Consciousness disorder course monitoring system

Abstract

The present invention relates to a consciousness disorder course monitoring system. A multi-dimensional data acquisition module is used to collect the patient's electroencephalogram (EEG), heart rate variability (HRV), and skin conductance (EDA) physiological signal data in real time through multiple sensors arranged on the patient's body, and pre-process the collected data through an edge computing device; an adaptive course prediction model is based on the physiological signal data collected by the multi-dimensional data acquisition module, and combined with the patient's historical course data, uses an adaptive algorithm to train a personalized course prediction model, and in the course of the patient's disease, the model parameters are adaptively updated through a machine learning algorithm to achieve dynamic prediction of the course of consciousness disorder; a continuous course monitoring module continuously obtains the patient's physiological signals through high-frequency sampling technology, and identifies key events in the course of the disease in real time through an abnormality detection algorithm, including sudden decrease in consciousness and epileptic seizures, generates early warning signals, and triggers an emergency response mechanism.

Description

Technical Field

[0001] The present invention relates to a disease course monitoring system, in particular to a consciousness disorder disease course monitoring system. Background Art

[0002] The current technology used for the course monitoring system of consciousness disorders, such as the system described in China Invention Publication No. CN110269611A, has made significant progress in EEG signal acquisition and analysis, but there are still some deficiencies and disadvantages in practical applications. First, in terms of the accuracy of data processing, although the invention of CN110269611A can collect human EEG signals through EEG electrodes and filter and digitize them, the core of its data analysis still relies on preset filters and traditional spectrum analysis methods. These methods have limitations in processing non-stationary and nonlinear EEG signals. EEG signals are highly dynamic and complex, and relying solely on fixed filtering processing may not fully capture the key features in the signal, especially in the early stages of patients' consciousness disorders or when the condition changes rapidly. For some nonlinear features, including small changes in electrical activity before epileptic seizures, traditional spectrum analysis may not be able to effectively detect, which can lead to misjudgment or missed judgment. In addition, the resolution and accuracy of digital processing are limited by the limitations of hardware equipment and sampling rate, which will lead to the loss of signal details or deviation of analysis results, affecting the accurate assessment of the patient's condition.

[0003] Secondly, the monitoring system in CN110269611A mainly relies on a single EEG signal. Although this single-modal signal analysis method can provide a certain degree of disease monitoring, it ignores the impact of other key physiological signals such as heart rate variability HRV and skin conductance EDA on the patient's overall condition. Consciousness disorders are usually accompanied by complex multi-system changes, and a single EEG signal cannot fully reflect the patient's health status. HRV signals can provide information on the function of the autonomic nervous system, while EDA signals are closely related to emotions, stress, and sympathetic nerve activity. The comprehensive analysis of these signals can more accurately reflect the patient's physiological state and level of consciousness. The system in CN110269611A lacks the ability to integrate and analyze multimodal physiological signals, and cannot comprehensively monitor the patient's overall health status, thereby affecting the accuracy of disease judgment and the formulation of treatment plans.

[0004] Finally, the overall intelligence level of the system still has room for improvement. Although the system of CN110269611A has achieved continuous monitoring and early warning of the patient's degree of consciousness disorder to a certain extent, it relies on traditional data analysis methods and early warning mechanisms and fails to make full use of modern artificial intelligence technologies, including machine learning, deep learning, and Bayesian networks. These technologies can better process large-scale, multimodal data, identify complex disease patterns, predict disease trends, and optimize decision-making processes based on real-time data. The lack of intelligence means that the system has certain limitations in processing complex data and responding to emergencies. It is impossible to give full play to the potential value of the data and it is difficult to achieve precision medicine in the true sense. Summary of the invention

[0005] The purpose of the present invention is to provide a system for monitoring the course of consciousness disorders, thereby solving some of the drawbacks and shortcomings pointed out in the background technology.

[0006] The present invention solves the above-mentioned technical problems by adopting the following technical solutions:

[0007] The consciousness disorder course monitoring system includes a multi-dimensional data acquisition module, an adaptive course prediction model, a continuous course monitoring module, and a course phase distribution and state transition model:

[0008] S1. The multi-dimensional data acquisition module is used to collect the patient's electroencephalogram (EEG), heart rate variability (HRV), and skin conductance (EDA) physiological signal data in real time through multiple sensors set on the patient's body, and pre-process the collected data through edge computing equipment;

[0009] S2. The adaptive disease course prediction model is based on the physiological signal data collected by the multi-dimensional data acquisition module, and combined with the patient's historical disease course data, uses an adaptive algorithm to train a personalized disease course prediction model. During the course of the patient's disease, the model parameters are adaptively updated through a machine learning algorithm to achieve dynamic prediction of the progression of the course of consciousness disorder;

[0010] S3. The continuous course of disease monitoring module continuously acquires the patient's physiological signals through high-frequency sampling technology to ensure the continuity and integrity of course of disease monitoring, and identifies key events in the course of disease in real time through anomaly detection algorithms, including sudden loss of consciousness and epileptic seizures, generates early warning signals, and triggers emergency response mechanisms;

[0011] S4. The disease course stage distribution and state transition model is obtained by analyzing the disease course data obtained by the continuous disease course monitoring module.

[0012] Furthermore, the EEG acquisition utilizes high-density electrode array technology to obtain electrical activity waveforms through head-mounted EEG sensors, heart rate variability HRV is acquired through chest or wrist sensors, and skin conductance EDA is acquired through palm or finger sensors. The overall description of the signal can be expressed as:

[0013]

[0014] Among them, Φ(t) is the overall fusion of multimodal signals, α n and κ n Represent the amplitude of EEG and EDA, ω n and are their angular frequencies, φ n and θ n are phase, β n is the time decay factor, which comprehensively represents the changes in physiological signals of the state of consciousness;

[0015] The edge computing device processes the multimodal data in real time, extracts characteristic signals related to the state of consciousness, and performs abnormal waveform detection.

