Abstract
This paper presents an experiment-based comparison of Absolute Threshold (AT) and Non-linear Energy Operator (NEO) spike detection algorithms in Intra-cortical Brain-Machine Interfaces (iBMIs). Results show an average increase in decoding performance of approximately 5% in Monkey A across 28 sessions recorded over 6 days and approximately 2% in Monkey B across 35 sessions recorded over 8 days when using NEO over AT. To the best of our knowledge, this is the first ever reported comparison of spike detection algorithms in an iBMI experimental framework involving two monkeys. Based on the improvements observed in an experimental setting, backed by previously reported improvements in simulation studies, we advocate switching from the current state-of-the-art spike detection technique, AT, to NEO.
Introduction
Intra-cortical Brain-Machine Interfaces (iBMIs) aim to restore movement for patients with spinal cord injuries by recording neural "spikes" (action potentials) and translating them into control commands. The critical first step in this process is spike detection: distinguishing neural activity from background noise. While the Absolute Threshold (AT) method is the current standard, it suffers from high sensitivity to threshold settings and significant hardware overhead for on-chip implementations. This study demonstrates that the Non-linear Energy Operator (NEO) offers a more robust and hardware-efficient alternative, providing a 2% to 5% boost in decoding accuracy and greater stability across experimental sessions.
Experimental Framework
The researchers utilized a substantial real-world dataset involving two monkeys performing a behavioral navigation task. Signals were recorded from the motor cortex (96 electrodes for Monkey A; 64 for Monkey B) while the subjects used a joystick to navigate a robotic platform toward a target.
Raw neural data was bandpass filtered (300–3000 Hz) before spike detection. To isolate the impact of the detection algorithm, the team employed Linear Discriminant Analysis (LDA) to map detected spikes to four discrete movements (forward, left, right, and stop). The simplicity of LDA ensures that performance gains are attributable to the spike detection method rather than the complexity of the decoder.
Spike Detection: AT vs. NEO
The standard AT method detects spikes when signal amplitude crosses a threshold based on the noise's standard deviation. However, this method is highly sensitive; minor threshold adjustments can lead to significant false positives or missed spikes.
In contrast, NEO calculates the "energy" of the signal by considering both amplitude and frequency. By multiplying the signal by itself and subtracting the product of adjacent samples, NEO acts as a pre-emphasis filter that boosts the Signal-to-Noise Ratio (SNR). This energy-based approach creates a clearer separation between signal and noise, making the system less dependent on precise threshold tuning.
Key Results
- Decoding Accuracy: NEO consistently outperformed AT across both subjects. Monkey A showed an average improvement of 5.29%, while Monkey B improved by 2.16%.
- Consistency and Reliability: NEO exhibited lower standard deviation in decoding accuracy across sessions (e.g., 8.09% vs. 10.72% for Monkey A), indicating superior day-to-day reliability.
- Hardware Efficiency: NEO implementations are highly amenable to on-chip processing, consuming as little as 40–50 nW per channel. This is orders of magnitude more efficient than the memory-intensive median-based calculations required for AT, making NEO ideal for future wireless implants.
Conclusion and Future Work
This study provides the first experimental evidence in a two-monkey framework that NEO is a superior spike detection algorithm for iBMIs. While these results were obtained through offline analysis, they strongly suggest that NEO can improve the performance and longevity of clinical neural prosthetics. Future work will focus on closed-loop trials to evaluate real-time performance and exploring the interaction between NEO and more sophisticated decoding models like Kalman filters or Deep Neural Networks.
Summary updated