Synthetic-data-driven LSTM framework for tracing cardiac pulsation in optical signals

Area of Science:

• Biomedical Optics
• Artificial Intelligence in Healthcare
• Signal Processing

Background:

• Optical monitoring techniques like near-infrared spectroscopy (NIRS), photoplethysmography (PPG), and diffuse correlation spectroscopy (DCS) are crucial for assessing cardiac pulsations.
• These methods are frequently compromised by motion artifacts and noise, limiting their clinical utility.
• Existing denoising methods often require extensive annotated clinical datasets, which are difficult to obtain.

Purpose of the Study:

• To develop a novel, synthetic-data-driven framework for denoising pulsatile optical waveforms.
• To enhance the accuracy of cardiac pulsation monitoring by mitigating motion artifacts and noise.
• To create a flexible AI model adaptable to various optical monitoring devices and noise conditions.

Main Methods:

• A long short-term memory (LSTM) neural network was trained using physiologically realistic synthetic pulsatile signals.
• Generated signals were intentionally corrupted with parameterized artifacts to simulate real-world noise.
• The LSTM model was applied to experimental NIRS, PPG, and DCS data, and its performance was compared to wavelet and temporal derivative distribution repair (TDDR) filters.

Main Results:

• The LSTM-based framework effectively traced and denoised pulsatile optical waveforms, outperforming traditional wavelet and TDDR filters.
• Heart rate (HR) extraction from the processed signals showed high agreement with electrocardiogram (ECG) measurements (mean absolute error = 0.59 bpm).
• The model demonstrated superior recovery of beat-to-beat waveform morphology.

Conclusions:

• Synthetic data generation combined with LSTM networks offers a powerful approach to denoise optical cardiac monitoring signals.
• This method overcomes the limitations of requiring annotated clinical datasets for training AI models.
• The proposed framework shows significant potential for improving the reliability and accuracy of non-invasive cardiac monitoring across diverse applications and environments.