[0016] Furthermore, the initial construction of the model using the adaptive algorithm is based on the patient's baseline data and historical disease course data, and an adaptive filtering algorithm is applied to determine the initial state of the model through the following formula:

[0017]

[0018] Among them, Ψ(t) is the output signal of the personalized disease course prediction model, Γ i (s) is the weight function of the ith physiological signal, Φ i (s) is the adaptive coefficient related to the signal, T i is the periodic parameter of the signal, φ i is the phase, ∈ i (s) is the noise term, α i is the quadratic time coefficient related to signal attenuation;

[0019] During the course of the patient's disease, the dynamic parameter update module collects and inputs new physiological data in real time, minimizes the prediction error, and dynamically adjusts the model parameters using a variable step function and nonlinear least squares optimization method;

[0020] Finally, by integrating deep learning techniques, including convolutional neural network (CNN) and long short-term memory (LSTM), the model is used for feature extraction and time series analysis.

[0021] Furthermore, the anomaly detection algorithm is based on multi-level feature extraction. By introducing the multi-level feature extraction algorithm in the edge computing device, the continuously collected physiological signals are analyzed in real time. Specifically, the following steps are included:

[0022] S1. Use wavelet transform and Fourier transform to extract feature information of different frequency bands. Its mathematical expression is:

[0023]

[0024] Among them, Θ(f,t) is the extracted feature signal, x(t′) is the input original physiological signal, is the wavelet function, s(f) is the frequency-related scale parameter, and α is the time decay factor;

[0025] S2, the abnormal event recognition is to analyze the extracted characteristic signals in real time through the convolutional neural network CNN and the isolation forest algorithm IsolationForest to identify key abnormal events;

[0026] S3. The emergency response mechanism adopts the Bayesian Updating method to dynamically adjust the threshold of anomaly detection according to real-time data. The formula is expressed as:

[0027]

[0028] Among them, τ(t) is the adaptive threshold at time t, γ is the learning rate, μ and σ are the mean and standard deviation of the real-time data respectively;

[0029] Furthermore, the warning signal generation and emergency response mechanism generates multi-level warning signals based on the severity of abnormal events. The generation process of the warning signal is expressed as:

[0030]

[0031] Among them, S(t) is the warning signal strength at time t, δ k is the warning weight coefficient, Θ k (t) is the kth characteristic signal, I{·} is the indicator function, when the characteristic signal exceeds the adaptive threshold τ k The warning is triggered.

[0032] Furthermore, the disease course stage distribution and state transition model includes a disease course stage division module, a state transition model and a disease course prediction and clinical decision support module, and is characterized by:

[0033] S1. The disease course stage division module dynamically divides the disease course into multiple stages including stable stage, deterioration stage and recovery stage based on the multimodal physiological signal data collected in real time and using the time series analysis method;

[0034] S2. The state transition model analyzes the state transition probability between each disease course stage through the hidden Markov model HMM to deal with the uncertainty in the disease course data; through Bayesian network modeling, the correlation between the stages is updated according to the real-time data to optimize the disease course prediction;

[0035] S3. The disease course prediction and clinical decision support module provides personalized disease course prediction based on the analysis of the state transition model and provides real-time decision support for clinicians.

[0036] Furthermore, the disease stage division module includes an adaptive time series analysis module, a multi-level feature recognition module and a dynamic stage division module, and is characterized in that:

[0037] The adaptive time series analysis module uses a dynamic analysis method based on the physiological signals collected in real time to identify the time variation pattern of the physiological signals in real time; through the adaptive function, it is expressed as:

[0038]

[0039] Where F(t) is the output signal of time series analysis, α(s) and β(s) are adaptive weight functions, x(s) is the input physiological signal, ω is the frequency parameter, and λ is the time attenuation factor;

[0040] The multi-level feature recognition algorithm identifies the key features of various physiological signals such as spectral changes in EEG signals, time domain and frequency domain features in HRV, and conductance fluctuations in EDA through the following formula:

[0041]

[0042] Among them, G(t) is the output signal of multi-level feature recognition, γ i (s) is the weight function of the i-th layer, f i (x(s)) is the feature extraction function of the i-th layer, θ i and φ i is the phase and frequency offset, μ i is the time attenuation factor of the characteristic signal;

[0043] Based on the results of time series analysis and multi-level feature recognition, the disease course stages are divided by the following formula:

[0044]

[0045] Among them, D(t) is the result of dynamic stage division, F j (t) and G j (t) are the output signals of time series analysis and feature recognition, respectively, k is the penalty coefficient of the kth stage.

[0046] Furthermore, the Hidden Markov Model HMM is used for multi-state analysis, and the patient's current disease stage and its transfer path are inferred through the observed physiological signal sequences including EEG, HRV, and EDA;

[0047] The Bayesian network is used for causal relationship modeling to describe the multivariate causal relationship between disease stages; the probability dependency relationship between nodes of the Bayesian network is optimized in real time through the following update formula:

[0048]

[0049] Among them, Π(t) is the dependency relationship between nodes in the Bayesian network, B i (s) is the weight function of the i-th node, f i (X(s)) is the characteristic function of the ith physiological signal, α i is the time decay coefficient obtained by real-time observation data.

[0050] Beneficial effects of the present invention:

[0051] First, by integrating the real-time acquisition and analysis of multimodal physiological signals, the system is not limited to the traditional single EEG signal monitoring, but combines multiple key physiological signals such as EEG, HRV and EDA to fully reflect the patient's state of consciousness and physiological changes. This multi-dimensional data integration and analysis significantly improves the accuracy of disease monitoring and enables medical staff to obtain more comprehensive patient health information.

[0052] Secondly, the system uses advanced algorithms such as adaptive time series analysis, multi-level feature recognition, and hidden Markov model HMM to dynamically analyze and predict the course of the disease. The system updates the division of the disease stages in real time, captures the subtle changes of patients with impaired consciousness at different stages, and provides data support for personalized treatment plans. In addition, the system uses Bayesian network modeling to model the causal relationship between disease stages, helping to identify potential causes and key disease nodes, making treatment plans more targeted. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 This is a flow chart of the consciousness disorder course monitoring system of the present invention.

