Opioid detection using EEG signals

Abstract

A System (1) for detecting the presence of a drug in a person (2) comprises: an EEG sensor (10) and an analysis unit (11) with a processor arrangement (110) and memory (111), the memory storing instructions that, when executed by the processor arrangement (110), cause the analysis unit (11) to perform the following steps: receive a signal (300-307) comprising a time series of values (310-313) using the EEG sensor (10); analyze the values (310-313) of the signal (300-307) to translate the signal (300-307) into a series of symbols (Si); generate a plurality of words (500-504), each word (500-504) comprising a plurality of consecutive symbols (Si); and provide the plurality of words (500-504) and / or one or more variables derived from the plurality of words (500-504) to a classifier (600) configured to indicate the presence of the drug in the person (2) based on the plurality of words (500-504).

Description

[0001] Opioid Detection using EEG Signals

[0002] Description

[0003] The invention relates to a system and a method for detecting the presence of a drug in a person, and to a computer-readable storage medium.

[0004] Electroencephalography (EEG) allows to record an electrogram of spontaneous electrical activity of the brain of a person. EEG is not only used in the context of epilepsy, but can also be used, e.g., to assess the depth of anesthesia in patients.

[0005] P. L. Purdon et al. describe “Electroencephalogram signatures of loss and recovery of consciousness from propofol” in Proc. Natl. Acad. Sci., 110(12), E1142- E1151, 2013. The authors recorded high-density EEGs in humans during gradual induction of and emergence from unconsciousness with propofol. During profound unconsciousness, alpha amplitudes were found to be maximal at low-frequency peaks, whereas during the transition into and out of unconsciousness, alpha amplitudes were maximal at low-frequency nadirs.

[0006] EEG has several advantages over other techniques, such as MRT, in that hardware costs can be low and the required hardware may have a small form factor. Further, EEG can be relatively tolerant of subject movement, silent and non-invasive.

[0007] Therefore, it is desirable to use EEG for additional purposes. EP 2 906 112 B1 describes systems and methods for monitoring and controlling a state of a patient and, more particularly, systems and methods for monitoring and controlling a state of a patient receiving a dose of anesthetic compounds or, more colloquially, receiving a dose of anesthesia. Therein, sets of EEG data are analyzed using thresholding to generate binary time series from the sets of EEG data and using a burst suppression probability algorithm implementing a state-space model to generate a time-series of burst suppression probabilities from the binary time series, wherein burst suppression probabilities are related to the patient brain's instantaneous likelihood of being in a suppressed state. Further, signature profiles consistent with the administration of at least one drug having anesthetic properties are determined using the time-series of burst suppression probabilities.

[0008] Burst suppression is an EEG pattern that is characterized by periods of high-voltage electrical activity alternating with periods of no or little activity in the brain. The pattern is found in patients with inactivated brain states, such as from general anesthesia or coma.

[0009] As an example, it can be possible to monitor the effect of propofol using the method described in EP 2 906 112 B1 . However, an opioid, such as remifentanil, only leads to burst suppression at extremely high doses, much higher than usually applied to patients. Therefore, the known method is limited to certain drugs and dose rates.

[0010] It is an object of the present invention to expand the range of applications of EEG.

[0011] This object is achieved by a system having the features of claim 1 .

[0012] Accordingly, a system for detecting the presence of a drug in a person comprises an EEG (Electroencephalography) sensor and an analysis unit. The analysis unit comprises a processor arrangement and memory. The memory stores instructions that, when executed by the processor arrangement, cause the analysis unit to perform the following steps: receive a signal comprising a time series of values using the EEG sensor; analyze the values of the signal to translate the signal into a series of symbols. The instructions, when executed by the processor arrangement, further cause the analysis unit to generate a plurality of words, wherein each word comprises a plurality of (e.g., consecutive) symbols; and to provide the generated plurality of words and / or one or more variables derived from the plurality of words (500-504) to a (e.g., trained) classifier configured to indicate the presence of the drug in the person based on the plurality of words. This is based on the finding that there is a highly significant correlation between the presence of remifentanil or other opioids in the blood of the person and the occurrence of specific words (e.g., among a number of possible words). Therefore, the classifier can perform a simple determination of a proportion or percentage of one or more predetermined words, or it can comprise a trained machine learning module. By means of the described system, the range of applications of EEG can be strongly expanded to the monitoring of opioids and other drugs. The one or more variables derived from the plurality of words may indicate a distance of the consecutive symbols of the respective word and / or an integral of the signal between the two consecutive symbols. The symbols may indicate zero crossings (alternatively or in addition the direction of the crossing of zero or another value).

