Risk warning system for in-hospital transfer of critical patients based on adaptive decision tree

Abstract

The application discloses a critical illness patient hospital transfer risk early warning system based on an adaptive decision tree, and relates to the technical field of electroencephalogram monitoring.The system comprises: an electroencephalogram measurement module, which is used for acquiring multi-lead electroencephalogram signals before a patient is taken off line, and extracting homologous artifact features and preliminary screening abnormal events representing electroencephalogram physiological states; an electroencephalogram calculation module, which generates an adaptive abnormal threshold through calculation, and generates a high-credibility abnormal load in combination with the propagation compatibility of electroencephalogram waves among leads and the artifact coupling relationship; a state updating module, which updates based on the abnormal residual amount of a previous time window and the high-credibility abnormal load of a current time window; and a risk decision module, which outputs an electroencephalogram residual risk warning grade before hospital transfer through an adaptive decision tree based on the measured electroencephalogram signals.The application is based on measured electroencephalogram physiological signals, starts from the space-time propagation characteristics, effectively suppresses artifacts, and realizes accurate evaluation of the electroencephalogram residual risk of a patient in a situation of off-line transfer.

Description

Technical Field

[0001] This invention relates to the field of electroencephalogram (EEG) signal monitoring technology, specifically to an early warning system for the risk of in-hospital transport of critically ill patients based on adaptive decision trees. Background Technology

[0002] In recent years, continuous electroencephalography (EEG) monitoring has played an increasingly important role in assessing the physiological electrical activity of brain neurons and detecting occult brain injuries such as non-convulsive epilepsy. However, critically ill patients often require intra-hospital transport (e.g., to the radiology department for imaging examinations), during which time EEG monitoring leads must usually be temporarily removed. This results in patients being in an EEG monitoring blind spot for a prolonged period during transport, and if abnormal neuronal discharges occur during this time, they face an extremely high risk of irreversible brain damage. Therefore, measuring the last EEG signal before weaning from ventilator is of significant clinical importance in assessing the risk of residual EEG activity in patients without monitoring.

[0003] However, existing EEG monitoring equipment has certain limitations in handling such weaning and transport scenarios. During the preparation period for removing physical leads, the touching by medical staff, the removal of leads, and the repositioning of the patient inevitably introduce strong mechanical contact artifacts and electromyographic interference into the EEG signal acquisition circuit. Traditional EEG monitoring equipment triggers alarms instantly based on fixed signal amplitudes, failing to physiologically distinguish between genuine neuronal discharges with spatial conduction characteristics in brain regions and multi-channel synchronous spikes caused by physical interference from leads. Once the physical connection is disconnected, traditional equipment immediately terminates monitoring, failing to combine abnormal discharge loads before weaning with brain physiological electrical activity to transform them into a continuous physiological risk assessment for future areas without monitoring blind spots. This easily leads to false alarms before weaning or missed alarms during transport.

[0004] Therefore, there is an urgent need for a system that overcomes the mechanical and electromyographic artifacts and abnormal neuronal discharges unique to the physical disconnection process when measuring electroencephalographic signals, so as to achieve accurate early warning of the risk of residual EEG after the patient leaves the physical monitoring environment. Summary of the Invention

[0005] To address the problems in existing EEG signal measurement, this invention provides an early warning system for the risk of in-hospital transport of critically ill patients based on adaptive decision trees.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] In the first aspect, this application discloses an intra-hospital transfer risk warning system for critically ill patients based on adaptive decision trees, including an electroencephalogram (EEG) measurement module, an EEG calculation module, a state update module, and a risk decision module.

[0008] The EEG measurement module acquires multi-lead EEG signals from the EEG monitoring device before monitoring stops and divides them into multiple sequentially arranged time windows. It then extracts homologous artifact features and preliminary screening abnormal events from the effective leads within each time window.

[0009] The EEG calculation module generates an adaptive abnormality threshold for the current time window based on the homology artifact features of the current time window and the abnormal residual amount of the previous time window. This threshold is then used to screen for initial abnormal events and obtain candidate abnormal EEG events. A brain region lead topology map is generated based on the effective leads, and the propagation morphology compatibility of the candidate abnormal EEG events between leads is calculated. Based on the propagation morphology compatibility, a protective correction is performed to obtain corrected homology artifact features, and the coupling relationship between the candidate abnormal EEG events and the corrected homology artifact features is calculated. This generates a high-confidence abnormality load for the current time window.

[0010] The state update module controls the decay of the abnormal residual amount in the previous time window and adds the high-confidence abnormal load of the current time window until all time windows are traversed, and then obtains the final abnormal residual amount that represents the electroencephalographic state.

[0011] The risk decision-making module outputs the risk warning level of EEG residue before in-hospital transfer based on the final abnormal residual amount through an adaptive decision tree.

[0012] Secondly, this application discloses a method for early warning of risks in the intra-hospital transfer of critically ill patients based on adaptive decision trees, which is applied to the aforementioned early warning system for early warning of risks in the intra-hospital transfer of critically ill patients based on adaptive decision trees, and includes the following steps:

[0013] The multi-lead EEG signals of the EEG monitoring device before monitoring was stopped were acquired and divided into multiple sequentially arranged time windows. Homologous artifact features and preliminary screening of abnormal events in the effective leads within each time window were extracted.

[0014] Based on the homology artifact features of the current time window and the abnormal residual amount of the previous time window, an adaptive abnormality threshold for the current time window is generated, and then the initial screening of abnormal events is filtered to obtain candidate abnormal EEG events; a brain region lead topology map is generated based on the effective leads, and then the propagation morphology compatibility of the candidate abnormal EEG events in the leads is calculated; based on the propagation morphology compatibility, a protection correction is performed to obtain the corrected homology artifact features, and the coupling relationship between the candidate abnormal EEG events and the corrected homology artifact features is calculated, thereby generating a high-confidence abnormality load for the current time window;

[0015] The residual abnormality from the previous time window is attenuated and superimposed with the high-confidence abnormality load of the current time window until all time windows are traversed, and the final residual abnormality representing the electroencephalological state is obtained.

[0016] Based on the final abnormal residual amount, an adaptive decision tree is used to output the risk warning level of EEG residuals before in-hospital transfer.

[0017] Thirdly, this application discloses a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the method for risk warning of intra-hospital transfer of critically ill patients based on adaptive decision trees as described in the second aspect.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0019] 1. This application enhances the resistance to physical interference in EEG physiological signal acquisition under complex clinical procedures, effectively curbing false alarms. It changes the traditional mechanical judgment method of EEG monitoring equipment that relies on fixed potential thresholds to trigger alarms. For operations unique to critically ill patients preparing for weaning and transport, such as lead touching and body repositioning, this system dynamically senses the intensity of local interference in the signal loop and adaptively adjusts the validity screening threshold of physiological signals. This mechanism, while reducing the probability of false alarms, adaptively isolates non-physiological composite artifacts introduced by mechanical contact and changes in the physical environment from the source of physiological signal acquisition, solving the device defect of high false positive alarms during the weaning preparation period from the physical measurement level.

