Unsupervised heart-rate estimation in wearables with Liquid states and a probabilistic readout

Area of Science:

• Biomedical Engineering
• Computational Neuroscience
• Machine Learning

Background:

• Heart-rate estimation is crucial for wearable health monitoring devices.
• Existing methods often require extensive data annotation or lack personalization.
• Spiking neural networks offer potential for low-power, high-accuracy computation.

Purpose of the Study:

• To develop a novel machine learning technique for accurate heart-rate estimation from electrocardiogram (ECG) data.
• To leverage spatio-temporal ECG signal properties within a Liquid State Machine (LSM) framework.
• To enable personalized, low-energy heart-rate monitoring on wearable devices.

Main Methods:

• Encoding ECG signals into spike trains to excite recurrently connected spiking neurons in an LSM.
• Utilizing a novel learning algorithm and an unsupervised readout based on Fuzzy c-Means clustering.
• Employing particle swarm optimization for neuron subset selection and validating with CARLsim simulator.

Main Results:

• The proposed method achieves high accuracy in heart-rate estimation across diverse subjects, including those with cardiac irregularities.
• Demonstrated a significantly low energy footprint, suitable for extended wearable device battery life.
• Validated personalization by learning directly from individual ECG signals without costly annotations.

Conclusions:

• The developed machine learning approach shows strong potential for integration into future wearable devices for efficient and accurate heart-rate monitoring.
• The technique offers a personalized, low-power alternative to existing heart-rate estimation methods.
• This work highlights the utility of spiking neural networks and novel learning algorithms in biomedical signal processing.