[0013] Analyzing the values of the signal may comprise comparing the values of the signal with predefined symbols to translate the signal into the series of symbols (e.g., being a time series of symbols). The series of symbols may comprise the same number of symbols as (or a smaller number than) the signal comprises values.

[0014] The EEG sensor may be a frontal EEG sensor for sensing frontal cortical activity. It has been found that by measuring the frontal cortical activity, e.g., with an EEG sensor configured to be adhered or otherwise mounted to the forehead, an improved precision in the results can be obtained. The EEG sensor has a plurality of electrodes. For example, the EEG sensor has at least or exactly three electrodes.

[0015] The signal may be a or comprise a raw signal. This allows a simple analysis. Alternatively, or in addition, the signal may be or comprise a pre-processed signal. The pre-processed signal may be obtained by pre-processing the raw signal. This allows to improve the precision of the results.

[0016] Pre-processing the raw signal may comprise band-pass filtering to obtain one or more frequency bands. For example, the frequency bands may be the a, [31 , |32, 0, 5 and y bands or a selection of one or more thereof. The pre-processed signal may comprise or be one or more of the frequency bands, particularly the a, [31, [32, 0, 6 and / or y bands. In this manner, optimized results for the specific drugs can be obtained.

[0017] Pre-processing the raw signal may comprise determining at least one signal envelope (e.g., of the raw signal or of one or more, e.g., all, of the frequency bands. The pre-processed signal may be or comprise the at least one signal envelope. It has been found that for specific drugs, such as remifentanil, using the envelope allows to improve the precision of the measurement. Alternatively, or in addition, pre-processing the raw signal may comprise determining at least one phase signal.

[0018] Each of the predefined symbols may be assigned to a respective range of signal values. By this, the complexity may be reduced, and the processing time may be increased. In addition, this allows to obtain meaningful results already with a comparably little number of samples in the available data.

[0019] The width of at least one of the ranges may equal one standard deviation (or a plurality of standard deviations) of the corresponding signal. By this, the complexity of the analysis can be reduced while maintaining the information content of the data.

[0020] For example, at least or exactly two, three, four or even more different symbols are predefined. That is, in a very simple example, only two different symbols are defined, and each value of the signal is assigned to one of the two symbols. For example, the two symbols could be “high” for a value above zero, and “low” for values below zero.

[0021] In an example, each of the plurality of words comprises the same number of consecutive symbols, e.g., two, three, four or more consecutive symbols. This allows a simple analysis that yields significant results.

[0022] The classifier may be configured to determine the number of occurrences of the words. It has been found that the occurrence of specific words correlates with the presence of certain drugs in the person, in particular remifentanil.

[0023] Alternatively, or in addition, the classifier may be configured to determine an entropy of the plurality of words, in particular by calculating the Renyi entropy. It has also been found that the entropy of the plurality of words correlates with the presence of certain drugs in the person, in particular remifentanil, and allows particularly significant results.

[0024] The classifier may comprise a classifier module that has been trained with a plurality of words generated using signals obtained from persons without medication as a first dataset, and with a plurality of words generated using signals obtained from persons administered with an opioid, particularly remifentanil; as a second dataset. Alternatively, or in addition, the classifier module may have been trained with a plurality of words generated using signals obtained from persons without medication as a first dataset, and with a plurality of words generated using signals obtained from persons administered with a narcotic agent, particularly propofol, as a second dataset. Alternatively, or in addition, the classifier module may have been trained with a plurality of words generated using signals obtained from persons administered with an opioid, particularly remifentanil, as a first dataset, and with a plurality of words generated using signals obtained from persons administered with a narcotic agent, particularly propofol, as a second dataset. This allows a high significance of the respective determination.

[0025] According to an aspect, a method for detecting the presence of a drug in a person is provided. The method may use the system of any aspect or embodiment described herein. The method comprises the following steps, e.g., performed by an analysis unit communicatively coupled with an EEG sensor (e.g., of the system of any example described above): receiving a signal comprising a time series of values using the EEG sensor; analyzing the values of the signal to translate the signal into a series of symbols (e.g., comparing the values of the signal with predefined symbols to translate the signal into a time series of symbols); generating a plurality of words, each word comprising a plurality of (e.g., consecutive) symbols; and providing the plurality of words (and / or one or more variables derived from the plurality of words) to a classifier configured to indicate the presence of the drug in the person based on the plurality of words. Regarding the advantages of the method, reference is made to the description of the system above.

