Source-linked AI summary
Real-time ultra-low power ECG anomaly detection using an event-driven neuromorphic processor
Felix Christian Bauer, Dylan Richard Muir, Giacomo Indiveri
TL;DR
Extended ECG monitoring can produce large volumes of mostly non-pathological data, making review burdensome while portable devices also face energy demands. This paper proposes continuous neuromorphic ECG anomaly detection and reports reliable detection with estimated system power below one milliwatt.
Problem
Extended monitoring generates large volumes of mostly non-pathological data for professionals to search, while recording, storage, and transmission create energy demands that conflict with portable devices’ long battery life.
Source: §1 · p. 2Method
The system continuously processes ECG signals using a spiking recurrent network and trained event-based readouts to detect anomalous patterns and produce a binary output.
Source: §1 · p. 3Results
91.3% of anomalous test segments were detected, 2.4% of normal heartbeats were falsely indicated, and estimated total system power was below 722.1 µW.
Source: §3.1 · p. 9, §3.2 · p. 10Takeaways & Limitations
The authors conclude that DYNAP-SE-based ECG anomaly detection is feasible for an always-on, ultra-low-power monitoring system, with remaining system parts realizable using existing circuitry.
Source: §5 · p. 13The spiking network ran on neuromorphic hardware, but event conversion, readout, and binary output were simulated, and the ECG input was already amplified and filtered.
Source: §4.3 · p. 12
Abstract
from arXiv · showhide
Accurate detection of pathological conditions in human subjects can be achieved through off-line analysis of recorded biological signals such as electrocardiograms (ECGs). However, human diagnosis is time-consuming and expensive, as it requires the time of medical professionals. This is especially inefficient when indicative patterns in the biological signals are infrequent. Moreover, patients with suspected pathologies are often monitored for extended periods, requiring the storage and examination of large amounts of non-pathological data, and entailing a difficult visual search task for diagnosing professionals. In this work we propose a compact and sub-mW low power neural processing system that can be used to perform on-line and real-time preliminary diagnosis of pathological conditions, to raise warnings for the existence of possible pathological conditions, or to trigger an off-line data recording system for further analysis by a medical professional. We apply the system to real-time classification of ECG data for distinguishing between healthy heartbeats and pathological rhythms. Multi-channel analog ECG traces are encoded as asynchronous streams of binary events and processed using a spiking recurrent neural network operated in a reservoir computing paradigm. An event-driven neuron output layer is then trained to recognize one of several pathologies. Finally, the filtered activity of this output layer is used to generate a binary trigger signal indicating the presence or absence of a pathological pattern. We validate the approach proposed using a Dynamic Neuromorphic Asynchronous Processor (DYNAP) chip, implemented using a standard 180 nm CMOS VLSI process, and present experimental results measured from the chip.
1 Introduction
Extended ECG monitoring can reveal early indicators, but it produces large volumes of mostly non-pathological data and conflicts with the need for portable, long-battery-life devices. The paper proposes continuous, low-power neuromorphic processing to identify suspicious patterns and support preliminary diagnosis and targeted follow-up.
1 Introduction
- Extended monitoring creates large volumes of mostly non-pathological ECG data that trained professionals must search, while high-accuracy recording and data storage or transmission consume energy needed for portable, long-life devices.Early detection of some conditions may support timely intervention.Source: §1 · p. 2
- The proposed system continuously processes signals locally in real time, aiming to reduce stored or transmitted data and facilitate later professional analysis.The paper describes a wearable use case in which suspicious rhythms trigger recording or transmission while most components remain in hibernation.Source: §1 · p. 2, Figure 1 · p. 3
- The authors identify a gap between conventional ECG classifiers that are generally unsuitable for ultra-low-power hardware and SNN approaches that had been implemented only in simulation on conventional computers.They also note cloud processing’s transmission power demand and potential mobility constraints.Source: §1 · pp. 2–3
- The system converts analog ECG signals into event trains, processes them with a spiking recurrent network, and uses trained event-based readouts to produce a binary pathological-event signal.The readout detects specific anomalous patterns from filtered neuron activity.Source: Figure 2 · p. 3, §1 · p. 4
- The work implements the recurrent network on DYNAP-SE and reports real-time, ultra-low-power analysis of two-channel ECG, while estimating complete-system power using existing implementations for other stages.The authors describe this as the first full-system description of a neuromorphic ECG anomaly detector and among the first real-world DYNAP-SE applications.Source: §1 · p. 4
2 Methods
The methods implement the ECG detector on DYNAP-SE, prepare a labeled MIT-BIH evaluation stream, expand it through a recurrent reservoir, and train readouts on filtered neural activity.
