Methods for training a model that can be used to calculate an injury perception index

Abstract

Method for training a model usable for calculating a nociception index. Method for training a model (M2) usable for calculating a nociception index (qNOX) related to nociceptive effects during a general anesthesia procedure, comprising: obtaining clinical data related to a plurality of previous anesthesia procedures during a training phase separate from the actual use of the model (M2) during an anesthesia procedure; deriving training data (TD) from the clinical data; deriving reference data (RD) from the clinical data; and training the model (M2) using the training data (TD) as input data for the model (M2) and using the reference data (RD) as output data for the model (M2), wherein the training comprises adjusting the model (M2) in dependence on the training data (TD) and the reference data (RD). Herein, the reference data (RD) is derived from the clinical data using an equation comprising a mathematical term whose value is non-linearly variable as a function of a concentration value related to a drug concentration in the patient's body during the anesthesia procedure.

Description

[0001] describe

[0002] The present invention relates to a method for training a model according to the preamble of claim 1, the model being used to calculate a nociception index related to the nociception effect during a general anesthesia procedure, the present invention relates to a processing system, and to a monitoring device for calculating a nociception index related to the nociception effect during a general anesthesia procedure.

[0003] In this method, clinical data related to multiple previous anesthesia procedures are acquired during a training phase, separate from the actual use of the model during the anesthesia process. Training data is obtained from the clinical data. Additionally, reference data is obtained from the clinical data. After obtaining the training and reference data, the model is trained using the training data as input and the reference data as output, wherein the model is adjusted based on the training and reference data during training.

[0004] Anesthesia is defined as a drug-induced state in which a patient loses consciousness, loses the ability to feel pain, or becomes unresponsive to any other stimulus. To achieve these goals, anesthesiologists can use different types of drugs, primarily sedatives and analgesics, to enable patients to undergo surgery and other procedures without experiencing the pain and discomfort they would otherwise endure.

[0005] Analgesia is achieved through the administration of analgesics. The need for analgesics varies from patient to patient. Therefore, continuous, preferably non-invasive, monitoring of the analgesic effect is required. Nociception and pain perception define the need for analgesia to achieve pain relief. Automatic responses such as tachycardia, hypertension, emotional sweating, and tearing, while not specific, are considered signs of nociception and therefore inadequate analgesia.

[0006] When a sufficient dose of a hypnotic is administered, the resulting loss of consciousness ensures that the patient does not perceive stimuli, but the autonomic nervous system and bodily responses do not necessarily cease. When a sufficient dose of analgesics is administered, nociceptive stimuli are blocked, and the autonomic nervous system and bodily responses are inhibited. However, analgesics do not necessarily cause loss of consciousness or memory loss.

[0007] Anesthesia can generally be considered a dynamic process in which the effects of anesthetic drugs are counteracted by the varying intensities of stimuli occurring during surgery. When this balance is disrupted, the patient may develop different depths of anesthesia without the anesthesiologist's awareness, potentially leading to intraoperative consciousness. Therefore, one of the goals of modern anesthesia is to ensure an appropriate level of consciousness that prevents sensory overload without unconsciously overloading the patient, which can increase postoperative complications. Several widely used clinical methods exist for assessing levels of consciousness during general anesthesia, including the Observer's Assessment of Alertness and Sedation Scale (OAAS) and the Ramsey Sedation Scale. However, the disadvantages of using clinical scales in the operating room are that they cannot be used continuously and are cumbersome to administer. Furthermore, they require patient cooperation, which can be difficult in some cases. This has led to research into automated assessment of levels of consciousness.

[0008] Recently, several automated devices have become available to provide objective quantification of a patient's level of consciousness. The most popular method is the analysis of electroencephalography (EEG), in which scalp EEG is recorded and subsequently processed by algorithms that map the EEG signals to an exponent—typically in the range of 0 to 100. EEG-based methods are well-established for assessing brain activity by recording and analyzing peripheral biopotential signals generated in the cerebral cortex using electrodes attached to the skin on the surface of the skull. For decades, EEG-based methods have been widely used in basic research on the neurological system and in the clinical diagnosis of various neurophysiological diseases and conditions.

[0009] For example, the article “Monitoring hypnotic effect and nociception with two EEG-derived indices, qNOX and qCON, during general anesthesia” by EW Jensen et al., Scandinavian Journal of Anesthesiology, 2014; 58: 933-941, describes a method for calculating the nociception index related to the nociception effect during general anesthesia using models such as quadratic models or so-called ANFIS models (ANFIS stands for Adaptive Neuro Fuzzy Inference System).

[0010] A method for calculating the harm perception index is also disclosed in WO2017 / 012622 A1.

[0011] The nociceptive index qNOX is defined as representing the probability of a response to a noxious stimulus from low to high, and its value can be assumed to be between 0 and 100. To calculate the nociceptive index, a model is employed that uses EEG data as input and provides the value of the nociceptive index qNOX as output. The quality of the information provided by the nociceptive index in this paper depends on the quality of the model; therefore, the model is trained during the initial training phase before actual use using a large amount of clinical data, such as clinical data involving multiple anesthesia procedures performed on multiple patients (e.g., hundreds of patients). Training data, based on clinical data, particularly training data related to EEG data obtained during several anesthesia procedures, is used as input to the model, and reference data, also derived from the clinical data, is used as output. The model is adjusted during training such that, when using the input training data, the model reliably predicts values ​​at least close to the reference data.