[0026] The classifier may comprise a classifier module that has been trained with a plurality of words generated using signals obtained from persons without medication as a first dataset, and with a plurality of words generated using signals obtained from persons administered with an opioid, particularly remifentanil; as a second dataset. Alternatively, or in addition, the classifier module may have been trained with a plurality of words generated using signals obtained from persons without medication as a first dataset, and with a plurality of words generated using signals obtained from persons administered with a narcotic agent, particularly propofol, as a second dataset. Alternatively, or in addition, the classifier module may have been trained with a plurality of words generated using signals obtained from persons administered with an opioid, particularly remifentanil, as a first dataset, and with a plurality of words generated using signals obtained from persons administered with a narcotic agent, particularly propofol, as a second dataset. The method may comprise the training of the classifier module, in advance of the detection of the presence of the drug in the person.

[0027] According to an aspect, computer-readable storage medium is provided. The computer- readable storage medium stores instructions that, when executed by a processor arrangement of an analysis unit (e.g., of the system described above), cause the analysis unit to perform any example of the method described above. Regarding the advantages, again reference is made to the description of the system above.

[0028] The idea underlying the invention shall subsequently be described in more detail by referring to the embodiments shown in the figures. Herein:

[0029] Fig. 1 shows a system for detecting the presence of a drug in a person;

[0030] Fig. 2 shows a raw signal obtained using an EEG sensor, and various pre- processed signals with different frequency bands thereof;

[0031] Fig. 3 shows one of the pre-processed signals of Fig. 2 together with an envelope of the pre-processed signal;

[0032] Fig. 4 shows a part of the pre-processed signal of Fig. 3, wherein values thereof are each assigned to one of a plurality of predetermined symbols;

[0033] Fig. 5 illustrates the training of a classifier and its application to test data;

[0034] Fig. 6 illustrates steps of the training of the classifier of Fig. 6;

[0035] Fig. 7 illustrates steps of an optimization procedure for the analysis of the EEG signals;

[0036] Fig. 8A shows parts of pre-processed signals taken from a person with and without administration of a drug;

[0037] Fig. 8B shows the parts of the pre-processed signals of Fig. 5A, wherein, additionally, a random number signal is illustrated; and

[0038] Fig. 9 shows an example of 64 possible words made of three symbols that can have four values.

[0039] Subsequently, particularly a system for detecting the presence of a drug in a person will be described. The embodiments described herein shall not be construed as limiting for the scope of the invention. Fig. 1 illustrates a system 1 configured for detecting the presence of a drug in a person 2. The system 1 comprises an electroencephalography sensor, EEG sensor 10 and an analysis unit 11 coupled to the EEG sensor 10.

[0040] The EEG sensor 10 of the present example is configured for sensing frontal cortical activity. For this purpose, the EEG sensor 10 is formed so as to be attached to the forehead 20 of the person 2. The EEG sensor 10 comprises a plurality of electrodes 100A, 100B, 100C, namely, at least two electrodes 100A-100C. In the example of Fig. 1 , the EEG sensor 10 comprises (exactly) three electrodes 100A-100C. Each of the electrodes 100A-100C is provided with an adhesive layer and can be adhered to the forehead 20 of the person 2. The EEG sensor 10 is elongate. The electrodes 100A-100C are arranged in a row. The EEG sensor 10 is made of a flexible material. Here, each of the electrodes 100A-100C of the EEG sensor 10 are mounted on a respective pad. The pad of a first electrode 100A of the electrodes 100A-100C is connected with the pad of a third electrode 100C of the electrodes 100A-100C via the pad of second electrode 100B of the electrodes 100A-100C. In the example of Fig. 1 , the first electrode 100A is configured to be attached to a central part of the person’s 2 forehead 20. The third electrode 100C forms an end of the EEG sensor 10 and is configured to be attached to a side of the person’s 2 forehead 20. The second sensor 100B is configured to be attached to the forehead 20 at a location between the two other electrodes 100A, 100B.

[0041] The EEG sensor 10 is operatively connected with the analysis unit 11. Here, each of the electrodes 100A-100C is electrically connected with the analysis unit 11. The analysis unit 11 is configured to receive EEG signals from the EEG sensor 10, e.g., as a voltage signal between two of the electrodes 100A-100C. Alternatively, the EEG sensor 10 can be connected to the analysis unit 11 in any other wired or wireless manner.

