Adaptive threshold calibration for spike detection

Abstract

Provided is a method and system for detecting one or more spikes in an electrophysiological signal corresponding to neurological activity. The method includes receiving and enhancing the electrophysiological signal using a signal to noise ratio (SNR) enhancement technique to obtain an enhanced signal. The method further includes calculating a standard deviation of the enhanced signal based on a distribution constant corresponding to the SNR-enhanced signal. The method further includes calculating a threshold based on the standard deviation. The method further includes detecting one or more spikes based on the threshold.

Description

ADAPTIVE THRESHOLD CALIBRATION FOR SPIKE DETECTIONTechnical Field

[0001] The embodiments disclosed herein relate to spike detection of electric signals and in particular spikes of electrophysiological signals corresponding to activity.Introduction

[0002] Neural spikes, as captured in electrophysiological recordings, particularly through extracellular methods, are crucial for understanding the complex dynamics of the brain. Unlike action potentials that occur within individual neurons, spikes in these recordings often represent the combined activity of multiple neurons firing near-simultaneously. This collective neural activity is vital for interpreting the functional aspects of brain networks.

[0003] Generally, spike detection relies on setting a threshold level, with neural activity above this threshold identified as a spike. In existing systems, particularly those with many electrodes and corresponding channels, also known as high-channel recording systems, setting this threshold in the process of spike detection presents unique challenges.

[0004] In some existing systems that perform spike detection and spike sorting, this threshold level is set manually. Thus, in systems with a high number of electrodes, this approach becomes increasingly complex and burdensome. The threshold for spike detection needs to be individually defined for each electrode, accommodating variations that can occur between different individuals, over time, and across the areas of the high channel count pelectrode array. It is also necessary to record the data with high sampling rate (i.e. 2* spike activity highest frequency content, ~15kHz), digitize the recordings, transmit the data to external unit, find the threshold value and transmit the threshold information back to the integrated circuit (IC) to set the threshold value. All of these steps need to be repeated (i.e. periodically) to account for the changes in the threshold value over time. Therefore, even if after setting the threshold value there is no further need for constant recording with high bandwidth or even digitization, implanted ICs require the capability to be repeatedly calibrated. For existing manual calibration systems, this recalibration necessity restricts the design process particularly for implantable applications where access and power consumption after implantation are restricted.

[0005] Furthermore, in existing systems, available power is divided across each channel. Therefore, in a high channel count system, particularly in an implantable application where power consumption is a critical constraint, there is very limited power budget per channel. Due to the power-noise trade-off, this restriction on power consumption results in large integrated input referred noise of the recording frontends, in existing systems. This along with the environment background noise, results in operating conditions of low Signal-to-Noise Ratio (SNR) recording, making the detection of genuine spikes (i.e. spikes corresponding to activity) difficult particularlyin existing systems, using the raw data of the bioelectrical signal (post amplification and bandpass filtering to isolate spiking activity).

[0006] In some existing systems, SNR enhancement techniques, such as the Nonlinear Energy Operator (NEO) or Energy of Derivative (ED) are applied to attenuate the low-frequency and low amplitude content of the signal (often where most of the noise resides) and accentuate the high-frequency and high amplitude content, which is indicative of spiking activity.

[0007] Even with SNR enhancement techniques, however, the threshold must be determined. In some existing systems, the threshold is based on an average of the signal (whether raw or enhanced). Averaging the signal, however, often produces an overly aggressive threshold that may exclude activity spikes or an under aggressive threshold that may include noise.

[0008] Accordingly, there is a need for spike detection systems and methods that overcome the difficulties described above.Summary

[0009] According to some embodiments, there is a method for detecting spikes in an electrophysiological signal corresponding to neurological activity. The method includes receiving and enhancing the electrophysiological signal using an SNR enhancement technique to obtain an enhanced signal. The method further includes calculating the standard deviation of the enhanced signal based on a distribution constant corresponding to the signal to noise ratio (SNR) enhancement technique. The method further includes calculating a threshold based on the calculated standard deviation. The method further includes detecting spikes based on the calculated threshold.

[0010] The SNR enhancement technique may be an energy of derivative (ED) technique.

[0011] Receiving the electrophysiological signal may include amplifying or filtering the electrophysiological signal.

[0012] According to some embodiments, there is a system for detecting spikes in an electrophysiological signal corresponding to neurological activity. The system includes an electrode configured to receive the electrophysiological signal to obtain a raw signal. The system further includes a signal enhancement module configured to enhance the raw signal using an SNR enhancement technique to obtain an enhanced signal. The system further includes a standard deviation calculator configured to calculate the standard deviation of the enhanced signal based on a distribution constant corresponding to the signal to noise ratio (SNR) enhancement technique. The system further includes a threshold calculator configured to calculate a threshold based on the calculated standard deviation. The system further includes a spike detection module configured to detect spikes based on the calculated threshold.

[0013] According to some embodiments, there is a method for detecting spikes in an electrophysiological signal corresponding to neurological activity. The method includes receiving and enhancing the electrophysiological signal using a SNR enhancement technique to obtain an enhanced signal. The method further includes calculating a standard deviation of the enhanced signal based on a distribution constant corresponding to the SNR enhanced signal. The method further includes calculating a threshold based on the standard deviation. The method further includes detecting one or more spikes based on the threshold.

[0014] The SNR enhancement technique may be an energy of derivative (ED) technique.

[0015] Receiving the electrophysiological signal may include amplifying or filtering the electrophysiological signal.

[0016] The electrophysiological signal may be amplified and passed through a band pass filter.

[0017] The electrophysiological signal may include local field potentials (LFPs).

[0018] The LFPs may be removed from the electrophysiological signal by a high pass pole.

[0019] The SNR enhancement technique may be a nonlinear energy operator (NEO) technique.

[0020] The one or more spikes may be sparse relative to background noise of the electrophysiological signal.

[0021] The standard deviation may be representative of the variation of the background noise.

[0022] The standard deviation may be calculated in the analog domain.

[0023] The standard deviation may be calculated in the digital domain.

[0024] The background noise may follow a Gaussian distribution with a mean of zero.

[0025] The distribution constant may be 0.234.

[0026] Calculating the threshold may include setting the threshold to a predetermined multiple of the standard deviation by a predetermined scaling factor.

[0027] The predetermined scaling factor may be 5.

[0028] The predetermined scaling factor may be 2 or 3.

[0029] Detecting the one or more spikes may include comparing the enhanced signal to the threshold.

[0030] The one or more spikes may be detected where the enhanced signal exceeds the threshold.

[0031] The electrophysiological signal may be received by an electrode.

[0032] The electrode may be a passive electrode.

[0033] The electrode may be an active electrode.

[0034] The electrode may include a high pass filter.

[0035] The electrode may include an amplifier.

[0036] The amplifier may have a bandwidth between 300 Hz to 8 kHz to remove the LFPs while passing the one or more spikes.

[0037] The amplifier may be a two-stage capacitively-coupled operational transconductance amplifier (OTA).

[0038] The amplifier may include a first stage including a low-noise OTA.

[0039] The amplifier may include a second stage including a low-power OTA.

[0040] The low-power OTA may be a current mirror OTA.

[0041] The ED may be derived from squaring a derivative of the electrophysiological signal.

[0042] The derivative may be calculated by a differentiator.

[0043] The differentiator may be implemented using passive components.

[0044] The differentiator may include a clock frequency selected based on a Nyquist rate of the electrophysiological signal.

