SYSTEM AND METHOD FOR MEASUREMENT AND ASSESSMENT OF DEPTH OF ANESTHESIA IN AN ANIMAL SUBJECT BASED ON ELECTROENCEPHALOGRAPHY

Abstract

The present invention provides a system for implementing a logistic regression classification mechanism to measure and assess a depth of anesthesia of an animal subject based on electroencephalography (EEG), which includes a signal pre-processor, an epoch generator, a feature extractor, a classifier, and a predictor. Related method of how to pre-process the raw data of EEG signal, epoch generation thereof, feature extraction from each epoch, classification based on extracted features, and prediction of different states of the animal subject based on a prediction decision mechanism is also provided. Classification accuracy of the present invention for 1-second and 10% overlapping epochs is about 94% with an average total system delay of about 12 μs and low on-chip power consumption. The present system is entirely optimized, which leads to a 100% accurate channel prediction after a 7-second run-time on average.

Claims

1. A system for measuring and assessing anesthesia of an animal subject based on electroencephalography of the animal subject, the system comprising: a signal pre-processor comprising at least two filters in different filtering frequency and a down-sampler for removing unwanted signals and noise from incoming signal stream of the system to generate an incoming signal for subsequent epoch generation; an epoch generator for generating an epoch signal containing 1-second and 10% overlapping epochs comprising a two-input multiplexer, an address generator and a memory, the two-input multiplexer receiving the incoming signal from the signal pre-processor and also a selector input from the address generator to feed an input signal stream to the memory after a counter value of a counter at a relatively higher frequency from the address generator reaches 500; the address generator having two counters at different frequencies and generating two counter signals with write address and read address, respectively, to be fed to the memory, and also a control signal to be fed directly for subsequent feature extraction; the memory receiving the input signal stream from the two-input multiplexer and two counter signals with the write address and read address, respectively, from the address generator and then generating the epoch signal containing the 1-second and 10% overlapping epochs for subsequent feature extraction; a feature extractor comprising a derivative calculator and a variance calculator, the derivative calculator receiving the epoch signal containing 1-second and 10% overlapping epochs from the memory of the epoch generator and calculating a mean of accumulated squared-differences among different epochs; the variance calculator receiving absolute value of each of the epochs accumulated and determining a mean of the accumulated epochs, obtaining a deviation of an epoch by subtracting the mean from one of the absolute values of the epoch, calculating a squared deviation followed by determining square root of an average squared deviation for subsequent classification; a classifier comprising two cascaded units for expanding features extracted by the feature extractor by double the number of the features followed by feature mapping to set a classification boundary, and subsequently using an output of the feature mapping to obtain a sigmoid function as a decision boundary in order for subsequent prediction; a predictor comprising a predictor circuit for accumulating the classifier's outputs, constraining a decision value of the classifier's outputs between 0 and 1, and determining level of anesthesia of the animal subject in terms of the constrained decision value based on the animal subject's real-time electroencephalogram.

2. The system of claim 1, wherein a first filter of the signal pre-processor has a filtering frequency to compensate the power-line interference of the incoming signal from the electroencephalography of the animal subject.

3. The system of claim 1, wherein a second filter of the signal pre-processor has a filtering frequency comparable to an average frequency of a wide electroencephalography frequency region in a range of 0-250 Hz of the incoming signal from the animal subject.

4. The system of claim 1, wherein the incoming signal is down sampled at least ten times after being subject to the second filter using the down-sampler.

5. The system of claim 1, wherein the epoch generator creates a sliding window to divide each electroencephalography channel into short-overlapped epochs. It is optimally designed to consume lower resources and the smallest possible size of memory unit (RAM).

6. The system of claim 1, wherein the derivative calculator of the feature extractor comprises a multiplier to square the differences among different epochs before calculating the mean of the accumulated squared-differences, and further comprises an accumulator and divisor for calculating the mean.

7. The system of claim 1, wherein the variance calculator comprises a multiplier, an accumulator and a divisor circuit to consecutively calculate the average squared deviation; the variance calculator further comprises a register to store the average squared deviation for being fed to the classifier subsequently.

