Using pseudo reference noise signals to remove latent noise from data signals and identify data sources

Abstract

The present disclosure describes various embodiments of systems, apparatuses, and methods for removing artifacts from data signals. One such method comprises obtaining a data signal across a plurality of data channels, wherein the data signal has artifacts across one or more channels; generating a pseudo reference noise signal representing noise activity across a plurality of noise channels via filtering of the data signal; analyzing the data signal with the pseudo reference noise signal to identify noise components that exist within both the data signal and the reference signal; scaling the noise components to project upon the data signal across the plurality of channels; and / or cleaning the data signal by subtracting the scaled noise components from the data signal across individual ones of the data channels.

Description

USING PSEUDO REFERENCE NOISE SIGNALS TO REMOVE LATENT NOISE FROM DATA SIGNALS AND IDENTIFY DATA SOURCESCROSS-REFERENCE TO RELATED APPLICATION

[0001] This application claims priority to co-pending U.S. provisional application entitled, “Using Pseudo Reference Noise Signals to Remove Latent Noise from Data Signals and Identify Data Sources,” having application number 63 / 580,664, filed September 5, 2023, which is entirely incorporated herein by reference.STATEMENT REGARDING FEDERALLY SPONSORED RESEARCH OR DEVELOPMENT

[0002] This invention was made with Government support under U01 AG061389 awarded by the National Institutes of Health. The Government has certain rights in this invention.TECHNICAL FIELD

[0003] The present disclosure is generally related to techniques for removing artifacts from data signals.BACKGROUND

[0004] Compared to functional brain imaging modalities which record relatively slow blood-oxygen-level-dependent signals (e.g., fMRI, fNIRS), Electroencephalography (EEG) records electrocortical dynamics with high temporal resolution. Additionally, in high-density applications (roughly 100+ channels), EEG also has the ability to localize brain sources with reasonable spatial resolution. Lastly, EEG is both portable and relatively cheap, making it ideal for studying theneural control of whole-body movement, both in laboratory settings and in the natural environment. In recent years, researchers have analyzed human brain activity during activities such as walking, running, bicycling, and table tennis. These papers highlight EEG’s potential to elucidate the neuromechanics of whole-body movement.

[0005] Electroencephalography (EEG) is very susceptible to artifacts because the signals it attempts to measure on the scalp are so small (tens of microvolts, uV). Although improvements to the EEG hardware (actively amplified electrodes) have helped provide cleaner raw EEG signals, motion-related artifacts are still a major obstacle to recording / analyzing brain activity during whole body movement. This limits both scientific research as well as commercial applications (e.g., brain computer interfaces for exoskeletons, virtual reality, and neuro-rehabilitation).

[0006] Depending on the recording environment, the EEG equipment being used, the task being studied, and individual variation in each participant’s behavior, a wide variety of artifacts hinder the ability to isolate electrocortical sources. Some artifacts are internal (or biological) in origin, such as muscle contractions, eye blinks, and eye saccades. Other artifacts come from the external environment, such as line noise artifacts and interference from various electrical equipment. Still, other artifacts originate from more complex mechanisms, such as motion artifacts, which are largely the result of cable sway. For example, as EEG cables sway through the air, they interact with each other and background electromagnetic fields. Inductive coupling and electromagnetic radiation in the environment can have a relatively large effect on EEG signal quality given the small voltage of electrocortical signals as they appear at the scalp (roughly 20 uV). Active electrodes can mitigate motion artifacts by amplifying the EEG signals prior to transmission and digital sampling. However, motion artifacts are still problematic for recording high fidelity EEG signals.

[0007] There are many software approaches to remove EEG artifacts, but no single method reliably cleans all types of artifacts and many methods are not well- suited for real-time implementation . Independent component analysis (ICA) is one of the most common approaches for blind EEG source separation and artifact removal. ICA can effectively remove eye blinks, muscle, and line noise. It can also extract high-quality, independent brain sources from mixed EEG data. However, ICA is computationally slow, requiring hours of computation on a modern home computer or work computer (non-supercomputer). Further, ICA generally requires a large amount of data to ensure a good decomposition. A general recommendation is to record at least 30 minutes of high-density EEG (100+ channels) at a sampling frequency of at least 500 Hz when attempting to separate sources with ICA. Generally speaking, ICA is not well-suited for real-time cleaning, although attempts have been made to remove eye blinks in real-time with ICA. Similarly, ensemble empirical mode decomposition is not further discussed due to its high computational cost.

[0008] When reference noise recordings are available, Adaptive Filtering is a popular approach for removing artifacts in real-time. Adaptive Filtering works by scaling one or more reference noise signals to optimally fit onto a single corrupted data signal of interest (minimizing the mismatch) and then subtracting the best fit. Thus, although often implemented recursively (online), Adaptive Filtering is based on the concept of linear regression, which can easily be solved (offline). Adaptive Filtering has been used to remove eye artifacts. Eye artifacts, in particular, are well suited for Adaptive Filtering because it is relatively easy to obtain accurate recordings of eye artifacts with only 2-4 electrooculogram electrodes. One limitation of Adaptive Filtering, however, is that it assumes noise sources project onto the EEG sensors and noise sensors nearly identically, with no significant difference other thansimple scaling (linear mixing). Thus, in certain scenarios, Adaptive Filtering may exhibit poor performance and need to be modified. For example, in Kilicarslan (2019), the authors reported poor performance when they tried to remove EEG motion artifacts using an inertial measurement unit (IMU) attached to the head (reference noise signals) and traditional Adaptive Filtering. See Kilicarslan, A.; Contreras-Vidal, J.L., “Characterization and Real-Time Removal of Motion Artifacts from EEG Signals,” J. Neural Eng. (2019). To achieve better performance, the authors modified the approach by using Volterra series expansion terms to capture nonlinear effects in the IMU signals. However, to implement the approach in Kilicarslan (2019), users must a priori know the frequency of the motion artifact and assume it remains the same throughout the data collection. Thus, the approach in Kilicarslan (2019) may not be easily generalizable across tasks, types of artifacts, or types of noise sensors.

[0009] When reference noise recordings are not available, Artifact Subspace Reconstruction (ASR) and Auto-Canonical Correlation Analysis (Auto-CCA) are two popular options. ASR is included by default with EEGLAB as an offline cleaning option, but it was first developed for real-time cleaning as part of BCILAB. The burst repair feature of ASR is based on principal component analysis. ASR has been shown to help with muscle and eye artifacts. ASR does not require reference noise signals; however, it does require clean EEG data for calibration. The user can supply the calibration data themselves (e.g., by recording clean resting-state data prior to recording the task of interest), or they can choose to automatically extract calibration data from contaminated data (assuming clean segments exist). Canonical correlation analysis (CCA) is a computationally efficient statistics method that can be used for blind source separation. Researchers previously applied CCA to EEG data and aslightly lagged (but otherwise identical) EEG dataset. This approach is often referred to as 1 -sample lag CCA, but we prefer the term Auto-CCA (auto-correlation extended to CCA). Auto-CCA was developed based on the idea that a small (e.g., 1- sample) shift represents a negligible phase change for low-frequency content but a significant phase change for high-frequency content. Therefore, relatively low- frequency Auto-CCA components should have strong correlation while relatively high-frequency Auto-CCA components should have weak correlation. Auto-CCA has been shown to remove high-frequency-muscle artifacts. Theoretically, Auto-CCA can also remove low-frequency artifacts such as motion artifacts and eye artifacts by rejecting high-correlation components, but users should exercise caution to avoid accidental deletion of brain activity (also low-frequency I high-correlation). Although ASR, Auto-CCA, and Adaptive Filtering are useful, there is still a need for improved cleaning performance and the ability to remove all types of artifacts with a single approach.SUMMARY

[0010] The present disclosure describes various embodiments of systems, apparatuses, and methods for removing artifacts from data signals. One such method comprises obtaining a data signal across a plurality of data channels, wherein the data signal has artifacts across one or more channels; generating a pseudo reference noise signal representing noise activity across a plurality of noise channels via filtering of the data signal; analyzing the data signal with the pseudo reference noise signal to identify noise components that exist within both the data signal and the reference signal; scaling the noise components to project upon the data signal across the plurality of channels; and / or cleaning the data signal bysubtracting the scaled noise components from the data signal across individual ones of the data channels.

[0011] Also disclosed herein is a system comprising a plurality of electroencephalography (EEG) electrodes positioned facing towards a head of a subject, wherein the plurality of EEG electrodes are configured to record a data signal representing brain activity of the subject across a plurality of EEG channels. The system further comprises a computing device that stores a signal cleaning program including computer-executable instructions configured to perform operations comprising: obtaining a recorded data signal representing the brain activity across the plurality of EEG channels, wherein the recorded data signal has artifacts across one or more EEG channels; generating a pseudo reference noise signal representing the noise activity across the plurality of noise channels via filtering of the recorded data signal; analyzing the recorded data signal with the pseudo reference noise signal to identify noise components that exist within both the recorded data signal and the pseudo reference noise signal; scaling the noise components to project upon the recorded data signal across the plurality of EEG channels; and / or cleaning the recorded data signal by subtracting the scaled noise components from the recorded data signal across individual ones of the EEG channels.