[0045] According to some embodiments, there is a system for detecting spikes in an electrophysiological signal corresponding to neurological activity. The system includes an electrode configured to receive the electrophysiological signal to obtain a raw signal. The system further includes a signal enhancement module configured to enhance the raw signal using a SNR enhancement technique to obtain an enhanced signal. The system further includes a standard deviation calculator configured to calculate a standard deviation of the SNR-enhanced signal based on a distribution constant corresponding to the signal to noise ratio (SNR) enhanced signal. The system further includes a threshold calculator configured to calculate a threshold based on the standard deviation. The system further includes a spike detection module configured to detect one or more spikes based on the threshold.

[0046] The electrophysiological signal may be amplified and passed through a band pass filter.

[0047] The electrophysiological signal may include local field potentials (LFPs).

[0048] The LFPs may be removed from the electrophysiological signal by a high pass pole.

[0049] The SNR enhancement technique may be a nonlinear energy operator (NEO) technique.

[0050] The one or more spikes may be sparse relative to background noise of the electrophysiological signal.

[0051] The standard deviation may be representative of the variation of the background noise.

[0052] The standard deviation may be calculated in the analog domain.

[0053] The standard deviation may be calculated in the digital domain.

[0054] The background noise may follow a Gaussian distribution with a mean of zero.

[0055] The distribution constant may be 0.234.

[0056] The threshold may be a predetermined multiple of the standard deviation by a predetermined scaling factor.

[0057] The predetermined scaling factor may be 5.

[0058] The predetermined scaling factor may be 2 or 3.

[0059] The spike detection module may detect the one or more spikes by comparing the enhanced signal to the threshold.

[0060] The one or more spikes may be detected where the enhanced signal exceeds the threshold.

[0061] The electrode may be a passive electrode.

[0062] The electrode may be an active electrode.

[0063] The electrode may include a high pass filter.

[0064] The electrode may include an amplifier.

[0065] The amplifier may have a bandwidth between 300 Hz to 8 kHz to remove the LFPs while passing the one or more spikes.

[0066] The amplifier may be a two-stage capacitively-coupled operational transconductance amplifier (OTA).

[0067] The amplifier may include a first stage including a low-noise OTA.

[0068] The amplifier may include a second stage including a low-power OTA.

[0069] The low-power OTA may be a current mirror OTA.

[0070] The ED may be derived from squaring a derivative of the electrophysiological signal.

[0071] The derivative may be calculated by a differentiator.

[0072] The differentiator may be implemented using passive components.

[0073] The differentiator may include a clock frequency selected based on a Nyquist rate of the electrophysiological signal.

[0074] Other aspects and features will become apparent, to those ordinarily skilled in the art, upon review of the following description of some exemplary embodiments.Brief Description of the Drawings

[0075] The drawings included herewith are for illustrating various examples of articles, methods, and apparatuses of the present specification. In the drawings:

[0076] Figure 1 is a flow diagram of a Brain-Computer Interface (BCI) application, according to an embodiment;

[0077] Figure 2 is a flow diagram of a method for detecting spikes of Figure 1 , according to an embodiment;

[0078] Figure 3 is example graphs of performance of the spike detection of Figure 2 for high and low SNR recordings using raw and SNR enhanced signals which shows sensitivity improvements especially in low SNR conditions, according to an embodiment;

[0079] Figure 4 is simulation results for sensitivity versus false positive rate receiver operating characteristic (ROC) generated spike detection with and without ED operator of Figure 2, according to an embodiment;

[0080] Figure 5A is a distribution of an ED enhanced signal of Figure 2, according to an embodiment;

[0081] Figure 5B is a feedforward structure for computing the distribution constant of Figure 5A, according to an embodiment;

[0082] Figure 6 is a distribution of ED signal of Figure 5A for a Gaussian noise input with various standard deviations, according to an embodiment;

[0083] Figure 7A is a block diagram of a system for detecting spikes of Figure 1 , according to an embodiment;

[0084] Figure 7B is a block schematic of a system for detecting spikes of Figure 1 , according to an embodiment;

[0085] Figure 8 is a block diagram of a Top-level implementation of an ED-based adaptive threshold generation IC of Figures 7A and 7B, according to an embodiment;

[0086] Figure 9 is a block diagram schematic with transistor sizing and simulation results showing the closed-loop gain and phase margin of a two stage amplifier of Figures 7A and 7B, according to an embodiment;

[0087] Figure 10A is a schematic of a circuit-level implementation of a switched-capacitor- based differentiator of Figures 7A and 7B, according to an embodiment;

[0088] Figure 10B is a phase plot of a circuit-level implementation of a switched-capacitor- based differentiator of Figure 10A, according to an embodiment;

[0089] Figure 11 is a simulation result showing the functionality of the pair of opposite phase differentiators of Figures 7A and 7B, according to an embodiment;

[0090] Figure 12A is a schematic along with the transistor sizings of a Gilbert cell of Figures 7A and 7B, according to an embodiment;

[0091] Figure 12B is a simulation results showing the functionality of a Gilbert cell of Figure 12A, according to an embodiment;

[0092] Figure 13 is schematic diagram of a detailed implementation of subtraction and threshold voltage calculation in a feedback loop of Figures 7A and 7B and the simulation result showing the variations at Gilbert cell of Figure 12A’s low and high impedance outputs, according to an embodiment;

[0093] Figure 14A is a circuit-level implementation of a switched-capacitor based rolling average of Figures 7A and 7B, according to an embodiment;

[0094] Figure 14B is simulation results for varying input DC levels and input pulse signal with varying duty cycles input to the switched-capacitor based rolling average of Figure 14A, according to an embodiment;

[0095] Figure 14C is a flow schematic indicating the operation of a resistance capacitor structure (RC) without a switched capacitor of Figure 14A, according to an embodiment;

[0096] Figure 15A is experimental measurement results of input referred noise of the two- stage amplifier of Figures 7A and 7B, according to an embodiment;

[0097] Figure 15B is experimental measurement results of gain and bandwidth the two-stage amplifier of Figures 7A and 7B, according to an embodiment;

[0098] Figure 16 is a dynamic response of the spike detection system of Figures 7A and 7B to Gaussian noise input, showing the automatically calculated standard deviation and the adaptive threshold over a 20-second interval, according to an embodiment;

[0099] Figure 17 is experimental measurement results of the spike detection system of Figures 7A and 7B to input noise with varying standard deviation, ED output with the automatically generated threshold, in-loop comparator output indicating 1s density, and identification of falsely detected spikes, according to an embodiment;

[0100] Figure 18 is experimental measurement results showing the functionality of the spike detection of Figures 7A and 7B in the presence of LFP signal ED-based spike detection and prevention of over-counting extended spikes due to an implemented refractory period, according to an embodiment;

[0101] Figure 19A is experimental measurement results of performance evaluation of spike detection of Figures 7A and 7B in a low-noise scenario using prerecorded neural data, according to an embodiment; and

[0102] Figure 19B is experimental measurement results of performance evaluation of spike detection of Figures 7A and 7B in a high-noise scenario using prerecorded neural data, according to an embodiment.Detailed Description

[0103] Various apparatuses or processes will be described below to provide an example of each claimed embodiment. No embodiment described below limits any claimed embodiment and any claimed embodiment may cover processes or apparatuses that differ from those described below. The claimed embodiments are not limited to apparatuses or processes having all of the features of any one apparatus or process described below or to features common to multiple or all of the apparatuses described below.