8. The system of claim 1, wherein the two cascaded units of the classifier comprises a feature mapping unit for said expanding and feature mapping to enhance classification accuracy by setting the classification boundary with an increase in classifier variance, and an exponential and reciprocal computation circuit for receiving the output of the feature mapping unit to determine the sigmoid function as the decision boundary for being fed to the subsequent predictor.

9. The system of claim 1, wherein the predictor is configured to accumulate the level of anesthesia in terms of percentage and constrain an output value as a decision score between 0 and 1 with respect to the percentage of the anesthesia of the animal subject, wherein a 100% anesthesia corresponds to the decision score of 1, while a 0% anesthesia corresponds to the decision score of 0.

10. A method for determining depth of anesthesia of an animal subject from a transient behavior of an electroencephalography using machine learning techniques thereof, the method comprising: recording electroencephalographic (EEG) signal of the animal subject and filtering thereof within a relatively lower frequency band to remove unwanted signals and noise; down-sampling the filtered EEG signal to at least ten times for reducing data size and accelerating subsequent processing without losing essential features of the EEG for subsequent classification; generating epochs from the down-sampled EEG signals comprising segmenting each EEG channel with a shortened signal time and overlapping one EEG signal with a preceding EEG signal thereof; extracting two features from each epoch selected from derivative and variance thereof; mapping the derivative and variance features to determine a classification boundary followed by using an output of the mapping to determine a sigmoid function as a decision boundary between awake and anesthetized states: if the output value is lower than 0.5, the epoch is classified as awake; otherwise, the epoch is classified as anesthetized; accumulating the output value of consecutive awake and anesthetized epochs, respectively, and constraining each of the output value between 0 and 1 in order to predict a likelihood of a successive awake or anesthetized epoch.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Embodiments of the invention are described in more details hereinafter with reference to the drawings, in which:

[0031] FIG. 1 schematically depicts an overall structure of the present system;

[0032] FIG. 2 schematically depicts the structure of a signal pre-processor according to an embodiment of the present invention;

[0033] FIG. 3A schematically depicts the structure of a first filter (IIR filter) of the signal pre-processor as shown in FIG. 2;

[0034] FIG. 3B schematically depicts the structure of a second filter (FIR filter) of the signal pre-processor as shown in FIG. 2;

[0035] FIG. 3C schematically depicts the structure of a down sampler of the signal pre-processor as shown in FIG. 2;

[0036] FIG. 4 schematically depicts the structure of an epoch generator according to an embodiment of the present invention;

[0037] FIG. 5 schematically depicts the structure of a feature extractor according to an embodiment of the present invention;

[0038] FIG. 6A schematically depicts the structure of a derivative calculator of the feature extractor as shown in FIG. 5;

[0039] FIG. 6B schematically depicts the structure of a variance calculator of the feature extractor as shown in FIG. 5;

[0040] FIG. 7 schematically depicts the circuitry structure of a classifier according to an embodiment of the present invention;

[0041] FIG. 8 schematically depicts the circuitry structure of a predictor according to an embodiment of the present invention;

[0042] FIG. 9A shows the comparison between a model simulation and output of the present system in an anesthetized animal model: top panel is spectrogram; middle panel is electroencephalogram; bottom panel shows prediction of superimposed waveforms of simulated and hardware outputs;

[0043] FIG. 9B shows the comparison between a model simulation and output of the present system in a conscious (awake) animal model: top panel is spectrogram; middle panel is electroencephalogram; bottom panel shows prediction of superimposed waveforms of simulated and hardware outputs;

[0044] FIG. 9C shows the comparison between a model simulation and output of the present system in an animal model during transition (conscious/anesthetic cycle): top panel is spectrogram; middle panel is electroencephalogram; bottom panel shows prediction of superimposed waveforms of simulated and hardware outputs;

[0045] FIG. 10A shows a time-coded scatter plots of features for transition cases as shown in FIG. 9C, where the subjects pass from anesthetized to awake state;

[0046] FIG. 10B shows a time-coded scatter plots of features for transition cases as shown in FIG. 9C, where the subjects pass from awake to anesthetized state.