[0054] Figure 2 This is a flow module diagram for the initial construction of the model using an adaptive algorithm in the present invention.

[0055] Figure 3 This is a module diagram of the anomaly detection algorithm of the present invention.

[0056] Figure 4 This is a flow chart of the stage distribution and state transition model of the disease course of the present invention. DETAILED DESCRIPTION

[0057] The specific implementation modes of the present invention will be described in detail below in conjunction with the accompanying drawings.

[0058] Combined with Figure 1 The consciousness disorder course monitoring system consists of four main components, each of which targets different aspects and needs of course monitoring: First, the multi-dimensional data acquisition module uses a variety of sensors implanted in the patient, including electroencephalogram (EEG) sensors, heart rate variability (HRV) sensors, and skin conductance (EDA) sensors, to collect physiological signals in real time. These data are then transmitted to the edge computing device for preliminary filtering and preprocessing, including denoising, normalization, and feature extraction, to optimize data quality and reduce the computational burden of subsequent processing.

[0059] Next, the adaptive course prediction model uses the preprocessed physiological signal data, combined with the patient's historical medical records and course data, and uses adaptive machine learning algorithms, including support vector machines or deep learning networks, to train a personalized course prediction model. This model can update its parameters in real time as the patient's disease progresses, such as adjusting weights through online learning strategies to adapt to changes in the patient's status and the input of new data, thereby providing dynamic and accurate course predictions.

[0060] The continuous course monitoring module uses high-frequency sampling technology to continuously capture the patient's physiological signals to ensure the continuity and integrity of the data. The module uses real-time anomaly detection algorithms to analyze this data and identify key events in the course of the disease in real time, such as a sharp change in the level of consciousness or an epileptic seizure. Once these key events are detected, the system will immediately generate an early warning signal and trigger an emergency response mechanism, such as notifying medical staff for emergency intervention.

[0061] Finally, the disease stage distribution and state transition model identifies the various stages and state transitions in the disease process by analyzing the detailed data provided by the continuous monitoring module. This model uses statistical analysis and pattern recognition techniques, such as hidden Markov models or sequence analysis, to determine the different stages and transition points of the disease process, helping the medical team better understand the patient's disease progression and develop more targeted and effective treatment plans.

[0062] Embodiment 1:

[0063] In this embodiment, the consciousness disorder course monitoring system is used to conduct 24-hour uninterrupted monitoring and prediction for patients with consciousness disorders. The monitoring equipment includes a head-mounted EEG sensor with a high-density electrode array, which is used to collect electroencephalograms in real time and record the brain wave activities of patients in different states of consciousness; a heart rate variability HRV sensor is installed on the chest to continuously monitor heart rate changes; and a skin conductance EDA sensor is worn on the palm to reflect the activation of the autonomic nervous system by measuring skin conductance. The collected physiological signals can be expressed by the following formula:

[0064]

[0065] Φ(t) represents the fusion output of multimodal signals at time t, which is a comprehensive description of EEG, HRV and EDA signals. N represents the number or components of the signal; α n and κ n Respectively represent the amplitude of EEG and EDA signals, ranging from 0.5 to 2.0; ω n and are the angular frequencies of EEG and EDA signals, ranging from 0.1 to 10.0 rad / s; φ n and θ n Respectively represent the phase of EEG and EDA signals, ranging from 0 to 2π; β n It is the time attenuation factor, which reflects the attenuation trend of the signal over time and ranges from 0.001 to 0.01.

[0066] When t=10 seconds, the parameter values ​​collected are: α1=1.2, κ1=0.8, ω1=1.5, β1=0.005. Substitute into the formula and calculate the signal fusion value at this moment:

[0067]

[0068] After calculation, the value of Φ(10) is about 1.65. This fusion signal value will be used for the next step of model prediction and monitoring.

[0069] The embodiment further processes the multimodal signals collected in real time through edge computing devices, and the formulas for feature extraction and anomaly detection are:

[0070]

[0071] In this formula, Ω(t) represents the fusion output of the extracted characteristic signal of the state of consciousness and the detected abnormal waveform; It is the time first-order derivative of the multimodal signal, reflecting the speed of signal change; It is the time second derivative of the HRV signal, which is used to capture the acceleration changes of the signal, which is usually related to the rapid changes in the condition; and η is the time constant of abnormality detection, which determines the detection sensitivity of abnormal signals, and its value range is usually between 0.001 and 0.01.

[0072] At time t = 10 seconds, The calculated result is 0.05, the second derivative of the HRV signal The calculation result is -0.02, and the abnormality detection time constant η is set to 0.005. Substitute into the formula for calculation:

[0073]

[0074] Calculate the integral term

[0075]

[0076] Next, calculate This section:

[0077] Since exp(-0.005·t′) is a function that decays exponentially with time, we can get the approximate calculation result as follows:

[0078]

[0079] So, combining all the calculation results, the final value of Ω(10) is:

[0080] Ω(10)≈0.5+(-0.0049)≈0.4951

[0081] The calculated Ω(10) value shows that at the time point of 10 seconds, the fusion value of the extracted consciousness state characteristic signal and the detected abnormal waveform is close to 0.4951.

[0082] Next, the system compares this value with a preset abnormal threshold value, which is set to 0.4. Since Ω(10)=0.4951 is greater than 0.4, the system detects the existence of an abnormal state.