[0012] Therefore, the quality of training depends on the quality of the reference data. Thus, it is necessary to obtain reference data that can be used during model training so that the model can accurately predict the nociceptive index when used later, where the nociceptive index reflects the probability of response to harmful stimuli during subsequent anesthesia.

[0013] The object of this invention is to provide a method and processing system for training a model that can be used to calculate a nociceptive index, the method and processing system being able to appropriately train the model using training data and reference data.

[0014] Another object of the present invention is to provide a monitoring device for calculating a nociceptive index related to nociceptive effects during general anesthesia using a model.

[0015] This objective is achieved by a method including the features of claim 1.

[0016] Therefore, the reference data is derived from clinical data using equations that include mathematical terms whose values ​​are non-linearly variable as a function of concentration values ​​related to drug concentrations in the patient during anesthesia.

[0017] In particular, the concentration value can be a function of an exponential function.

[0018] The mathematical term can be defined as a·f1(CeRemi), where a is a coefficient, f1 defines the function, and CeRemi is the concentration value.

[0019] Concentration values ​​can specifically refer to the site-effect concentration of remifentanil, which is used as an analgesic during anesthesia.

[0020] Clinical data are obtained from multiple prior anesthesia procedures. These prior procedures can be performed, for example, using target-controlled infusion (TCI) as described in WO2014 / 173558 A1. During such procedures, a large amount of data related to the anesthesia process is recorded, such as EEG data, data related to events (such as stimulating events during surgery), and concentration values ​​in different compartments of the patient (e.g., calculated in target-controlled infusion). In particular, effector site concentrations in the patient's brain are also obtained during anesthesia. Training and reference data are obtained using this previously obtained clinical data, which can be used to train the model.

[0021] In this paper, reference data should represent the actual index of harm perception that should be (precisely) output by the model when the corresponding training data is fed to it. The training data in this paper specifically refers to EEG data, as it will later be input into the model during the actual anesthesia procedure. Reference data is derived from additional data from clinical data associated with the EEG data.

[0022] To obtain reference data, it has been found in this paper that concentration values ​​related to drug concentrations in the patient's body, such as analgesic concentrations, particularly the site-effect concentration of analgesics in the patient's brain, should be considered in a non-linear manner by using a non-linear function for calculating the reference data. This function can be, in particular, an exponential function, causing low concentration values ​​to significantly increase the reference value, wherein the function decays non-linearly with increasing concentration values, such that large concentration values ​​contribute close to zero to the reference value.

[0023] By using a nonlinear function that takes concentration values ​​into account to obtain reference data, a more reliable reference can be obtained for training the model.

[0024] Specifically, the following equation can be used to calculate the reference data:

[0025] RD=a·f1(CeRemi)+b·qCON+f3(Resp)

[0026] Where RD is the reference data, a and b are coefficients, qCON is the consciousness index, and f3 is the function.

[0027] Resp is a patient response parameter. qCON is also calculated from clinical data using another model, as described in WO2017 / 012622A1. Resp is a parameter that reflects the intensity of movement as a stimulus outcome.

[0028] f3(Resp) can be calculated, for example, using the following equation:

[0029]

[0030] in

[0031]

[0032] To obtain reference data, the coefficients a and b can be iteratively set experimentally. For example, the coefficients can be initially set to a specific set of values, with the model trained using reference data obtained in this way. The coefficients can then be adjusted to improve the model's training. This can be iterated multiple times.

[0033] The clinical data involved multiple previous anesthesia procedures. In the anesthesia procedures discussed in this paper, concentration values ​​typically varied over time, reflecting the anesthesia process achieved through the infusion of appropriate anesthetics such as remifentanil or propofol at time-varying doses. For each anesthesia procedure for which clinical data was obtained, a reference curve over time could be calculated, allowing the model to be trained by applying the reference curves associated with the anesthesia procedure stored in the clinical data and the corresponding EEG data.

[0034] Each reference curve in this article can be scaled to fall within the range of 0 to 100 and / or smoothed by applying moving average techniques—particularly exponential moving averages—(e.g., using a factor of 0.9 for previous values ​​and a factor of 0.1 for new values).

[0035] The model used to calculate the nociception index can be a fuzzy logic model (particularly the so-called ANFIS model) or a quadratic model, such as those described in WO2017 / 012622A1. Specifically, the model can be defined using a system of equations for calculating the nociception index based on input data obtained from electroencephalogram (EEG) signals, where training data is used as input data during training, and the coefficients are adjusted to define the model. Once training is complete, the model is frozen and can be used by monitoring equipment in real time during actual anesthesia to calculate the nociception index during anesthesia, providing information about nociception during general anesthesia.

[0036] The processing system configured to execute software code implementing the methods described above can, for example, be separate from the monitoring device. Such a processing system could be, for example, a general-purpose computer system located at the site of the developer of the monitoring device, whereby, after the model is completed, it can be installed on the monitoring device and used on-site in the hospital during the anesthesia process to monitor the anesthesia procedure.