[0020] 2. This application accurately isolates mechanical spurs in leads, truly identifying the physical spatial conduction pathways of neuronal electrical activity in the brain. It overcomes the limitations of traditional monitoring equipment that only compares the amplitude of single-channel signals, deeply integrating the natural spatial conduction patterns of abnormal neuronal discharges (such as epileptic waves) within physical brain regions. By capturing the waveform delay and morphological evolution characteristics of EEG signals between different physical leads, the system can improve the ability to distinguish between "multi-lead instantaneous synchronous physical spurs caused by external mechanical contact" and "abnormal physiological discharges with real neural pathway conduction characteristics."

[0021] 3. This application objectively quantifies the temporal continuity of abnormal brain electrophysiological states, filling the physical gap in monitoring during the weaning period. For the special situation of no monitoring equipment connection during weaning transport, a dynamic measurement mechanism for the "residual effect" of abnormal brain electrophysiology is constructed. By continuously tracking the physiological decay process of abnormal load, this system successfully transforms the last segment of physical EEG signal before weaning into an objective quantification of the continuity of brain physiological states in the future weaning blind zone, overcoming the limitation of traditional equipment that "loses diagnostic capability upon physical disconnection." Attached Figure Description

[0022] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts. Wherein:

[0023] Figure 1 This is a system framework diagram for the present invention;

[0024] Figure 2 This is a system flowchart for the present invention;

[0025] Figure 3 This is a flowchart illustrating the propagation morphology compatibility of this invention.

[0026] Figure 4 This is a flowchart of the risk warning method for intra-hospital transfer of critically ill patients based on adaptive decision trees, as described in Example 2. Detailed Implementation

[0027] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0028] In existing technologies, traditional EEG monitoring systems for acute and critical illnesses mostly use static absolute amplitude thresholds to detect and identify abnormal electrophysiological waves, making it difficult to balance the preservation of real abnormal neuronal discharges with the suppression of artifacts from physical mechanical contact. While real-time EEG screening technology can capture sudden abnormalities in a continuous physical connection, it is prone to generating false alarms when the underlying signal quality is compromised due to the lack of effective artifact coupling removal mechanisms and quantitative models of physiological effect continuation in the offline state. Especially in the physical environment of the preparation period for offline transport, which is full of strong mechanical disturbances such as body position changes and lead insertion and removal, existing systems usually directly perform linear discrimination on mixed signals containing multi-channel instantaneous synchronous jumps, resulting in severe aliasing of real brain region spatial conduction characteristics and physical interference signals.

[0029] To address the aforementioned issues, research has identified the need to establish a closed-loop feedback mechanism for spatiotemporal propagation compatibility and adaptive artifact suppression to optimize the EEG measurement process. Particularly in scenarios involving the weaning of critically ill patients from ventilators and in environments with extremely poor physical interference, the system needs to automatically sense the coupling strength of local mechanical artifacts to guide and dynamically adjust the physiological threshold for initial screening of EEG abnormalities. Simultaneously, a computational model for the accumulation and temporal decay of EEG abnormalities needs to be established to remove irregular synchronization interference from surface physical contact, enabling effective focusing and enhancement of abnormal physiological signals with genuine spatial conduction characteristics in deeper layers. Based on this, this proposal suggests an adaptive discrimination-based EEG measurement and evaluation system. By establishing a feedback loop between low-level artifact perception and high-level signal screening, it achieves the removal of mechanical synchronization artifacts from multi-lead physiological sequences with strong physical interference and aliasing, tracks the spatial propagation patterns of genuine brain regions, dynamically quantifies the residual decay effect of physiological abnormalities, and ultimately produces objective quantitative evaluation results with high physiological confidence.

[0030] After introducing the basic concept of the present invention, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0031] Example 1

[0032] like Figure 1 As shown in the figure, this embodiment introduces an intra-hospital transfer risk warning system for critically ill patients based on adaptive decision tree, including an EEG measurement module 100, an EEG calculation module 200, a status update module 300, and a risk decision module 400.

[0033] The EEG measurement module 100 is used to acquire multi-lead EEG signals from the EEG monitoring device before monitoring is stopped and divide them into multiple sequentially arranged time windows, extracting homologous artifact features and screening abnormal events in the effective leads within each time window.

[0034] Specifically, in this embodiment, when the EEG monitoring device detects a stopped monitoring state, a recording termination state, a lead removal state, or a state where monitoring is about to end, it automatically locks a segment of continuous collectable EEG signals before the monitoring stops, which serves as a withdrawal window for subsequent EEG residual risk analysis before in-hospital transfer.

[0035] Specifically, the moment when the EEG monitoring device can obtain the last reliable EEG sampling data is taken as the time to stop monitoring, denoted as . ,in This is the endpoint of the withdrawal window on the timeline, and also the starting point for the system to backtrack and determine the withdrawal window. If the times corresponding to lead removal, recording termination, monitoring cessation, or monitoring about to end are inconsistent, the time of the last reliable EEG sampling will be used as the starting point. Starting from that moment and working backwards, the duration of valid EEG signals that continuously met the conditions of complete sampling, valid leads, no prolonged saturation, no prolonged missing signals, and no severe distortion before monitoring ceased is recorded as follows: ,in The duration of a valid electroencephalogram (EEG) signal is measured in seconds.

[0036] Preset minimum evacuation window length and maximum evacuation window length ,in This is the minimum amount of valid data required to ensure risk assessment. Used to limit the maximum analysis range for forward backtracking. In this embodiment, Set to 60 seconds. Set to 600 seconds. .

[0037] The specific preset validity conditions are as follows:

[0038] like This indicates that there was insufficient effective EEG data before monitoring was stopped, and the system entered a low-confidence output branch.

[0039] like Mark the data validity Set to 1, and take the smaller of the effective EEG duration and the maximum allowed recall duration as the actual withdrawal window length. ,Right now ,in This is the actual evacuation window length, in seconds. Monitoring will then cease. The continuously acquireable EEG signals within the range are defined as the withdrawal window W. Signals within the withdrawal window maintain overall acquisition continuity, but short-term quality degradation may occur in local time periods. The EEG measurement module will further perform a fine-grained window-by-window validity assessment within the withdrawal window.