[0012] In one or more aspects of such systems, apparatuses, and methods, the noise components are identified using Canonical Correlation Analysis (CCA); the data signal represents brain activity across a plurality of electroencephalography (EEG) channels, wherein the plurality of data channels comprise the plurality of electroencephalography (EEG) channels; after cleaning the data signal, noise sources that overlap with data of interest are deleted without deleting the data ofinterest; the artifacts comprise at least one of motion artifacts, line noise, muscle artifacts, or eye artifacts; and / or a pseudo reference noise signal is generated by temporal filtering of the data signal.

[0013] In one or more aspects, such systems, apparatuses, and methods further involve or perform determining an optimal scaling factor that explains how each noise component projects onto each of the plurality of data channels, wherein the noise components are scaled using the optimal scaling factor.

[0014] Also disclosed herein is a method comprising obtaining, by at least one computing device, a data signal across a plurality of data channels; generating, by the at least one computing device, a pseudo reference noise signal representing noise activity across a plurality of noise channels via filtering of the data signal; analyzing, by the at least one computing device, the data signal with the pseudo reference noise signal to identify components that exist within both the data signal and the reference signal; and / or selecting, by the at least one computing device, data components of interest.

[0015] Also disclosed herein is a system comprising a plurality of electroencephalography (EEG) electrodes positioned facing towards a head of a subject, wherein the plurality of EEG electrodes are configured to record a data signal representing brain activity of the subject across a plurality of EEG channels. The system further includes a computing device that stores a signal cleaning program including computer-executable instructions configured to perform operations comprising: obtaining a recorded data signal representing the brain activity across the plurality of EEG channels, wherein the recorded data signal has artifacts across one or more EEG channels; generating a pseudo reference noise signal representing the noise activity across the plurality of noise channels viafiltering of the recorded data signal; analyzing the recorded data signal with the pseudo reference noise signal to identify components that exist within both the recorded data signal and the pseudo reference noise signal; and / or selecting data components of interest.

[0016] In one or more aspects of such systems, apparatuses, and methods, the pseudo reference noise signal is generated by temporal filtering of the data signal; the components are identified using Canonical Correlation Analysis (CCA); the data components of interest are selected based on at least one of frequency content, projection onto data channels, or a strength of the relationship between components as they exist in the data signal and reference signal; the data signal is an electroencephalography (EEG) signal; and / or the data components of interest comprise brain components.

[0017] Other systems, methods, features, and advantages of the present disclosure will be or become apparent to one with skill in the art upon examination of the following drawings and detailed description. It is intended that all such additional systems, methods, features, and advantages be included within this description and be within the scope of the present disclosure.BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Many aspects of the present disclosure can be better understood with reference to the following drawings. The components in the drawings are not necessarily to scale, emphasis instead being placed upon clearly illustrating the principles of the present disclosure. Moreover, in the drawings, like reference numerals designate corresponding parts throughout the several views.

[0019] FIG. 1 outlines a general framework for a signal cleaning algorithm that requires noise sensor hardware to take a set of noise signals and intelligently clean a set of signals of interest.

[0020] FIG. 2 outlines a general framework for an exemplary signal cleaning algorithm that does not require noise sensor hardware to intelligently clean a set of signals of interest in accordance with embodiments of the present disclosure.

[0021] FIG. 3 illustrates an exemplary signal cleaning algorithm to identify brain sources within a data signal in accordance with various embodiments of the present disclosure.

[0022] FIG. 4 depicts a schematic block diagram of a computing device that can be used to implement various embodiments of the present disclosure.

[0023] FIG. 5 shows a flow chart (on the left side) for a signal cleaning algorithm that requires noise sensor hardware to take a set of noise signals and intelligently clean a set of signals of interest and a flow chart (on the right side) for an exemplary signal cleaning algorithm that does not require noise sensor hardware to intelligently clean a set of signals of interest in accordance with embodiments of the present disclosure.

[0024] FIG. 6 shows a graphical user interface (GUI) implementation in accordance with various embodiments of the present disclosure.

[0025] FIG. 7 shows photographic images of (A) a design phantom head mold that is fitted with 10 antennae inside the mold to act as brain source and introduce artifacts (eye blink / saccade, neck muscle artifacts, facial muscle artifacts; (B) that is then filled with ballistics gelatin to mimic the physical and electrical properties of the human head; (C) that is fitted with conductive and a wig over the phantom head toact as scalp and hair; and (D) was secured to a robotic platform to induce motion artifacts during validation of an exemplary method of the present disclosure.

[0026] FIG. 8 provides an overview of validation of a newly developed signal cleaning method (also referred to as “iCanClean”) with (A) a flow chart summarizing a process used to validate iCanClean and compare it to other cleaning algorithms; (B) a summary of the conditions that were tested and which conditions had reference noise recordings available that could be used to assist cleaning; (C) a summary of cleaning algorithms tested and the main parameters that were varied; and (D) a summary how a Data Quality Score was calculated.

[0027] FIG. 9 shows a chart of summarized quantitative results showing iCanClean and a plurality of other cleaning method’s best Data Quality Score for six conditions tested in accordance with the present disclosure.

[0028] FIGS. 10A-10D present a qualitative (visual) assessment of the Brain condition before and after each processing method (iCanClean, Adaptive Filtering, Auto-CCA, ASR), using their respective ideal settings (maximized Data Quality Score after cleaning).

[0029] FIGS. 11A-11 D present a qualitative (visual) assessment of the Brain + Eyes condition before and after each processing method (iCanClean, Adaptive Filtering, Auto-CCA, ASR), using their respective ideal settings (maximized Data Quality Score after cleaning).

[0030] FIGS. 12A-12D presents a qualitative (visual) assessment of the Brain + Neck Muscles condition before and after each processing method (iCanClean,Adaptive Filtering, Auto-CCA, ASR), using their respective ideal settings (maximized Data Quality Score after cleaning).

[0031] FIGS. 13A-13D present a qualitative (visual) assessment of the Brain + Walking Motion condition before and after each processing method (iCanClean, Adaptive Filtering, Auto-CCA, ASR), using their respective ideal settings (maximized Data Quality Score after cleaning).

[0032] FIGS. 14A-14D present a qualitative (visual) assessment of the Brain + Facial Muscles condition before and after each processing method (iCanClean, Adaptive Filtering, Auto-CCA, ASR), using their respective ideal settings (maximized Data Quality Score after cleaning).

[0033] FIGS. 15A-15D present a qualitative (visual) assessment of the Brain + All Artifacts condition before and after each processing method (iCanClean, Adaptive Filtering, Auto-CCA, ASR), using their respective ideal settings (maximized Data Quality Score after cleaning).

[0034] FIG. 16 demonstrates the finding of brain components with iCanClean in phantom data.DETAILED DESCRIPTION

[0035] The present disclosure describes various embodiments of systems, apparatuses, and methods for removing artifacts from data signals, such as electroencephalography (EEG) data signals. As discussed, artifacts are a major obstacle to recording / analyzing brain activity during whole body movement. Depending on the specific scenario, different types of artifacts can include line noise artifacts, motion artifacts, muscle artifacts, and eye artifacts. This limits both basic research as well as commercial applications (e.g., brain computer interfaces, neurorehabilitation, etc.).

[0036] To overcome this limitation, novel systems and methods for cleaning EEG data have been developed. In accordance with embodiments of the present disclosure, an exemplary signal cleaning algorithm does not require dedicated "noise sensors" to assist in cleaning the data sensors. Rather, the corrupted EEG data signals themselves can be used as surrogate reference noise signals. For example, a copy of the corrupted EEG can be filtered in the frequency domain to remove brain content. Then, canonical correlation analysis can be applied to find noise sources in common to the original (noisy) EEG signals and the newly created "reference noise signals." Once noise sources are identified, the data can be cleaned. Thus, an exemplary signal cleaning algorithm can be performed without the need for specific noise sensor hardware. This means the cleaning method can now be more widely adopted. Additionally, the exemplary signal cleaning algorithm has the added benefit of identifying brain sources in the EEG data.