[0083] Embodiment 2:

[0084] Combined with Figure 2 In the embodiment, in the consciousness disorder course monitoring system, a personalized course prediction model is constructed based on the patient's baseline data and historical course data. The baseline data includes the EEG, HRV and EDA physiological signals collected by the patient in a healthy state, while the historical course data includes the changes in the physiological signals recorded by the patient during the previous consciousness disorder. Using these data, an adaptive filtering algorithm is applied to construct a patient's course prediction model, and the initial state is determined by the following formula:

[0085]

[0086] Ψ(t) is the output signal of the personalized disease course prediction model at time t, Γ i (s) is the weight function of the ith physiological signal including EEG, HRV or EDA, which indicates the relative importance of the signal to the prediction of the disease course, ranging from 0.1 to 1.0; Φ i (s) is the adaptive coefficient related to the signal, representing the characteristics of the signal changing over time, and its value range is between 0.5 and 2.0; T i is the periodic parameter of the signal, ranging from 1 to 20 seconds; φ i is the phase, ranging from 0 to 2π; ∈ i (s) is the noise term, which is used to simulate the measurement noise and has a value range of -0.05 to 0.05; α i It is the quadratic time coefficient related to signal attenuation, and its value range is between 0.001 and 0.01.

[0087] Select three physiological signals: EEG (i=1), HRV (i=2) and EDA (i=3). At time t=5 seconds, set the following parameter values: for EEG signal, Γ1(s)=0.8, Φ1(s)=1.2, T1=8 seconds, α1=0.005; for HRV signal, Γ2(s)=0.5, Φ2(s)=1.0, T2=12 seconds, α2=0.003; for EDA signal, Γ3(s)=0.7, Φ3(s)=0.9, T3=15 seconds, α3=0.007. Noise term ∈ i (s) takes 0.02, -0.01, and 0.03 for the three signals respectively.

[0088] Substituting these parameters into the formula, calculate for t = 5 seconds:

[0089] For EEG signal (i=1):

[0090]

[0091] When calculating this part of the integral term, it can be approximated:

[0092]

[0093] For HRV signal (i=2):

[0094]

[0095] A similar calculation method yields:

[0096] Ψ2(5)≈0.41

[0097] For EDA signal (i=3):

[0098]

[0099] The calculation results are:

[0100] Ψ3(5)≈0.54

[0101] Finally, the results of each part are added together to obtain the output signal of the personalized disease course prediction model at t = 5 seconds:

[0102] Ψ(5)≈Ψ1(5)+Ψ2(5)+Ψ3(5)=0.72+0.41+0.54=1.67

[0103] The value of Ψ(5) = 1.67 will be used as the initial prediction output of the model at the time point t = 5 seconds to further guide the dynamic monitoring and prediction of the patient's disease course.

[0104] The embodiment adopts a dynamic parameter update strategy based on variable step function and nonlinear least squares optimization method to optimize model parameters by minimizing prediction error. It is assumed that new physiological data are collected at time t=15 seconds, including the latest readings of EEG signals, HRV signals and EDA signals, and these data are used to update the previously constructed disease course prediction model. The mathematical representation of the update process is as follows:

[0105]

[0106] Among them, Ω(t) represents the current model parameter vector at time t; η(t) is the time-dependent adaptive learning rate function, ranging from 0.001 to 0.01; ζ(s) is the nonlinear variable step function that affects parameter update, and its value range is between 0.0001 and 0.001.

[0107] It is assumed that at the time point t = 15 seconds, the obtained physiological signal causes the model output Ψ(15) to be 1.85, and Ω(t-1) is the model parameter vector at the previous time point, which is set to [0.9, 0.7, 0.5]. The adaptive learning rate η(15) is set to 0.005, and the nonlinear variable step function ζ(s) is set to 0.0005 at t = 15 seconds.

[0108] First calculate the integral term

[0109]

[0110] Since exp(-0.0005·s 3) is a function that decays exponentially with time, and the integral is calculated using the numerical integration method. After numerical integration, the integral value is 12.3.

[0111] Next, calculate the gradient of the model parameters For the derivative of the above integral, the value of this gradient calculation is [0.03, 0.02, 0.01]. The updated model parameter vector Ω(15) is calculated as follows:

[0112] Ω(15)=[0.9,0.7,0.5]-0.005·[0.03,0.02,0.01]

[0113] The calculations show that:

[0114] Ω(15)=[0.9-0.00015,0.7-0.0001,0.5-0.00005]=[0.89985,0.6999,0.49995]

[0115] This updating process is performed at every time point in the course of the patient's disease, allowing the model to dynamically adapt to new physiological signal inputs. At this time, the updated model parameter vector Ω(15) = [0.89985, 0.6999, 0.49995] will reflect the patient's current physiological state and be used for subsequent disease course prediction.

[0116] The last embodiment introduces integrated deep learning technology and uses CNN to extract features from EEG signals, HRV signals, and EDA signals collected in real time. At time t = 20 seconds, a high-dimensional feature vector representing the patient's current state of consciousness is extracted through multi-layer convolution. These features are then input into LSTM to analyze the changing trend of the signal in the time series. LSTM can effectively memorize data dependencies over a long time span and predict the course of the patient's disease at several future time points. In order to further improve the prediction accuracy of the model, the model output is personalized and tuned through a multi-objective optimization mechanism. The optimization formula used is as follows:

[0117]

[0118] Among them, Δ(t) is the model output after personalized tuning; π represents the multi-objective optimization strategy; R(π,S) is the reward function at time s, ranging from 0.1 to 1.0; λ is the regularization coefficient, ranging from 0.001 to 0.01; is the gradient modulus of the state value function, which is used to measure the rate of change of the value of the model in different states; γ is the discount factor, which ranges from 0.8 to 1.0; V(s′) is the value function of the future state; T′ is the periodic parameter of the future state, which ranges from 5 to 20 seconds.

[0119] Setting parameters: R(π,S)=0.8,λ=0.005, The integral of is 0.02 at s = 20 seconds, γ = 0.9, V(s′) = 1.5, T′ = 10 seconds. Substituting these parameters into the formula, the output of the personalized tuned model Δ(20) at time t = 20 seconds is calculated:

[0120] First, calculate The result is 0.02:

[0121]

[0122] After simplification, the expression is:

[0123]

[0124] Continue to calculate the above formula and set The mean value in the interval is 0.5, so the integral is approximately:

[0125]

[0126] Therefore, at the time point t = 20 seconds, the output Δ(20) of the model after personalized tuning is approximately 29.5. This result reflects the output state of the disease course prediction model at time t = 20 seconds after integrating deep learning technology and multi-objective optimization.

