TL;DR
This paper introduces ERAASR, a novel algorithm that effectively removes large electrical stimulation artifacts from multielectrode array recordings. By exploiting the shared structure of artifacts across channels, pulses, and trials, ERAASR allows for the simultaneous recording of neural activity and electrical stimulation, crucial for understanding brain function and developing neural prosthetics.
What Problem Does This Paper Address?
In neuroscience research and clinical applications like deep brain stimulation and neural prosthetics, electrical stimulation is a powerful tool to probe and modulate neural circuits. However, this stimulation generates significant electrical artifacts that overwhelm the delicate signals of neuronal spiking activity recorded by nearby electrodes. These artifacts make it incredibly difficult, if not impossible, to accurately detect and analyze the neural responses that occur during or immediately after stimulation. Existing methods often discard large portions of the recording data or rely on hardware that is expensive and difficult to scale. This paper addresses the critical need for a robust, software-based method to remove these artifacts, enabling researchers to observe neural activity in real-time with stimulation.
Methods
The core of ERAASR is a sequential application of Principal Component Analysis (PCA) and linear regression, applied across different dimensions of the recorded data. The authors treat the multi-channel, multi-pulse, and multi-trial recordings as a data tensor.
- Channel-wise Cleaning: PCA is used to identify the dominant artifact patterns shared across all recording channels. These components are then used to reconstruct and subtract the artifact from each channel, while strategically excluding neighboring channels to preserve local neural signals.
- Pulse-wise Cleaning: The data is then rearranged to analyze the artifact structure across individual pulses within a stimulation train. Again, PCA and regression are used to identify and remove pulse-specific artifact components.
- Trial-wise Cleaning: Finally, the process is repeated to remove residual artifact structure that is consistent across multiple trials of the same stimulation pattern.
- Post-stimulation Transient Cleaning: A separate step addresses slower, lingering artifact transients that occur after stimulation ceases.
This multi-stage approach leverages the inherent similarities in stimulation artifacts across different recording dimensions.
Datasets
The algorithm was validated using electrophysiological recordings from macaque dorsal premotor cortex. The data was collected using a 24-channel linear multielectrode array ("V probes") during an arm-movement task, with single-electrode electrical stimulation. The authors collected data across 9 recording and stimulation sessions. No new benchmark datasets were introduced; rather, existing experimental data was used to demonstrate the algorithm's effectiveness.
Key Insights and Ideas
- Artifacts have shared structure: The fundamental insight is that electrical stimulation artifacts, unlike neural spiking, exhibit a high degree of similarity across simultaneously recorded channels, across individual pulses in a train, and across repeated trials. This shared structure is the key to their removal.
- PCA as a universal artifact detector: PCA is remarkably effective at isolating the dominant, shared artifact patterns from the more localized and variable neural signals.
- Exploiting spatial and temporal locality: By strategically excluding neighboring channels, pulses, or trials during artifact reconstruction, the algorithm minimizes the risk of removing genuine neural activity, which is inherently localized in space and time.
- Artifacts as a localization tool: The paper demonstrates that the spatial distribution of stimulation artifacts can be used to infer the relative positions of recording electrodes, a novel application of artifact analysis.
Key Results
- ERAASR successfully removed large electrical stimulation artifacts (up to ±7500 µV) that were orders of magnitude larger than neural spiking waveforms (~100 µV).
- After artifact removal, channels with no spontaneous activity showed quiescent signals, while active channels revealed evoked spikes that closely resembled spontaneous spiking waveforms.
- The algorithm allowed for the recovery of neural signals occurring during the stimulation period, a significant improvement over methods that discard data.
- The artifact amplitude model accurately predicted artifact behavior, with a squared correlation of R² = 0.91, and the fitted probe locations aligned well with experimental observations.
- The cleaned signals showed very little residual artifact, and the RMS voltage of channels post-cleaning was comparable to pre-stimulation RMS.
- Waveform correlation analysis confirmed that evoked spikes during stimulation closely matched spontaneous spikes.
Limitations and Future Work
- Linearity Assumption: The method critically relies on the assumption that stimulation artifact linearly superimposes with neural spiking. This limits its direct applicability to stimulation configurations and amplitudes that do not saturate the recording amplifier. The authors suggest hybrid approaches combining hardware artifact cancellation with their software method for broader applicability.
- Sequential Cleaning: The sequential nature of the cleaning (channels, then pulses, then trials) might not be optimal. The authors suggest that tensor decomposition methods could potentially offer a more integrated approach to identifying and removing shared structure across all dimensions simultaneously.
- Greedy Removal: The algorithm's greedy approach to removing shared structure could inadvertently distort or remove genuine spiking activity, especially with higher stimulation currents that might lead to more synchronous evoked spiking. They mention experimenting with matched filters for spike detection but did not explore this further for higher currents.
- Future Directions: The authors highlight the potential for their method to enable precise causal links between neural activity and behavior, and to aid in the design of advanced neural prosthetics. They also suggest that a joint estimation of artifact and spiking, similar to methods used in other fields, could be adapted for more robust artifact removal.
Summary updated