[0037] The monitoring device for calculating the nociceptive index related to nociceptive effects during general anesthesia includes a processor device configured to calculate the nociceptive index during the actual anesthesia procedure using a model and input data obtained from electroencephalogram (EEG) signals acquired during general anesthesia, wherein the value of the nociceptive index is obtained as output from the model. The processor device described herein is configured to modify the nociceptive index value obtained from the model using additional information obtained from the EEG signals to obtain a corrected value for the nociceptive index.

[0038] The model provides a measurement of the nociceptive index. In this paper, to further improve the reliability of the nociceptive index and its ability to predict the probability of a patient's response to stimuli during surgery, the nociceptive index value is corrected by considering additional information also obtained from the EEG signal after it is received from the model.

[0039] The processor device can be specifically configured to calculate the value of the nociceptive index in real time during the actual anesthesia procedure. The monitoring device may include, for example, a display on which the (corrected) value of the nociceptive index is displayed in real time during the anesthesia procedure to represent the actual probability of the patient's response to the stimulus.

[0040] In one implementation, the processor device is configured to modify the value of the nociceptive index obtained from the model by using information related to at least one of the electrooculogram obtained from the electroencephalogram signal, the burst inhibition rate obtained from the electroencephalogram signal, and the near burst inhibition index obtained from the electroencephalogram signal.

[0041] For example, an electrooculogram (EOG) can be used to calculate the amount (count) of peak values ​​above a specific amplitude threshold within a window of the EEG signal. The amount of EOG signal obtained from the EEG signal is correlated with the patient's level of consciousness, making high EOG values ​​usable for correcting for higher nociceptive index values.

[0042] The burst suppression ratio is associated with deep anesthesia and reflects the ratio of blank periods to rapid activity periods in the EEG signal. The burst suppression ratio can be used to correct the nociceptive index using the following equation:

[0043] qNOXcor=max(0,1-BSR / 30)*qNOX+min(1,BSR / 30)*(41-0.41 BSR)

[0044] Where qNOXcor is the correction value for the nociceptive index, max() and min() correspond to the maximum and minimum operators respectively, and BSR is the burst suppression rate. As can be observed, for BSR values ​​of 30 or greater, the value of the qNOX index depends only on the BSR parameter. In one implementation, BSR correction is limited to reducing the qNOX value. An increase in the qNOX value due to burst suppression is generally not expected.

[0045] The near burst suppression ratio (NBS) indicates whether an EEG exhibits a pattern similar to burst suppression, but the signal amplitude between bursts is insufficient to be considered a suppressed EEG. This pattern typically appears slightly before burst suppression is detected. The near burst suppression ratio can be calculated, for example, as the standard deviation of energy in the 11 Hz to 22 Hz band over a 10-second interval. In one implementation, the output of qNOX after BSR correction can be modified based on the near burst suppression parameter.

[0046] In one implementation, exponential correction is used to reduce the qNOX value during the period approaching burst suppression. NBS correction is limited to reducing the qNOX value. The correction applies to values ​​where the NBS parameter is higher than 0.8 and the EOG count is less than 60.

[0047] The value of the harm perception index can be corrected, for example, using the following formula based on the proximity burst suppression index:

[0048] qNOX = exp(a + b * qNOX - cNBS)

[0049] Where cNBS is the near burst suppression index, and a and b are experimentally defined coefficients.

[0050] In one implementation, in addition to modifying the nociceptive index (as the model output) by taking into account information obtained from the EEG signal, the value of the nociceptive index can be further modified by applying at least one of a scaling operation and a smoothing operation. Within a scaling range, the value of the nociceptive index is scaled to fall within, for example, a range between 25 and 100. This scaling can be performed, for example, using the following formula system:

[0051] qNOX = 1.3751 * qNOX - 28.7514 for 80 ≤ qNOX < 94

[0052] qNOX = 0.4 * qNOX + 21 for 15 ≤ qNOX < 35

[0053] qNOX = 0.1333 * qNOX + 25 (for qNOX, the value is 15)

[0054] The model used to calculate the injury perception index can be, in particular, a model defined in the initial training phase using the methods described above for training models.

[0055] The concepts underlying the present invention will then be described in more detail with reference to the embodiments shown in the accompanying drawings. In this document:

[0056] Figure 1 A schematic diagram of the setup during anesthesia is shown;

[0057] Figure 2 It shows Figure 1 The function diagram of the settings;

[0058] Figure 3 A functional diagram of a model used to model drug dose distribution in a patient's body is shown.

[0059] Figure 4 A schematic diagram of the model used to calculate the consciousness index (qCON) is shown;

[0060] Figure 5 A schematic diagram of the model used to calculate the nociceptive index (qNOX) is shown;

[0061] Figure 6 A schematic diagram is shown for correcting the value of the harm perception index output by the model;

[0062] Figure 7 A schematic diagram is shown during the training of the model used to calculate the nociceptive index;

[0063] Figure 8 A graph of the function used to take concentration values ​​into account to derive reference data from clinical data is shown;

[0064] Figure 9 A schematic diagram of a monitoring device using a model for calculating the awareness index and the harm perception index is shown.

