Extracting wavelet based neural features from human intracortical recordings for neuroprosthetics applications

Abstract

Background: Understanding the long-term behavior of intracortically-recorded signals is essential for improving the performance of Brain Computer Interfaces. However, few studies have systematically investigated chronic neural recordings from an implanted microelectrode array in the human brain.

Methods: In this study, we show the applicability of wavelet decomposition method to extract and demonstrate the utility of long-term stable features in neural signals obtained from a microelectrode array implanted in the motor cortex of a human with tetraplegia. Wavelet decomposition was applied to the raw voltage data to generate mean wavelet power (MWP) features, which were further divided into three sub-frequency bands, low-frequency MWP (lf-MWP, 0-234 Hz), mid-frequency MWP (mf-MWP, 234 Hz-3.75 kHz) and high-frequency MWP (hf-MWP, >3.75 kHz). We analyzed these features using data collected from two experiments that were repeated over the course of about 3 years and compared their signal stability and decoding performance with the more standard threshold crossings, local field potentials (LFP), multi-unit activity (MUA) features obtained from the raw voltage recordings.

Results: All neural features could stably track neural information for over 3 years post-implantation and were less prone to signal degradation compared to threshold crossings. Furthermore, when used as an input to support vector machine based decoding algorithms, the mf-MWP and MUA demonstrated significantly better performance, respectively, in classifying imagined motor tasks than using the lf-MWP, hf-MWP, LFP, or threshold crossings.

Conclusions: Our results suggest that using MWP features in the appropriate frequency bands can provide an effective neural feature for brain computer interface intended for chronic applications.

Trial registration: This study was approved by the U.S. Food and Drug Administration (Investigational Device Exemption) and the Ohio State University Medical Center Institutional Review Board (Columbus, Ohio). The study conformed to institutional requirements for the conduct of human subjects and was filed on ClinicalTrials.gov (Identifier NCT01997125).

Keywords: Brain computer interface; Chronic decoding; Intracortical recordings; Mean wavelet power; Signal quality.

Conflict of interest statement

Competing interestsThe authors declare that they have no competing interests.

© The Author(s) 2018.

Figures

Fig. 1

a System setup. The participant was seated in front of a computer monitor, where imaginary movements provided to him via a virtual hand at rest condition at the lower left corner on the screen. b Experimental timeline for a single block. Here, it shows a block for Task 1 and 2, respectively. In Task1, there was only two hand movements, hand opening and hand closing. While in Task2 there were four cued hand movements within each block. In between the blocks within each task, there was usually a 2–3 min break

Fig. 2

Processing of raw signal into different neural features. Step 1: A 100 ms section of neural signal was selected from a larger raw voltage recording. Step 2: Conduct signal processing for this 100 ms raw signal. In method a), raw signal was decomposed into 11 wavelet scales to get the rectified wavelet coefficients of each scale; In method b), a high-pass filter and threshold of - 4.5 times of the RMS value was applied to detect the TCs within this 100 ms section of raw signal; In method c), a low pass filter was applied to get LFP of the raw signal; In method d), band pass filter and customized RMS values were calculated to get MUA of the raw signal. Step 3: The processed signal within this time window were then averaged over this 100 ms, respectively, to compose the related one data point in the averaged larger time series of an entire block. Step 4: Signal smoothing and standardization. To generate MWP feature time series, a 1-s moving average and a 15-s wide mean subtraction were applied to this averaged time series of the entire block. Afterwards, the processed time series were standardized and averaged accordingly across selected scales to produce a new time series for each channel. To generate other feature time series, the processed signal after Step 3 was applied 1-s moving average, a 15-s mean subtraction and standardization, sequentially (Please refer to the Methods section for more details)

Fig. 3

a Average impedance over time. Each data point here shows the average impedance across two different reference electrodes on the MEA. Most of the sharp decline happened within the first 400 days. In (b, c and d), the results were summarized from the data collected in Task 2. b Signal-to-Noise Ratio (SNR) over time. Each data point shown here is the average SNR value of an experimental day. The SNR was stable in general with an average value at 18.28 ± 0.25 dB throughout the length of the study, with a slight decline overall. c Peak-peak value of the detected TCs from the raw signal. Each data point presents the average peak-peak value of all the detected TCs from an experimental day. The Peak-peak value decreased the most during the first 400 days and the overall average throughout the study was 153.68 ± 17.92 μV. d RMS noise value of the raw signal. RMS value shared a similar temporal profile over the period of entire study compared to that of peak-peak value and SNR value. The overall average RMS noise was 18.78 ± 1.89 μV. In each figure, the line indicated the Loess regression result of the discrete values throughout the entire study, with 95% confidence intervals in each plot

Fig. 4

Brain signal modulated by the presentation of cues during Task 2. A representative snapshot of neural modulation as the participant imagined the cued hand movements during Task 2. Heat maps show lf-MWP, mf-MWP, hf-MWP, LFP and MUA features, and raster plot of TCs from raw recordings, respectively. Each time point is 100 ms. For TCs, each data point represents the total number of detected TCs within 100 ms

Fig. 5

Signal stability presented by the normalized signal strength. Within each time series, the data is normalized to its initial value. In Task 1, the normalized signals strength for mf-MWP, hf-MWP, MUA and TCs were stabilized after approximately 300 days of decline, while normalized signal strength for lf-MWP and LFP showed a better signal stability over time. MUA and mf-MWP signals shared a very similar profile in signal declining with most of their confidence interval ranges overlap with each other. In a longer study in Task 2, similar signal stability was observed overall, while TCs signal showed a more gradual decline after 800 days post-implantation. The band in each time series shows the range of its 95% confidence interval of a LOESS fit

Fig. 6

Correlation values between paired channels of neural features affected by the inter-electrode distance. We selected 3 days over the course of the study from Task 2 to investigate the effect of inter-electrode distance on neural features correlations. In each insertion, y-axis shows the correlation from 0 to 1 and x-axis shows the inter-elelctrode distance in mm. A data point represents the correlation value of two selected channels for a particular type of neural features time series. The solid line represents a 3rd polynomial fit of the data and dashed lines indicate its 95% confidence interval. In general, signal correlation decreases as the inter-elelctrode distance increases. Signal with higher frequency range has lower inter-electrode correlation, and TCs features has the lowest inter-electrode correlation values overall

Fig. 7

Averaged inter-signal correlation matrix. Inter-signal temporal correlation time series were averaged to give an overall correlation value for a given paired signal over the course of the study for Task 2 (For more details, please refer to Methods section and also Additional file 1: Figure S1). In general, MUA and mf-MWP features showed a very high level of correlation, with an average correlation value of 0.87 ± 0.19, compared to other paired neural features

Fig. 8

Overall classification accuracy of the decoder output when using different features as input across the entire study. a, b Each data point shows the overall accuracy from an experimental day, and the line is the LOESS regression of all the discrete data points across time. In Task 1 and Task 2, using mf-MWP and MUA features as input, it consistently generated the highest overall accuracy throughout the entire phase of the study (Statistical analysis indicates difference between the two overall accuracy time series were non-significant). c The averaged overall accuracy of the study was significantly higher (* indicates p < 0.001, n = 62 in Task1, and n = 64 in Task2) when using mf-MWP and MUA as input into the decoders, compared to those using lf-MWP, hf-MWP, LFP or TCs features as decoder input. Each error bar shows the standard deviation of the accuracy time series for a feature