joint optimization significantly improved VT assessment performance.
• This dynamic VT estimation method using ECGs has potential applications in daily fitness management and cardiovascular rehabilitation.