[0104] While the above description provides examples of one or more compositions, apparatus, methods, or systems, it will be appreciated that other compositions, apparatus, methods, or systems may be within the scope of the claims as interpreted by one of skill in the art.

[0105] Provided herein are systems and methods for spike detection of electric signals and in particular spikes of bioelectric signals corresponding to activity. The spikes are detected based on a threshold adaptively determined based on the standard deviation of the bioelectric signal enhanced via SNR enhancement techniques.

[0106] Spikes provide a window into the cooperative and synchronized activities of neuronal populations, revealing how groups of neurons communicate and process information in concert. The ability to accurately detect and analyze these spikes is therefore essential for decoding the intricate patterns of neural communication and understanding the emergent properties of neural networks. This focus on spikes reflects a broader perspective in neuroscience, emphasizing the importance of collective neural behavior over individual neuronal events in understanding brain function and disorders.

[0107] Referring to Figure 1 , shown therein is a flow diagram of a Brain-Computer Interface (BCI) application, according to an embodiment. The application 100 includes a signal acquisition module 102 configure to acquire a bioelectrical signal. The signal acquisition module 102 may be an implantable electrode or the like configured to monitor electrical signals corresponding to a specific activity such as neurological activity. The signal can be from any neural cell such as a retinal ganglion cell (RGC). The raw bioelectrical signal is provided to a spike detection module 104 for detecting spikes of the bioelectrical signal. Particularly, a detection threshold of the spikedetection module 104 is finely tuned to effectively capture spiking activities while excluding incidental peaks attributable to background noise. The spikes are analyzed by the signal processing module 106. The detection and analysis of the neural spikes enables the development and refinement of BCIs and other advanced medical technologies. In a general BCI application, accurate spike detection enables translating neural activity into commands such as via a control signal module 108 that controls external devices. The refinement of the BCI beneficially offers life-changing possibilities for individuals with mobility or communication impairments. BCI technology relies heavily on the precise detection of spike patterns, as these represent the collective neural activity that can be harnessed to operate prosthetics, computer interfaces, and other assistive devices. Beyond BCIs, the insights gained from spike analysis are instrumental in the development of neurotherapeutic strategies and rehabilitation techniques.

[0108] A key factor in enhancing the effectiveness of BCIs and advanced medical technologies is the ability to utilize a large number of recording electrodes. The multiplicity of more electrodes allows for a more comprehensive capture of neural activity, providing a more detailed representation of brain function. In BCIs, for instance, having a greater number of electrodes enhances the system’s ability to interpret neural signals accurately, leading to more precise control of external devices and a better user experience. Similarly, in medical diagnostics and treatment, multiple electrodes enable more complete monitoring and analysis of neural patterns. Detecting spikes via an integrated circuit at the electrode acquiring the signal beneficially avoids noise produced from transmitting the signal to a front end as well as the amplification of the same. It also beneficially avoids access and time synchronization issues associated with providing the signal to a front end.

[0109] Referring to Figure 2, shown therein is a flow diagram of a method 200 for detecting spikes, according to an embodiment.

[0110] At 204, the method 200 includes receiving a bioelectrical signal 202 to obtain a raw signal 206. At 204, the bioelectrical signal 202 may be amplified and passed through a band pass filter. It should be noted that the extracellular recording includes both spikes and Local Field Potentials (LFPs) which represent the summed electrical activity of a larger neuron population within a certain volume of brain tissue. LFPs typically exhibit lower frequencies compared to spikes, generally in the range of 0.5 Hz to 300 Hz. As a result, the high pass pole removes the LFP content of the recording signal even though the LFP signal may have larger amplitude in comparison with the spiking activities.

[0111] At 208, the method 200 includes enhancing the raw signal 206 to obtain the enhanced signal 210. The signal is enhanced by SNR enhancement techniques. The SNR enhancement techniques are applied to attenuate the low-frequency and low amplitude content of the signal (often where most of the noise resides) and accentuate the high-frequency and high amplitude content, which is indicative of spiking activity. By doing so, they effectively enhance the energycontent of spikes relative to the background noise. This results in a clearer distinction between neural spikes and noise, facilitating the task of defining the spike detection threshold.

[0112] SNR enhancement techniques can mitigate higher levels of integrated input-referred noise in the recording front-end, resulting in a lower SNR, which can significantly hinder the effectiveness of spike detection when relying solely on raw data, by boosting the signal component relative to the noise, making the spikes more prominent and distinguishable even in a low SNR setting.

[0113] In high channel count applications, where power constraints are a primary concern, each channel typically operates with a limited power budget. This restriction often leads to a higher level of integrated input-referred noise in the recording front-end, resulting in a lower SNR, which can significantly hinder the effectiveness of spike detection when relying solely on raw data. The amplified noise level can obscure the spikes, leading to increased false negatives (missed spikes) or false positives (noise mistaken for spikes). By contrast, SNR enhancement techniques can mitigate these issues by boosting the signal component relative to the noise, making the spikes more prominent and distinguishable even in a low SNR setting.

[0114] Such techniques may include NEO or ED. For example, NEO enhances the detection of energy transients in signals, such as neural spikes. NEO works by nonlinearly combining a signal with its first and second derivatives to accentuate high-energy features. For a continuous signal x(t), the NEO is defined as i| / NEo(x(t)) = x’(t)2- x(t)x”(t). This equation amplifies points in the signal where there is a significant energy difference, thereby enhancing sharp transients. Therefore, NEO is especially effective in suppressing low-frequency components prevalent in background noise, while enhancing high-frequency components associated with spikes.

[0115] The Energy of Derivative (ED) technique emphasizes the energy content of a signal’s temporal derivatives to enhance spike detectability. ED is based on the premise that neural spikes, due to their rapid onset and offset, exhibit considerable energy in their first derivatives. The energy of the derivative is calculated as ipED(x(t)) = x’(t)2. ED amplifies parts of the signal with rapid changes, such as the edges of a spike, aiding in distinguishing spikes from slower signal components.

[0116] The ED technique focuses on the square of the signal’s derivative eliminating the need for dual-threshold levels by ensuring that all data points are positive. This beneficially simplifies the thresholding process over, for example, other SNR enhancement techniques or using raw data and enables a single threshold level that can be determined in the analog domain.

[0117] This also provides a straightforward and resource-optimized entirely analog implementation beneficially avoiding the use of high resolution analogue to digital converters.

[0118] Referring to Figures 3 and shown therein are examples 300 of performance of spike detection for high and low SNR recordings using raw and SNR enhanced signals which showssensitivity improvements especially in low SNR conditions, according to an embodiment. These figures show how the ED operator is more necessary in case of high background noise.

[0119] Referring to Figure 4, shown therein are simulation results 400 for sensitivity versus false positive rate (FPR) receiver operating characteristic (ROC) generated spike detection with and without ED operator, according to an embodiment. The ED operator may be the ED operator 506 of Figure 5. The results show how the ED operator is more necessary in case of high background noise by raw signal standard deviation (RSTD) or ED STD methods with analog and digital implementation. The comparison in Figure 4 shows how the spike detection of the raw signal can be effective for low noise recording, but it will fail for recording with higher noise. The SNR enhancement performed by the ED operator makes it possible to detect the spike even in a noisy background. In the digital implementation for both methods, a constant number of previous samples are used to define the sigma of the raw data or sigma of the ED signal. However, for analog implementation all the data points from the start of simulation are considered to calculate the sigma at each point. As a result, with analog implementation, the spiking activities have minimal effect on the calculated standard deviation due to the sparse spiking activity. However, with digital implementation, the spiking activities will affect the calculation of the background noise and that is the reason that analog implementation has better performance than their digital counterparts. This issue is a well-studied challenge, and the proposed solution is to somehow remove the spiking activity from the background to avoid its effect on the background noise. However, this causes extra complexity in the process and needs a large buffer to keep the previous data points or interacting with memory blocks on the chip.

