Real-Time Ultra-Low Power ECG Anomaly Detection Using an Event-Driven Neuromorphic Processor.

Area of Science:

• Neuromorphic Engineering
• Biomedical Signal Processing
• Artificial Intelligence in Healthcare

Background:

• Accurate detection of pathological conditions via biological signals like ECGs is crucial but time-consuming and costly.
• Current methods involve extensive off-line analysis, inefficient storage of non-pathological data, and difficult visual searches for medical professionals.
• Infrequent pathological patterns in biological signals exacerbate diagnostic inefficiencies.

Purpose of the Study:

• To propose a compact, low-power neural processing system for on-line, real-time preliminary diagnosis of pathological conditions.
• To develop a system capable of raising warnings for potential pathologies or triggering off-line data recording.
• To apply the system for real-time ECG classification, distinguishing healthy heartbeats from pathological rhythms.

Main Methods:

• Utilized a spiking recurrent neural network operating in a reservoir computing paradigm.
• Encoded multi-channel analog ECG traces as asynchronous streams of binary events.
• Employed an event-driven neuron output layer trained for pathology recognition and validated on a Dynamic Neuromorphic Asynchronous Processor (DYNAP) chip.

Main Results:

• The system successfully performed real-time classification of ECG data.
• Demonstrated the ability to distinguish between healthy heartbeats and pathological rhythms.
• Generated a binary trigger signal indicating the presence or absence of pathological patterns, validated by experimental chip measurements.

Conclusions:

• The proposed compact, sub-mW neural system enables efficient on-line, real-time preliminary diagnosis of pathological conditions from ECG data.
• This approach significantly enhances diagnostic efficiency by reducing reliance on manual review and extensive data storage.
• The validated DYNAP chip implementation showcases the practical feasibility of neuromorphic systems for real-time biomedical signal analysis.