[0040] Within the evacuation window, the validity of each lead is determined to identify valid leads. All valid leads constitute the set of valid leads. A valid lead is a lead that simultaneously meets all of the following preset conditions within the evacuation window: the measured electrode impedance value does not exceed the preset impedance upper limit. (This embodiment uses) The proportion of saturated samples to total sampling points shall not exceed 0.15; the duration of the longest continuous signal loss shall not exceed 5 seconds; the signal distortion ratio shall not exceed 0.10, where distortion refers to sampling points where the signal amplitude exceeds the normal range of the EEG monitoring device or the waveform shows obvious clipping; and the broadband noise level shall not exceed the preset noise upper limit. (This embodiment takes as) Leads that meet all the above conditions are considered valid leads for subsequent analysis. The number of valid leads is denoted as M, where M is a non-negative integer. If M=0, the analysis proceeds to the low-confidence output branch.

[0041] This system will flag data validity when any of the following conditions are detected. Set to 0 and enter the low-confidence output branch: The duration of valid EEG data before monitoring stops is insufficient for the minimum withdrawal window length; there are no valid leads within the withdrawal window; the number of valid time windows is insufficient to meet the minimum requirement; or the quality factor of all valid leads is zero. Among these, This serves as a global data validity indicator; a value of 1 indicates valid data that meets the requirements for subsequent analysis, while a value of 0 indicates that it does not. After entering the low-confidence output branch, the EEG warning confidence state will be... Set to low confidence and classify EEG residual risk level Set to null, undetermined, or "do not output definitive risk level" status. In this case, the residual EEG risk level will not be set to low, medium, or high risk; if existing EEG signals or auxiliary analysis results are still displayed, these results are only used as auxiliary information and will not participate in the final risk level determination.

[0042] In this embodiment, the time closest to the stop monitoring time is used first. We need to analyze the entire time window to avoid missing any key abnormal EEG activity that may occur at the end of the withdrawal window.

[0043] Preset the length of each time window ,in The time window length, in seconds. Set to 10 seconds. ,and To ensure that at least one evacuation window is formed after passing the minimum evacuation window inspection. A complete time window. The number K of complete time windows that can be formed is calculated based on the actual evacuation window length, i.e. ,in This indicates rounding down, where K is the total number of complete time windows.

[0044] Only keep the one closest K complete time windows are used as the actual analysis windows. These windows are numbered sequentially from the 1st to the Kth time window, with the last time window directly adjacent to the stop monitoring time. .

[0045] Generate time window validity flags for each time window Where k is the time window number, , The value can be either 0 or 1. Specifically, if the effective sampling ratio is lower than 0.8, the saturated sample ratio exceeds 0.15, or the continuous missing duration exceeds 1 second within a certain time window, then that time window is marked as invalid. It is not included in the determination of certain risks; otherwise, it is marked as valid. .

[0046] Number of effective time windows That is, all The number of time windows, and the number of built-in minimum effective time windows. Comparison. In this embodiment, Let it be 3. If Enter the low-confidence output branch.

[0047] For each valid lead m ( The lead quality factor is generated by weighted summation based on the signal integrity rate, saturation ratio, distortion ratio, effective sampling ratio, and noise level within the actual analysis window. Among them, signal integrity rate and effective sampling ratio are positive indicators and are directly included in the summation; saturation ratio, distortion ratio, and noise level are negative indicators, and their complements relative to the preset upper limit are taken before being included in the summation (noise level is calculated as its value relative to the noise upper limit). (The complement of the ratio is involved). In this embodiment, the weights corresponding to signal integrity rate, saturation ratio, distortion ratio, effective sampling ratio, and noise level are 0.25, 0.20, 0.20, 0.20, and 0.15, respectively. Quality factor The higher the value, the more reliable the lead.

[0048] If the quality factors of all valid leads are zero, proceed to the low-confidence output branch. If at least one quality factor is greater than zero, adjust the weights of each lead. Let its quality factor be the proportion of the sum of the quality factors of all valid leads, so that higher-quality leads have a greater proportion in multi-lead fusion. The sum of the weights of all leads is kept to 1.

[0049] Specifically, EEG preprocessing and feature extraction are performed for each effective time window and each effective lead. Preprocessing includes the following steps: high-pass filtering with a cutoff frequency of 0.5Hz to remove baseline drift; notch filtering with a center frequency of 50Hz to suppress power line interference; and band-pass filtering with a passband of 0.5–70Hz to preserve the main EEG frequency bands and suppress high-frequency noise. In this embodiment, 256Hz is used for subsequent analysis. All the above filters are implemented using zero-phase-shift finite impulse response (ZPR) filters. Before performing ZPR filtering, the beginning and end of the signal to be filtered are mirror-extended to reduce filtering boundary effects. The mirror extension length is not less than half the filter order, and preferably not less than the number of sampling points corresponding to 1 second. After filtering, the extended portion is removed, and only the filtering result corresponding to the original time window is retained for subsequent analysis.

[0050] Within each lead m and each effective time window k, the following four types of local artifact features are extracted:

[0051] Electromyography artifact index : Perform spectral analysis on the signal within the time window to calculate the proportion of power in the 30–70Hz band to the total power in the 0.5–70Hz band. The higher this proportion, the stronger the electromyographic artifact.

[0052] Eye movement artifact index The detected amplitude exceeds [a certain value] in the low-frequency range of 0.5–4 Hz. The slow wave deflection was counted, and the number of occurrences was normalized by dividing by the time window length. The higher the value, the stronger the eye movement or blink artifact.

[0053] Motion artifact index The first-order difference is calculated point by point for the signal. The mean of the absolute values ​​of the first-order differences is then divided by the standard deviation of the signal amplitude within the same time window to obtain a ratio that reflects the severity of the signal change. The higher the ratio, the more likely the signal is to be affected by motion interference.

[0054] Electrode artifact index : The duration of the detected signal is less than And the amplitude exceeds The pulse-like abrupt changes (corresponding to instantaneous jumps caused by poor electrode contact) are counted, and their occurrence frequency is normalized by dividing by the time window length.

[0055] After truncating each of the four types of indicators to the range [0,1], a weighted sum is calculated according to preset weights to generate the local artifact intensity of lead m within the time window k. In this embodiment, the weights of electromyography, eye movement, motor function, and electrode artifact indicators are 0.30, 0.25, 0.25, and 0.20, respectively.

[0056] Then, the synchronization jump ratio of near-zero delay in multi-lead connections is calculated. Specifically, in a length of... Peak jumps in the signals of each lead are detected within a sliding sub-window, with a sliding step size of 10 ms. For the preprocessed signal of lead m within time window k, the median of the signal amplitude within that time window is calculated. The robust amplitude scale is obtained by multiplying the median of the absolute deviation relative to the median by 1.4826. Local jump amplitude threshold. Set as and The larger one.