[0120] The other point that is visible from the simulation result is that for data from low-noise recording, the analog implementation of RSTD method has better performance of the digital implementation of EDSTD method. However, for high-noise recording, the performance of the RSTD method degrades and thanks to the SNR improvement feature of the ED operator, even the digital implementation of the EDSTD method has better performance. Based on the advantages and limitations outlined in the preceding sections, our goal is to achieve adaptive spike detection that not only operates robustly within the analog domain but also incorporates SNR enhancement techniques to ensure the precision and robustness of spike detection. “RSTD” may be understood as referring to a standard deviation of raw data to calculate the threshold. “EDSTD” may be understood as referring to using standard deviation of the output of an ED operator to calculate the threshold.

[0121] Referring back to Figure 2, the method 200 further includes, at 212, calculating the standard deviation of the enhanced signal 210 as Sigmaenhanced 214. Because spiking events are sparse relative to the background noise within the enhanced signal 210, the Sigmaenhanced 214 is selected as a representative of the background noise variation. Consequently, Sigmaenhanced 214predominantly reflects the fluctuations of the background activity, thereby providing a robust baseline against which spikes, characterized by their relatively higher amplitude, can be detected.

[0122] The process of calculating the standard deviation of the background noise can be executed in both analog and digital domains. However, these two approaches offer distinct advantages and challenges, especially in the context of high channel count recording applications. In digital implementation, there is digitizing of the data using Nyquist rate sampling rate (15kHz) and storage of a large enough amount of data to perform the digital computation. This method is acceptable if this large amount of data is needed to completely reconstruct the data in certain research applications. However, in real-life applications, especially when having a large number of channels, it is not practical to handle all this data. Therefore, the goal is to perform as much computing as possible in the analog domain and ideally extract the spiking activity which can trigger real-time application or trigger specific algorithms that have been implemented using spiking neural networks which can best imitate the biological activities. Therefore, while both analog and digital domains have their merits in adaptive spike detection, analog implementation emerges as a more suitable option for high-channel count applications considering the advantages listed below:

[0123] No Need for Data Storage: Analog systems inherently process signals in real-time, eliminating the need for large-scale data storage, which is a significant advantage in high-channel count applications.

[0124] Energy Efficiency: Analog processing can be more energy-efficient, a critical factor in implantable high-channel-count neural recording systems.

[0125] Real-Time Processing: The ability to process signals continuously and in real-time allows for lower latency in spike detection, which is crucial for applications requiring immediate response, such as in some BCI implementations.

[0126] Compactness: Analog circuits can be more compact than digital ones for simple calculations, beneficial in applications where channel area is a constraint.

[0127] Parallel Processing Capabilities: Analog circuits can handle parallel processing, which is advantageous when dealing with multiple channels of neural data simultaneously.

[0128] Considering that the spikes are sparse activities the target is to find the standard deviation of the background noise. It has been shown that measured background noise from actual recording have almost a Gaussian distribution. Consequently, when a data stream characterized by a Gaussian noise distribution is compared against its own standard deviation, the duty cycle (defined as the proportion of time the output signal is ’high’ or at a logic level of 1 , over the total observation time) would stabilize at a distribution constant value of 0.159xVDDComp. This value represents the percentage of time that the signal exceeds its own standard deviation within the Gaussian distribution.

[0129] Referring to Figures 5A and 5B, shown therein is a distribution 500 of an ED enhanced signal and a feedforward structure 502 for computing the corresponding distribution constant, according to an embodiment. A fundamental assumption in these techniques is that the background noise follows a Gaussian distribution with a mean of zero. This assumption is largely supported by empirical observations from actual neural recordings, which typically exhibit a Gaussian noise distribution, albeit with slightly broader tails as indicated in literature. Consequently, the noise in these recordings can be adequately characterized by its root mean square (rms) value, which, in the case of a Gaussian distribution, corresponds to its standard deviation.

[0130] To extend the principle of adaptive thresholding to an enhanced signal the distribution constant corresponding to the enhanced signal is determined. It is evident from distributions across different noise levels 504 that 23.4% of the data lies beyond SigmaED 508. In other words, if the output of the ED operator 506 and its computed SigmaED 508 is fed into a comparator 510 as shown in Figure 5B and the duty cycle 512 of the comparator’s output is measured, it will reach a steady-state value 514 of 0.234x DDcomP, following the dissipation of initial transients. This steady-state value 514 serves as a distribution constant.

[0131] Referring to Figure 6, shown therein is a distribution of ED signal for a Gaussian noise input with various standard deviations a, according to an embodiment. A prerequisite of determining the distribution constant of the enhanced signal as above is that the enhanced signal exhibits a constant distribution when the input has Gaussian distribution. As shown, the transformed distributions retain a similar shape when the x-axis of the distribution plots is normalized against SigmaED. This indicates that the condition is satisfied.

[0132] Referring back to Figure 2, the method 200, at 216, includes calculating the threshold 218. Calculating the threshold 218 comprises setting the threshold to a predetermined multiple of the Sigmaenhanced 214 by a predetermined scaling factor. This effectively filters out the majority of the background noise. For instance, implementing a threshold 218 of 5 times Sigmaenhanced 214 significantly minimizes the likelihood of false detections due to noise. In statistical terms, with a 5 times Sigmaenhanced 214 threshold 218, the probability of Gaussian noise erroneously triggering the spike detector is approximately 3x10-7. In biomedical applications a multiplier of between 2- 4 times of the Sigmaenhanced 214 balances the sensitivity to detect all spiking activity and robustness against noises.

[0133] The method 200, at 220, includes detecting spikes 222. Detecting spikes 222 comprises comparing the enhanced signal 210 to the threshold 218. Instances where the enhanced signal 210 exceeds the threshold 218 are captured as spikes 222.

[0134] Referring to Figures 7A and 7B shown therein is a block diagram and block schematic respectively of a system 700 for detecting spikes 222, according to embodiment. The system 700 is configured to implement the method 200 of Figure 2.

[0135] The system 700 may include an electrode 704. The electrode 704 obtains a bioelectrical signal 202. The electrode 704 may be in contact with or implanted in a subject. The electrode 704 may be passive (i.e. a conductor) or active (i.e. including an IC). In embodiments with an active electrode, the electrode may include a high pass filter or an amplifier. The amplifier may have a bandwidth between 300Hz to 8kHz to remove the LFP signal while passing the spiking activity.

[0136] The system 700 includes a signal enhancement module 708. The signal enhancement module receives a raw signal 206 output from the electrode 704 and applies SNR enhancement techniques to obtain an enhanced signal 210. The signal enhancement techniques may be NEO or ED as described at 208 of Figure 2.

[0137] The system 700 further includes a standard deviation calculator 712. The standard deviation calculator 712 calculates a Sigmaenhanced 214 based on the enhanced signal 210 and a distribution constant 724. The distribution constant 724 may be calculated via a feedforward structure such as the feed forward structure 502 of Figure 5B.