[0042] The analysis unit 11 , which can also be referred to as analyzer, comprises readout electronics 113 which are electrically connected with the EEG sensor 10, more specifically, with the electrodes 100A-100C. The readout electronics 113 comprise a voltmeter to measure a voltage between two of the electrodes 100A-100B. Further, the readout electronics 113 comprise an analog-to-digital converter, ADC. The ADC receives analog signals from the voltmeter and outputs corresponding digital signals.

[0043] The analysis unit 11 further comprises a processor arrangement 110 and memory 111. The processor arrangement 110 comprises one or more processors. The memory 111 is a computer-readable storage medium. The memory 111 stores instructions for execution by the processor arrangement 110. The processor arrangement 110 is configured to process the signals received from the EEG sensor 10. In the present example, the processor arrangement 110 is configured to receive the digital EEG signals. The analysis unit 11 further comprises a display 112. The processor arrangement 110 may be configured to display the EEG signals on the display 112. The processor arrangement 110 may also be configured to display results of the detection of the presence of a drug in the person 2 on the display 112, which will be described in further detail below.

[0044] Fig. 2 shows diagrams with examples of various signals 300-306 obtained using the EEG sensor 10. Each diagram shows an amplitude of the signal on the y axis versus the number of the sample on the x axis. The ADC is configured to sample the analog signal at a given sampling frequency. In the present example, the sampling frequency is 256 Hz. A number of samples divided by the sampling frequency equals a time span, so the depicted window corresponds to about 5 seconds. The amplitude can be in an arbitrary unit. In the present example, the amplitude is shown in microvolts, pV.

[0045] The uppermost diagram of Fig. 2 (labelled with “EEG”) shows an EEG signal 300 including the whole EEG frequency band, in the present example from 0.5 to 45 Hz (here, the resolution of the frequency is 0.5 Hz). The signal 300 may have been subject to artifact and noise removal. It contains all EEG information and therefore is also referred to as raw signal 300 in the following for ease of reference.

[0046] The other diagrams in Fig. 2 show band-pass filtered parts of the raw signal 300. Therefore, these other diagrams show various pre-processed signals 301-306 obtained from the raw signal 300.

[0047] The second diagram from the top shows a pre-processed signal 301 corresponding to the 5 band from 0.5 to 4 Hz.

[0048] The third diagram from the top shows a pre-processed signal 302 corresponding to the 0 band from 4.5 to 8 Hz.

[0049] The fourth diagram from the top shows a pre-processed signal 303 corresponding to the a band from 8.5 to 13 Hz. The fifth diagram from the top shows a pre-processed signal 304 corresponding to the 1 band from 13.5 to 21 Hz.

[0050] The sixth diagram from the top shows a pre-processed signal 305 corresponding to the 2 band from 21.5 to 30 Hz.

[0051] The seventh diagram from the top shows a pre-processed signal 306 corresponding to the y band from 30.5 to 45 Hz.

[0052] When the instructions stored on the memory 111 are executed by the processor arrangement 110, the instructions cause the analysis unit 11 to obtain the raw signal 300 by reading out the EEG sensor 10 with the readout electronics 113 (and optionally performing artifact and noise removal and / or filtering to remove frequencies outside of the a, 01, 02, 0, 6 and y frequency bands). Further, when the instructions stored on the memory 111 are executed by the processor arrangement 110, the instructions cause the analysis unit 11 to perform band-pass filtering operations on the raw signal 300 to obtain the various pre-processed signals 301-306 according to the different EEG frequency bands a, 01 , 02, 0, 5, y.

[0053] The analysis unit 11 is configured to determine a signal envelope of a signal 300-306 to obtain another pre-processed signal 307, e.g., as shown in Fig. 3.

[0054] Fig. 3 illustrates a diagram with the time on the x axis showing a part of the 01 frequency band signal 304 and the envelope of the signal 304. A further pre-processed signal 307 comprises the envelope. The envelope wraps around the signal 304. One maximum of the envelope wraps a plurality of maxima of the signal 304. The envelope wraps the signal 304 on its positive side.

[0055] The envelope may be calculated using the Hilbert transform. For example, the analysis unit 11 (by execution of corresponding instructions in the memory) is configured to calculate the modulus of the known Hilbert transform, which is defined using the Cauchy principal value (denoted by p.v ):

[0056] The analysis unit 11 may be configured to calculate the envelope (e.g., by calculating the respective Hilbert transform) to the raw signal 300 and / or to one or more, e.g., all, band-pass filtered frequency bands a, [31 , [32, 0, 5, y. According to an example, all 7 (raw signal plus the bands) signals are processed further as described and all 7 envelopes thereof, i.e. , 14 inputs to the further analysis in total. Alternatively, only one or some of these inputs are used for the further analysis, e.g., for a given drug a predefined input or set of inputs.