[0057] If a sampling point is a local peak point, and its relative to The absolute amplitude is not less than If the sampling point corresponds to a specific time, then the peak jump time of that lead is recorded. For each sliding sub-window, the peak jumps detected in each valid lead are first sorted according to their peak jump times, and then the leads with time differences less than the synchronization artifact detection threshold are selected. Peak jumps are grouped into the same synchronous jump group; only the peak jump with the largest amplitude is retained within the same synchronous jump group for the same lead. If a synchronous jump group contains peak jumps from two or more different valid leads, then a multi-lead synchronous jump is determined to have occurred within that sliding sub-window. The proportion of leads with synchronous jumps to the total number of valid leads M is calculated, and then the average value is taken over all sliding sub-windows as the proportion of multi-lead synchronous jumps within the current time window k. In this embodiment, Set to 50 milliseconds. Set to 5 milliseconds.

[0058] Specifically, the intensity of local artifacts is combined with the synchronous jump ratio of multiple leads to generate homologous artifact features for each lead within the current time window. :

[0059] ;

[0060] in In this embodiment, the local artifact weighting coefficient is used. Take as , This represents the proportional weighting of synchronous jumps. This formula reflects that the characteristics of homologous artifacts are composed of both single-lead local artifacts and multi-lead synchronous jumps.

[0061] Subsequently, based on lead weights We obtain the window-level homology artifact features of the k-th time window by weighted summation of the homology artifact features of each lead. .

[0062] To avoid extreme values ​​affecting subsequent judgments, robust normalization is performed on the window-level homogeneous artifact features within the effective time window. Normalization is based on the features within the effective time window. The median is used as the reference center, and the difference between the 75th quantile and the median is used as the scale reference; to avoid the normalization failure due to a scale of zero, the actual scale is the difference between the 75th quantile and the median and the lower limit of the normalization scale. The larger of the two (in this embodiment) (Take 0.01). Subtract the reference center from the features of each window-level homologous artifact, divide by this scale, and truncate to the range [0,1] to obtain the normalized window-level homologous artifact features. The closer the value is to 1, the more likely the time window is to be significantly affected by artifacts. Homologous artifact characteristics for each lead. Using the same reference center and scale as the window level, the same normalization is performed to obtain the lead-level normalized homologous artifact features. The value is truncated to [0,1] and used for artifact correction calculation in the subsequent EEG calculation module.

[0063] In this embodiment, a lenient initial screening of abnormal events is performed for each lead and each effective time window, using a joint rule-based detection method based on amplitude, duration, and frequency band energy. Specifically, spikes, spike-slow-wave complexes, abnormal high-frequency activity, and abnormal clustering activity are identified.

[0064] Among them, spikes are defined as waves with a duration between 70 and 200 milliseconds and a peak amplitude exceeding [a certain value]. Transient waveforms; spikes are transient waveforms with a duration between 20 and 70 milliseconds and a peak amplitude exceeding [a certain value]. The transient waveform; the spike-slow composite wave is a slow wave component that appears within 500 milliseconds after a spike or twist wave, with a duration greater than 200 milliseconds and a peak amplitude exceeding [missing value]. .

[0065] For abnormal high-frequency activity, short-time band power is calculated within the 30–70 Hz frequency band. In this embodiment, the short-time sliding window length is 200 milliseconds, and the sliding step size is 50 milliseconds. If the instantaneous power in the 30–70 Hz frequency band within a certain short-time sliding window is not less than three times the corresponding baseline power, it is identified as an initial screening abnormal high-frequency activity. The corresponding baseline power is preferentially calculated based on the low-artifact baseline segment within the withdrawal window. If the low-artifact baseline segment is insufficient, the robust statistical power of the same lead within the effective time window of the withdrawal window is used as the baseline power. If a stable baseline power still cannot be obtained, the baseline power calibrated at the device factory or the baseline power calibrated offline using historically labeled EEG data is used. For abnormal clustering activity, a clustering detection sliding window with a length of 2 seconds and a step size of 0.5 seconds is set. If the number of events satisfying any of the above initial screening rules within any clustering detection sliding window reaches 3 or more, it is identified as an initial screening abnormal clustering activity.

[0066] The thresholds mentioned above are all relatively low initial screening thresholds. Candidate events that satisfy any initial screening rule and their event-level abnormality intensity are all retained, forming an initial screening abnormal event set that has not been filtered by the final threshold. The final screening is completed by the adaptive threshold in the EEG calculation module.

[0067] For each initially screened abnormal event, calculate its event-level abnormality intensity. Specifically, the peak amplitude, frequency band energy, and duration of the event are each normalized by dividing by their respective reference values, and then weighted and summed using fixed preset weights to generate the final value. In this embodiment, the amplitude reference value is taken as... The reference value for frequency band energy is the average power of the 0.5–70Hz band within the current time window of the same lead multiplied by 0.1 seconds, and the reference value for duration is 0.1 seconds. The weights of the amplitude, frequency band energy, and duration items are 0.4, 0.35, and 0.25, respectively. Each weight is a fixed preset value and is not adjusted with changes in the signal.

[0068] Finally, the event-level abnormality intensities of all initial screening abnormal events within the same lead m and the same time window k are summed to obtain the cumulative initial screening abnormality intensity of the lead-level window.

[0069] The EEG calculation module 200 is used to generate an adaptive abnormality threshold for the current time window based on the homology artifact features of the current time window and the abnormal residual amount of the previous time window, and then screen the initial abnormal events to obtain candidate abnormal EEG events; generate a brain region lead topology map based on the effective leads, and then calculate the propagation morphology compatibility of the candidate abnormal EEG events between leads; perform protection correction based on the propagation morphology compatibility to obtain corrected homology artifact features, and calculate the coupling relationship between the candidate abnormal EEG events and the corrected homology artifact features, and then generate a high-confidence abnormality load for the current time window.

[0070] like Figure 2 As shown, in this embodiment, the normalized window-level homogeneous artifact features are first analyzed. Select from the effective time window Effective time windows within the lowest 30% quantile range form a low-artifact baseline segment set. When there are tandem values ​​among effective time windows at the lowest 30% quantile boundary, the effective time windows with the same boundary value are also included in the low-artifact baseline segment set.

[0071] If the low-artifact baseline segment is empty, the built-in default event-level anomaly threshold is used. (In this embodiment, the value is 0.01), default lead-level window anomaly scale (In this embodiment, we take 0.1) and the default anomalous residual normalization scale. (In this embodiment, we take 1), and mark the baseline sufficiency. Set to 0. Where, This indicates that both the base threshold and the normalized scale are generated from low-artifact baseline segments within the withdrawal window. This indicates that at least one base threshold or normalization scale uses the default value.