[0138] The system 700, further includes a threshold calculator 716. The threshold calculator 716 receives the Sigmaenhanced 214 from the standard deviation calculator 712 and calculates a threshold 218 based on the Sigmaenhanced 214. The threshold 218 is a multiple of the Sigmaenhanced 214. The multiplier (K) by which the threshold 218 is a multiple of the Sigmaenhanced 214 may be predetermined based on the application of the system. For example, in biomedical applications the multiplier may be 2-4.

[0139] The system 700, further includes a spike detection module 720. The spike detection 720 compares the threshold 218 to the enhanced signal 210 to detect spikes 222.

[0140] The system 700 further includes a comparator 726. The comparator block 726 provides a high gain and minimizes the difference between the two input terminals of the standard deviation calculator 712.

[0141] The system 700 further includes a weighted rolling average 728. The weighted rolling average 728 calculates an average of the output from the comparator 726.

[0142] The system 700 further includes a subtraction module 730. The subtraction module 730 calculates a subtraction of the output of the weighted rolling average 728 and the distribution constant 724.

[0143] The system 700 further includes a gain 732. The gain 732 multiplies the signal output by the subtraction module 730 by a factor K.

[0144] Referring to Figure 8, shown therein is a block diagram of a Top-level implementation of an ED-based adaptive threshold generation IC 800, according to an embodiment. The IC 800 is an embodiment of the system 700 of Figure 7.

[0145] The IC 800 includes a two-stage amplifier 802 including a low-noise and low-power bioamplifier which is implemented by a two-stage capacitively coupled structure which also implements a 300-8kHz bandpass filter.

[0146] Referring also to Figure 9, shown therein is a block diagram 900a of the two stage amplifier 802, schematic with transistor sizing 900b, and simulation results showing the closed- loop gain and phase margin 900c of the two stage amplifier, according to an embodiment. The amplifier 802 may be a two stage capacitively-coupled operational transconductance amplifier (OTA). The amplifier 802 includes a first stage 902 and a second stage 906. The first stage 902 of the amplifier focuses on signal amplification and filtering, employing a two-stage capacitively- coupled OTA. Although it is possible to design the recording front-end 904 to record both LFP and spikes, considering their different signal characteristics and frequency bands, it would be a suboptimal design to record both signals with the same front-end and it is more efficient to have dedicated circuit with different bandwidth for each application (i.e., 10-300Hz for LFP and 300- 8kHz for spikes). As the target application of this structure is to detect the spiking activity, this two-stage amplifier is tailored for spike activity recording and Low-Frequency Potential (LFP) signal rejection. In this embodiment, the constant gain is secured through a capacitive gain structure, the high pass pole is determined by the feedback RC network including resistor 926 and capacitor 928, and the low pass pole is derived from Gm / AMCL with AM = Cin / Cf, which are designed to provide a gain of 52dB and bandwidth of 300Hz-8kHz. Here, Gmmay be the Gmof the first stage 902, CL may be the load capacitance, Cm may be the capacitor 930, and Cf may be the capacitor 928. As the first stage 902 is more prone to noise, a low-noise (LN) OTA is used in this stage. Following this initial amplification, the second stage 906 employs a low-power OTA, as the signal is now more robust against noise factors.

[0147] The low power OTA of the second stage 906 is designed with a similar architecture and transistor sizing of the first stage 902. However, the size of the M0 transistor 908 has a width / length (W / L) of 1 m / 5pm, making the direct current (DC) bias current of the low-power OTA 906 (0.25pA) half of the low-noise OTA 902 (0.5pA), also referred to as the current mirror OTA. In this embodiment, the current mirror OTA 902 is selected for its robustness and potential to achieve a low Noise Efficiency Factor (NEF). Adhering to the gm / lD methodology, the integrated input referred noise of the OTA 902 is shown to be inversely proportional to the gmof transistors Mi 910 and M2912, and directly proportional to the gmof transistors M3914, M4918, M3b916, M4b920, Ms 922, and Me 924. The gm / lD ratio is maximized for the input transistors, which are set to operate in the subthreshold region, thus demanding a large W / L ratio. However, for M3-6(914, 918, 922, 924), M3b916, and M4b920, the gm / lD ratio is minimized by having a W / L ratio less thanunity. Given that flicker noise is inversely proportional to the transistor dimensions (W x|_), enlarging both W and L is a strategic choice to reduce flicker noise.

[0148] After the initial design, a noise summary tool is employed to pinpoint and attenuate the primary noise contributors. If thermal noise from the feedback resistor predominates, such as from feedback resistors 926 and 932, lowering the low cut-off frequency below the target can be advantageous. Should thermal noise from the input transistors be significant, such as from transistors 910 and 912, increasing the tail bias current or enlarging the W / L ratio of the transistors is beneficial. It will be appreciated that the gm / lD plot plateaus after some point and therefore, increasing the W / L of the transistor will not increase the gmproportionally. In cases where flicker noise is predominant, the transistor size may be increased, or a chopper structure may be employed.

[0149] Referring only to Figure 8, the system 800 further includes an ED calculator 804. The ED calculator 804 is an embodiment of the Signal Enhancement Module 708 of Figures 7A and 7B. The ED calculator 804 is responsible for computing an Energy of Derivative (ED) 806, which is derived from squaring the derivative of the output of the amplifier 802. As previously discussed, the purpose of this stage is to amplify the signal-to-noise ratio (SNR) of spike recordings. This is achieved by emphasizing the high-frequency components, representative of spiking activities, while suppressing the low-frequency components, associated with noise. In this embodiment, differentiation is executed through two discrete-time switched-capacitor based differentiators 810a and 810b, and the squaring function is realized using a Gilbert-cell multiplier 808. The differentiator 810a outputs the differentiation with a positive polarity, and the differentiator 810b outputs the differentiation with a negative polarity.

[0150] Referring also to Figures 10A and 10B, shown therein is a schematic and phase plot, respectively, of a circuit-level implementation of a switched-capacitor-based differentiator 1000, according to an embodiment. The differentiator 1000 is an embodiment of the discrete-time, switched-capacitor-based differentiator 810 of Figure 8. The derivative of the differentiator 1000 is implemented using passive components. The differentiator 1000 is, therefore, more power efficient in comparison with the OTA-based implementation 900a. The differentiator 1000 is also less sensitive to high frequency noise as this circuit inherently avoids amplification of high- frequency signal. The differentiator 1000 is also implemented in discrete fashion and therefore, is inherently stable. Also, the clock frequency is selected based on the signal Nyquist rate and therefore, it is not prone to high frequency noise saturation.

[0151] In this embodiment, during <pi phase 1004 the signal 1002 is sampled on capacitor Csi 1006 and during the <P2 phase 1008 the signal is sampled on capacitor Cs2 1010. During the <po phase 1012, the capacitors Csi 1006 and CS2 1010 are connected in series with opposite polarities which means that their voltages are subtracted from each other and is stored on the capacitor Co1014. The Mo transistor 1016 is turned on with a delay to ensure that the capacitor Co 1014 doesnot affect the subtraction and is turned off before the MSUB transistor 1018 to avoid the effect of its charge injection. Also, the MDtransistor 1020 is used in a dummy structure to absorb the charge injection of the Motransistor 1016. It should be noted that with this embodiment the differentiation is performed in a non-overlapping fashion.