[0037] For reference, a basic visual outline of an approach using dedicated noise sensor hardware is provided in FIG. 1. Here, a set of noise signals are acquired using the noise sensor hardware. For example, dual-layer EEG sensors may be used to record raw noise signals such as motion artifacts and line noise (in addition to EEG signals). Then, EEG electrodes can record corrupted brain activity from the subject (desired brain activity plus unwanted artifacts / noise from internal / external sources), while the noise electrodes attempt to record strictly noise signals (whether they be motion artifacts, line noise, eye artifacts, muscle artifacts, or other type of artifact). Thus, a noisy EEG signal from the EEG electrodes and a reference noise signal from the noise electrodes / sensors have in common the noise from internal or external sources. Such pre-existing techniques can remove a variety of artifacts that commonly contaminate EEG signals so long as the appropriate noise recording(s)is / are available (e.g., dual-layer sensor for motion artifact and line noise removal, EMG for removing contamination by muscle contractions, EOG for eye blink and saccade removal, etc.).

[0038] Unlike the pre-existing techniques, a cleaning signal algorithm of the present disclosure does not utilize noise sensor hardware to obtain a noise reference signal. Consider FIG. 2 that shows a general framework for an exemplary cleaning algorithm that does not require noise sensor hardware to intelligently clean a set of signals of interest in accordance with embodiments of the present disclosure. In this new technique, raw data is filtered to attenuate the data signal of interest (dashed box “A”) thereby generating its own reference noise signals. Thus, there is no need to purchase dedicated noise sensors. Both the old method (using reference noise signals from extra sensors) and the new method (filtering data signals to become their own noise reference) provide clean channel-level data at the end. However, the new method also happens to find clean data at the source-level (dashed box “B”).

[0039] While filtering does not remove all of the data signal of interest, it removes most of it, such that it is sufficient to produce “pseudo” or “self-generated” reference noise signals. For example, a strong bandstop filter from 5-50 Hz is sufficient to remove most brain activity from EEG signals. Then, when these pseudo or selfgenerated reference noise signals sources are utilized in the cleaning signal algorithm, we not only obtain clean data channels, but we also happen to find data sources (FIG. 2, dashed box B). In the case of EEG, the data source are brain sources. Importantly, very little data is required to find the brain sources. For example, just 3-minutes of data was sufficient to find the brain source in FIG. 3 and the calculation took only a couple seconds. If you tried to find brain sources withindependent component analysis (ICA), you would need much more data (30+ minutes) and it would take much longer to process (hours versus seconds).

[0040] In accordance with various embodiments of the present disclosure, an exemplary cleaning algorithm uses Canonical Correlation Analysis (CCA) to identify latent relationships between the pseudo reference noise signals and the EEG data signals and thereby find the noise components that are contaminating the EEG signals so that the noise components can subsequently be removed. In various embodiments, CCA can provide the noise components in a ranked order according to their correlation strength so that components above a certain threshold value can be marked for removal. In various embodiments, an exemplary computing system may be configured with a knob or other controller to select or adjust the threshold value as desired, although note there are other options for selecting a subset of CCA components for removal as will be subsequently discussed. Once a subset of components has been identified for removal with the exemplary cleaning algorithm, the noise components are scaled and then subtracted from the raw EEG signals resulting in clean EEG signals.

[0041] Canonical Correlation Analysis (CCA) is a statistical method that is useful for inferring information between two sets of variables via their cross-covariance matrices. CCA finds relationships between two sets of variables X and Y, where X and Y can each be composed of many data channels (columns) and many samples (rows). For an input of raw data matrices (X, Y), CCA returns linear mixtures of X and Y which are labeled II and V, respectively, where II and V are sets of components (subspaces of X and Y) that are maximally correlated to each other. Note that the number of components returned from CCA (i.e., the number of columns in II and V) depend on the dimensions of X and Y such that it equals the minimumrank of X or Y, whichever is smaller. CCA also returns weight matrices A, B which define how the component sets II and V were extracted from the raw data (U = XMCA, V = YMCB where XMC and YMC denote the mean centered versions of X and Y), along with a vector R which quantifies the correlation between U and V. In ranking the results, Ui is the mixture (or subspace) of X that is most correlated with a mixture (or subspace) of Y (Vi), II2 is the next mixture of X that is most strongly correlated with a mixture of Y (V2), and so on with every UV pair being independent from the others. In accordance with various embodiments of the present disclosure, CCA is used to find latent relationships between pseudo reference noise signals and data signals recorded via EEG electrodes. In doing so, CCA returns mixtures of the EEG channels (i.e., components or subspaces) that most strongly resemble mixtures of the noise channels. Since the resulting components are returned in ranked order, a subset of noise components can be easily identified by using a simple threshold (e.g., by marking all components with an R-squared value greater than some desired cutoff value as being components that should be removed). However, other approaches could be employed for selecting a subset of CCA components for removal such as examining the power spectra of the components (e.g., to verify a particular component’s power spectral profile resembles that of noise rather than brain activity) or by using a random resampling (e.g., bootstrapping) approach to determine correlation values that occur above chance level rather than a priori picking a particular value for a threshold.

[0042] Once a subset of noise components has been identified, their presence on the EEG channels (or other signals of interest) may be removed. Since the amplitudes of the noise components (U,V) returned by CCA are normalized by default and since the amount of noise present on each EEG channel varies, thenoise components must first be appropriately scaled to each EEG channel prior to deletion. Thus, to remove the noise components, an exemplary signal cleaning algorithm determines an optimal scaling factor that explains how each component (or a set of components) project(s) onto each noisy EEG channel (or other signal of interest) that is to be cleaned. This optimal scale factor can be determined in a number of ways, including for example by using the Moore-Penrose pseudoinverse to calculate the inverse mapping from noise components to channels, given the forward mapping from channels to components that CCA yielded (i.e., given weight matrices A, B such that U = XMCA and V = YMCB, use a pseudoinverse to solve for the inverse mapping). However, the inventor has found that using a more general least squares solution (without the additional minimum norm constraint of the pseudoinverse approach) outperformed a pseudoinverse when the number of noise channels is small compared to the number of EEG channels. Thus, in the figures of the present disclosure, an exemplary signal cleaning algorithm uses a least squares solution to solve for the noise projections such as is implemented by the MATLAB built-in functions for matrix division (mrdivide and mldivide). Once the optimal scaling factors have been calculated, the scaled (projected) version of the noise components is subtracted from each channel of the noisy EEG data. In various embodiments, U or V (or a mixture of the two) can be used as the noise components for subtraction as preferences dictate. In accordance with embodiments of the present disclosure, the exemplary signal cleaning algorithm makes use of the pseudo reference noise signal to calculate a set of independent, underlying noise sources / components. Then, the noise sources can be individually scaled and deleted (noise source by noise source and EEG channel by EEG channel). In this way, an exemplary algorithm, which utilizes CCA to first identify latent noise components, is better ableto remove artifacts than approaches that simply scale up a single raw noise channel (not component) paired to a single raw EEG channel.

[0043] Independent component analysis (ICA) can also be used to help clean EEG signals. However, ICA is computationally expensive (on the order of hours to process data, often necessitating the use of supercomputer clusters to speed up the process). Further, one can run the same exact ICA algorithm twice in a row and obtain different results because it is a numerically driven minimization algorithm with random starting points. ICA also only considers the information from a single set of data channels (in this case, the EEG signals and the noise signals would have to be lumped together as one set of signals).

[0044] Diversely, an exemplary signal cleaning algorithm, in accordance with embodiments of the present disclosure, exploits the known relationship between the EEG signals and the pseudo reference noise signals (EEG signals contain noise but pseudo reference noise signals do not contain brain activity), runs extremely fast (on the order of seconds to minutes for data that would otherwise take ICA hours), and will return the same results every time given the same input data and parameters. Specifically, an exemplary cleaning algorithm uses canonical correlation analysis (CCA) to identify independent subspaces within the noise that are highly correlated with subspaces in the EEG recordings (i.e., the noisy subspaces within the EEG data). As discussed, a subset of identified noise components may then be scaled to find how they project onto the EEG sensors, such that the scaled version of the independent noise components may be removed from the EEG signal data in the time domain. It is noted that using ICA and using an exemplary CCA-based signal cleaning algorithm of the present disclosure need not be mutually exclusive. In fact, one useful application of the exemplary signal cleaning algorithm is to use it toremove problematic artifacts from raw EEG signals before later being passed on to ICA. ICA would then take the clean EEG channel data and decompose them into underlying brain sources. Thus, for real-time applications (where ICA would be too slow), users can apply the exemplary signal cleaning algorithm on its own; meanwhile, for offline applications, ICA can be used alongside the exemplary signal cleaning algorithm.

[0045] Correspondingly, systems and methods of the present disclosure are not limited to being applied to only noisy EEG signals and can be applied in order to clean other types of noisy signals in general (e.g., to remove motion artifacts from EMG signals).

[0046] FIG. 4 depicts a schematic block diagram of a computing device 400 that can be used to implement various embodiments of the present disclosure. An exemplary computing device 400 includes at least one processor circuit, for example, having a processor 1302 and a memory 404, both of which are coupled to a local interface 406, and one or more input and output (I / O) devices 1308. The local interface 406 may comprise, for example, a data bus with an accompanying address / control bus or other bus structure as can be appreciated. The CPU can perform various operations including any of the various operations described herein.