[0057] Each of the signals comprise a time series of values. Hence, the instructions, when executed by the processor arrangement 110, cause the analysis unit 11 to receive a signal 300-307 comprising a time series of values 310-313 using the EEG sensor 10.

[0058] Fig. 4 shows a part of the signal 304 (samples 10 to 24). The signal 304 comprises the time series of values 311-313.

[0059] As illustrated by means of Fig. 4, the instructions, when executed by the processor arrangement 110, cause the analysis unit 11 to analyze the values 310-313 of the signal 304 to translate the signal 304 into a series of symbols, or, more specifically, to compare the values 310-313 of the signal 304 with predefined symbols Si to translate the signal 304 into a time series of symbols 400.

[0060] In Fig. 4, signal 304 is shown as an example, but any of the raw or pre-processed signals 300- 307 including the envelope signal 307 of any one of the signals 300-306 (e.g., all) may be processed accordingly.

[0061] Each of the predefined symbols Si is assigned to a respective range of signal values (here: amplitudes). In the specific example, the width of some of the ranges equals a standard deviation of the corresponding signal 304. That is, the symbols Si are defined based on the standard deviation of the signal 304.

[0062] In the example of Fig. 4, four symbols Si are predefined (other numbers, such as two, three, five or more are also conceivable). Here, these symbols Si are simply labelled with “1”, “2”, “3” and “4”. Two symbols (symbols 2 and 3) are defined to cover the respective range from zero (amplitude or envelope amplitude) to + / - “a” times the standard deviation (e.g., one standard deviation). The other two symbols are defined to cover the respective range above / below starting from + / - “a” times the standard deviation (e.g., one standard deviation).

[0063] Therefore, the three first values 311-313, although being different from one another, all lie within the range of symbol 1. Therefore, the first three values 311-313 are each assigned symbol 1 and the time series of symbols 400 starts with three times the symbol 1. The next value lies in the range of symbol 3, and the same holds for the subsequent two values, which are followed by a value corresponding to the symbol 2 and so forth.

[0064] For the further processing, the time series of symbols 400 is used.

[0065] Next, the instructions, when executed by the processor arrangement 110, cause the analysis unit 11 to generate a plurality of words 500-504, each word 500-504 comprising a plurality of consecutive symbols Si.

[0066] In Fig. 4 some of the words 500-504 are illustrated below the diagram and the time series of symbols 400. The words 500-504 have a predefined length. In the present example, the length is three, but other lengths are also possible, such as two, four, five or more.

[0067] The words 500-504 are defined with overlapping symbols Si. Here, subsequent words 500- 504 are each defined to start one symbol difference. Every word 500-504 starts (and ends) one symbol further than the directly preceding word 500-504. That is, a first word 500 starts with the first symbol Si in the time series of symbols 400 (and contains the first, second and third symbol Si), the second word 501 starts with the second symbol Si in the time series of symbols 400 (and contains the second, third and fourth symbol Si) and so forth. Therefore, the number of words 500-504 corresponds to the number of symbols Si in the time series of symbols 400 (disregarding an end of the time series). Alternatively, the words could be defined not to overlap, but to contain neighboring sets of symbols Si. In that case, the number of words would equal the number of symbols Si divided by the word length. As another example, the words could be shifted differently with respect to one another, e.g., the words could overlap by only one symbol Si, not by two.

[0068] Thus, the values of the signals are analyzed and translated into a series of symbols Si in accordance with a predetermined number N of different symbol Si, a length L of the words 500- 504 and a quantification a of the standard deviation.

[0069] When executed by the processor arrangement 110, the instructions may further cause the analysis unit 11 to determine one or more descriptors.

[0070] One example of such a descriptor is the number of occurrences of the various different possible words 500-504. The word count for every possible word 500-504 can be one descriptor. Another example of such a descriptor is an entropy of the generated plurality of words 500- 504. To determine the entropy, the Shannon entropy of the words 500-504 may be calculated:

[0071] Here, X is a random variable with the value range {x1 , x2, xn}, and P is the probability of the occurrence of a given symbol.

[0072] Alternatively, or in addition, the Renyi entropy of the words 500-504 may be calculated:

[0073] Therein, q is the order of the Renyi entropy and 0 < q < infinity, wherein q is not equal to 1. The limit of the Renyi entropy for q approaching 1 is the Shannon entropy.