[0152] The differentiator 1000 further includes an OR operator 1026. The OR operator 1026 generates a control signal. The control signal is input to a sampling transistor (M5) 1028. The transistor 1028 is closed in the <pi phase 1004 and the <p2phase 1008.

[0153] The differentiator 1000 further includes an inverter operator 1030. The inverter 1030 inverts a control signal 1032 provided as input to the transistor 1020, such that the control signal 1032, provided as input to the transistor 1016, and related to the signal 1012 shown in Fig. 10B, is the inverse of the control signal 1032 provided as input to the transistor 1020. The rising and falling edges of the signal 1032 occur while the <p0phase signal 1012 is high as shown in Figure 10B.

[0154] Using this embodiment also beneficially simplifies the biasing of the differentiator 1000 which is the input of the Gilbert cell 808 of Figure 8. The sampled voltage on the Cocapacitor 1014 is equal to: Vout = CM + (Vin[<p1] - Vin[<p2]) (0.5p / (0.5p + 0.2p)) where the CM voltage 1022 is the DC bias of the output 1024 and can be set to whatever is suitable for the next stage without affecting the performance of this stage.

[0155] Referring to Figure 11 , shown therein is a simulation result 1100 showing the functionality of the pair of opposite phase differentiators 1000 of Figure 10A. Minimizing the size of the Co 1014 to avoid attenuating the output voltage makes the storage capacitor (i.e., the Cocapacitor 1014) prone to fast discharging and clock feedthrough. Therefore, a pair of duplicate switched-capacitor-based differentiators are used where the <pi 1004 and <p21008 phases are opposite but the <p01012 is the same. As shown, the output of the two Differentiators 1000 have the same non-idealities which are canceled out on the difference of the two waveforms to obtain a clean discrete-form differentiation of the input voltage.

[0156] Referring back to Figure 8, the ED calculator 804 further includes a Gilbert cell squarer 808. The Gilber cell squarer 808 is used to implement the multiplication function of the ED calculator 804.

[0157] Referring also to Figures 12A and 12B, shown therein are a schematic along with the transistor sizings and simulation results showing the functionality, respectively, of a Gilbert cell 1200, according to an embodiment. The Gilbert cell 1200 may be the Gilber cell squarer 808 of Figure 8. The Gilbert cell 1200 may be a metal-oxide-semiconductor field-effect transistor (MOSFET)-based Gilbert cell structure. In this embodiment, the multiplication function of the ED calculator 804 is implemented by connecting the bottom differential pair (Mi 1202 and M21204) and the top differential pair (M3 1206 and Me 1208) to the same voltages Vi 1210 or V21212.

[0158] In general application of a Gilbert cell the bottom pair 1202, 1204 is connected to the radio frequency (RF) signal and the top pairs 1206 and 1207, and 1208 and 1209, are connected to the local oscillator (LO) signal the output is proportional to RFx[_O signal. In the present embodiment of the Gilbert cell 1200, the top pairs 1206 and 1207, and 1208 and 1209, are connected to V1 1210 and V2 1212 such that the output of the Gilbert cell 1200 with the structure is (Vi - V2)2. Considering that the Vi 1210 and V21212 voltages are proportional of dVinand -dVin, respectively, this subtraction and squaring also provides a gain of 4.

[0159] The simulation result 1201 of the Gilbert cell 1200 shows that the Gilbert cell 1200 can generate the ( i - V2)2waveform 1216 with good matching with the ideal square of the voltage difference 1218. It will be appreciated that the reason for asymmetry in this result is that the difference waveform 1216 is calculated based on non-overlapping sampling and as a result, the Voiff waveform 1216 is not symmetric.

[0160] Referring to Figure 13, shown therein is a schematic diagram of a detailed implementation of subtraction and threshold voltage calculation in a feedback loop and the simulation result showing the variations at the low and high impedance outputs of the Gilbert cell 1200 of Figure 12A, according to an embodiment. A distinction of this embodiment is the unique output characteristics of the Gilbert cell 1200, which yields an expression of (V1 - V2)2added to a constant voltage term (C) 1302. The constant voltage term 1302 is the DC operation point of the system.

[0161] From a statistical standpoint, the addition of a constant to every element within a dataset does not alter its standard deviation; hence, this constant voltage 1302 component does not affect the standard deviation of the ED signal. Nevertheless, the feedback loop in discussion is designed to set the negative input of a comparator 1304 to SigmaED +C. Therefore, if the resistor Ri 1308 were connected to ground as in a conventional structure, the resulting threshold voltage would be Kx(sigmaED + C), deviating from the intended value of C + K x sigmaED. To rectify this and achieve the desired threshold, Ri 1308 is not connected to ground but rather to the constant C voltage 1302.

[0162] It will be appreciated that the C voltage 1302 does not represent the mean of the ED signal, but rather indicates the DC shift, the minimum value or the operating point of this signal. As shown, the voltage across the diode-connected transistor, such as transistor 1312, within the Gilbert cell architecture reflects the extent of this upward shift, considering it is a node of low impedance. Simulation results 1301 corroborate that this voltage corresponds to the minimum of the ED voltage, not its average.

[0163] In pursuit of this design objective, connecting Ri 1308 directly to this node is avoided to prevent any interference with the operation of the Gilbert cell 1200 by the injection of Ri 1308 current. For optimal functioning of this buffer, the current passing through Ri 1308, which isdirected into the buffer 1310 , is substantially lower than the buffer’s bias current, adhering to the criterion lBias > 10 x |R1.

[0164] In order to mitigate the increase in the buffer’s power consumption, a low-power OTA is utilized for the buffer 1310 (also referred to as the OTA 1310) such that the current of Ri 1308 is drawn from the OTA output via a current mirror structure. This approach is crucial as it effectively isolates the current flow of the OTA 1310 from the current flow through Ri 1308. Consequently, the OTA 1310 is able to maintain its functionality with a minimal bias current. This configuration beneficially ensures efficient voltage buffering and also preserves the OTA’s performance without necessitating a significant increase in power consumption.

[0165] Referring only to Figure 8, the system 800 further includes a standard deviation calculator 812. The standard deviation calculator 812 is an embodiment of the standard deviation calculator 712 of Figure 7. The standard deviation calculator 812 receives the output 806 from the Gilbert cell 808, representing the computed ED. In this configuration, the feedback loop is designed to generate the voltage of SigmaED 814.

[0166] The output 806 from the Gilbert cell 808, representing the computed ED, is fed into the positive terminal of a comparator 816. This stage is followed by a latch which means there is no reset phase. Subsequently, a duty cycle of the output of the comparator 816 is determined by a weighted rolling average 818.

[0167] Referring also to Figure 14A, shown therein is a circuit-level implementation of a switched-capacitor based rolling average 1400, according to an embodiment. In some embodiments, the duty cycle of the standard deviation calculator 812 of Figure 8 includes a switched-capacitor-based weighted rolling average 1400. The switched-capacitor-based weighted rolling average 1400 quantifies the proportion of time during which the output of the comparator 816 remains high, starting from the circuit’s initial operation. This implies that rather than evaluating a fixed time window, the circuit retains all prior data, assigning greater significance to more recent inputs. This computation is fundamentally different from that of an integrator.