[0047] Stored in the memory 404 are both data and several components that are executable by the processor 402. In particular, stored in the memory 404 and executable by the processor 402 is a signal cleaning routine 412 in accordance with embodiments of the present disclosure. Also stored in the memory 404 may be a data store 414 and other data. The data store 414 can include data signal recordings, and potentially other data. In addition, an operating system may be stored in the memory 404 and executable by the processor 402. The I / O devices 408may include input devices, for example but not limited to, a keyboard, touchscreen, mouse, recording devices, and / or sensors, etc. Furthermore, the I / O devices 408 may also include output devices, for example but not limited to, a display, speaker, earbuds, audio output port, a printer, etc.

[0048] The following discussion shows how corrupt EEG signals can be temporarily converted to pseudo-reference noise signals to clean artifacts more generally (without the need for separate noise sensors). Using a phantom head apparatus, with ground-truth brain signals readily known, we collect EEG data corrupted by motion, muscle, eye, and line noise artifacts. We then perform multiple parameter sweeps on the EEG data and compare our newly developed method (also referred to as “iCanClean”) to competing methods (ASR, Auto-CCA, and adaptive reference filtering).

[0049] The acronym iCanClean stands for implementing Canonical correlation to Cancel Latent Electromagnetic Artifacts and Noise. iCanClean includes four main steps, as depicted on the left side of FIG. 5 (labeled as “A”). The first step is to identify candidate noise components by sending corrupt EEG data and reference noise data to canonical correlation analysis (CCA). In so doing, the iCanClean algorithm uses canonical correlation analysis (CCA) to identify latent relationships between raw (noisy) EEG data and reference noise data (e.g., dual-layer noise sensor recordings). Accordingly, CCA finds and returns, in ranked order, subspaces of the corrupt EEG channels that are most correlated with subspaces of the reference noise channels.

[0050] The second step is to select a subset of noise components for removal, for example, those with the strongest correlation. iCanClean marks CCA components for removal based on which components have the strongest correlation.The user selects the correlation threshold and decides whether the noise components should be constructed from mixtures of the EEG and / or mixtures of the noise channels.

[0051] The third step is to calculate the projection from the bad (noise) components to the EEG channels, for example, with linear regression. The fourth step is to directly subtract the projected noise components from the EEG channels. Because the noise components were calculated as a linear mixture of the EEG and / or noise channels and because the projection was calculated as a linear mixture of noise components, the iCanClean algorithm can be considered a spatial filter.

[0052] These steps can be applied to a large fixed-window or to a smaller moving-window (e.g., to deal with nonstationarity in the data).

[0053] When direct recording of reference noise signals with dedicated sensors is not possible, one alternative is to make use of pseudo-reference noise signals. As depicted on the right side of FIG. 5 (labeled as “B”), pseudo-reference signals can be created by taking contaminated EEG signals and applying a basic temporal filter to attenuate the majority of brain activity (e.g., 5-50 Hz notch filter). For example, a 5- 50 Hz notch filter will significantly remove brain activity from the EEG data while mostly sparing the artifacts. Once the pseudo-reference signals are created, the rest of the algorithm is the same. In various embodiments, the iCanClean algorithm is packaged into an EEGLAB plugin so others can easily implement an exemplary method on their own data. In various embodiments, the plugin includes a graphical user interface (GUI), as illustrated in FIG. 6, complete with the option for copying and / or filtering signals prior to finding noise sources in the EEG data (e.g., to create pseudo-reference noise signals).

[0054] At the top of the GUI implementation of FIG. 6, a button is provided for the user to recall their last used settings and another button is provided to revert any settings that may have changed back to their default values. The user first selects the corrupted data channels of interest (e.g., the desired signals to be cleaned, typically EEG channels). The user then selects channels to be used as a reference noise. The corrupted data and reference noise can be optionally preprocessed (filtered or re-referenced) prior to canonical correlation analysis (CCA). For example, in the case that the noise channels are connected to the same amplifier as the EEG system and share the same reference, temporarily referencing the noise channels to their own average could prevent accidental deletion of brain activity. Another optional section allows for the use of copying the CCA unmixing matrices from a calibration dataset and applying them to the current dataset, assuming the datasets are compatible (e.g., same exact set of channels). Candidate noise components are then identified in either a moving or a fixed window (user choice). CCA components with an R2correlation greater than a user-defined threshold are marked for removal. The user can choose to clean their data channels using mixtures of the data channels themselves (default), or using mixtures of the noise channels or, both. Similarly, the user can choose to clean the reference noise channels (if useful for subsequent processing), or the user can choose to skip cleaning the noise channels to save time (if it is not useful for subsequent processing). Additionally, in various embodiments, the GUI provides the user with other options such as visualizing the cleaning results and the ability to respect boundary events (e.g., breaks in the data) by applying iCanClean separately to each segment of data, no overlapping allowed).

[0055] For validation purposes, a phantom head apparatus was used to validate iCanClean. This allowed for ground-truth signals to quantify the cleaningperformance and to compare iCanClean with other real-time capable cleaning algorithms. Accordingly, the inventor made a custom head phantom, as illustrated in FIG. 7, based on the open EEG phantom project. The phantom head included antennae so that ground-truth electrical signals could be broadcast from inside the head. In particular, 10 antennae were placed inside the head to act as brain sources (ground-truth signals known). Additionally, 11 more antennae acted as artifactual sources. Two were for eye blinks (electrically coupled to a single source); one was for eye saccades; four were for neck muscle artifacts (bilateral trapezius and sternocleidomastoid); and four were for facial muscle artifacts (bilateral masseter and temporalis). The tip (T) and sleeve (S) portion of 3.5 mm TRS audio jacks was used to create electric dipoles for the antennae.

[0056] A mixture of ballistics gelatin (1 kg), deionized water (5 L), and salt (50 g NaCI) was created and poured into a plastic 3D-printed mold. The ballistics gelatin and salt mixture mimics the physical properties of the human body and allows for volume conduction of the electrical signals. A wig was added to the phantom (nonconductive) along with a layer of conductive fabric to mimic hair and scalp. The inventor placed the phantom on a robotic motion platform (hexapod) (NOTUS, Symetrie, France) to cause motion artifacts, as desired. Dual-layer EEG electrodes were placed on the head so that reference motion artifact signals could be recorded alongside traditional EEG. After gelling the scalp-facing electrodes, the outward facing (noise) electrodes were covered with conductive fabric and kept electrically isolated from the scalp-facing electrodes. EEG amplifiers were secured above the phantom, with cables hanging loosely but bundled. Eight external sensors (not shown) were placed over the neck and connected to the EEG amplifier. These sensors recorded neck electromyography (EMG) which could be used as referencenoise signals to remove neck muscle artifacts. Similarly, the dual-layer EEG sensors recorded motion and line noise artifacts.

[0057] Ground-truth brain signals for the head phantom were created using neural mass models. The inventor adjusted the parameters of the neural mass model to create signals that were similar to example human data. Specifically, the inventor randomized the input parameters to the neural mass model generator to calculate 10,000 candidate neural mass models. The power spectral density (PSD) of the output waveform was calculated for each set of input parameters and the power spectral profiles was compared to ICA components coming from clean resting state data (sitting eyes open). The 10 neural mass models were kept whose PSD curves best matched human data in the frequency band of 5-50 Hz (minimized sum of squares error). The ground-truth brain signals were broadcast into the head phantom with a digital-to-analogue converter (cDAQ-9178 chassis with Nl 9269 output modules, National Instruments, USA). Each of the ground truth signals was first normalized such that 99.5% of the data points were in the interval [-1 ,1]. The inventor then scaled the signal coming out of the digital-to-analogue converter (going into the phantom) so that the EEG signals measured at the scalp were in physiological range. Specifically, the inventor targeted 20 uV for the brain signals. This was done manually (iteratively adjusted and checked by eye).

[0058] Artifactual (contaminating) sources for the head phantom was created from example human data. Neck muscle sources are the same as in Richer (2020), which were created from direct bipolar electromyography (EMG) recordings of the trapezius and sternocleidomastoid muscle during a 1.5 m / s walking task. See Richer, N.; Downey, R.J.; Hairston, W.D.; Ferris, D.P.; Nordin, “A.D. Motion and Muscle Artifact Removal Validation Using an Electrical Head Phantom, Robotic MotionPlatform, and Dual Layer Mobile EEG,” IEEE Trans. Neural Syst. Rehabil. Eng. (2020). Meanwhile, facial muscle sources and eye blink / saccade sources were newly created for the present disclosure (voluntarily provided by the inventor). To create facial muscle sources, EEG signals were recorded during a trial which consisted of chewing food, swallowing food, and drinking water. To create eye sources, EEG was recorded during frequent blinking. Unlike the neck muscle sources, which were direct bipolar recordings, here example artifacts were extracted by sending mixed EEG data to independent component analysis. Known artifactual sources were broadcast into the head phantom using a digital-to-analogue converter, targeting various amplitudes at the scalp (150 uV neck muscles, 300 uV facial muscles, 150 uV eyes).