[0065] Figure 10 A quality analysis graph showing the nociceptive index and concentration during anesthesia is shown; and

[0066] Figure 11A , Figure 11B The mathematical formula for the ANFIS nonlinear model is shown.

[0067] Subsequently, methods and processing systems for defining a model—which can be used to calculate a nociceptive index related to nociceptive effects during general anesthesia—will be described in some embodiments, as well as monitoring equipment for calculating the nociceptive index during actual anesthesia procedures. The embodiments described herein should not be construed as limiting the scope of the invention.

[0068] Use the same reference numerals in all the accompanying figures as needed.

[0069] Models for calculating the nociceptive index are commonly used during general anesthesia, such as target-controlled infusion (TCI), and are subsequently used based on... Figures 1 to 3 A description is provided. However, using models, the nociceptive index can typically be calculated from EEG input data during any general anesthesia procedure.

[0070] Figure 1 A schematic diagram of a setup is shown, typically used for anesthesia, such as administering anesthetic drugs like propofol and / or remifentanil to patient P. In this setup, multiple devices are arranged on a support 1 and connected to patient P via different lines.

[0071] Specifically, infusion devices 31, 32, and 33, such as infusion pumps, particularly syringe pumps or volumetric pumps, are connected to patient P and used to intravenously inject different drugs (e.g., propofol, remifentanil, and / or muscle relaxants) into patient P through lines 310, 320, and 330 to achieve the desired anesthetic effect. Lines 310, 320, and 330 are connected, for example, to a single port providing access to the venous system of patient P, allowing the corresponding drug to be injected into the patient's venous system through lines 310, 320, and 330.

[0072] The stent 1 also holds the ventilation device 4 for providing artificial respiration to patient P when patient P is under anesthesia. The ventilation device 4 is connected to the mouthpiece 40 via tubing 400, thereby connecting the ventilation device 4 to patient P's respiratory system.

[0073] The stent 1 also holds an EEG monitor 5, which is connected via a tubing or tubing bundle 500 to an electrode 50 attached to the patient’s head for monitoring brain activity during anesthesia.

[0074] Additionally, the control device 2 is held by the bracket 1 and includes a measuring device 20 connected via a tubing 200 to a connector 41 of the mouthpiece 40. The control device 2 is used to control the infusion operation of one or more of the infusion devices 31, 32, and 33 during anesthesia, so that the infusion devices 31, 32, and 33 administer anesthetic drugs to the patient P in a controlled manner to achieve the desired anesthetic effect. This will be explained in more detail below.

[0075] Measuring device 20 is used to measure the concentration of one or more anesthetic drugs in the breath of patient P. For example, measuring device 20 can measure the concentration of propofol in the exhalation of patient P. For this purpose, measuring device 20 can measure continuously, for example, over a predetermined number of respiratory cycles (inspiration and exhalation), such as six respiratory cycles, so that the measured concentration in the breath of patient P can be appropriately averaged over the respiratory cycles. Alternatively, measuring device 20 can also measure the concentration of, for example, propofol only during the exhalation phase, where a suitable triggering mechanism can be used to trigger the measurement, or measuring device 20 can measure the propofol concentration continuously.

[0076] Control device 2 may be adapted to provide information about drug concentrations measured in the patient P's respiration or in other compartments of the patient P, or about the drug's effects in the patient's brain compartment. Such information may be output via monitor 6 attached to support 1, allowing personnel such as anesthesiologists to monitor drug concentrations and related effects in the patient P during anesthesia.

[0077] Figure 2 A functional diagram of the control loop for controlling the infusion operations of infusion devices 31, 32, and 33 during anesthesia is shown. This control loop can, in principle, be configured as a closed loop, where the operation of infusion devices 31, 32, and 33 is automatically controlled without user interaction. However, advantageously, the system is configured as an open-loop system, where user interaction is required at specific times, particularly before administering the drug dose to the patient, to manually confirm the operation.

[0078] Control device 2—also referred to as "infusion manager"—is connected to support 1, which serves as a communication link to infusion devices 31, 32, and 33, which are also attached to support 1. Control device 2 outputs control signals to control the operation of infusion devices 31, 32, and 33, which inject a predetermined dose of medication into patient P according to the received control signals.

[0079] EEG readings of patient P are acquired via EEG monitor 5, and the concentrations of one or more drugs in patient P's respiration are measured via measuring device 20. The measurement data obtained by EEG monitor 5 and measuring device 20 are fed back to control device 2, which adjusts its control operation accordingly and outputs modified control signals to infusion devices 31, 32, and 33 to achieve the desired anesthetic effect.

[0080] The measuring device 20 may, for example, consist of a so-called IMS monitor, used to measure the drug concentration in the patient P's breath via so-called ion mobility spectrometry. Other sensor technologies may also be used.