[0168] In this embodiment, a switched-capacitor-based weighted rolling average structure is implemented to overcome this limitation. Considering just the first two capacitors, Ci 1402 and C21404. During the 1 phase 1406, the input 1408 is sampled onto a 30fF capacitor 1402. In the2phase 1410, Ci 1402 and C21404 are connected to equalize their voltages. The discrete-time operation of this circuit is expressed as Y(n) = (Ci x x(n; + C2x Yn-i) / (C1 + C2). This expression represents the weighted rolling average of previous samples, enabling the circuit output to fluctuate around the target average. An additional RC output stage including resistor 1412 and capacitor 1414 calculates the average of the signal 1416 that is already stabilized to its mean value. Therefore, the RC stage in this embodiment aims to reduce or smooth the ripple in the output of the Ci 1402 and C21404 capacitors.

[0169] Referring to Figure 14B shown therein are simulation results for varying input DC levels and input pulse signals with varying duty cycles input to the switched-capacitor based rolling average 1400, according to an embodiment. Shown are simulation results under two conditions: (top 1401) when the input is a constant DC level (0.2V- 0.4V- 0.6V- 0.8V) and (bottom 1403) when the input is a pulse waveform with varying duty cycles (20%-40%-60%-80%). The simulations show that the circuit accurately computes the average, with the switched-capacitor arrangement results in minor deviations around the target value and the subsequent stage applying a low-pass filter to produce a near-DC output voltage.

[0170] Referring to Figure 14C shown therein is a flow schematic indicating the operation of a resistance capacitor structure (RC) 1418 without a switched capacitor, according to an embodiment. For a basic implementation, an RC structure would suffice to achieve this operation. However, the output 1420 from this RC structure 1418, which feeds into a subtraction circuit, has a relatively stable value. Given that the comparator’s output 1422 alternates between 0 and VDD from rail to rial, to minimize fluctuations at the RC output 1420, a small pole, necessitating a large resistor 1424 and capacitor 1426 (time constant > 100ms), is needed. In the conventional structure, this has been accomplished using a capacitor coupled to a gmcell to generate a high resistive value of 1 / gm. Yet, to attain the necessary time constant, such a configuration would demand a large capacitor 1424 (>10-100pF), which is impractical for on-chip implementation, given the area limitations in high-channel-count recording systems.

[0171] Referring to Figure 8 alone, the output of the weighted rolling average 818 is subtracted from a preset distribution constant voltage level 820 of 0.234xVDD by a subtraction module 824. The output of the subtraction module 824 is multiplied by a factor K1 by a gain module 826. The output of the gain module 826 is directed to the negative input 822 of the comparator 816. The architecture of the feedback loop inherently adjusts the voltage at the negative input 822 of the comparator 816 to sigmaED voltage 814. This adjustment of the closed-loop system is owing to the high gain provided by the comparator 816. The adjustment is such that the duty cycle of the comparator’s output stabilizes at the distribution constant voltage of 0.234xVDD.

[0172] The IC 800 further includes a threshold calculator 828 and a spike detection module 830. The calculator 828 and the module 830 operate substantially similarly to the threshold calculator 716 and the spike detection module 720 of Figure 7B, respectively. The calculator 828 represents the gain of the threshold signal with respect to the calculated standard deviation. The module 830 is the final comparator that compares the output of ED operator with the automatically generated threshold signal and outputs the detected spikes.

[0173]

[0174] Measurement Validation:

[0175] Referring to Figures 15A and 15B, shown therein are experimental measurement results of input referred noise 1500 and gain and bandwidth 1502, respectively of the two-stage amplifier 802 of Figure 8, according to an embodiment. The results show the measured integrated input referred noise and voltage gain of the input two-stage amplifier 802, confirming an Infinitely iterated rippled noise (URN) of 8.82pV in the band of interest, mid-band gain of 51.45 dB and a bandwidth of 237.5Hz- 8.27kHz.

[0176] Referring to Figure 16, shown therein is a dynamic response 1600 of the spike detection system 720 of Figure 7A to Gaussian noise input, showing the automatically calculated standard deviation and the adaptive threshold over a 20-second interval, according to an embodiment. The results show a response of the spike detection system 720 to a Gaussian noise with temporally varying standard deviation fed to the recording front-end. The automatically calculated sigmaED and threshold voltages for a Gaussian noise are shown with the temporally varying standard deviation input signal. The results indicate that the threshold is set at a voltage that is effectively higher than the noise activity except in very few points (6 points in 20s of data) validating the performance of the sigma detection loop.

[0177] Referring to Figure 17, shown therein are experimental measurement results 1700 of the spike detection system 720 of Figure 7A to input noise with varying standard deviation 1702, ED output with the automatically generated threshold 1704, in-loop comparator (i.e., the comparator 726) output indicating 1s density 1706, and identification of falsely detected spikes 1708, according to an embodiment. The results show responses to Gaussian noise characterized by a temporally varying standard deviation fed to the recording front-end which indicate the feedback loop’s capability to ascertain sigmaED, which fluctuates in accordance with the amplitude of the background noise. Furthermore, the in-loop comparator’s output was analyzed, showing the proportionality between the incidence of ’ones’ in the comparator’s output and the standard deviation of the ED signal. The final graph depicts instances of falsely detected spikes, which dominantly occur under conditions of exceedingly high noise levels. These results further clarify the functionality of the sigma detection loop.

[0178] Referring to Figure 18, shown therein are experimental measurement results 1800 showing the functionality of the spike detection system 720 of Figure 7A in the presence of LFP signal ED-based spike detection and prevention of over-counting extended spikes due to an implemented refractory period, according to an embodiment. Specifically shown are the measurement outcomes when pre-recorded neural signals, encompassing both local field potentials (LFP) and action potentials (spikes), are input into the system. The top plot 1802 displays the combined LFP and spike signal. The subsequent plot 1804 demonstrates the efficacy of the bandpass amplifier in isolating the spikes by attenuating the LFP components. The third plot 1806 illustrates the spikes being accurately identified using the ED method with the automatically generated threshold indicating the detection cutoff. The final plot 1808 showcasesthat spikes with extended duration are not falsely counted multiple times, owing to the refractory period mechanism integrated within the spike detection algorithm. This refractory period ensures that once a spike is detected, a certain time elapses before another spike can be identified, thereby preventing overcounting and improving the fidelity of spike detection.

[0179] Referring to Figures 19A and 19B, shown therein are experimental measurement results of performance evaluation of spike detection in a low-noise scenario 1900 and a high noise scenario 1901 , respectively, using prerecorded neural data, according to an embodiment. The results showcase the efficacy of the ED operator, such as ED operator 506 of Figure 5, in SNR enhancement and spike identification. The experimental measurement results include a comparative evaluation of the presented spike detection circuit in environments with different levels of background noise, using a dataset of prerecorded neural signals.

[0180] Referring specifically to Figure 19A, the results represent conditions with low background noise. The low-noise scenario 1900 shows the original input signal 1902 as the ground truth provided by the QQ Standard Spike Dataset from the University of Leicester. The following plot reveals the recording output 1904, where the low noise level is evident. Subsequently, 1906 shows the ED output along with the automatically generated threshold, demonstrating the system’s capability to detect spikes against the background noise. In the final plot 1908, the detected spikes are illustrated, indicating a successful identification process.