[0072] If the low artifact baseline segment is not empty, first use the flag to apply the default parameters. Initialize to 0. For the event-level baseline anomaly threshold, a sample set is constructed by collecting the event-level anomaly intensities of all initially screened anomaly events from the low-artifact baseline segment. The sample size is not less than the minimum baseline event count. Then, existing robust statistical methods are used to generate event-level basic anomaly thresholds. Specifically, the median of the sample set is used as the central reference, and the median absolute deviation is used as the dispersion reference. When the median absolute deviation is 0 or lower than the preset stability lower limit (0.01 in this embodiment), the preset stability lower limit is used to participate in the threshold generation.

[0073] For each lead m, the window-level anomaly intensity normalization scale If the number of available samples in the low-artifact baseline segment of the lead is not less than And the 90th quantile is greater than the lower stability limit. Then the 90th quantile value is taken as Otherwise adopt and will Set to 1. For window-level anomaly residual normalization scale If the number of usable samples with the cumulative initial screening abnormal intensity in each lead window of the low artifact baseline segment is not less than And the 95th percentile is greater than Then the 95th quantile value is taken as Otherwise adopt and will Set to 1. In this embodiment, Take it as 3. Set it to 0.001.

[0074] After generating the basic threshold and normalized scale as described above, if ,Will Set to 1; if ,Will If set to 0, subsequent processing will continue, but the final output will be an additional downgraded confidence state.

[0075] Before entering the first time window, the abnormal residual amount Initialize to 0, normalize abnormal residuals Initialize to 0, set the initial adaptive threshold Initialization is based on the basic anomaly threshold. And set the basic abnormal residual decay time constant. ,in This represents the base rate of natural decay of the anomalous residue when no new high-confidence anomalous loads are introduced, expressed in seconds. In this embodiment, Set to 30 seconds, of which The high-confidence anomalous load will also be initialized as... This is used to calculate the load increase trend in the first time window.

[0076] Specifically, a directed lead adjacency graph is generated based on the valid EEG leads, serving as a brain region lead topology graph. Each node in the graph represents a valid lead, and directed edges represent the directional relationship that an abnormal event in one lead may propagate to another lead. The edge weight represents the relative reliability of this directional relationship in the propagation analysis.

[0077] The number of effective leads in actual in-hospital EEG monitoring varies depending on the scenario. Simplified bedside monitoring typically has 2–8 leads, while standard EEG requires 19 or more. If the number of effective leads (M) is less than 2, inter-lead propagation edges cannot be formed. In this case, inter-lead propagation morphology compatibility calculation is not performed, and the propagation morphology compatibility of all time windows is set to 0 by default. However, single-lead focal repetitive abnormality protection and abnormal residual quantity calculation are still allowed. When there are fewer than 2 effective leads, the requirements for high-risk propagation evidence cannot be met because inter-lead propagation compatibility cannot be calculated. However, a medium-risk result can be output based on focal repetitive abnormalities and residual duration, with a maximum output of medium risk.

[0078] If the number of valid leads M is no less than 2, candidate directed edges are automatically generated based on the spatial location of the leads and their functional correlation with brain regions. Specifically, the corresponding standardized scalp spatial coordinates are obtained for each valid lead. and brain region labels Standardized scalp spatial coordinates can be provided by the International 10-20 or a lead location table preset by the EEG monitoring device; brain region labels include at least one of the frontal, central, temporal, parietal, occipital, and midline regions.

[0079] For any two distinct valid leads i and j, calculate their standardized spatial distance. .like Not greater than the spatial adjacency distance threshold Or, leads i and j belong to the same or adjacent brain regions and Not greater than the functional adjacency distance threshold If the two leads are approximately from the same origin in the left and right hemispheres, then lead i and lead j are identified as candidate related lead pairs. In this embodiment, Take it as 0.35 times the standardized scalp radius. It is taken as 0.55 times the standardized scalp radius.

[0080] For each candidate associated lead pair (i,j), two opposite-direction candidate directed edges are established simultaneously, one from i to j and the other from j to i. The actual effective propagation direction is then determined solely by the order of event arrival times. For simplified bedside monitoring scenarios with no more than four effective leads, bidirectional candidate directed edges are established between every pair of all effective leads.

[0081] Set initial edge weights for each candidate directed edge. The initial edge weights are related to... Proportional; for candidate edges that simultaneously satisfy functional correlation or left-right homology, multiply the spatial weights by a functional enhancement factor. Functional relationships include those between leads in the same brain region, those between leads in adjacent brain regions, and those considered to have spatial propagational relationships in clinical EEG monitoring; left-right homology relationships include those between leads in the left and right hemispheres that correspond in name or are approximately mirror images of each other in spatial location. These relationships can be determined by a pre-set lead topology table on the device or a lead relationship table configured by the department. In this embodiment, Set it to 0.35. The initial edge weight is set to 0.5. The initial edge weights of all candidate directed edges are normalized so that the sum of the initial edge weights of all candidate directed edges is 1. If the monitoring device cannot provide standardized scalp spatial coordinates, the 10-20 template coordinates corresponding to the lead name are used; if the lead name cannot be mapped to template coordinates either, bidirectional candidate directed edges are established between each pair of valid leads, and uniform edge weights are used.

[0082] For each directed edge (i,j), initialize the following reference propagation parameters. The initial state uses relaxed parameters to accommodate a wider range of propagation delays and morphological changes in the early stages: reference arrival delay and reference arrival delay standard deviation All are initialized to the maximum allowable delay. Half of (in this embodiment) This makes the initial matching window wider; refer to the mean vector of morphological changes. Initialized as a four-dimensional zero vector, its four components correspond to amplitude variation, duration variation, frequency band energy variation, and polarity-phase consistency, respectively; reference morphological variation covariance matrix. Initialize to diagonal elements of Diagonal matrix (in this embodiment) (Set to 4.0) to make the initial morphological distribution wider.

[0083] In this embodiment, for the k-th time window ( Based on the current normalized window-level homogeneous artifact features Normalized anomaly residual amount compared to the previous time window Generate the initial adaptive threshold for the current time window.

[0084] The core rule is: when the homologous artifact characteristics of the current time window increase, the event filtering threshold is increased to reduce the probability of artifacts being misjudged as abnormal events; when the abnormal residual amount of the previous time window increases, the event filtering threshold is decreased to enhance the sensitivity to continuous abnormal activity.