DETAILED DESCRIPTION OF THE INVENTION

[0047] In the following description, systems, devices, methods of, and the likes are set forth as preferred examples. It will be apparent to those skilled in the art that modifications, including additions and/or substitutions may be made without departing from the scope and spirit of the invention. Specific details may be omitted so as not to obscure the invention; however, the disclosure is written to enable one skilled in the art to practice the teachings herein without undue experimentation.

[0048] It should be apparent to practitioner skilled in the art that the foregoing and subsequent examples of the system and method are only for the purposes of illustration of working principle of the present invention. It is not intended to be exhaustive or to limit the invention to the precise forms disclosed.

[0049] Turning to FIG. 1, the present system basically includes a signal pre-processor 100, an epoch generator 200, a feature extractor 300, a classifier 400, and a predictor 500. When the EEG signal generated from the brain of a subject is detected by an EEG electrode or alike (not shown in FIG. 1), it is initially pre-processed by the signal pre-processor 100, then processed by the following sequence of different levels in the present system: epoch generator, feature extractor, classifier and predictor. The signal inputted to and after processed by the present system will be outputted as a score of anesthetic state of the subject denoting the depth of anesthesia (DOA) of the subject.

[0050] Turning to FIG. 2, as a first level of the present system, the signal pre-processor includes two filters and a down sampler. A first filter 101, selected from a 50 Hz band-stop Infinite Impulse Response (IIR) filter, initially receives the EEG signal detected by an EEG electrode or alike from the subject. A 50 Hz band-stop IIR filter is selected for the first filter of the pre-processor is because the EEG collected from the brain of a small animal, e.g., mouse, has a very low voltage amplitude. The afore-mentioned band-stop frequency of the first filter can vary depending on the voltage amplitude of the EEG of a designated animal. Detailed circuitry diagram of the first filter is shown in FIG. 3A.

[0051] A second filter 102, selected from a low-pass equiripple Finite Impulse Response (FIR) filter, is configured in the signal pre-processor subsequent to the first filter because the anesthetic effect on brain signal of small animal is only significant in the low-frequency region, and the low-frequency region is usually less than 250 Hz. In certain embodiments of the present invention, the incoming signal to the FIR filter is already passed through another low-pass filter with a cut-off frequency of 250 Hz. Employing the FIR filter with the same cut-off frequency is for reducing further unnecessary data from the signal, according to the Nyquist-Shannon sampling theorem which states that sampling at a rate which is twice as the highest significant frequency can represent the signal without any loss of information. Hence, in those embodiments, the sampling frequency can be up to 500 Hz. Detailed circuitry diagram of the second filter is shown in FIG. 3B.

[0052] A down sampler 103 is configured subsequent to the second filter in the signal pre-processor for down sampling the EEG signal received by the signal pre-processor. Taking a mouse EEG signal as an example, the sampling frequency is 5 KHz. According to the Nyquist-Shannon sampling theorem and the selected FIR filter cut-off frequency, the mouse EEG signal from this mouse model has to be down sampled 10 times (r=10) according to the architecture of the down sampler as shown in FIG. 3C. In FIG. 3C, the down sampler includes a count limited up counter which counts from 1 to 10, and repeats at the frequency of 5 KHz. When the counter value is equal to 1, the value of the input signal is stored in a register and kept as the output for the subsequent cycles.