[0074] The memory 111 further stores a classifier 600 (which will be described in further detail with reference to Fig. 5 below). The instructions, when executed by the processor arrangement 110, cause the analysis unit 11 to provide the plurality of generated words 500-504 to the classifier 600, wherein the classifier 600 is configured to indicate the presence of the drug in the person 2 based on the plurality of words 500-504. The words 500-504 contain information about electrical activity of the brain of the person 2 indicative of the presence of specific drugs. The classifier 600 has been trained with training data taken under the presence of the specific drugs in the person, and with training data with no drugs present in the person.

[0075] As shown in Fig. 5, a database is provided containing a training set of EEG data and a testing set of EEG data.

[0076] For the present examples, frontal EEG from 168 patients was recorded and analyzed: 81 in awake state before any drug was infused (ND, no drug), 59 only under propofol (PROP) and 28 only under remifentanil (REMI). There is no overlap between patients and groups. A subset of patients (n = 84) was equalized to balance the number of patients of each group (n = 28). A single window (10 seconds) of each patient of this subset was randomly selected as training set. The testing set comprised 15,929 windows of ND, 7,289 of PROP and 1 ,855 of REMI, from all the patients (n = 168). The training set is analyzed as described above and as will be further described below with reference to Fig. 6 (abbreviated by “SD”, symbol dynamics, in Fig. 5). The symbol dynamics yields descriptors, e.g., one or more, e.g., all of the descriptors referred to above.

[0077] These descriptors are provided to one or more classifier modules 601 , 602, 603. Examples of classifier modules 601-603 perform a linear descriptor analysis LDA, a quadratic descriptor analysis QDA and a support vector machine SVM. The classifier 600 may comprise one, two or all of the classifier modules 601-603. Other classifier modules are also conceivable to be applied additionally or instead. For example, a classifier module of the classifier 600 could be any linear or non-linear classifier, or any algorithm to be used for classification (e.g., quadratic, decision tree, ANFIS model or the like). The classification module of the classifier 600 may be an artificial intelligence module. The classification module of the classifier 600 may be configured for machine learning.

[0078] At least one or at least two inputs may be provided to the classifier module(s) 601-603. In an example, the following input is provided to the classifier module(s) 601-603: The entropy / entropies (of an amplitude signal or an envelope signal or both), of one or more of the raw EEG signal 300 or any of the frequency components, signals 301-307, and / or being the entropy computed with a combination on N (number of different symbols), L (word length) and a (standard deviation) parameters. As a further input, the following may be provided to the classifier module(s) 601-603: The occurrence(s) of one or more given word(s), e.g., with the same definitions (amplitude / envelope, raw EEG / frequency components, N, L, a combination as for the entropy).

[0079] Notably, instead of or in addition to words comprising consecutive symbols Si, one or more variables derived from the plurality of words may be provided to the classifier 600. For example, the symbols may indicate zero crossings (or a crossing of another pre-defined value) and the one or more variables may indicate the distance between the consecutive symbols or the area under the signal between the consecutive symbols. Optionally, descriptors as described above may be determined based on such variables, and then provided to the classifier module(s) 601-603. Alternatively or in addition, the variables may be directly provided to the classifier module(s) 601-603.

[0080] In the training, the classifier module 601-603 is also provided with the information of the category of the training data, here ND, REMI or PROP. The training can be performed with a plurality of words 500-504 generated using signals 300- 307 obtained from persons 2 without medication and with a plurality of words 500-504 generated using signals 300-307 obtained from persons 2 administered with an opioid, particularly remifentanil; and / or with a plurality of words 500-504 generated using signals 300- 307 obtained from persons 2 without medication and with a plurality of words 500-504 generated using signals 300-307 obtained from persons 2 administered with a narcotic agent, particularly propofol; and / or with a plurality of words 500-504 generated using signals 300-307 obtained from persons 2 administered with an opioid, particularly remifentanil, and with a plurality of words 500-504 generated using signals 300-307 obtained from persons 2 administered with a narcotic agent, particularly propofol.

[0081] Thus, the analysis unit 11 is configured to perform a method comprising the following steps: receiving a signal 300-307 comprising a time series of values 310-313 using the EEG sensor 10; comparing the values 310-313 of the signal 300-307 with predefined symbols Si to translate the signal 300-307 into a time series of symbols 400; generating a plurality of words 500-504, each word 500-504 comprising a plurality of consecutive symbols Si; and providing the plurality of words 500-504 to a classifier 600 configured to indicate the presence of the drug in the person 2 based on the plurality of words 500-504.

