ResNet-BiLSTM: A Multiscale Deep Learning Model for Heartbeat Detection Using Ballistocardiogram Signals

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Heartbeat detection from ballistocardiogram (BCG) signals is challenged by respiratory and motion artifacts.
• Existing methods often fail to capture long-term rhythm features, limiting beat-to-beat interval accuracy.
• Advanced signal processing is crucial for reliable J-peak detection in BCG signals.

Purpose of the Study:

• To develop a deep learning framework for robust beat-to-beat detection using BCG signals.
• To leverage multiscale features for improved J-peak identification and interval accuracy.
• To evaluate the model's performance across different sleep postures and time scales.

Main Methods:

• A deep learning framework combining ResNet and bidirectional long short-term memory (BiLSTM) was developed.
• The model utilizes multiscale features encompassing both morphological and rhythm characteristics of BCG signals.
• Leave-one-out cross-validation was employed with BCG data from 21 healthy subjects.

Main Results:

• The proposed multiscale deep learning model demonstrated robust performance across various sleep postures.
• Achieved averaged absolute error (Eabs) of 9.92 ms and averaged relative error (Erel) of 2.67 ms for heartbeat intervals compared to R-R intervals.
• Outperformed state-of-the-art detection protocols in accuracy and reliability.

Conclusions:

• A novel multiscale deep learning model for BCG-based heartbeat detection was successfully designed and validated.
• The integration of multiscale features significantly enhances detection performance over existing methods.
• Future work will focus on assessing the model's efficacy in clinical settings for patients with cardiovascular disorders.