[0081] Control device 2 uses a pharmacokinetic-pharmacodynamic (PK / PD) model to control the infusion operation of one or more infusion devices 31, 32, 33. This pharmacokinetic-pharmacodynamic model is a pharmacological model used to model the processes by which a drug acts on a patient P. These processes include the reabsorption, distribution, biochemical metabolism, and excretion of the drug within the patient P (referred to as pharmacokinetics) and the drug's effects on the organism (referred to as pharmacodynamics). Preferably, a physiological PK / PD model with N compartments is used, where the transfer rate coefficient has been experimentally measured beforehand (e.g., in a proband study) and is therefore known. To simplify the PK / PD model, it is preferable to use no more than four to five compartments.

[0082] Figure 3 A schematic functional diagram of the setup of this PK / PD model p is shown. Logically, the PK / PD model p divides the patient P into different compartments A1 to A5, such as a plasma compartment A1 corresponding to the patient P's blood flow, a lung compartment A2 corresponding to the patient P's lungs, a brain compartment A3 corresponding to the patient P's brain, and other compartments A4 and A5 corresponding to, for example, muscle tissue or fat and connective tissue. The PK / PD model p considers the volume V of the different compartments A1 to A5. Lung V plasma V brain V i V j and the transfer rate constant K, which represents the transfer rate between plasma compartment A1 and other compartments A2 to A5. PL K LP K BP K PB K IP ;K PI K JP K PJ Assume that drug dose D is injected into plasma compartment A1 via infusion device 33, and plasma compartment A1 connects to other compartments A2 to A5 such that exchanges between compartments A2 to A5 always occur through plasma compartment A1. The PK / PD model p is used to predict the concentration of injected drug Clun in different compartments A1 to A5 as a function of time. g C p lasma, Cbrain, Ci, Cj.

[0083] During general anesthesia, for example, through the use of control device 2 and control in the sense of target-controlled infusion (TCI) as described above, it is generally desirable to provide an accurate assessment of the patient's anesthetic status. Therefore, based on information obtained during the general anesthesia process, particularly EEG signals obtained from EEG monitor 5, an index reflecting the patient's level of consciousness and level of nociceptive perception during the anesthesia process should be calculated.

[0084] This is Figure 4 and Figure 5 The diagram illustrates this schematically. Specifically, based on EEG signals obtained during general anesthesia, the first model M1 is used to calculate the consciousness index qCON (…). Figure 4 ), and the second model M2 is used to calculate the harm perception index qNOX ( Figure 5 (e.g., as described in WO2017 / 012622A1). Therefore, during general anesthesia, models M1 and M2 are used to calculate the consciousness index qCON, reflecting the level of consciousness, and the nociceptive index qNOX, reflecting the probability of response to noxious stimuli, from the input EEG signal. Both indices typically have values ​​ranging from 0 to 100 (where higher values ​​represent increased consciousness and nociceptive perception, respectively).

[0085] To calculate the nociceptive index qNOX, EEG data is fed into model M2, and the value of the nociceptive index qNOX is obtained as the output of model M2. This paper also proposes using a correction function C to correct the value of the nociceptive index qNOX, obtaining a corrected value qNOXcor for the nociceptive index.

[0086] Now refer to Figure 6 To correct the injury perception index, different correction functions C1 to C5 can be used.

[0087] That is, in the first correction function C1, scaling can be used to scale the value of the nociceptive index to a value between 25 and 99, therefore values ​​below 25 are not allowed. For scaling, the following equation can be used, for example:

[0088] qNOX = 1.3751 * qNOX - 28.7514 for 80 ≤ qNOX < 94

[0089] qNOX = 0.4 * qNOX + 21 for 15 ≤ qNOX < 35

[0090] qNOX = 0.1333 * qNOX + 25 for qNOX < 15

[0091] In the second correction function C2, correction can be performed based on the so-called electrooculographic EOG, which calculates the EOG count from the EEG signal as a peak value above a specific threshold within a window of the EEG signal. A large EOG count can indicate increased alertness, which can be used to correct the nociceptive index.

[0092] In the third correction function C3, the burst suppression rate, representing the state of deep anesthesia, can be considered to specifically correct the nociceptive index to a lower value. Within the correction function C3, the following equation can be specifically used:

[0093] qNOXcor=max(0,1-BSR / 30)*qNOX+min(1,BSR / 30)*(40-0.41 BSR)

[0094] Where qNOXcor is the correction value for the nociceptive index, max() and min() correspond to the maximum and minimum operators respectively, and BSR is the burst suppression rate. As can be observed, for BSR values ​​of 30 or greater, the value of the qNOX index depends only on the BSR parameter. In one implementation, BSR correction is limited to reducing the qNOX value. An increase in the qNOX value due to burst suppression is generally not expected.

[0095] In the fourth correction function C4, the proximity burst suppression index (NBS) can be used to correct the nociceptive index value. The proximity burst suppression index indicates whether the EEG exhibits a pattern similar to burst suppression, but the signal amplitude between bursts is insufficient to be considered a suppressed EEG. This pattern typically appears slightly before burst suppression is detected. The proximity burst suppression index can be calculated, for example, as the standard deviation of energy in the 11Hz to 22Hz frequency band over a 10-second interval. In one implementation, the output of qNOX after BSR correction can be modified based on the proximity burst suppression parameter.