[0053] Turning to FIG. 4, an epoch generator 200 is configured to consume lower resources and reduce the size of the memory 203 to the smallest. The epoch generator 200 creates a “sliding window” for 1-second and 10% overlapping epochs. As shown in FIG. 4, after pre-processing the incoming signal through the first filter, second filter and down sampler of the pre-processor sequentially according to certain embodiments of the present invention, a downsampler's output, DS(i), is fed as one of the three inputs (d1) of a two-input multiplexer 201 of the epoch generator 200. A second input (d0) is a feedback signal, epoch(i), from the memory 203 (e.g., RAM). An address generator 202 determines a selector input (sel) of the multiplexer 201, where there are two inputs of the address generator: one at 250 KHz (counter_fast) and the other at 500 Hz (counter_slow), in which counter_slow is an up counter that starts counting from 51 to 500 at 500 Hz frequency. The multiplexer's output is initially set as DS(i) until the counter value reaches 500. Once the multiplexer's output is equal to 500, counter_fast is activated and starts counting until the end of that cycle. During that time, the multiplexer's output is selected as epoch(i) and the counter_fast completes the counting from 1 to 500 at 250 KHz frequency. The output of the counter_slow and counter_fast are generated by the address generator as write address (w) and read address (r), respectively, which are further fed to the memory 203. The address generator 202 also generates a control signal, ctrl(i), for subsequent operations/processing. The output (out) of the multiplexer is an input data stream (in) of the memory. The output of the memory, epoch(i), together with ctrl(i), are fed into the feature extractor of the present system for further processing.

[0054] Turning to FIG. 5, because two sets of features, derivative and variance, are needed to be extracted for the subsequent classification, both epoch(i) and control signal, ctrl(i), outputted from the epoch generator 200 are simultaneously fed into the feature extractor 300. The feature extractor 300 includes two distinct units: derivative calculator 301 and variance calculator 302, to respectively determine derivative, dif(i), and variance, var(i). Detailed circuitry diagrams of the derivative calculator 301 and variance calculator 302 are shown in FIG. 6A and FIG. 6B, respectively.

[0055] Turning to FIG. 6A, the derivative feature of the epoch signal is determined by the derivative calculator 301. A measure of a first-order derivative represents the rate of change in neural activity with respect to time. The neural activity increases in awake (conscious) state compared to anesthetic state. Determination of this derivative is a first feature to differentiate between the two states in terms of their neural activity variation, which is determined by calculating the mean of the accumulated squared-difference between each sample and its preceding one according to the following equation (1):

[00001]$\begin{array}{cc}{X}_l\left(i\right)=\mathrm{mean}\left[{\left(\frac{{\mathrm{dx}}_i}{\mathrm{dt}}\right)}^2\right]& \left(1\right)\end{array}$

where X.sub.1(i) is the derivative feature for i.sup.th EEG epoch x.sub.i(t).

[0056] The first derivative is squared before computing the mean to avoid the result be zero and give a single concrete positive measure. To implement in the derivative calculator, the following equation (5) is given:

derivative,dif(i)=mean[(epoch(i)−epoch(i−1)).sup.2]  (5)

[0057] In FIG. 6A, a delay is used, along with a subtractor, to calculate the difference between two epochs. After that, a multiplier is used to perform the square operation. An accumulator and a divisor are used to calculate the mean thereafter. Finally, with the help of the control signal, ctrl(i), the derivative, dif(i), is stored in a register for subsequent classification.

[0058] Turning to FIG. 6B, a variance of a series of epochs of both states is determined by the square root of the squared epoch according to equation (2):

[00002]$\begin{array}{cc}\sigma =\sqrt{\frac{{.Math.}_{i=1}^N{\left(x_i^{+}-{\mu }^{+}\right)}^2}{N}}& \left(2\right)\end{array}$

where x.sub.i.sup.+=√{square root over (x.sub.i.sup.2)} and μ+ is an average of x.sub.i.sup.+.Math.σ and N represents population variance and size of each successive epoch.

[0059] Considering the absolute value of epoch amplitude, mean, μ, can be determined by equation (6):

[00003]$\begin{array}{cc}\mathrm{mean},\mu =\frac{1}{n}\underset{i=1}{\overset{n}{.Math.}}.Math.\mathrm{epoch}\left(i\right).Math.;& \left(6\right)\end{array}$

and variance, var(i), can be determined by equation (7):

[00004]$\begin{array}{cc}\mathrm{Variance},\mathrm{var}\left(i\right)=\sqrt{\frac{1}{n}\underset{i=1}{\overset{n}{.Math.}}{\left(.Math.\mathrm{epoch}\left(i\right).Math.-\mu \right)}^2}& \left(7\right)\end{array}$

[0060] In FIG. 6B, the absolute value of each epoch is accumulated and divided by the epoch's total sample number to determine the mean. The mean is subtracted from epoch(i) to determine its deviation. An average of the squared deviation is determined by consecutively using a multiplier, an accumulator, and a divisor circuit. Finally, the square root of the average squared deviation is obtained, which is variance, var(i). Variance, var(i), is stored in a register for subsequent classification.