[0085] The core formula for this closed-loop mechanism is: ;

[0086] in, This means truncating x to the range [a, b]. This is the initial adaptive threshold before the rate of change constraint for the k-th time window. The basic event-level exception threshold, This is the artifact adjustment coefficient. This is the residual adjustment coefficient. This is the lower limit allowed by the initial adaptive threshold. This is the upper limit allowed by the initial adaptive threshold. In this embodiment, Let it be 1.5. Let it be 1.5. Take as , Take as .in , .

[0087] To avoid abrupt changes in the threshold between adjacent time windows, a rate of change constraint is further applied to this threshold. The specific formula is as follows:

[0088] ;

[0089] Then limit the results to Within the range. For adaptive anomaly threshold, To preset a safety ratio, to meet In this embodiment, The value is set to 0.3. This mechanism constitutes an important closed-loop adaptive mechanism in this embodiment: the stronger the artifact, the more conservative the screening; the higher the residue, the more sensitive the screening.

[0090] In this embodiment, an adaptive anomaly threshold for the current time window is used. The initial screening of abnormal events within each lead is performed. Only event-level abnormalities are considered. Events that reach or exceed the current adaptive threshold, i.e. Only then were they retained as candidate abnormal EEG events.

[0091] Each candidate abnormal EEG event must include at least the event start time. Peak time Peak amplitude Duration Frequency band energy Polarity marking (Value +1 indicates positive polarity, -1 indicates negative polarity), phase characteristics (Instantaneous phase at peak time, in radians) and event-level anomaly intensity The phase features can be obtained using existing methods for extracting the instantaneous phase of EEG signals. In this embodiment, the preprocessed signals of the leads containing candidate abnormal EEG events are first bandpass processed according to the event's dominant frequency band, and then the analytical signal is obtained through Hilbert transform. The instantaneous phase corresponding to the event's peak moment is taken as the location feature. When the event's dominant frequency band cannot be stably determined, a preset analysis frequency band corresponding to the event type can be used to extract the instantaneous phase.

[0092] Count the number of candidate abnormal events for each lead m within the current time window k, and sum the event-level abnormality intensities of these events to generate the lead-level window cumulative candidate abnormality intensity. .

[0093] To facilitate subsequent calculations of the near-zero delay coupling between artifacts and anomalies, the current time window is uniformly divided into multiple sub-windows, each with a length of [missing information]. ,in This represents the length of the sub-window, in seconds. In this embodiment, Set to 1 second, each time window contains Each child window. ( The summation of candidate anomaly intensities for lead m within a given window is denoted as the sub-window level candidate anomaly intensity. Simultaneously, within each sub-window, the local artifact intensity is calculated using the same four-category artifact index extraction and weighting method as at the time window level. This intensity is then combined with the synchronization jump ratio of multi-leads within that sub-window to generate sub-window-level homologous artifact features. These features are then normalized using the same reference center and scale as at the window level to obtain sub-window-level normalized homologous artifact features. This is used for subsequent artifact-anomaly coupling analysis.

[0094] Further statistics are compiled on the total number of candidate abnormal events for all leads within the current time window. .like This indicates that there are no candidate abnormal events that have passed the adaptive threshold screening within the current time window. At this point, the propagation morphology compatibility will be considered. Localized repetitive abnormality protection value Effective compatibility Window-level coupling index and high-confidence anomalous loads All values ​​are set to 0, and the abnormal residual decay time constant of this time window is denoted as... During this time window, the inter-lead reference propagation parameters are not updated; all reference delay and reference morphology parameters remain in the state of the previous time window. Subsequently, the process transitions to the state update module, which updates the parameters according to a unified formula for updating abnormal residual values. and .

[0095] Specifically, when there are candidate abnormal events in the current time window ( If the set of guided edges is not empty, perform candidate exception event matching for each guided edge.

[0096] For a directed edge from lead i to lead j, find matching event pairs between candidate events in lead i and lead j. A matching event pair must satisfy three conditions simultaneously: first, the two events originate from the starting lead i and ending lead j of the directed edge, respectively; second, the peak time of the event in the ending lead is equal to the peak time of the event in the ending lead. No earlier than the peak time of the initiation lead event Furthermore, the time difference between the two is not less than the synchronization determination threshold. (In this embodiment) Take 5 milliseconds, the time difference is less than (Those that are considered synchronous events rather than directed propagation events are not included); third, the peak time difference between the two does not exceed the maximum allowable delay preset. (This example uses 50 milliseconds).

[0097] If an edge has multiple matching event pairs, the one with the closest current arrival delay to the reference arrival delay of the previous time window will be selected first. The event pair is selected as the optimal matching event pair; if two or more event pairs are exactly equal in distance to the reference delay, then the event pair with the larger sum of the event-level anomaly intensities is further selected. The current waveform propagation arrival delay is calculated based on the optimal matching event pair. This is the difference in peak time between the optimal matching events in the endpoint lead and the origin lead, and the specific formula is as follows:

[0098] ;in and These are the peak times of the events in the starting and ending leads of the optimal matching event pair, respectively. Only when the propagation direction, maximum delay limit, and compatibility with the previous reference delay are simultaneously satisfied (i.e., ... Only directed edges that are valid leads are included in the set of valid leads for the current time window. .

[0099] If the set of adjacency edges of the guide is empty, or if no valid guide edges are formed in the current time window, then the propagation pattern compatibility of the current time window will be checked. Set it to 0, and continue to perform focal recurrence anomaly protection and subsequent residual calculation.

[0100] In this embodiment, for each valid lead edge (i,j) within the current time window, the time delay soft-match value is calculated, and the waveform morphology change vector between leads is extracted.

[0101] The time delay soft match value is used to measure the arrival delay of the current waveform propagation. Reference arrival delay compared to the previous time window The degree of consistency between the two. Using the difference between the two as input, and referencing the standard deviation of arrival delay. With stable terms The sum is used as a scale (to avoid abnormal fluctuations in matching values ​​when the standard deviation is too small, in this embodiment...) (Taken to 0.5 milliseconds), the value range generated by Gaussian kernel mapping is: Time delay soft matching value The specific formula is as follows:

[0102] ;

[0103] The smaller the difference, The closer to 1, the larger the difference, the closer to 0.

[0104] For the optimal matching event pair of each valid lead edge, the morphological change of the endpoint lead event relative to the starting lead event is extracted to form a four-dimensional waveform morphological change feature vector. The four components are as follows: amplitude variation component, which is the logarithmic ratio of the absolute values ​​of the peak amplitudes of the endpoint lead and the origin lead; duration variation component, which is the logarithmic ratio of the durations of the endpoint lead and the origin lead; frequency band energy variation component, which is the logarithmic ratio of the frequency band energy of the endpoint lead and the origin lead; and polarity-phase consistency component, which is the cosine of the product of the polarity markings of the origin and endpoint leads multiplied by the difference in their peak instantaneous phases (the value ranges from -1 to 1, with values ​​close to 1 indicating consistent polarity and close phase, consistent with typical propagation characteristics). In the calculation of the above three logarithmic ratio components, stabilizing terms are introduced into the numerator and denominator to avoid the denominator being too small (in this embodiment, the stabilizing terms corresponding to amplitude, duration, and frequency band energy are respectively taken as...). , , ).