[0096] For example, the following formula can be used to reduce the qNOX value by applying exponential correction during the period close to the burst suppression period:

[0097] qNOX = exp(a + b * qNOX - cNBS)

[0098] Where cNBS is the actual value close to the burst suppression index, and a and b are experimentally defined coefficients.

[0099] In the fifth correction function C5, smoothing can be used, for example, based on the signal quality index SQI. This smoothing can be performed, for example, by a moving average.

[0100] The model M2 used to calculate the nociceptive index can be, for example, a quadratic model or a fuzzy logic model, particularly an ANFIS model, which will be described in more detail later with different examples. Typically, model M2 can be represented by a system of equations containing multiple coefficients that are appropriately defined during the initial training phase by training model M2 so that when input data from EEG signals is fed, model M2 reliably provides the output of the nociceptive index, such as... Figure 5 As shown.

[0101] like Figure 7 As shown, in this paper, training is usually performed by using training data TD as input and reference data RD as output. Model M2 is fitted to the training data TD and reference data RD during the adjustment process using a large amount of clinical data used to obtain the training data TD and reference data RD.

[0102] Clinical data is typically obtained during anesthesia, for example, by using the methods described above. Figures 1 to 3 The described target-controlled infusion (TCI) procedure for general anesthesia aims to obtain clinical data during the anesthesia process. During this procedure, data is recorded in relation to, for example, EEG signals, concentration values ​​in different compartments obtained from a PK / PD model, events occurring during the procedure such as surgical events (e.g., the timing of incisions indicating stimulation), the patient's status such as signs of awakening, and all other data available during the anesthesia process. Therefore, in clinical data, EEG signals are correlated with the actual drug concentrations in the patient's different compartments (particularly at the patient's effector site, typically the brain) and the patient's status and response to stimuli occurring during the procedure.

[0103] Therefore, the target output that model M2 should produce can be obtained from clinical data as a reference for the value of the nociceptive index.

[0104] This paper proposes using the following equation to obtain reference data for the harm perception index used to train the model:

[0105] RD=a·f1(CeRemi)+b·qCON+f3(Resp)

[0106] Where a and b are coefficients, f1 is an exponential function, CeRemi is the concentration at the site of effect of the infused anesthetic (e.g., remifentanil), qCON is the consciousness index, f3 is a function, and Resp is the patient response parameter.

[0107] Figure 8 The exponential function f1 is shown.

[0108] qCON is also calculated from clinical data using model M1, as described in WO2017 / 012622A1, for example.

[0109] Resp is a parameter that reflects the intensity of motion as a stimulus outcome.

[0110] f3(Resp) can be calculated, for example, using the following equation:

[0111]

[0112] in

[0113]

[0114] Alternatively, the reference data RD can be defined using the following equation:

[0115] Ref = ARef + RqCON

[0116] In this paper, ARef represents the difference between the maximum drug dose and the actual drug dose at each time interval. RqCON is derived from the qCON exponent, where RqCON is set to 0 for values ​​between 0 and 20, 1 for values ​​between 20 and 40, 2 for values ​​between 40 and 60, 3 for values ​​between 60 and 80, and 4 for values ​​between 80 and 100.

[0117] In addition, this paper may consider events of noxious stimulation. If a noxious stimulation occurs and causes the patient to move, the reference value is set to 10; otherwise, the reference value remains at the value provided by the above equation.

[0118] To obtain a reference, the value obtained in this way is scaled to the range between 0 and 100 by multiplying the obtained value by a factor of 10.

[0119] The reference data RD obtained in this way is used to train model M2. In this paper, the reference data RD is typically provided as a time-varying reference curve, each curve relating to a specific anesthesia procedure and to EEG data acquired during that procedure. Therefore, the training data is derived from the corresponding EEG data, and the model is trained such that model M2 predicts the reference data RD (at least approximately) when fed the training data TD. This is applied to large amounts of clinical data involving, for example, hundreds of patients, so that model M2 is fitted to reliably predict the value of the nociceptive index when fed input data derived from EEG signals.

[0120] Once training is complete, model M2 is frozen and mounted on monitoring device 7 for use in actual anesthesia procedures, such as... Figure 1 and Figure 2The setup is shown below. Monitoring device 7 is functionally connected to EEG monitor 5, so it can receive EEG data as input during the actual anesthesia process, enabling the nociceptive index qNOX to be calculated in real time during the anesthesia process.

[0121] like Figure 9 As shown, the monitoring device 7 includes a processor device 17 and a memory device 71. A model M1 for calculating the consciousness index qCON and a model M2 for calculating the nociceptive index qNOX are stored in the memory device 71 and used to calculate the corresponding indices in real time during the anesthesia process. The monitoring device 7 includes, for example, a display device 72, so that the corresponding indices can be displayed to the user in real time.

[0122] Figure 10 The curves showing the nociceptive index qNOX over time and the corresponding curves showing drug concentrations at the effector site are illustrated, for example, the concentration of remifentanil in the patient's brain during TCI anesthesia calculated, for example, using a PK / PD model. When reference data is derived from clinical data, the reference data can have, for example, the concentration of remifentanil in the patient's brain calculated using a PK / PD model. Figure 10 The shape of the qNOX curve shown.