[0061] Turning to FIG. 7, the two features, variance, var(i), and derivative, dif(i), extracted by the feature extractor 300 and stored in the respective register of the variance calculator 302 and derivative calculator 301 are inputted into the classifier 400 to implement a logistic regression classification which is an efficient and hardware-friendly machine learning algorithm for multiclass classification. The classifier 400 comprises two main cascaded units. A first unit is a feature mapping unit for expanding all features by double the number of the features, where the features doubled include original features and corresponding squared values. By feature mapping, classification accuracy is enhanced due to a better classification boundary produced by an increase in classifier variance. A second unit is a sigmoid function that is fed by the output of the feature mapping unit to serve as a decision boundary, which is implemented by an exponential and reciprocal computation circuit for subsequent prediction. The sigmoid function, ƒ(x), is according to equation (3):

[00005]$\begin{array}{cc}f\left(x\right)=\frac{1}{1+e^{-\left(\beta x\right)}}& \left(3\right)\end{array}$

[0062] In FIG. 7, W.sub.1-W.sub.5 are regression weights which are multiplied with the inputs and their squared values by the feature mapping unit before being fed to the exponential and reciprocal computation circuit. If the output of the classifier, cls(i), is lower than 0.5, the epoch will be classified as awake (conscious) state; otherwise, anesthetized state.

[0063] To get a highly accurate and fast prediction of EEG channels, a prediction mechanism based on accumulation of classifier outputs for consecutive EEG epochs is implemented in the predictor 500. In particular, the predictor 500 is configured to accumulate the classifier's outputs and constrains the result between 0 and 1. It initially starts with a decision value of 0.5 (i.e., 50%), meaning both awake and anesthetized classes have the same probability. The classifier's output of each received segment leads the prediction decision either up or down according to equation (4):

P.sub.pred(i)=satlin[P.sub.pred(i−1)+(2ƒ(x.sub.i)−1)B]  (4)

where i=1: number of epochs, P.sub.pred(0)=0.5, B=0.05 (jump bound), and satlin is a saturating linear transfer function to keep the predictor output out(i) bounded within 0 and 1.

[0064] Successive anesthetized or awake segments accumulate the prediction probability (P.sub.pred) upward or downward, respectively, with scaled steps. A confident prediction decision is achieved after crossing the anesthetized (75%) or awake (25%) decision threshold. By this mechanism, a confident prediction decision of 100% channel prediction accuracy for all datasets can be guaranteed.

[0065] Turning to FIG. 8, the predictor 500 includes a predictor circuit configured to determine the level of anesthesia according to equation (8) which is derived from equation (4):

out(i)=satlin[out(i−1)+(2*cls(i)−1)*0.05]  (8)

where satlin is the saturating linear transfer function which helps follow the transient behavior of the EEG instantly, i.e., to track immediately if a subject is waking up from anesthesia. This function also helps obtain a meaningful interpretation of the output level in terms of decimal values of percentages; out(i) denotes the measured DOA.

EXAMPLE

Example 1

[0066] Table 1 below summarizes performance of the present system by using multiple classes of testing datasets to measure the accuracy of the present system. In this example, all datasets have been pre-processed to generate 1-second epochs with 10% overlapping before being subject to classification.