[0127] Embodiment 3:

[0128] Combined with Figure 3 In the consciousness disorder course monitoring system, wavelet transform and Fourier transform are used to perform real-time analysis on continuously collected physiological signals to extract characteristic information of different frequency bands. The mathematical formula is:

[0129]

[0130] Θ(f,t) is the extracted feature signal; x(t′) is the input original physiological signal, including EEG, HRV, and EDA; is a wavelet function used for time-frequency analysis of local signals; s(f) is a frequency-related scale parameter that determines the compression or expansion of the wavelet; α is a time attenuation factor that ranges from 0.001 to 0.01.

[0131] At t = 30 seconds, the EEG signal x(t′) collected from the patient is a combination of high-frequency signals, whose main frequency components are in the range of 8 to 12 Hz. In order to extract the characteristic information of different frequency bands, the wavelet function is set The scale parameter s(f) is 0.1, and the time decay factor α is chosen to be 0.005.

[0132] Substituting these parameters into the formula for calculation, setting x(t′)=sin(2π·10·t′) as the main frequency component of the EEG signal, the calculation is as follows:

[0133]

[0134] set up The wavelet function is in the form of Then the integral can be expressed as:

[0135]

[0136] The result obtained by numerical calculation integration is Θ(10,30)≈0.85. This value Θ(10,30) represents the characteristic information of the EEG signal collected at the time point t=30 seconds when the frequency is 10 Hz.

[0137] Next, the embodiment uses convolutional neural network CNN and isolation forest algorithm IsolationForest to perform real-time analysis on the extracted feature signals to identify key abnormal events that occur during the patient's disease course. CNN is responsible for learning complex patterns and features from feature signals, while isolation forest algorithm is used to detect abnormal data points. In order to improve the accuracy of anomaly detection, the following optimization function is used to train and optimize the model:

[0138]

[0139] Φ(X) is the loss function of anomaly detection; W and b are the weight and bias of the model respectively; y i is the target label of the i-th sample, ranging from -1 (normal) to 1 (abnormal); λ is the regularization parameter, ranging from 0.001 to 0.01, which is used to prevent overfitting; R(X,W) is the regional constraint function, which helps limit the excessive complexity of the model in the feature space Ω.

[0140] The feature signal matrix X extracted by CNN i Contains 5 sample data, each of which has 3 main features: the amplitude of the EEG signal, the frequency of the HRV signal, and the conductance value of the EDA signal. Set the target marker y i They are 1, -1, 1, -1, and 1 respectively, indicating that three samples are marked as abnormal and the remaining two are normal.

[0141] Set the initial weight W to 0.5, -0.3, 0.8, the bias b to 0.1, and the regularization parameter λ to 0.005. Substitute these parameters into the optimization function for calculation:

[0142]

[0143] Calculate the logarithmic loss term for each sample, set the sample data of the first row of the feature matrix X to 0.6, 0.2, -0.4, then the linear combination result is:

[0144] W T X1+b=0.5×0.6+(-0.3)×0.2+0.8×(-0.4)+0.1=0.3-0.06-0.32+0.1=0.02

[0145] Then calculate the logarithmic part of the loss function:

[0146]

[0147] Then calculate the regularization term λ|W| 2 :

[0148] 0.005×(0.5 2 +(-0.3) 2 +0.8 2 )=0.005×(0.25+0.09+0.64)=0.005×0.98=0.0049

[0149] For the regional constraint function R(X,W), set its integral in the feature space Ω to 0.03. Then the value of the overall loss function is:

[0150]

[0151] By minimizing the loss function, the weights W and bias b are updated to optimize the anomaly detection model so that it can better identify key abnormal events in the course of the disease.

[0152] The next step of the embodiment adopts the Bayesian update method to dynamically adjust the threshold of abnormal detection according to real-time data to ensure that an early warning can be issued in time when the condition changes suddenly. The formula is as follows:

[0153]

[0154] In this formula, τ(t) is the adaptive threshold at time t; γ is the learning rate, ranging from 0.01 to 0.1; μ and σ are the mean and standard deviation of the real-time data, respectively; x(t) is the physiological signal data collected at time t.

[0155] At t = 40 seconds, the patient's physiological signal data x(t) = 1.8 is collected. At the previous time point t-1 = 39 seconds, the abnormal detection threshold τ(39) is 0.6. The mean μ of the real-time data is set to 1.5, the standard deviation σ = 0.3, and the learning rate γ = 0.05.

[0156] Substituting these parameters into the Bayesian update formula, the adaptive threshold τ(40) at t = 40 seconds is calculated:

[0157] First, calculate the Gaussian distribution term:

[0158]

[0159] Next, the new threshold is calculated using the Bayesian update formula:

[0160] τ(40)=0.6+0.05·(0.32-0.6)=0.6+0.05·(-0.28)=0.6-0.014=0.586

[0161] The anomaly detection threshold at t=40 seconds is updated to τ(40)=0.586.

[0162] The embodiment then evaluates in real time whether these characteristic signals exceed corresponding thresholds, and generates multi-level warning signals according to the severity of the anomaly, using the formula:

[0163]

[0164] In this formula, S(t) is the strength of the warning signal generated at time t; δ k is the warning weight coefficient of the kth feature signal, ranging from 0.1 to 1.0, reflecting the importance of the signal to the overall warning; Θ k (t) is the kth characteristic signal, including the real-time value of EEG, HRV or EDA; I{·} is the indicator function, when the characteristic signal exceeds the adaptive threshold τ k The warning is triggered.

[0165] Assume that at a certain time t = 50 seconds, the following physiological signals are monitored:

[0166] The EEG signal Θ1(t) is 1.7, the threshold τ1 is 1.5, and the warning weight coefficient δ1 is 0.7;

[0167] The HRV signal Θ2(t) is 0.9, the threshold τ2 is 1.0, and the warning weight coefficient δ2 is 0.5;

[0168] The EDA signal Θ3(t) is 2.0, the threshold τ3 is 1.8, and the warning weight coefficient δ3 is 0.8.