[0082] After the training, the testing data can be applied to the trained classifier 600. In Fig. 5 the classification module 601 is illustrated as an example.

[0083] Fig. 6 illustrates how the performance of the classification may be determined. The training data set is optionally provided to a Hilbert transform module to determine the envelope(s). The selected inputs are provided to a symbol dynamics analysis module which also receives the number of symbols N, the word length L and the quantification value, here the standard deviation a. By performing the analysis iteratively with variations of these parameters, the best parameters can be determined. Then, the descriptors obtained by the symbol dynamics analysis are, optionally after undertaking a log transformation, fed to a discriminant function. This may comprise a selected classifier module 601-603. The result is used to determine a true positive rate (TPR), i.e., sensitivity, a true negative rate (TNR), i.e., specificity and / or an area under curve (AUG). According to Fig. 7 the training set is used to determine the best descriptors and the best classification model(s), depending on the highest TPR, TNR and / or AUC results. Those can vary between the different drugs.

[0084] When the best descriptors and the best classification model(s), the testing data is analyzed to obtain the actual results. These results are then used to evaluate the quality of the overall analysis.

[0085] Fig. 8A shows, for comparison, signals from a person with no drug ND, a person with remifentanil REMI and a person with propofol PROP.

[0086] When applying a symbol number N of 2, here: the symbol “high” for values above 0 and the symbol “low” for values below 0 and a word length of 2, only 4 different words can occur: high- high, high-low, low-high and low-low.

[0087] By performing the analysis as described above, the analysis unit 11 provided the following word counts:

[0088] As can be seen, these have significant differences for the three cases.

[0089] To compare with a random number signal (RNG) as additionally illustrated in Fig. 8B, the word counts have been determined to be 648, 641 , 640 and 627.

[0090] In each case, two parameters have been used.

[0091] For detecting remifentanil (or no drug), for example, the first parameter was the Renyi entropy with the order of 2. For this, the envelope ENV of the [31 frequency band has been used as input (with a symbol number of 2 and a word length of 2). Further, the number of occurrences of the word high-low from the [31 envelope has been used, leading to the sensitivity, the specificity and the area under curve indicated on the right. AMP stands for amplitude, no envelope).

[0092] These results show that non-linear parameters based on symbolic dynamics of the EEG, in both amplitude and envelope domains, can identify the remifentanil presence in the frontal EEG of a patient and distinguish it from the propofol effect.

[0093] The classifier 600 may provide an output indicating the result of the classification, i.e., whether or not the drug (e.g., remifentanil or propofol) is present in the person. The output may be provided on the display 112.

[0094] As an example, this allows to determine whether or not a patient is already in a brain state which is comparable to a state under an opioid, or to check whether or not opioid effects are still active in a patient, prior to starting a therapy. Further, the system and method described may be used to indicate if a patient correctly declared himself (or has been declared) as sober (e.g., of an opioid such as remifentanil).

[0095] While a word length of 2 and a symbol number of 2 leads to 4 different words, a word length of 3 and a symbol number of 4 leads to 64 different words. Fig. 9 illustrates all different words for such an example. Depending on the available statistics and the drug, these parameters can yield improved results. The best parameters are determined as described above.

[0096] The idea of the invention is not limited to the embodiments described above but may be implemented in a different fashion. List of Reference Numerals

[0097] System

[0098] 10 EEG sensor

[0099] 100A-100C Electrode

[0100] 11 Analysis unit

[0101] 110 Processor arrangement

[0102] 111 Memory

[0103] 112 Display

[0104] 113 Readout electronics

[0105] 2 Person

[0106] 20 Forehead

[0107] 300 Raw signal

[0108] 301-307 Pre-processed signal

[0109] 310-313 Value

[0110] 400 Time series of symbols

[0111] 500-504 Word

[0112] 600 Classifier

[0113] 601-603 Classifier module

[0114] Si Symbol a, p1, p2, 0, 5, Y Frequency band

Claims

Claims1. System (1) for detecting the presence of a drug in a person (2), comprising: an EEG sensor (10) and an analysis unit (11) with a processor arrangement (110) and memory (111), the memory storing instructions that, when executed by the processor arrangement (110), cause the analysis unit (11) to perform the following steps: o receive a signal (300-307) comprising a time series of values (310-313) using the EEG sensor (10); o analyze the values (310-313) of the signal (300-307) to translate the signal (300- 307) into a series of symbols (Si); o generate a plurality of words (500-504), each word (500-504) comprising a plurality of consecutive symbols (Si); and o provide the plurality of words (500-504) and / or one or more variables derived from the plurality of words (500-504) to a classifier (600) configured to indicate the presence of the drug in the person (2) based on the plurality of words (500-504).