[0123] The training of the nonlinear model is advantageously performed using a large amount of clinical data obtained during multiple previous anesthesia procedures, particularly TCI anesthesia procedures. The training, completed in the initial preprocessing step, defines the parameters of the model, and then, after training and freezing the model, when the input is presented to the model, it can predict various indices during the actual anesthesia procedure.

[0124] As mentioned above, for processing, a nonlinear model with the shape of a fuzzy logic model or a quadratic equation model can be used. However, other nonlinear models can also be used.

[0125] The following text provides details about the ANFIS model and the quadratic equation model through examples.

[0126] ANFIS model :

[0127] The fuzzy logic model can be, for example, a so-called ANFIS model. In this case, the system uses an ANFIS model to combine parameters to define the qCON and qNOX exponents. The parameters extracted from the EEG signal are used as input to the Adaptive Neural Fuzzy Inference System (ANFIS).

[0128] ANFIS is a hybrid of fuzzy logic systems and neural networks. ANFIS does not employ any mathematical functions to control the relationship between inputs and outputs. Instead, it uses a data-driven approach, where training data determines the system's behavior.

[0129] Figure 11A and Figure 11B The five layers of ANFIS shown have the following functions:

[0130] - Each cell in layer 1 stores three parameters to define the bell-shaped membership function. Each cell is connected to exactly one input cell and the membership degree of the obtained input value is calculated.

[0131] - In layer 2, each rule is represented by a unit. Each unit is connected to those units from the premises of the rule in the previous layer. The input to the unit is the membership degree, which is multiplied to determine the degree to which the represented rule is implemented.

[0132] - In layer 3, for each rule, there exists a unit that calculates its relative degree of realization using a normalization equation. Each unit is connected to all rule units in layer 2.

[0133] The cells in layer 4 are connected to all input cells and exactly one cell in layer 3. Each cell computes the output of the rule.

[0134] The output unit in layer 5 calculates the final output by summing all the outputs from layer 4.

[0135] The standard learning process from neural network theory is applied to ANFIS. Backpropagation is used to learn the premise parameters, i.e., membership functions, and least squares estimation is used to determine the coefficients of the linear combination in the results of the rules. The learning process has two paths. In the first path, the forward path, the input pattern is propagated, and the optimal result parameters are estimated through an iterative least mean square process, while the premise parameters are fixed for the current loop using the training set. In the second path (backward path), the pattern is propagated again, and in this path, backpropagation is used to modify the premise parameters while the result parameters remain fixed. The process is then iterated over the desired number of epochs. If the premise parameters are appropriately chosen initially based on expert knowledge, one epoch is usually sufficient because the LMS algorithm determines the optimal result parameters in one path, and if the premise is not significantly changed by using gradient descent, the LMS calculation of the result will not produce another result. For example, in a 2-input, 2-rule system, rule 1 is defined as:

[0136] If x is A and y is B, then f I =p I x+q I y+r I

[0137] Where p, q, and r are linear, they are called outcome parameters or outcome only. Because higher-order sugeno fuzzy models introduce enormous complexity and offer almost no obvious advantages, the first-order f is the most common.

[0138] The input to the ANFIS system is fuzzified into multiple predetermined classes. The number of classes should be greater than or equal to two. The number of classes can be determined using different methods. In traditional fuzzy logic, classes are defined by experts. This method can be applied if it is obvious to the experts where the boundary markers between two classes can be placed. ANFIS optimizes the location of the boundary markers; however, if the initial values ​​of the parameters constraining the classes are close to their optimal values, the gradient descent method will reach its minimum more quickly. By default, the initial ANFIS boundary markers are selected by dividing all the data from minimum to maximum intervals into n equidistant intervals, where n is the number of classes. The number of classes can also be selected by plotting the data in a histogram and by visually determining the appropriate number of classes through various clustering methods or Markov models using sorting methods such as those performed by FIR. ANFIS is chosen by default for this invention, and it is shown that more than three classes during the validation phase lead to instability; therefore, two or three classes are used.

[0139] The number of classes and the number of inputs both increase the complexity of the model, i.e., the number of parameters. For example, in a system with four inputs, each input can be fuzzified into three classes, consisting of 36 premise (non-linear) parameters and 405 outcome (linear) parameters, which can be calculated using the following two formulas:

[0140] Prerequisite = Number of classes × Number of inputs × 3

[0141] Result = Number of classes / Number of inputs × (Number of inputs + 1)

[0142] The number of input-output pairs should typically be much larger than the number of parameters (at least 10 times) to obtain meaningful solutions for the parameters.

[0143] A useful tool for ensuring stability is experience gained through the following: working with a neurofuzzy system such as ANFIS on a specific dataset and testing with extreme data, for example, obtained through simulation.

[0144] ANFIS uses the root mean square error (RMSE) to validate training results, and the RMSE validation error can be calculated from a set of validation data after each training epoch. An epoch is defined as one update to both the premise parameters and the outcome parameters. Increasing the number of epochs generally reduces the training error.