2.1 The DYNAP-SE system
- DYNAP-SE is a mixed-signal asynchronous 180-nm chip with four cores of 256 AdExp neurons; device mismatch is exploited to vary neuron parameters and increase signal dimensionality.The development kit provides a PC interface for configuring connections, sending input events, and reading neural activity.Source: §2.1 · pp. 4–5, Figure 3 · p. 5
2.2 ECG data set and conversion to spikes
- Evaluation used 48 half-hour two-channel excerpts from 47 subjects, retaining normal rhythms and five prevalent anomalies that cover 95.6% of recording time.Training used randomly drawn individual beats; validation and test used 5–10 contiguous same-label beats, with channels sigma-delta encoded into up- and down-events.Source: §2.2 · pp. 5, 7
2.3 sRNN architecture
- A 128-neuron input expansion feeds a 512-excitatory/128-inhibitory recurrent reservoir, whose random connections and hardware mismatch project inputs into a high-dimensional state.Recurrent connections let the reservoir state encode recent input history; reservoir weights are randomly initialized and learning occurs only in the output layer.Source: §2.3 · p. 7
2.4 Learning and readout
- Readout units sum filtered neural activity and detect anomalies above fixed thresholds; least-squares weights and validation-set thresholds are fitted independently for each anomaly type.The target is 1 when the corresponding anomaly is present and 0 otherwise; validation thresholds minimize false negatives and false positives.Source: §2.4 · pp. 8–9
3 Results
The network flags anomalous ECG segments without requiring classification of the specific anomaly type, detecting 91.3% of anomalous test segments. The paper also estimates total system power below 722.1 µW.
3.1 Network performance
- 91.3% of anomalous test segments were detected, while 2.4% of normal heartbeats were falsely flagged—about one false positive every 31.5 seconds.The task treats any readout unit's anomaly indication as a positive detection, regardless of whether it matches that unit's target label.Source: §3.1 · p. 9
- Three of five anomaly types had all segments detected, one had 88% sensitivity, and premature ventricular contraction sensitivity was 71.43%.Specificity exceeded 99% for every anomaly type; left bundle branch block beat and atrial premature beat produced two and one false positives, respectively, over the test set.Source: §3.1 · p. 9, Readout · p. 10, Table 3 · p. 10
- PPV was lower than corresponding sensitivities, reaching a low of 45.5% for premature ventricular contraction, while NPV was high for all conditions.The reported metrics quantify binary anomaly detection, not correct identification of the specific anomaly type.Source: §3.1 · p. 9, Readout · p. 10
3.2 Power consumption
- Estimated total system power was less than 722.1 µW, including 286.1 µW dynamic and 230 µW static power for the DYNAP-SE sRNN.The complete-system figure combines chip measurements and estimates for other processing stages.Source: §3.2 · p. 10, Table 4
4 Discussion
The discussion reports reliable detection for most evaluated anomalies but identifies premature ventricular contraction as difficult, while emphasizing application-dependent trade-offs in sensitivity, specificity, and false alarms. It also outlines feasibility constraints and possible hardware and learning approaches for completing the system.
4.1 Performance metric
- Sensitivity is evaluated over same-label contiguous rhythm segments, whereas specificity uses normal heartbeats and can be expressed as mean time between false positives.A missed pathological heartbeat is not considered problematic if another heartbeat in the same segment is detected.Source: §4.1 · p. 11
4.2 Network performance
- Premature ventricular contraction is the exception to reliable detection of most presented anomalies, and detection difficulty varies by anomaly type.The authors report no apparent correlation between individual subjects and sensitivity for this label.Source: §4.2 · p. 11
- Overall specificity and mean time between false positives are lower than per-anomaly values, and low PPV means positive results support further analysis rather than definite diagnosis.True positives and false negatives are counted by multi-beat segments, unlike true negatives and false positives, which are counted by individual beats.Source: §4.2 · pp. 11–12
- The accuracy–false-alarm trade-off depends on the use case: considering only atrial premature beats yields an MTBFP above 19 minutes with the current setup.Threshold changes can trade sensitivity against false alarms, and a diagnosing professional may review two rather than 70 to 80 rhythms per recorded minute.Source: §4.2 · p. 12
4.3.1 Preprocessing and conversion to events
- The spiking network was implemented on neuromorphic hardware, but event conversion, readout, and binary output were simulated, and the ECG was already amplified and filtered.The authors describe existing low-power amplifiers, filters, and asynchronous sigma-delta encoders as potential preprocessing implementations.Source: §4.3 · p. 12, §4.3.1 · p. 12
4.3.2 Readout
- Readout units were not implemented as processor neurons because of limited fan-in and shared synaptic weights; an estimated readout implementation would consume less than 90 µW.Larger fan-in hardware, variable selection, and regularization are discussed as ways to address the readout constraint; alternative readout learning models may also yield more suitable weights.Source: §4.2 · p. 12, §4.3.2 · p. 13
5 Conclusion
The paper proposes an always-on, sub-milliwatt neuromorphic system for real-time ECG anomaly detection. It implements the spiking recurrent network on DYNAP-SE and reports reliable detection, with the remaining system stages described as feasible using existing circuitry.
5 Conclusion
- The proposed always-on system performs real-time anomaly detection on multi-channel ECG with mean power below one milliwatt.The spiking recurrent neural network is implemented on a DYNAP-SE chip.Source: §5 · p. 13
- The DYNAP-SE implementation is reported to reliably detect anomalous ECG patterns.Source: §5 · p. 13
- The remaining system components are described as realizable with existing circuitry in a wearable health-monitoring device.Source: §5 · p. 13