[0105] To avoid excessive differences in dimensions and numerical ranges among different components, the amplitude variation component, duration variation component, and frequency band energy variation component are truncated with preset upper and lower limits before calculating morphological evolution compatibility, and scale normalization is performed based on the reference morphological change distribution; the polarity-phase consistency component is itself limited to a preset range. After the above processing, all four components participate in the morphological evolution compatibility calculation in a dimensionless form, thereby reducing the impact of scale inconsistencies among different components on the results.

[0106] Subsequently, the reference morphological change distribution formed after the end of the previous time window is used, including the reference morphological change mean vector. and reference morphological change covariance matrix Calculate the Mahalanobis distance between the current waveform morphology change feature vector and the reference mean, under the reference covariance matrix after regularization. Regularization is achieved by adding regularization terms (regularization coefficients) to the diagonal of the covariance matrix. In this embodiment, the value is set to 0.1 to avoid matrix degradation or irreversibility. A morphological evolution compatibility value ranging from (0,1] is generated using a monotonically decreasing Gaussian mapping, with the specific formula being:

[0107] ;

[0108] in The Mahalanobis distance is the mapping scale parameter, which is set to 3.0 in this embodiment. The smaller the Mahalanobis distance, the higher the morphological evolution compatibility.

[0109] In this embodiment, for each valid lead edge (i,j), the edge-level propagation morphological compatibility is generated by combining the time delay soft-match value and the morphological evolution compatibility. The time delay soft-match value reflects whether the propagation arrival time is reasonable, while the morphological evolution compatibility reflects whether the waveform changes before and after propagation are reasonable. The edge-level propagation morphological compatibility is generated by multiplying these two values ​​by their respective exponential weights, using the following formula:

[0110] ;

[0111] in To match the exponential weights for time delay, This represents the weighting of the morphological evolution index. In this embodiment, Set it to 1.0. Set it to 1.5. This indicates that in this embodiment, the consistency of morphological evolution has a greater impact on the determination of propagation authenticity than time delay matching. Different embodiments can adjust the relative magnitudes of these two factors according to the clinical scenario. If an edge simultaneously possesses stable propagation time delay and consistent waveform evolution characteristics, its edge-level propagation morphological compatibility is high; if either aspect is significantly mismatched, the compatibility decreases.

[0112] like Figure 3 As shown, in this embodiment, when there are valid lead edges in the current time window, a weighted average is generated based on the edge-level propagation morphology compatibility of each valid lead edge, according to the edge weight of each edge, to produce window-level propagation morphology compatibility. The higher the edge weight, the greater its contribution to the window-level propagation morphology compatibility. If the sum of the edge weights involved in the calculation is zero, then... Set to 0. This process is used to determine whether abnormal events in the current time window exhibit a propagation pattern that conforms to the lead topology and waveform evolution rules.

[0113] Some true EEG abnormalities may be confined to a single lead or a few leads, and may not necessarily form a stable inter-lead propagation path. To prevent such focal repetitive abnormalities from being completely suppressed by propagation compatibility mechanisms, a focal repetitive abnormality protection mechanism is established.

[0114] For each valid lead m, backtracking forward from the current time window in the valid time window sequence, we determine whether there are candidate abnormal EEG events for each valid lead within each valid time window, and count the number of valid time windows with consecutive candidate abnormal EEG events. If an invalid time window occurs or if no candidate abnormal EEG event is found in a valid lead within a valid time window, the continuous counting is interrupted.

[0115] When the number of valid time windows for consecutive candidate abnormal EEG events is less than the preset number of focal repetition windows, the valid lead is determined not to meet the focal repetition abnormality condition, and the lead-level focal repetition abnormality protection value for that valid lead is set to 0. (Preset number of focal repetition windows) The value is set to 3; in other implementations, the effective time window can be set to 2 to 5 depending on the length of the time window and the monitoring requirements.

[0116] When the number of valid time windows with consecutive candidate abnormal EEG events reaches the preset number of focal repetition windows, the valid lead is determined to meet the focal repetition abnormality condition. The system extracts the cumulative candidate abnormality intensity of the lead-level windows within the aforementioned consecutive valid time windows, and calculates the arithmetic mean of the cumulative candidate abnormality intensity of the lead-level windows to obtain the average abnormality intensity of the valid lead; then, the average abnormality intensity is normalized according to the window-level abnormality intensity of the corresponding lead. Perform normalization and limit the normalization result to the range of 0 to 1.

[0117] The system further generates a repeatability contribution value based on the number of effective time windows containing consecutive candidate abnormal EEG events. Specifically, when the number of effective time windows containing consecutive candidate abnormal EEG events reaches a preset number of focal repetition windows, the repeatability contribution value takes a preset initial contribution value; as the number of effective time windows containing consecutive candidate abnormal EEG events continues to increase, the repeatability contribution value increases accordingly; when the number of effective time windows containing consecutive candidate abnormal EEG events reaches a preset upper limit for continuous repetition window saturation... When the time is set, the continuous contribution value is 1. In this embodiment, the preset initial contribution value is 0.2, and the preset upper limit of continuous repetition window saturation is 10.

[0118] The normalized average abnormality intensity is multiplied by the repetitive persistence contribution value, and the result is limited to between 0 and 1 to obtain the lead-level focal repetitive abnormality protection value for that effective lead. The window-level focal repetitive abnormality protection value for the current time window is the maximum value among the lead-level focal repetitive abnormality protection values ​​for all effective leads.

[0119] Subsequently, the compatibility of propagation modes will be considered. and focal repetitive abnormality protection value The larger one is considered as the effective compatibility. This is used for subsequent artifact protection correction and anomalous residue calculation. The design intent of the focal repetitive anomaly protection mechanism is to prevent focal anomalies from being completely ignored, allowing them to contribute to the calculation of anomalous residue and intermediate risk identification.

[0120] In this embodiment, the homologous artifact features are protectively modified based on effective compatibility. The core idea is that when anomalous activity within a certain time window exhibits high propagation morphology compatibility, or strong focal repetitive anomaly protection, these anomalies should not be completely suppressed simply because the homologous artifact features are high. Therefore, the suppressive effect of homologous artifact features on anomalies is appropriately reduced based on effective compatibility.