[0145] quadratic model

[0146] Alternatively, a quadratic equation model can be used for models M1 and M2. In this case, the system uses a quadratic model to combine the parameters used to define the qCON and qNOX exponents. The parameters extracted from the EEG signal are used as inputs to the quadratic model.

[0147] The output exponent is derived from a quadratic generalized model that uses data extracted from EEG as input. Such a model includes: independent coefficients called the intercept, a linear term for each input, a squared term for each input, and interaction terms between each pair of entries. The model can be represented as:

[0148]

[0149] in:

[0150] Intercept: intersection point or constant term.

[0151] Input: Input model.

[0152] Output: Model output.

[0153] n: The number of inputs to the model.

[0154] a: Linear term.

[0155] b: Square term

[0156] c: Interaction items between inputs.

[0157] List of reference numerals

[0158] 1. Bracket

[0159] 2. Control equipment

[0160] 20 Measuring equipment

[0161] 200 pipeline

[0162] Infusion equipment 31, 32, 33

[0163] 310, 320, 330 pipelines

[0164] 4. Ventilation equipment

[0165] 40 Mouth Components

[0166] 400 pipeline

[0167] 41 Connector

[0168] 5 EGG monitors

[0169] 50 electrodes

[0170] 500 pipeline

[0171] 6 Display devices

[0172] 7. Monitoring equipment

[0173] 70 Processor Devices

[0174] 71 Memory devices

[0175] 72 Display devices

[0176] A1-A5 compartments

[0177] C Correction Mode

[0178] D. Drug dosage

[0179] M1 and M2 models

[0180] p model

[0181] Patient P

[0182] qCON Consciousness Index

[0183] qNOX Harm Perception Index

[0184] qNOXcor is the corrected qNOX exponent.

[0185] RD Reference Data

[0186] TD training data

Claims

1. A method for training a model (M2) that can be used to calculate a nociceptive index (qNOX) related to nociceptive effects during general anesthesia, the method comprising: Clinical data related to multiple previous anesthesia procedures were obtained during a training phase separate from the actual use of the model (M2) during the anesthesia procedure. Training data (TD) is obtained from the clinical data. Reference data (RD) were obtained from the clinical data, and The model (M2) is trained using the training data (TD) as input data and the reference data (RD) as output data, wherein the training includes adjusting the model (M2) based on the training data (TD) and the reference data (RD). Its features are, The reference data (RD) is obtained using mathematical terms. The equation The mathematical term derived from the clinical data The value of is non-linearly variable as a function of the concentration of the drug in the patient's body during anesthesia. Wherein, the function of the concentration value is an exponential function, such that the low value of the concentration value significantly increases the reference value, wherein, It is a coefficient. Define the function, It is the concentration value, where, It is a coefficient. It is a consciousness index. It is a function. These are patient response parameters, and the model (M2) includes a set of coefficients for calculating the nociceptive index (qNOX) based on input data derived from electroencephalogram (EEG) signals, wherein the coefficients are adjusted during the training period to define the model.

2. The method according to claim 1, characterized in that, The concentration value varies over time within a specific anesthetic procedure from among the multiple previous anesthetic procedures.

3. The method according to claim 1 or 2, characterized in that, In order to obtain the reference data (RD), a reference curve is calculated over time for at least one subset of the plurality of previous anesthesia procedures.

4. The method according to claim 3, characterized in that, The reference curve is subjected to at least one of the following: scaling to a range between 0 and 100; and smoothing by applying a moving average technique.

5. The method according to claim 1 or 2, characterized in that, The model is a fuzzy logic model or a quadratic equation model.

6. The method according to claim 3, characterized in that, The model is a fuzzy logic model or a quadratic equation model.

7. The method according to claim 4, characterized in that, The model is a fuzzy logic model or a quadratic equation model.

8. A processing system configured to execute software code for implementing the method of any one of the preceding claims.

9. A monitoring device (7) for calculating a nociceptive index (qNOX) related to nociceptive effects during general anesthesia, said monitoring device (7) comprising: A processor device (70) is configured to calculate the nociceptive index (qNOX) during the actual anesthesia process using a model (M2) and input data obtained from electroencephalogram (EEG) signals acquired during the general anesthesia process, wherein the value of the nociceptive index (qNOX) is obtained as an output from the model (M2). Its features are, The processor device (70) is configured to modify the value of the nociceptive index (qNOX) obtained from the model (M2) using information related to at least one of the electrooculogram (EEG) obtained from the EEG signal, the burst inhibition rate obtained from the EEG signal, and the near burst inhibition index obtained from the EEG signal, to obtain a corrected value for the nociceptive index (qNOX), wherein the model (M2) is defined by the method of any one of claims 1 to 7 during a training phase prior to the actual anesthesia procedure.

10. The monitoring device (7) according to claim 9, characterized in that, The processor device (70) is configured to calculate the value of the nociceptive index (qNOX) in real time during the actual anesthesia process.

11. The monitoring device (7) according to claim 9 or 10, characterized in that, The processor device (70) is configured to modify the value of the injury perception index (qNOX) obtained from the model (M2) by applying at least one of a scaling operation and a smoothing operation.

Images