[0169] First calculate whether each signal triggers an alert:

[0170] For the EEG signal: Θ1(t)=1.7, which exceeds the threshold τ1=1.5, thus triggering an early warning.

[0171] For HRV signal: Θ2(t) = 0.9, does not exceed the threshold τ2 = 1.0, and does not trigger an early warning.

[0172] For the EDA signal: Θ3(t)=2.0, which exceeds the threshold τ3=1.8, thus triggering an early warning.

[0173] Next, the warning signal strength is calculated according to the warning weight coefficient:

[0174] S(50)=0.7·1.7·1+0.5·0.9·0+0.8·2.0·1

[0175] The calculation results are:

[0176] S(50)=1.19+0+1.6=2.79

[0177] Therefore, at t=50 seconds, the warning signal strength S(50)=2.79 generated by the system reflects the overall abnormality of the patient at that moment.

[0178] Embodiment 4:

[0179] Embodiment In the consciousness disorder course monitoring system, an adaptive time series analysis module is constructed. Based on the real-time collected physiological signals, including EEG, HRV and EDA, a dynamic analysis method is used to identify the time variation patterns of these signals. Through the adaptive function:

[0180]

[0181] F(t) is the output signal of time series analysis, which is used to identify the stage of the disease; α(s) and β(s) are adaptive weight functions with a value range of 0.1 to 1.0, which are used to adjust the weights of the derivative term and the signal term in the analysis; x(s) is the input physiological signal; ω is the frequency parameter with a value range of 0.5 to 2.0 rad / s; λ is the time attenuation factor with a value range of 0.001 to 0.01, which is used to attenuate the historical influence of the signal.

[0182] At t = 60 seconds, set the adaptive weight function α(s) = 0.8 and β(s) = 0.5, the frequency parameter ω = 1.0 rad / s, and the time decay factor λ = 0.005. The derivative of the EEG signal at time s is And x(s)=1.5 appears as a relatively stable signal within the range of s.

[0183] Calculate the time series analysis output signal F(60) at t = 60 seconds:

[0184]

[0185] For the derivative

[0186]

[0187] For the signal Assuming the mean value of sin(1.0·s) in this interval is 0.5, the simplified calculation is:

[0188]

[0189] Therefore, the final time series analysis output signal is:

[0190] F(60)≈0.829+19.44=20.27

[0191] The value of the time series analysis output signal F(60) is used to identify the stage of the patient's disease at that moment. If the system continues to detect an upward trend in the F(t) value over the next period of time, it means that the patient's condition has entered a more serious stage. The system will mark this stage and recommend increasing the monitoring frequency or adjusting the treatment plan.

[0192] Next, the embodiment identifies the spectrum changes in the EEG signal, the time domain and frequency domain features in the HRV, and the conductance fluctuations in the EDA, and calculates the output signal of the multi-level feature recognition by the following formula:

[0193]

[0194] G(t) is the output signal of multi-level feature recognition; γ i (s) is the weight function of the i-th layer, ranging from 0.1 to 1.0, which is used to adjust the importance of different features in the recognition process; f i (x(s)) is the feature extraction function of the i-th layer, which is the spectrum analysis of EEG signals, the time domain or frequency domain analysis of HRV signals, or the conductance change analysis of EDA signals; θ i and φ i is the phase and frequency offset, θ i The value range is 0.01 to 0.1, φ i The value range of μ is 0 to 2π; i is the time attenuation factor of the characteristic signal, ranging from 0.001 to 0.01.

[0195] At t = 80 seconds, the spectral changes in the patient's EEG signal, the time domain features in the HRV signal, and the conductance fluctuations in the EDA signal are identified. The three main feature layers are set as:

[0196] Layer 1: EEG spectrum analysis, setting γ1(s) = 0.8, f1(x(s)) = sin(2π·10s) to indicate that the main frequency of the EEG signal is 10 Hz, θ1 = 0.05, μ1=0.005;

[0197] Layer 2: HRV time domain analysis, set γ2(s) = 0.6, f2(x(s)) = x(s) represents the direct time domain value of the HRV signal, θ2 = 0.03, μ2=0.007;

[0198] Layer 3: EDA conductivity fluctuation analysis, set γ3(s) = 0.9, f3(x(s)) = cos(2π·0.2s) to represent the conductivity fluctuation of the EDA signal, θ3 = 0.02, μ3=0.003.

[0199] Substituting these parameters into the formula, the multi-level feature recognition output signal G(80) at t = 80 seconds is calculated:

[0200] For the feature identification items of layer 1 EEG signal:

[0201]

[0202] For the feature identification items of the second layer HRV signal:

[0203]

[0204] For the feature identification items of the 3rd layer EDA signal:

[0205]

[0206] The time decay factor is then applied to each layer:

[0207] G(80)≈(2.5·e -0.005·80 )+(1.8·e -0.007·80 )+(3.2·e -0.003·80 )

[0208] The specific calculation is:

[0209] G(80)≈(2.5·e -0.4 )+(1.8·e -0.56 )+(3.2·e -0.24 )≈(2.5·0.67)+(1.8·0.57)+(3.2·0.79)The final result is:

[0210] G(80)≈1.675+1.026+2.528≈5.229

[0211] When G(80)=5.229, the system recognizes that the patient's physiological signal at this moment has a significant change, indicating further development of the disease or potential risk of deterioration.

[0212] After the system performs time series analysis and multi-level feature recognition on the patient's EEG, HRV, and EDA signals, the embodiment further divides the disease course stages according to these analysis results, using the following formula for dynamic stage division:

[0213]

[0214] D(t) is the result of the disease stage division at time t; F j (t) and G j (t) are the output signals of time series analysis and multi-level feature recognition, and the signal of the jth layer respectively; ρ k is the penalty coefficient of the kth stage, usually ranging from 0.01 to 0.1, and is used to adjust the sensitivity to signal differences when dividing the stages; M represents the number of signal layers analyzed.