TABLE-US-00001 TABLE 1 Duration Total C.C..sup.# Class C.P.D.{circumflex over ( )} Set Class (s) Epochs Epochs Accuracy Time (s)  1 Anesth. 600 666 644 96.70% 6.5  2 Anesth. 600 666 594 89.19% 6.4  3 Anesth. 600 666 635 95.35% 6.8  4 Anesth. 600 666 645 96.85% 6.5  5 Anesth. 600 666 638 95.80% 6.6  6 Anesth. 600 666 623 93.54% 7.9  7 Anesth. 600 666 640 96.10% 6.2  8 Anesth. 600 666 577 86.64% 7.3  9 Anesth. 600 666 543 81.53% 8.1 10 Anesth. 600 666 592 88.89% 7.3 11 Awake 600 666 650 97.60% 4.6 12 Awake 600 666 654 98.20% 4.6 13 Awake 600 666 655 98.35% 4.6 14 Awake 600 666 654 98.20% 4.5 15 Awake 600 666 651 97.75% 5 16 Awake 600 666 619 92.94% 10 17 Awake 600 666 622 93.39% 10 18 Awake 600 666 598 89.79% 11 19 Awake 600 666 647 97.15% 4.6 20 Awake 600 666 652 97.90% 4.6 21 Trans. 400 446 412 92.38% 14.9 22 Trans. 400 446 433 97.09% 7.9 23 Trans. 400 446 426 95.52% 6.6 24 Trans. 400 446 412 92.38% 12.4 Keys: .sup.#Correctly Classified Epochs {circumflex over ( )}Confident Prediction Decision Time “Anesth.”: Anesthetized “Trans.”: Transition, i.e., From Anesthetized to Awake to Anesthetized

[0067] In cases of completely anesthetized subjects, the classification accuracy is about 92% on average; those from awake subjects result in about 96% average classification accuracy; those from transition subjects result in about 94% average classification accuracy. To reach 100% channel prediction accuracy, the present system takes about 7 seconds on average. The accuracy found in this example suggests that the DOA obtained by the present system in multiple classes is comparable to a clinical-level accurate DOA, or even more accurate.

Example 2

[0068] Table 2 below summarizes the resource utilization of the present invention incorporated into a conventional FPGA (Xilinx Artix-7 FPGA is selected in this example)

TABLE-US-00002 TABLE 2 Resource Available Utilized Utilization Rate LUT 53200 18522 34.82%  LUTRAM 17400 33 0.19% FF 106400 876 0.82% BRAM 140 2 1.43% DSP 220 54 24.55%  Keys: “LUT”: Look Up Table “LUTRAM”: LUT Random-Access Memory “FF”: FlipFlop “BRAM”: Block RAM “DSP”: Digital Signal Processing Blocks

[0069] The results from Table 2 suggests that the present invention consumes lower level of resources in different aspects compared to some conventional hardware-implemented DOA systems, such as Saadeh et al. (2019), in which it requires six feature extraction and uses the fine-decision-tree classification algorithm (requiring 26,520 FFs; 50,111 LUTs) for measuring DOA. In contrast, the present invention uses a simple logistic regression machine learning algorithm for classification; the present invention only requires 876 FFs and 18,522 LUTs in FPGA, because only two features are required to be extracted, and the features selected are hardware-friendly and mathematically uncomplicated. One more advantage of the present invention over the conventional DOA measurement system is a relatively lower on-chip power consumption (only 0.446 watts including 0.338 watts of dynamic and 0.108 watts of static power, respectively) because a 28 mm CMOS chip is used. An average total system delay of the datasets is about 12 μs which is mainly due to the inherent properties of the FIR filter used in the signal pre-processor. This latency is within the tolerance of measuring EEG of small animal. If this system latency needs to be shortened to fit other models or purposes, it can be further reduced by using an alternative analog equivalent to the FIR filter of the signal pre-processor.