2. System (1) according to claim 1 , characterized in that analyzing the values (310-313) of the signal (300-307) comprises comparing the values (310-313) of the signal (300-307) with predefined symbols (Si) to translate the signal (300-307) into the series of symbols (400).

3. System (1) according to claim 1 or 2, characterized in that the EEG sensor (10) is a frontal EEG sensor (10) for sensing frontal cortical activity.

4. System (1) according to one of the preceding claims, characterized in that the signal (300-307) is a raw signal (300) or a pre-processed signal (301-307) obtained by preprocessing the raw signal (300).

5. System (1) according to claim 4, characterized in that pre-processing the raw signal (300) comprises band-pass filtering to obtain a plurality of frequency bands (a, [31 , |32, 0, 5, y), wherein the pre-processed signal (301-307) comprises one or more of the frequency bands (a, [31 , |32, 0, 5, y).

6. System (1) according to claim 4 or 5, characterized in that pre-processing the raw signal (300) comprises determining at least one signal envelope, wherein the pre-processed signal (307) comprises the at least one signal envelope.

7. System (1) according to one of the preceding claims, characterized in that each of the predefined symbols (Si) is assigned to a respective range of signal values.

8. System (1) according to claim 7, characterized in that the width of at least one of the ranges equals to a standard deviation or a multiple of the standard deviation of the corresponding signal (300-307).

9. System (1) according to one of the preceding claims, characterized in that two, three, four or more different symbols (Si) are predefined.

10. System (1) according to one of the preceding claims, characterized in that each of the plurality of words (500-504) comprises the same number of two, three or more consecutive symbols (Si).

11. System (1) according to one of the preceding claims, characterized in that the classifier (600) is configured to determine the number of occurrences of the words (500-504).

12. System (1) according to one of the preceding claims, characterized in that the classifier (600) is configured to determine an entropy of the plurality of words (500-504), in particular by calculating the Renyi entropy.

13. System (1) according to one of the preceding claims, characterized in that the classifier(600) comprises a classifier module (601-603) that has been trained with:a plurality of words (500-504) generated using signals (300-307) obtained from persons (2) without medication and with a plurality of words (500-504) generated using signals (300-307) obtained from persons (2) administered with an opioid, particularly remifentanil; and / or a plurality of words (500-504) generated using signals (300-307) obtained from persons (2) without medication and with a plurality of words (500-504) generated using signals (300-307) obtained from persons (2) administered with a narcotic agent, particularly propofol; and / or a plurality of words (500-504) generated using signals (300-307) obtained from persons (2) administered with an opioid, particularly remifentanil, and with a plurality of words (500-504) generated using signals (300-307) obtained from persons (2) administered with a narcotic agent, particularly propofol.

14. Method for detecting the presence of a drug in a person (2), comprising the following steps, performed by an analysis unit (11) communicatively coupled with an EEG sensor (10): receiving a signal (300-307) comprising a time series of values (310-313) using the EEG sensor (10); analyzing the values (310-313) of the signal (300-307) to translate the signal (300- 307) into a series of symbols (Si); generating a plurality of words (500-504), each word (500-504) comprising a plurality of consecutive symbols (Si); and providing the plurality of words (500-504) and / or one or more variables derived from the plurality of words (500-504) to a classifier (600) configured to indicate the presence of the drug in the person (2) based on the plurality of words (500-504).

15. Method according to claim 14, wherein the classifier (600) comprises a classifier module (601-603) that has been trained with: a plurality of words (500-504) generated using signals (300-307) obtained from persons (2) without medication and with a plurality of words (500-504) generated using signals (300-307) obtained from persons (2) administered with an opioid, particularly remifentanil; and / or a plurality of words (500-504) generated using signals (300-307) obtained from persons (2) without medication and with a plurality of words (500-504) generated using signals (300-307) obtained from persons (2) administered with a narcotic agent, particularly propofol; and / ora plurality of words (500-504) generated using signals (300-307) obtained from persons (2) administered with an opioid, particularly remifentanil, and with a plurality of words (500-504) generated using signals (300-307) obtained from persons (2) administered with a narcotic agent, particularly propofol.

16. Computer-readable storage medium storing instructions that, when executed by a processor arrangement (110) of an analysis unit (11), cause the analysis unit (11) to perform the method according to claim 14 or 15.

Images