[0215] The physiological signal was analyzed at t = 100 seconds and the following output was obtained:

[0216] Time series analysis output: F1(100)=2.5, F2(100)=3.0, F3(100)=1.8;

[0217] Multi-level feature recognition output: G1(100)=5.2, G2(100)=4.5, G3(100)=2.7;

[0218] Penalty coefficients ρ1=0.05, ρ2=0.07, ρ3=0.1.

[0219] First calculate F j (t)·G j The cumulative sum of (t):

[0220]

[0221] The specific calculation is:

[0222] =13.0+13.5+4.86=31.36

[0223] Next, calculate the integral term In the previous time series analysis and feature identification process, and The difference integral result is:

[0224] j = 1 layer: the integral result is 1.2;

[0225] j = 2 layers: the integral result is 0.8;

[0226] j=3 layers: the integral result is 1.5.

[0227] Therefore, the penalty term considering the penalty coefficient is:

[0228] ρ1×1.2+ρ2×0.8+ρ3×1.5=0.05×1.2+0.07×0.8+0.1×1.5

[0229] The specific calculation is:

[0230] =0.06+0.056+0.15=0.266

[0231] Then, the score of the dynamic phase division is calculated:

[0232]

[0233] The patient's disease course stage is divided according to the stage k corresponding to the maximum score. k The selection is set according to the experience values ​​of different stages of the disease. At this time, the system recognizes that the patient is in the second stage of the disease. This stage is related to the aggravation of consciousness disorder and requires enhanced monitoring and adjustment of treatment plans.

[0234] Embodiment 5:

[0235] Combined with Figure 4 In the consciousness disorder disease course monitoring system, the Hidden Markov Model HMM is used to perform multi-state analysis on its physiological signal sequence (including EEG, HRV and EDA) to infer the patient's current disease course stage and future transition path. The state transition probability matrix P(t) is represented by the following dynamically adjusted formula:

[0236]

[0237] P(t) is the state transition probability matrix at time t; A(s) is the state transition weight matrix, which represents the transition weights between different states, and its elements range from 0 to 1; λ is the time attenuation factor, ranging from 0.001 to 0.01; ω is the frequency parameter, ranging from 0.5 to 2.0 rad / s; φ is the phase offset, ranging from 0 to 2π.

[0238] When the time is set to t = 120 seconds, the physiological signal sequence observed is as follows:

[0239] 1. EEG signals show high-frequency activity;

[0240] 2.HRV signal shows significant low-frequency fluctuations;

[0241] 3. EDA signal shows a sharp rise in skin conductance.

[0242] These signals indicate that the patient is in the transition period of the disease process. Set the element value of the state transition weight matrix A(s) to:

[0243]

[0244] Set the time attenuation factor λ = 0.005, the frequency parameter ω = 1.0rad / s, and the phase offset

[0245] Substituting these parameters into the state transition probability matrix formula, calculate the state transition probability matrix P(120) at t = 120 seconds:

[0246] First, calculate the integral term:

[0247]

[0248] Compute the elements of the matrix individually:

[0249]

[0250] Assume that P is obtained by numerical integration 11 (120)≈0.12.

[0251] The same method is used to calculate other elements:

[0252] P 12 (120)≈0.18,P 21 (120)≈0.24,P 22 (120)≈0.06

[0253] Therefore, the state transition probability matrix P(120) is:

[0254]

[0255] When the system detects that the combined pattern of EEG, HRV, and EDA signals shows that the patient has moved from a stable state to an unstable state, the system determines the next state the patient is about to enter based on the calculated P(120) value.

[0256] The last embodiment uses a Bayesian network to represent the dependencies between different physiological signals and disease stages through nodes and edges. Each node represents a physiological signal or disease stage, and the direction of the edge represents the causal relationship. The probability dependency between nodes of the Bayesian network is optimized in real time through the following update formula:

[0257]

[0258] Π(t) represents the dependency relationship between nodes in the Bayesian network at time t; B i (s) is the weight function of the ith node, reflecting the importance of the node in the network, with a value range of 0.1 to 1.0; f i (X(s)) is the characteristic function of the i-th physiological signal, including the characteristic signals of EEG, HRV, and EDA; α i is the time decay coefficient, ranging from 0.001 to 0.01.

[0259] At t = 150 seconds, there are the following physiological signal data:

[0260] 1.EEG signal: frequency characteristic f1(X(s))=sin(2π·10s);

[0261] 2.HRV signal: time domain feature f2(X(s))=X(s);

[0262] 3. EDA signal: conductivity fluctuation characteristic f3(X(s))=cos(2π·0.5s).

[0263] Set the weight function of the corresponding node in the Bayesian network to:

[0264] B1(s)=0.8·s, B3(s)=0.9·log(s+1)

[0265] The time attenuation coefficients are set to α1=0.003, α2=0.005, and α3=0.007.

[0266] Substituting these parameters into the formula, the output signal Π(150) of the dependency relationship between the Bayesian network nodes at t = 150 seconds is calculated:

[0267] For the dependency terms of the first node (EEG signal):

[0268]

[0269] here, And the result obtained by numerical integration is set to 0.12.

[0270] For the dependency items of the second node (HRV signal):

[0271]

[0272] in, Set the integral result to 0.09.

[0273] For the dependencies of the third node (EDA signal):

[0274]

[0275] in, Set the integral result to 0.15.

[0276] Finally, the node dependency Π(150) of the Bayesian network is calculated as:

[0277] Π(150)=0.12·0.09·0.15≈0.00162

[0278] At t=150 seconds, the system identified a lower dependency value Π(150), indicating that the causal relationship between the physiological signals was weak and the disease state was in a stable stage.