[0059] For validation purposes, we tested 6 conditions: Brain, Brain + Motion Artifacts, Brain + Eye Artifacts, Brain + Neck Muscle Artifacts, Brain + Facial Muscle Artifacts, and Brain + All Artifacts (i.e., all artifacts simultaneously present). The Brain condition represents relatively clean EEG data at rest, with no other artifacts purposefully introduced. The Brain condition, however, still contains small amounts of 60 Hz line noise which could not be avoided. Meanwhile, the inventor experimentally manipulated (imposed) motion, muscle, and eye artifacts with the phantom apparatus. For the conditions which involved motion of the robotic platform (Brain + Motion and Brain + All Artifacts), the same walking trajectory of 1.5 m / s, as in Richer (2020), was used. For the conditions which did not involve motion (Brain, Brain + Neck Muscles, Brain + Facial Muscles, and Brain + Eyes), the platform was stationary with the motors turned off. For conditions where muscle and / or eye artifacts were not involved, the associated sources were simply not broadcast to the antennae (i.e., turned off electrically by setting the output gain to zero).

[0060] EEG signals were recorded with a custom-made 128+128 channel duallayer system (ActiveTwo, BioSemi, Netherlands)]. The dual-layer system records raw noise signals (motion and line noise artifact) alongside traditional EEG electrodes. EEG signals were also recorded from 8 extra electrodes connected to the same EEG system. These extra electrodes were placed over the four neck muscle antennae (2 per muscle; 1 superior, 1 inferior) to serve as neck EMG sensors. Extra electrodes were not placed over the facial muscle antennae or eye antennae of the phantom because we prioritized having reference noise signals available for removing neck muscle artifacts. To promote cable sway and increase motion artifact during the Brain + Motion and Brain + All Artifacts conditions, the electrode cables were hung loosely. The digital-to-analog converter and EEG system were synchronized with trigger events at the beginning and end of each trial. This allowed us to directly compare the 10 ground-truth brain signals, which were broadcast into the head phantom with the 128-channel EEG data recorded at the scalp.

[0061] iCanClean was tested against competing methods in a large parameter sweep, as demonstrated in FIG. 8. Under part A of FIG. 8, a flow chart is shown depicting and overview of a cleaning algorithm validation with raw EEG data being imported and high-pass filtered to remove large DC (0 Hz) offsets and slow drifts (eegfiltne w function, 1 Hz highpass = -6 dB at 0.5 Hz). No channels were rejected or re-referenced. Minimally processed data were then sent through various cleaning algorithms, each with their own parameter sweep. After each preprocessing iteration, the cleaning effectiveness was quantified by a Data Quality Score, using ground truth knowledge of the underlying brain sources.

[0062] The various cleaning algorithms included four methods: iCanClean, ASR,Auto-CCA, and Adaptive Filtering, with variations on each method and theirparameters. An IMU-based Filtering approach was also tested for removing motion artifacts. For iCanClean, two main parameters of interest were varied. First, the R2threshold cutoff was varied from 0 (max cleaning) to 1 (no cleaning). The R2parameter adjusts the overall aggressiveness of the cleaning by determining how many noise components to remove. Second, the type of reference noise signals was varied. The choices varied between signals recorded by dual-layer EEG sensors (useful for removing motion and line noise artifacts), signals recorded by neck muscle electromyography (useful for removing neck muscle artifacts), and pseudoreference signals created from bandpass filtered EEG (useful for removing artifacts when reference noise signals are otherwise not available). For demonstration purposes, the inventor also tested the effect of temporarily re-referencing the EEG and noise signals prior to iCanClean (useful when noise signals are initially recorded with the same reference as the EEG signals). The inventor also tested cleaning the EEG data using mixtures of the EEG channels themselves versus using mixtures of the noise channels versus using mixtures of both.

[0063] Referring back to FIG. 8, part B of FIG. 8 presents a summary of the conditions that were tested and which conditions had reference noise recordings available that could be used to assist cleaning and part C of FIG. 8 presents a summary of cleaning algorithms tested and the main parameters that were varied. For Adaptive Filtering, an R2threshold was varied from 0 (max cleaning) to 1 (no cleaning) as well as the reference noise signal type, similar to iCanClean. The R2threshold was used to determine which noise channels were sufficiently correlated to each EEG channel prior to performing linear regression. Typically for Adaptive Filtering, all noise channels would be included in the regression problem (i.e., always set R2threshold = 0 for max aggressiveness). However, a range of thresholds wasimplemented to provide the fairest comparison possible (i.e., to demonstrate iCanClean’s improved performance is not due to the R2thresholding aspect).

[0064] For the Brain + Motion condition, an extension to Adaptive Filtering, described in Kilicarslan (2019), which cleans EEG data using accelerometer signals from an inertial measurement unit (IMU) attached to the head was also tested. See Kilicarslan, A.; Contreras-Vidal, J.L., “Characterization and Real-Time Removal of Motion Artifacts from EEG Signals,” J. Neural Eng. (2019)

[0065] For Auto-CCA, the R2threshold (0-1), the lag amount (1 , 2, 3, or 4 samples at 512 Hz), and whether high or low correlation Auto-CCA components were removed were varied. Note for Auto-CCA that R2values of 0 and 1 can flip roles (max cleaning versus no cleaning), depending on the rejection setting (high / low correlation). In existing literature, typically only low-correlation components are removed and only a lag of 1 is used. Here, the inventor also tested removing high- correlation components and tested multiple lags for completeness.

[0066] For ASR, the burst criterion value (1-250 standard deviations), which adjusts the overall aggressiveness of the cleaning, was varied. The inventor also tested directly providing a clean dataset to ASR (external calibration) versus asking ASR to automatically find clean sections from the same data to be cleaned (auto calibration). For the external dataset, ASR was provided with clean, stationary data from the Brain (only) condition (i.e., the ideal scenario with no eye, muscle, or movement artifacts).

[0067] For I Mil-based Filtering, an equivalent offline version of the algorithm described in Kilicarslan (2019), including the 2nd order Volterra expansion terms and cascade filtering approach was implemented. For the parameter sweep, the filterbank frequencies (base frequency, 0.85-1.05 Hz; number of harmonics, 1-4) and the optimization type (minimize L2 norm versus Linfinsty norm) was varied.

[0068] The cleaning performance was quantified with a Data Quality Score, based on correlation between EEG channels and ground-truth brain sources. Correspondingly, part D of FIG. 8 presents a summary how the data quality score was calculated based on Pearson’s R2correlation. We first time-warped the EEG data to the ground-truth brain sources using event markers (start and stop of each trial). We then calculated Pearson's R2correlation matrix between the EEG channels and the ground truth sources (128 EEG channels by 10 brain sources). The raw score was calculated by summing the R2correlation across all ground-truth brain sources (to calculate, for each EEG channel, the percent variance which is explained by the brain sources). We then averaged across all channels to yield a scalar value that quantifies how well the EEG data can be explained by a linear mixture of the ground-truth brain sources (0-100%, higher is better). Note that a score of 100% is achieved when the time series of the EEG channels are strictly mixtures of the 10 brain sources and nothing else (no other extraneous sources to introduce noise). However, note that a score of 100% could also be achieved in the case where 1 or more brain sources are accidentally deleted during cleaning, so long as no artifacts remain on the EEG channels. Therefore, a correction factor (penalty term) is introduced to emphasize we do not want to accidentally delete brain activity during cleaning.

[0069] The correction factor was calculated as follows. For each ground-truth brain signal, we used linear regression to first determine the best possible reconstruction that could be produced using linear mixtures of the minimally processed EEG data (Pre). After cleaning, the best possible reconstruction wasrecalculated for each ground-truth brain signal using mixtures of cleaned EEG data (Post). Potential deletion of each ground-truth brain source was quantified using the ratio of the best reconstructions (variance accounted for Post divided by variance accounted for Pre). A ratio of 1 indicates no brain source removal, whereas a ratio of 0.75 would indicate 25% removal of brain activity (Post relative to Pre). After calculating the ratio for each of the 10 brain sources, the minimum value across all brain sources was taken. We multiplied this correction factor to the raw score to calculate the final (corrected) data quality score.

[0070] The quantitative results for iCanClean, ASR, Auto-CCA, and Adaptive Filtering were analyzed by comparing each algorithm’s best cleaning performance (i.e. , Data Quality Score) for each condition. Then, time-series snapshots and power spectral density plots were constructed to qualitatively compare the algorithms at their optimal settings.