[0070] FIGS. 9A-9C provides spectrogram, electroencephalogram and prediction by superimposing output waveforms of a simulation and a hardware output in three different classes, respectively. In the top panel of each of FIGS. 9A-9C, the corresponding spectrogram provides a way to observe a temporal EEG in a spectral domain. For example, the dominant power of anesthetized EEG is in the lower frequency spectrum (top panel in FIG. 9A). On the other hand, the power is distributed in the higher frequency oscillation region in awake EEG cases (top panel in FIG. 9B). The spectrogram from the transition EEG cases is especially remarkable in terms of giving class alteration boundary (marked by two arrows in top and middle panels of FIG. 9C). From the superimposed output waveforms (bottom panel in each of FIGS. 9A-9C) which provides a mean of graphical comparison between simulated and hardware-generated results, respectively, it can be seen that the two output traces are extensively identical in all three classes, suggesting that the present system is capable of measuring DOA substantially in line with the simulated result from a corresponding software running the same algorithm.

[0071] To illustrate a gradient shift of properties in feature space, FIGS. 10A and 10B provide time-coded scatter plots of features by taking the transition cases according to FIG. 9C as an example. Corresponding dotted line in FIGS. 10A and 10B represent the decision boundary separating two classes (anesthetized and awake). From the plots in FIGS. 10A and 10B, it is evident that the measures of the associated features according to the present invention go through a gradient transition from one class to another in those transition cases, which validates the incremental DOA measurement according to the present invention.

[0072] Table 3 below summarizes the difference(s) between the conventional DOA measurement methods and the present invention.

TABLE-US-00003 TABLE 3 No. of Classif- Power Signal features Classification ication Require- Used Subject extracted Method Accuracy Sensitivity Specificity Hardware? CMOS? ment Liu et al. EEG Human 1 Random Forest 70.78% Not Not No Not Not (2018) available available available available Nagaraj et al. EEG Human 6 SVM 81.18% 81.30% 81.06% No Not Not (2018) available available Shalbaf et al. EEG Human 4 ANFIS-LH 93.00% Not Not No Not Not (2018) available available available available Ha et al. EEG, Human 10 DNN Not Not Not Yes 65 mm  1.0 V (2018) NIRS available available available Khan et al. EEG, Human 7 S. Decision Tree 79.00% 83.40% 74.6% Yes 65 mm  1.0 V (2018) EMG Saadeh et al. EEG Human 6 F. Decision Tree 92.20% 91.90% 92.06% Yes 65 mm 0.90 V (2019) Yoon et al. EEG Animal 1 Mod. Shamnon Not Not Not No Not Not (2011) available available available available available Kortelainen EEG Animal 1 Bayesian Info. Not Not Not No Not Not et al. (2012) available available available available available Xu et al. EEG Animal 1 Lempel-Ziv Not Not Not No Not Not (2005) available available available available available Present EEG Animal 2 Logistic 94.00%.sup.# or 92.06%.sup.# 95.06% Yes 28 mm 0.95 V Invention Regression 100.00%{circumflex over ( )} .sup.#Classification for 1-second Epoch {circumflex over ( )}Channel prediction after an average 7 second run-time

[0073] It can be seen that most of the conventional DOA measurement methods are for human subjects; their classification accuracy only ranges from 70% to 93%. There is no hardware-implemented DOA measurement for smaller animal subjects, but only software-based conventional methods are available with relatively low classification accuracy. Most of the conventional DOA measurement methods with hardware implementation use 65 mm CMOS chips, whereas only 28 mm CMOS chip is used in the present invention to enable a better FPGA performance.

[0074] The foregoing description of the present invention has been provided for the purposes of illustration and description. It is not intended to be exhaustive or to limit the invention to the precise forms disclosed. Many modifications and variations will be apparent to the practitioner skilled in the art.

[0075] The embodiments were chosen and described in order to best explain the principles of the invention and its practical application, thereby enabling others skilled in the art to understand the invention for various embodiments and with various modifications that are suited to the particular use contemplated. It is intended that the scope of the invention be defined by the following claims and their equivalence.

INDUSTRIAL APPLICABILITY

[0076] The present invention provides a hardware-implemented DOA measurement based on EEG of small animal with high classification accuracy and channel prediction accuracy within a tolerable system latency, which has potentials in applying to veterinary medicine and surgery requiring anesthesia to an animal subject during surgical operation. It also has potentials to other operations at other settings requiring anesthesia or in observing change in other physiological parameters during transition from anesthesia to conscious state or vice versa of a subject.