[0279] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A system for monitoring the course of consciousness disorders, including a multi-dimensional data acquisition module, an adaptive course prediction model, a continuous course monitoring module, and a course phase distribution and state transition model, characterized by: The multi-dimensional data acquisition module is used to collect the patient's electroencephalogram (EEG), heart rate variability (HRV), and skin conductance (EDA) physiological signals in real time through multiple sensors installed on the patient's body, and pre-process the collected data through edge computing equipment; The adaptive disease course prediction model is based on the physiological signals collected by the multi-dimensional data collection module, and is combined with the patient's historical disease course data, including the changes in the physiological signals recorded during the patient's previous consciousness disorder period, and uses an adaptive algorithm to train a personalized disease course prediction model. During the course of the patient's disease, the model parameters are adaptively updated through a machine learning algorithm to achieve dynamic prediction of the progression of the course of consciousness disorder; The continuous course of disease monitoring module continuously acquires the patient's physiological signals through high-frequency sampling technology to ensure the continuity and integrity of course of disease monitoring, and identifies key events in the course of disease in real time through anomaly detection algorithms, including sudden loss of consciousness and epileptic seizures, generates early warning signals, and triggers emergency response mechanisms; The disease course stage distribution and state transition model identifies various stages and state transitions in the disease course by analyzing the patient's physiological signals obtained by the continuous disease course monitoring module; The EEG acquisition uses high-density electrode array technology to obtain electrical activity waveforms through head-mounted EEG sensors, heart rate variability HRV is acquired through sensors on the chest or wrist, and skin conductance EDA is acquired through sensors on the palm or fingers. The overall description of the signal can be expressed as: Among them, Φ(t) is the fusion of EEG and EDA signals, α n and κ n Represent the amplitude of EEG and EDA, ω n and are their angular frequencies, φ n and θ n are phase, β n is the time decay factor, which comprehensively represents the changes in physiological signals of the state of consciousness; The edge computing device processes the above Φ(t) data and HRV signal in real time, extracts characteristic signals related to the state of consciousness, and performs abnormal waveform detection. The formulas for feature extraction and abnormality detection are: In this formula, Ω(t) represents the fusion output of the extracted characteristic signal of the state of consciousness and the detected abnormal waveform; It is the time first-order derivative of the fused signal, reflecting the speed of signal change; It is the time second-order derivative of the HRV signal, which is used to capture the acceleration change of the signal; η is the time constant of anomaly detection, which determines the detection sensitivity of abnormal signals and ranges from 0.001 to 0.

01.

2. The consciousness disorder course monitoring system according to claim 1, characterized in that: The initial construction of the adaptive disease course prediction model is based on the patient's baseline data, including the EEG, HRV and EDA physiological signals and historical disease course data collected by the patient in a healthy state, and an adaptive filtering algorithm is applied to determine the initial state of the model through the following formula: Among them, Ψ(t) is the output of the personalized disease course prediction model, Γ i (s) is the weight function of the ith physiological signal including EEG, HRV or EDA, Φ i (s) is the adaptive coefficient related to the physiological signal, T i is the periodic parameter of the physiological signal, φ i is the phase, ∈ i (s) is the noise term, α i is the quadratic time coefficient related to the attenuation of physiological signals; During the course of the patient's disease, the dynamic parameter update module collects and inputs new physiological signals in real time, minimizes the prediction error, and dynamically adjusts the model parameters using a variable step function and nonlinear least squares optimization method.

3. The consciousness disorder course monitoring system according to claim 1, characterized in that: The anomaly detection algorithm is based on multi-level feature extraction. By introducing a multi-level feature extraction algorithm in an edge computing device, the continuously collected physiological signals are analyzed in real time. Specifically, the following steps are included: The wavelet transform WaveletTransform and Fourier transform FourierTransform methods are used to extract the characteristic information of different frequency bands, and its mathematical expression is: Among them, Θ(f,t) is the extracted feature signal, x(t ′ ) is the input original physiological signal, including EEG, HRV, EDA, is the wavelet function, s(f) is the frequency-related scale parameter, and α is the time decay factor; The extracted feature signals are analyzed in real time through the convolutional neural network (CNN) and the isolation forest algorithm (IsolationForest) to identify key abnormal events; The emergency response mechanism adopts the Bayesian Updating method to dynamically adjust the threshold of anomaly detection according to real-time data. The formula is expressed as: Among them, τ(t) is the adaptive threshold at time t, γ is the learning rate, μ and σ are the mean and standard deviation of the real-time data respectively; Furthermore, the warning signal generation and emergency response mechanism generates multi-level warning signals based on the severity of abnormal events. The generation process of the warning signal is expressed as: Among them, S(t) is the warning signal strength at time t, δ k is the warning weight coefficient, Θ k (t) is the kth characteristic signal, I{·} is the indicator function, when the characteristic signal exceeds the adaptive threshold τ k The warning is triggered.

4. The consciousness disorder course monitoring system according to claim 1, characterized in that: The disease course stage distribution and state transition model includes a disease course stage division module, a state transition model and a disease course prediction and clinical decision support module, and is characterized by: The disease course stage division module is based on real-time collected physiological signals, including EEG, HRV and EDA, and uses time series analysis methods to dynamically divide the disease course into multiple stages including stable period, deterioration period and recovery period; The state transition model analyzes the state transition probability between each disease course stage through the hidden Markov model HMM to deal with the uncertainty in the disease course data; through Bayesian network modeling, the correlation between stages is updated according to real-time data to optimize the disease course prediction; The disease course prediction and clinical decision support module provides personalized disease course prediction based on the analysis of the state transition model and provides real-time decision support for clinicians.

5. The consciousness disorder course monitoring system according to claim 4, characterized in that: The Hidden Markov Model HMM is used for multi-state analysis, and the patient's current disease stage and its transfer path are inferred through the observed physiological signal sequences including EEG, HRV, and EDA; The Bayesian network is used for causal relationship modeling to describe the multivariate causal relationship between disease stages; the probability dependency relationship between nodes of the Bayesian network is optimized in real time through the following update formula: Among them, Π(t) is the dependency relationship between nodes in the Bayesian network, B i (s) is the weight function of the i-th node, f i (X(s)) is the characteristic function of the ith physiological signal, α i is the time decay coefficient obtained by real-time observation data.

Images