[0071] By testing many parameters for iCanClean and competing methods (ASR, Auto-CCA, and reference Adaptive Filtering), the best outcome (best Data Quality Score) was acquired for each cleaning algorithm tested, condition by condition. As depicted in FIG. 9, the iCanClean algorithm at its optimal settings consistently outperformed other real-time capable EEG cleaning algorithms at their respective optimal settings.

[0072] All algorithms except ASR slightly cleaned the Brain condition, which had mild line noise artifact contamination (unintended) but no purposefully imposed (experimentally manipulated) artifacts. This was the only condition where iCanClean was not the top performer, but iCanClean still performed well (2nd place) and the difference in the Data Quality Scores is negligible (59.1% iCanClean vs. 59.5% Adaptive Filtering, illustrated in FIG. 9.

[0073] FIGS. 10A-10D present a qualitative (visual) assessment of the Brain condition before and after each processing method (iCanClean, Adaptive Filtering, Auto-CCA, ASR), using their respective ideal settings (maximized Data Quality Score after cleaning). The top row of FIGS. 10A and 10B show a time series plot of raw EEG data for the Brain condition. Brain signals were sent to the phantom starting at time 0.5 seconds. The second row of FIGS. 10A and 10B shows a time series plot of ideally cleaned data. The third row shows a time series plot of the relative difference (what was removed during cleaning), and FIGS. 10C and 10D show power spectral density plots of the associated time series data (Raw = red, Clean = blue, and Difference = black). Note that Adaptive Filtering appears to have accidentally deleted some brain activity in addition to the line noise, as can be seen in the difference plot.

[0074] All algorithms improved the Brain + Eyes condition, but iCanClean (58.6%) and Auto-CCA (58.5%) were the best performers, as illustrated in FIG. 9. Correspondingly, FIGS. 11A-11 D present a qualitative (visual) assessment of the Brain + Eyes condition before and after each processing method (iCanClean, Adaptive Filtering, Auto-CCA, ASR), using their respective ideal settings (maximized Data Quality Score after cleaning). The top row of FIGS. 11 A and 11 B show a time series plot of raw EEG data for the Brain + Eyes condition. Brain and eye signals were sent to the phantom starting at time 0.5 seconds. The second row of FIGS. 11 A and 11 B show a time series plot of ideally cleaned data. The third row shows a time series plot of the relative difference (what was removed during cleaning), and FIGS. 11 C and 11 D show power spectral density plots of the associated time series data (Raw = red, Clean = blue, and Difference = black). Note that Adaptive Filtering appears to have accidentally deleted some brain activity, as can be seen in thedifference plot. Also note that the shape of the eye blink artifacts that were removed are distorted (more sinusoidal) for Adaptive Filtering (see Difference plot around time = 2.5 seconds). This is due to Adaptive Filtering using band pass filtered EEG channels as the pseudo-reference noise channels (no electrooculogram signals present). Here, iCanClean used the same pseudo-reference noise channels as Adaptive Filtering, but iCanClean was able to extract a more accurate estimate of the eye blink artifacts. Surprisingly, ASR did little to remove eye artifacts

[0075] All algorithms improved the Brain + Neck Muscles condition, but iCanClean (58.1 %) and Adaptive Filtering (56.4%) were the best performers, as illustrated in FIG. 9. Correspondingly, FIGS. 12A-12D present a qualitative (visual) assessment of the Brain + Neck Muscles condition before and after each processing method (iCanClean, Adaptive Filtering, Auto-CCA, ASR), using their respective ideal settings (maximized Data Quality Score after cleaning). The top row of FIGS. 12A and 12B show a time series plot of raw EEG data for the Brain + Neck Muscles condition. Brain and neck muscle signals were sent to the phantom starting at time 0.5 seconds. The second row of FIGS. 12A and 12B shows a time series plot of ideally cleaned data. The third row shows a time series plot of the relative difference (what was removed during cleaning), and FIGS. 12C and 12D show power spectral density plots of the associated time series data (Raw = red, Clean = blue, and Difference = black). Note that iCanClean was the only algorithm to simultaneously remove neck muscle artifacts and 60 Hz line noise. All other algorithms left the line noise in place.

[0076] All algorithms improved the Brain + Walking Motion condition, but iCanClean (54.3%) clearly outperformed the rest, as illustrated in FIG. 9 (next best score = 43.3%). Correspondingly, FIGS. 13A-13D present a qualitative (visual)assessment of the Brain + Walking Motion condition before and after each processing method (iCanClean, Adaptive Filtering, Auto-CCA, ASR), using their respective ideal settings (maximized Data Quality Score after cleaning). In FIGS. 13A and 13B, the top row shows time series plot of raw EEG data for the Brain + Walking Motion condition. Brain signals were sent to the phantom at time 0.5 seconds, and similarly, the hexapod motion platform began to move (walking motion) starting at time 0.5 seconds. The second row of FIGS. 13A and 13B shows a time series plot of ideally cleaned data. The third row shows a time series plot of the relative difference (what was removed during cleaning), and FIGS. 13C and 13D show power spectral density plots of the associated time series data (Raw = red, Clean = blue, and Difference = black). Note that iCanClean not only removed motion artifacts but also 60 Hz line noise and other electromagnetic interference from the motors on the hexapod motion platform. Auto-CCA removed milder motion (low amplitude, consistent frequency) artifacts along with line noise (as can be seen in the Difference plot); however, Auto-CCA was not able to remove more severe artifacts (large amplitude fluctuations remain in the ‘Clean’ plot, especially prior to 2 seconds). Meanwhile, ASR was only responsive to cleaning the worst contaminated section (first 2 seconds) and largely was not helpful for removing motion artifacts. iCanClean was not sensitive to the sub-type of motion artifact; it performed well at all time points.

[0077] All algorithms improved the Brain + Facial Muscles condition, but iCanClean (57.9%) and ASR (54.3%) were the best performers, as illustrated in FIG. 9. Correspondingly, FIGS. 14A-14D present a qualitative (visual) assessment of theBrain + Facial Muscles condition before and after each processing method (iCanClean, Adaptive Filtering, Auto-CCA, ASR), using their respective ideal settings(maximized Data Quality Score after cleaning). The top row of FIGS. 14A and 14B shows a time series plot of raw EEG data for the Brain + Facial Muscles condition. Brain and facial muscle signals were sent to the phantom starting at time 0.5 seconds. The second row of FIGS. 14A and 14B shows a time series plot of ideally cleaned data. The third row shows a time series plot of the relative difference (what was removed during cleaning), and FIGS. 14C and 14D show a power spectral density plots of the associated time series data (Raw = red, Clean = blue, and Difference = black). Note that iCanClean not only removed facial muscle artifacts but also 60 Hz line noise. Adaptive Filtering with pseudo-reference noise signals (bandstop filtered EEG) accidentally deleted brain activity, while iCanClean used the same pseudo-reference noise signals but kept brain activity intact.

[0078] When all artifacts were simultaneously involved, iCanClean offered a substantial improvement over all other methods. Starting from a score of 15.7% (before cleaning), the Brain + All Artifacts condition improved to 55.9% after iCanClean. Meanwhile, it only improved to 27.6%, 27.2%, and 32.9% after ASR, Auto-CCA, and Adaptive Filtering. For context, the Brain (only) condition scored 57.2% without cleaning, which is a reasonable value to target. Correspondingly, FIGS. 15A-15D present a qualitative (visual) assessment of the Brain + All Artifacts condition before and after each processing method (iCanClean, Adaptive Filtering, Auto-CCA, ASR), using their respective ideal settings (maximized Data Quality Score after cleaning). The top row of FIGS. 15A and 15B shows a time series plot of raw EEG data for the Brain + All Artifacts condition. Brain, eye, neck muscle, and facial muscle signals were sent to the phantom starting at time 0.5 seconds, and similarly, the phantom began to move (walking motion) at time 0.5 seconds. The second row of FIGS. 15A and 15B shows a time series plot of ideally cleaned data.The third row of FIGS. 15A and 15B shows a time series plot of the relative difference (what was removed during cleaning), and FIGS. 15C and 15D show power spectral density plots of the associated time series data (Raw = red, Clean = blue, and Difference = black). Note how the power spectra for the Brain + All Artifacts data (after iCanClean) resembles the already clean data in the Brain (only) condition (compare the blue curve in FIGS. 15C and 15D with the blue curve in FIGS. 10C and 10D.

