Deep learning for predicting respiratory rate from biosignals.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Deep learning models are increasingly used in body sensor networks for bio-signal prediction.
• Evaluating novel deep learning models for bio-signal prediction accuracy is crucial due to recent innovations.

Purpose of the Study:

• To evaluate the performance of various deep learning models for respiratory rate prediction using bio-sensor data.
• To compare the effectiveness of different deep learning architectures and window sizes for respiratory rate estimation.

Main Methods:

• Three bio-sensor datasets (ECG, PPG, sEMG) were utilized.
• Evaluated deep learning models included LSTM, Bi-LSTM, attention-based LSTMs, CNN-LSTM, and Convolutional-LSTM.
• Models were assessed using 32s and 64s windows, with Mean Absolute Error (MAE) as the performance metric.

Main Results:

• The 64s window provided more accurate predictions than the 32s window.
• Bidirectional LSTM (Bi-LSTM) with Bahdanu Attention demonstrated superior performance across most datasets.
• The best performance achieved was an MAE of 0.24 ± 0.03 for PPG and ECG data using Bi-LSTM with Bahdanau attention.

Conclusions:

• Deep learning models, particularly Bi-LSTM with attention, are effective for respiratory rate prediction from bio-signals.
• The choice of model architecture and data window size significantly impacts prediction accuracy.
• These findings support the use of advanced deep learning techniques for non-invasive respiratory monitoring.