[0181] Referring specifically to Figure 19B, the results depict a scenario 1901 when the background noise is significantly higher. The higher background noise poses a challenge for spike detection. The initial plot 1910 once again shows the input signal which is the ground through to validate the effectiveness of the ED and adaptive threshold loop in identifying the spiking activity. The recording output 1912, with increased noise level, underscores the difficulty in setting an appropriate threshold for the raw data. Nevertheless, the application of the ED operator 506 transforms the signal in such a manner that threshold setting becomes feasible, as shown in the third plot 1914. The ED operator 506 effectively enhances the spikes, making them distinguishable from the noisy background. The final plot 1916 in the sequence illustrates the detected spikes, confirming that, despite the presence of substantial noise, the system is able to accurately identify spikes, thereby validating the effectiveness of the ED operator 506 in high- noise conditions. The post processing of 5 data epochs (each comprising 100-ms windows of prerecorded neural signals) shows the spike detection circuit achieves a 97% sensitivity and 0.4% false alarm rate in low noise conditions and 92% sensitivity and 3.8% false alarm rate in high noise condition.

[0182] While the above description provides examples of one or more apparatus, methods, or systems, it will be appreciated that other apparatus, methods, or systems may be within the scope of the claims as interpreted by one of skill in the art.

Claims

Claims:1 . A method for detecting one or more spikes in an electrophysiological signal corresponding to neurological activity, the method comprising: receiving and enhancing the electrophysiological signal using a signal to noise ratio (SNR) enhancement technique to obtain an enhanced signal; calculating a standard deviation of the enhanced signal based on a distribution constant corresponding to the SNR-enhanced signal; calculating a threshold based on the standard deviation; and detecting the one or more spikes based on the threshold.

2. The method of claim 1 , wherein the SNR enhancement technique is an energy of derivative (ED) technique.

3. The method of claim 1 , wherein receiving the electrophysiological signal comprises one or more of amplifying and filtering the electrophysiological signal.

4. The method of claim 1 , wherein the electrophysiological signal is amplified and passed through a band pass filter.

5. The method of claim 1 , wherein the electrophysiological signal includes local field potentials (LFPs).

6. The method of claim 5, wherein the LFPs are removed from the electrophysiological signal by a high pass pole.

7. The method of claim 1 , wherein the SNR enhancement technique is a nonlinear energy operator (NEO) technique.

8. The method of claim 1 , wherein the one or more spikes are sparse relative to background noise of the electrophysiological signal.

9. The method of claim 8, wherein the standard deviation is representative of the variation of the background noise.

10. The method of claim 9, wherein the standard deviation is calculated in the analog domain.11 . The method of claim 9, wherein the standard deviation is calculated in the digital domain.

12. The method of claim 9, wherein the background noise follows a Gaussian distribution with a mean of zero.

13. The method of claim 1 , wherein the distribution constant is 0.234.

14. The method of claim 1 , wherein calculating the threshold comprises setting the threshold to a predetermined multiple of the standard deviation by a predetermined scaling factor.

15. The method of claim 14, wherein the predetermined scaling factor is 2 or 3.

16. The method of claim 1 , wherein detecting the one or more spikes comprises comparing the enhanced signal to the threshold.

17. The method of claim 16, wherein the one or more spikes are detected where the enhanced signal exceeds the threshold.

18. The method of claim 1 , wherein the electrophysiological signal is received by an electrode.

19. The method of claim 18, wherein the electrode is a passive electrode.

20. The method of claim 18, wherein the electrode is an active electrode.21 . The method of claim 20, wherein the electrode includes a high pass filter.

22. The method of claim 20, wherein the electrode includes an amplifier.

23. The method of claim 22, wherein the amplifier has a bandwidth between 300 Hz to 8 kHz to remove the LFPs while passing the one or more spikes.

24. The method of claim 22, wherein the amplifier is a two-stage capacitively-coupled operational transconductance amplifier (OTA).

25. The method of claim 24, wherein the amplifier includes a first stage including a low-noise OTA.

26. The method of claim 25, wherein the amplifier includes a second stage including a low- power OTA.

27. The method of claim 26, wherein the low-power OTA is a current mirror OTA.

28. The method of claim 2, wherein the ED is derived from squaring a derivative of the electrophysiological signal.

29. The method of claim 28, wherein the derivative is calculated by a differentiator.

30. The method of claim 29, wherein the differentiator is implemented using passive components.

31. The method of claim 29, wherein the differentiator includes a clock frequency selected based on a Nyquist rate of the electrophysiological signal.

32. A system for detecting one or more spikes in an electrophysiological signal corresponding to neurological activity, the system comprising: an electrode configured to receive the electrophysiological signal to obtain a raw signal; a signal enhancement module configured to enhance the raw signal using a SNR enhancement technique to obtain an enhanced signal; a standard deviation calculator configured to calculate a standard deviation of the enhanced signal based on a distribution constant corresponding to the SNR enhanced signal; a threshold calculator configured to calculate a threshold based on the standard deviation; and a spike detection module configured to detect the one or more spikes based on the threshold.

33. The system of claim 32, wherein the SNR enhancement technique is an energy of derivative (ED) technique.

34. The system of claim 32, wherein the receiving the electrophysiological signal comprises one or more of amplifying and filtering the electrophysiological signal.

35. The system of claim 32, wherein the electrophysiological signal is amplified and passed through a band pass filter.

36. The system of claim 32, wherein the electrophysiological signal includes local field potentials (LFPs).

37. The system of claim 36, wherein the LFPs are removed from the electrophysiological signal by a high pass pole.

38. The system of claim 32, wherein the SNR enhancement technique is a nonlinear energy operator (NEO) technique.

39. The system of claim 32, wherein the one or more spikes are sparse relative to background noise of the electrophysiological signal.

40. The system of claim 39, wherein the standard deviation is representative of the variation of the background noise.41 . The system of claim 40, wherein the standard deviation is calculated in the analog domain.

42. The system of claim 40, wherein the standard deviation is calculated in the digital domain.

43. The system of claim 40, wherein the background noise follows a Gaussian distribution with a mean of zero.

44. The system of claim 32, wherein the distribution constant is 0.234.

45. The system of claim 32, wherein the threshold is a predetermined multiple of the standard deviation by a predetermined scaling factor.

46. The system of claim 45, wherein the predetermined scaling factor is 2 or 3.

47. The system of claim 32, wherein the spike detection module detects the one or more spikes by comparing the enhanced signal to the threshold.

48. The system of claim 47, wherein the one or more spikes are detected where the enhanced signal exceeds the threshold.

49. The system of claim 32, wherein the electrode is a passive electrode.

50. The system of claim 32, wherein the electrode is an active electrode.51 . The system of claim 50, wherein the electrode includes a high pass filter.

52. The system of claim 50, wherein the electrode includes an amplifier.

53. The system of claim 52, wherein the amplifier has a bandwidth between 300 Hz to 8 kHz to remove the LFPs while passing the one or more spikes.

54. The method of claim 52, wherein the amplifier is a two-stage capacitively-coupled operational transconductance amplifier (OTA).

55. The method of claim 54, wherein the amplifier includes a first stage including a low-noise OTA.

56. The method of claim 55, wherein the amplifier includes a second stage including a low- power OTA.

57. The method of claim 56, wherein the low-power OTA is a current mirror OTA.

58. The system of claim 33, wherein the ED is derived from squaring a derivative of the electrophysiological signal.

59. The system of claim 58, wherein the derivative is calculated by a differentiator.

60. The system of claim 59, wherein the differentiator is implemented using passive components.

61. The system of claim 59, wherein the differentiator includes a clock frequency selected based on a Nyquist rate of the electrophysiological signal.

Images