[0079] Compared to other popular cleaning algorithms, such as Adaptive Filtering, Auto-CCA, and ASR, an exemplary novel method of the present disclosure has several theoretical advantages. Unlike Adaptive Filtering, which only considers one EEG channel at a time, iCanClean searches for latent relationships that exist between the entire set of EEG channels and the entire set of reference (or pseudoreference) noise channels. Thus, iCanClean can identify noise components as they exist in the subspace of all EEG channels, which should theoretically lead to better cleaning performance (i.e., better Data Quality Scores). Unlike Auto-CCA, which searches for relationships between EEG data and a lagged version of itself, iCanClean uses canonical correlation analysis (CCA) to search for relationships between EEG signals and reference (or pseudo-reference) noise signals. Thus, iCanClean exploits an extra set of valuable information not utilized by Auto-CCA, which should theoretically lead to better cleaning performance and reduce the likelihood of accidentally removing brain activity. Unlike ASR, which requires clean example data (user provided or automatically determined), iCanClean does not depend on clean example data. iCanClean can theoretically be applied directly to contaminated data, thereby saving time (unnecessary data collection).

[0080] Unlike Adaptive Filtering, Auto-CCA, and ASR, an exemplary novel method of the present disclosure provided consistently high Data Quality Scores, regardless of the type or number of artifacts present (i.e., for all conditions). Of particular note was the Brain + All Artifacts condition where iCanClean was the only algorithm to clean the data to satisfactory levels (i.e., similar score to data from the Brain condition). Thus, iCanClean shows promise as an all-in-one solution for removing a wide variety of artifacts.

[0081] There are some other patterns in the results worth noting. First, there was a benefit to cleaning EEG data using mixtures of the EEG channels themselves. Whereas Adaptive Filtering cleans EEG data using strictly mixtures of reference noise signals, iCanClean can also use mixtures of the EEG channels, if desired. As anticipated, an improved performance was witnessed when iCanClean was set to clean the corrupted EEG channels using mixtures of (raw) EEG channels themselves rather than using mixtures of reference (or pseudo-reference) noise channels. For example, iCanClean improved the Brain + All Artifacts condition to 55.9% using mixtures of EEG channels versus 39.0% using mixtures of noise channels. Second, there was a benefit to using CCA to identify common information between corrupt EEG data and reference noise signals prior to linear regression. Even when iCanClean was set to clean the corrupted EEG channels using only mixtures of noise channels (similar to Adaptive Filtering), iCanClean still achieved higher scores than Adaptive Filtering. For example, for the Brain + All Artifacts condition, iCanClean achieved a score of 39.0% using only mixtures of pseudoreference noise channels, and Adaptive Filtering with the same pseudo-reference noise channels scored 32.9%. Third, there was a benefit to using reference noise signals to guide CCA toward finding noise components (i.e., the iCanCleanapproach) as opposed to applying CCA to a lagged copy of the EEG signals for blind separation (i.e., the Auto-CCA approach). Besides better Data Quality Scores in general, the advantage of iCanClean over Auto-CCA is readily evident in the number of components that were removed for the Brain + Neck Muscle condition. A total of 4 independent neck muscle sources were sent into the head phantom. Therefore, in the ideal scenario, a cleaning algorithm is expected to identify exactly 4 distinct artifactual sources (or 5 with line noise). When 8-channel neck EMG were available as the reference noise signals, iCanClean removed a total of 4 components and achieved a score of 56.4%. When no neck EMG signals were available (and pseudo reference noise signals were used instead), iCanClean removed 6 components and achieved a score of 58.1 %. Meanwhile, Auto-CCA removed 43 components to achieve a score of 54.8%. This emphasizes that Auto-CCA may be more likely to accidentally delete brain activity, or to delete little at all, because Auto-CCA does not exploit reference noise signals to direct CCA toward finding the underlying artifacts.

[0082] There were also unexpected results. First, Auto-CCA did surprisingly well with eye artifact removal. Auto-CCA typically aims to remove low-correlation Auto- CCA components such as muscle artifacts. Here, for sake of completeness, we also tested removing high-correlation Auto-CCA components. Auto-CCA achieved good results removing high-correlation eye blink artifacts at the optimal settings. Note, however, that Auto-CCA components containing brain activity will have relatively high correlation, similar to eye artifacts. Therefore, users should be cautious when rejecting high-correlation components with the Auto-CCA approach. Second, it was surprising that the pseudo-reference version of iCanClean worked so well. It was expected that dedicated noise sensors such as dual-layer EEG would better capture motion artifacts than pseudo-reference signals, but this may not always be the case.For the Brain + Neck Muscles condition, using neck EMG reference electrodes scored 56.4% while the pseudo-reference version of iCanClean scored 58.1 %. For the Brain + Walking Motion condition, using dual-layer noise sensors scored 45.6% while the pseudo-reference version of iCanClean scored 54.3%.

[0083] One major implication to the findings of the present disclosure is that iCanClean may advance mobile brain imaging research by alleviating the need for accurate noise sensors and, perhaps, the need for noise sensors at all. Second, due to its minimal processing time, iCanClean may advance the feasibility of brain- computer-interfaces for mobile applications (e.g., thought-control of virtual reality, thought-control of exoskeletons, and neurofeedback for physical rehabilitation).

[0084] On the first point, one downside to Adaptive Filtering is its reliance on accurate noise sensors. When saving cost is a major consideration, an inertial measurement unit (IMU) on the head is a popular choice of noise sensor. IMU technology is cheap and readily available in a small form factor. Intuition suggests that signals from an IMU placed on the head should contain information about the motion artifacts of the electrodes placed on the head. However, in the inventor’s experience, IMU signals are not strongly related to the motion artifacts that appear on EEG electrodes. Note that motion artifacts are largely due to cable sway, which is not directly related to the movement of the head. Indeed, motion artifacts can be induced simply by moving the cables themselves, even if the head is stationary and the electrodes do not move relative to the scalp. Considering the relatively poor quality of IMU signals as reference noise signals, additional modifications to Adaptive Filtering may need to be employed to remove motion artifacts. For example, in Kilicarslan (2019), the IMU-based algorithm employs narrow filter banks, a Volterra series expansion, an H-infinity update law, and cascade filtering. Thenarrow filter banks decompose a single accelerometer signal into multiple frequency bands of interest, tied to the fundamental movement frequency and its harmonics (apriori supposedly known information). The Volterra series expansion turns each original reference signal into multiple new signals, hopefully capturing nonlinear effects. The H-infinity filtering update law solves the linear regression problem according to the Linfinity norm rather than L2 for robustness purposes. The cascade filtering solution to the regression problem helps prevent overfitting. The inventor implemented an equivalent offline version of the IMU-based filtering algorithm in Kilicarslan (2019). It was found that the IMU-based algorithm did not provide suitable cleaning performance, despite the various modifications to Adaptive Filtering. Meanwhile, it was showed that iCanClean can remove motion artifacts (and other artifacts in general), using only pseudo-reference signals (no dedicated noise sensors). Thus, iCanClean may advance mobile brain imaging research by alleviating the need for accurate noise sensors and, perhaps, the need for noise sensors at all. The method in Kilicarslan (2019) also requires a priori knowledge of the fundamental movement artifact frequency and requires the frequency to remain constant throughout the data collection. This is unrealistic for most real-world applications since individuals frequently speed up and slow down while walking. iCanClean does not have these limitations to real-world implementation.

[0085] On the second point, although the method in Kilicarslan (2019) is presented as being real-time capable, the offline equivalent version was found to be very slow, particularly when the Linfinity norm (rather than the L2 norm) was used. For example, the IMU-based algorithm took 128 hours to clean 5 minutes of the Brain + Walking Motion data, and it worsened the Data Quality Score from 34.1 % (raw) to 2.5% (Linfinity, 1 harmonic, 3 IMU gravity corrected signals become 42 after Volterra2nd order expansion, cascade filtering solution). Although implemented offline in the present disclosure, the iCanClean algorithm is computationally efficient and has potential for real-time cleaning applications. Thus, iCanClean may advance the feasibility of brain-computer-interfaces as an all-in-one cleaning approach by removing multiple types of artifacts, simultaneously, in a computationally efficient manner, without requiring dedicated hardware to record reference noise signals.

[0086] When reference noise signals contain brain activity, average rereferencing to temporarily remove brain activity may generally be a good idea. For example, our EMG recordings shared a reference electrode with the EEG system, so the raw EMG recordings naturally contain some brain activity. If CCA were used to search for latent relationships between the raw EMG and EEG recordings, it could accidentally delete brain activity (because brain activity would be in common to both the reference noise signals and the contaminated EEG). By average re-referencing the EMG sensors to themselves, the common activity across all EMG sensors is removed and one rank of data is temporarily deleted prior to CCA. The present results showed this to improve the Data Quality Score. New users can view an optional plot with the iCanClean GUI that shows the R2correlation as a function of the CCA component number (“Plot R2correlation” checkbox in FIG. 6). This may make it easier to more quickly determine the optimal range of R2values for a particular setup (type of noise sensors, number of sensors, the task being studied, etc.). Users may also want to vary the window size for their application. When using CCA to identify the noise components, the user has a choice as to what sections of data (time windows) to include. The noise components can be calculated on the entire set of data or with a moving window. Assuming the data is stationary, longer windows should be better able to find and delete latent noise components byexploiting more total data. Meanwhile, shorter windows could better handle non- stationary data.

[0087] In a previous work, the inventor showed that a moving-window version of iCanClean with 120+120 dual-layer EEG sensors can improve the ICA decomposition of motion artifacted data, for both table-tennis and uneven terrain treadmill walking. Based on Gonsisko (2023), it appears that a shorter time window (2-4 seconds) is better for removing motion artifacts with dual-layer EEG in human data, as opposed to a longer window. See Gonsisko, C.B.; Ferris, D.P.; Downey, R.J., “ICanClean Improves Independent Component Analysis of Mobile Brain Imaging with EEG,’’ Sensors (2023). It is suspected that this is due to non- stationarity in the data. Performance may be improved by incorporating larger windows of data. If there proves to be a benefit to using more data, then it is possible that CCA could be implemented recursively in the future, given recursive CCA algorithms exist in literature. This would reduce the computational cost of calculating CCA on a very large (and ever growing) window of data for real-time applications. Thankfully, the performance using a moving 2-second window is already relatively good. This makes iCanClean really easy to implement in real-time and minimizes computational cost. In various embodiments, removal of other types of artifacts may also be performed. For example, there may be significant potential for iCanClean to improve EEG studies that simultaneously record fMRI or apply transcranial magnetic stimulation (TMS), both of which induce large artifacts on the EEG recordings. For example, the work in Chowdhury (2014), Steyrl (2017), van der Meer (2016), Steyrl(2019) may be improved by applying iCanClean. Additionally, iCanClean’s can also potentially identify brain components from noisy EEG data, similar to ICA, but with less total data needed. See Chowdhury, M.E.H.; Mullinger, K.J.; Glover, P.; Bowtell,R, “Reference Layer Artefact Subtraction (RLAS): A Novel Method of Minimizing EEG Artefacts During Simultaneous FMRI,” Neuroimage (2014); Steyrl, D.; Krausz, G.; Koschutnig, K.; Edlinger, G.G.; Muller-Putz, G.R., “Reference Layer Adaptive Filtering (RLAF) for EEG Artifact Reduction in Simultaneous EEG-FMRI,” J. Neural Eng. (2017); van der Meer, J.N.; Pampel, A.; Van Someren, E.J.W.; Ramautar, J.R.; van der Werf, Y.D.; Gomez-Herrero, G.; Lepsien, J.; Hellrung, L.; Hinrichs, H.; Moller, H.E.; et al., “Carbon-Wire Loop Based Artifact Correction Outperforms Postprocessing EEG / FMRI Corrections-A Validation of a Real-Time Simultaneous EEG / FMRI Correction Method,” Neuroimage (2016); Steyrl, D.; Muller-Putz, G.R., “Artifacts in EEG of Simultaneous EEG-FMRI: Pulse Artifact Remainders in the Gradient Artifact Template Are a Source of Artifact Residuals after Average Artifact Subtraction,” J. Neural Eng. (2019). In FIG. 3, the finding of brain components with iCanClean in human data is demonstrated, and in FIG. 16, the finding of brain components with iCanClean in phantom data is demonstrated.

[0088] Here, in FIG. 16, ground truth brain sources are recovered from the Brain + All Artifacts condition. Accordingly, pseudo reference signals are used in iCanClean (band stop 4-50 Hz), but instead of using iCanClean to clean the channels, it was checked if any of the iCanClean components resembled the groundtruth brain components. Plotted is the R2 correlation between each ground truth brain component (10 total, labeled by anatomical location in the phantom head) and the iCanClean component that best matches it. Note that these values have been normalized according to the best that any linear unmixing algorithm could do (e.g. ICA). For example, the best possible R2 correlation for the left hand is 63.1% (strictly limited to using linear mixtures of the EEG channels), and iCanClean found a candidate component with an R2 correlation of 58.3%. After normalizing, the groundtruth recovery value for this brain source is 92.4% (58.3 / 63.1 = 92.4%). Overall, iCanClean did surprisingly well as it reconstructed 6 out of 10 brain components with a score >85% and 9 out of 10 <75%. Identifying brain sources helps explain why the pseudo-reference version performed so well. Note also that iCanClean auto sorted all the brain components to the end of the list. Out of 128 total components, the best matching components were #119-128 (i.e., the last 10).

[0089] In brief, iCanClean is a promising, all-in-one, real-time capable cleaning algorithm which could facilitate the analysis of EEG data corrupted by a wide variety of artifacts. It is computationally efficient and it can simultaneously remove multiple types of artifacts.

[0090] Certain embodiments of the present disclosure can be implemented in hardware, software, firmware, or a combination thereof. If implemented in software, signal cleaning logic or functionality, in accordance with embodiments of the present disclosure, are implemented in software or firmware that is stored in a memory and that is executed by a suitable instruction execution system. If implemented in hardware, the signal cleaning logic or functionality can be implemented with any or a combination of the following technologies, which are all well known in the art: a discrete logic circuit(s) having logic gates for implementing logic functions upon data signals, an application specific integrated circuit (ASIC) having appropriate combinational logic gates, a programmable gate array(s) (PGA), a field programmable gate array (FPGA), etc.

[0091] It should be emphasized that the above-described embodiments are merely possible examples of implementations, merely set forth for a clear understanding of the principles of the present disclosure. Many variations and modifications may be made to the above-described embodiment(s) without departingsubstantially from the principles of the present disclosure. All such modifications and variations are intended to be included herein within the scope of this disclosure.

Claims

CLAIMSTherefore, at least the following is claimed:1 . A method comprising: obtaining, by at least one computing device, a data signal across a plurality of data channels, wherein the data signal has artifacts across one or more channels; generating, by the at least one computing device, a pseudo reference noise signal representing noise activity across a plurality of noise channels via filtering of the data signal; analyzing, by the at least one computing device, the data signal with the pseudo reference noise signal to identify noise components that exist within both the data signal and the reference signal; scaling, by the at least one computing device, the noise components to project upon the data signal across the plurality of channels; and cleaning, by the at least one computing device, the data signal by subtracting the scaled noise components from the data signal across individual ones of the data channels.

2. The method of claim 1 , wherein the noise components are identified using Canonical Correlation Analysis (CCA).

3. A method of claim 1 , wherein the data signal represents brain activity across a plurality of electroencephalography (EEG) channels, wherein the plurality of data channels comprise the plurality of electroencephalography (EEG) channels.

4. The method of claim 1 , further comprising determining an optimal scaling factor that explains how each noise component projects onto each of the plurality of data channels, wherein the noise components are scaled using the optimal scaling factor.

5. The method of claim 1 , wherein after cleaning the data signal, noise sources that overlap with data of interest are deleted without deleting the data of interest.

6. The method of claim 1 , wherein the artifacts comprise at least one of motion artifacts, line noise, muscle artifacts, or eye artifacts.

7. The method of claim 1 , wherein a pseudo reference noise signal is generated by temporal filtering of the data signal.

8. A system comprising: a plurality of electroencephalography (EEG) electrodes positioned facing towards a head of a subject, wherein the plurality of EEG electrodes are configured to record a data signal representing brain activity of the subject across a plurality of EEG channels; and a computing device that stores a signal cleaning program including computerexecutable instructions configured to perform operations comprising: obtaining a recorded data signal representing the brain activity across the plurality of EEG channels, wherein the recorded data signal has artifacts across one or more EEG channels; generating a pseudo reference noise signal representing the noise activity across the plurality of noise channels via filtering of the recorded data signal; analyzing the recorded data signal with the pseudo reference noise signal to identify noise components that exist within both the recorded data signal and the pseudo reference noise signal; scaling the noise components to project upon the recorded data signal across the plurality of EEG channels; and cleaning the recorded data signal by subtracting the scaled noise components from the recorded data signal across individual ones of the EEG channels.

9. The system of claim 8, wherein the noise components are identified using Canonical Correlation Analysis (CCA).

10. The system of claim 8, wherein the operations further comprise determining an optimal scaling factor that explains how each noise component projects onto each of the plurality of EEG channels, wherein the noise components are scaled using the optimal scaling factor.

11. The system of claim 8, wherein the artifacts comprise at least one of motion artifacts, line noise, muscle artifacts, or eye artifacts.

12. The system of claim 8, wherein the pseudo reference noise signal is generated by temporal filtering of the data signal.

13. A method comprising: obtaining, by at least one computing device, a data signal across a plurality of data channels; generating, by the at least one computing device, a pseudo reference noise signal representing noise activity across a plurality of noise channels via filtering of the data signal; analyzing, by the at least one computing device, the data signal with the pseudo reference noise signal to identify components that exist within both the data signal and the reference signal; and selecting, by the at least one computing device, data components of interest.

14. The method of 13, wherein the pseudo reference noise signal is generated by temporal filtering of the data signal.

15. The method of 13, wherein the components are identified using Canonical Correlation Analysis (CCA).

Images