[0121] For each lead m, the lead-level normalized homology artifact features are corrected using effective compatibility, as shown in the following formula:

[0122] ;

[0123] in To correct the characteristics of artifacts of the same origin, This refers to the normalized homology artifact characteristics of the lead levels before correction. For effective compatibility, The minimum artifact retention ratio is set at 0.2 in this embodiment. Higher effective compatibility results in lower correction values ​​for homologous artifact features, but not lower than the original value. To ensure that the risk of artifacts is not completely eliminated, the sub-window level normalized homogeneous artifact features are also optimized to support subsequent artifact-anomaly coupling analysis. The same correction coefficients are used to perform the same correction, resulting in the corrected homologous artifact features for each lead and each sub-window.

[0124] In this embodiment, within the current time window k, the near-zero delay coupling index of artifacts and anomalies is calculated based on the corrected sub-window level homologous artifact feature sequence and the sub-window level candidate anomaly intensity sequence.

[0125] Preset maximum correlation delay range (The unit is the number of sub-windows, which is taken as 2 in this embodiment), and convert it into a sub-window offset range. For each allowed integer offset, determine the set of sub-window indices that still have overlapping samples. If the number of overlapping sub-windows at a certain offset is less than the minimum number of related samples... (In this embodiment, it is set to 4), then this offset does not participate in the correlation calculation; if none of the offsets meet the minimum sample size requirement, then the coupling index of lead m is... Set to 0. Even if there is a valid offset, if the variance of the corrected artifact sequence or candidate anomaly intensity sequence is too low (the lower limit of variance in this embodiment is set to 0), the error will be considered as follows: (If the value is below the lower limit, it indicates that the correlation calculation is unreliable). Similarly, Set it to 0.

[0126] When the computational conditions are met, for each effective offset, the Pearson correlation coefficient is calculated between the corrected sub-window level homologous artifact feature sequence and the sub-window level candidate anomaly intensity sequence within the overlapping interval. The maximum positive correlation value under all effective offsets is taken as the lead-level coupling index. If the correlation coefficients under all offsets are non-positive, then Set to 0. A higher value indicates that candidate anomalies are more likely to occur simultaneously with artifacts, and therefore more susceptible to artifact contamination. Finally, according to lead weights... The window-level coupling index is obtained by weighted summation of the coupling indices of each lead. .

[0127] In this embodiment, a highly reliable anomalous load is generated based on the lead-level coupling index, propagation morphology compatibility, and window-level focal repetitive anomaly protection value.

[0128] To prevent isolated but significant candidate EEG abnormalities from being completely suppressed due to insufficient propagation morphology compatibility or the lack of focal repetition evidence, a protection clause for isolated severe abnormalities is set. Specifically, when the lead-level normalized candidate abnormality intensity of lead m... Reaching the preset severe abnormal intensity threshold (In this embodiment, the value is taken as 1.0), and the coupling index of this lead is... Not exceeding the preset isolation judgment coupling limit (In this embodiment, the value is set to 0.3). When lead m is identified as having an isolated severe anomaly within the current time window, a minimum credible anomaly contribution limit is set for it. (In this embodiment, it is set to 0.05), so that such isolated severe anomalies can still enter the anomaly residual amount update process.

[0129] A baseline-corrected sigmoid-type smoothing mapping function is uniformly used to generate the artifact suppression factor, propagation enhancement factor, and focal protection enhancement factor. This mapping function is defined as follows: ,in For the standard sigmoid function , A positive parameter controls the steepness of the mapping. x is non-negative and does not exceed 1. This mapping guarantees that the output is confined to the range [0,1], with the output being 0 when the input is 0 and monotonically approaching 1 as the input increases. The artifact suppression factor is determined by the coupling exponent via... Generation (stronger coupling, stronger inhibition), the propagation enhancement factor is determined by the propagation morphology compatibility through The focal protection enhancement factor is generated (the higher the compatibility, the stronger the enhancement), which is derived from the focal repetitive abnormality protection value through... Generation (used in cases where a stable propagation path cannot be formed but continuous focal anomalies exist). In this embodiment, Take it as 5. Take as , Let it be 4.

[0130] The high-confidence anomalous load at the lead level is determined by three parts, as shown in the following formula:

[0131] ;

[0132] Based on this, for leads that meet the above-mentioned isolated severe anomaly protection conditions, the lead-level high-confidence anomaly load is taken as the result calculated by the above formula and the lower limit of the lowest-confidence anomaly contribution. The larger of the two, that is: when ≥ and ≤ hour, for For leads that do not meet this condition, this lower limit is not applied.

[0133] in, The cumulative candidate abnormality intensity of the m-th lead within the k-th time window is calculated. For lead-level coupling index, For window-level propagation compatibility, This is the protection value for window-level localized recurring anomalies.

[0134] In this formula, the artifact suppression factor uses the lead-level coupling index. This is because different leads are contaminated to varying degrees by artifacts, requiring differentiated inhibition to be applied lead by lead; window values ​​are used for propagation enhancement factors and focal protection enhancement factors. and This is because the propagation morphology compatibility and focal repetition anomaly protection values ​​are used to judge the global anomaly propagation pattern within the current time window and are applicable to all leads within the window. When both the propagation morphology compatibility and focal repetition anomaly protection values ​​are 0, and the current lead does not meet the above-mentioned isolated severe anomaly protection conditions, the lead-level high-confidence anomaly load is 0, indicating that candidate anomaly events with neither inter-lead propagation morphology nor repetitive focal characteristics, and which do not reach the intensity of severe anomalies, do not form an anomaly residual load; if the isolated severe anomaly protection conditions are met, the lead-level high-confidence anomaly load is not less than Subsequently, based on the lead weights The high-confidence anomalous load of each lead is obtained by weighted summation and fusion. .

[0135] The state update module 300 is used to control the decay of the abnormal residual amount in the previous time window and to superimpose the high-confidence abnormal load of the current time window until the final abnormal residual amount representing the EEG physiological state is obtained after traversing all time windows.

[0136] Specifically, in this embodiment, based on the effective compatibility of the current time window Abnormal clustering degree Highly reliable abnormal load rising trend and continuous stable window ratio Generate the abnormal residual decay time constant for the current time window. .

[0137] Abnormal clustering degree The total number of candidate abnormal events in all leads within the current time window Divide by the upper limit of the clustering reference And truncate to [0,1] to obtain (in this embodiment) Take as ).

[0138] Highly reliable abnormal load upward trend The increase in high-confidence anomalous load in the current time window relative to the previous time window (i.e. (Take the non-negative part) divided by the load rise reference scale And truncate to [0,1] to obtain (in this embodiment) (Set to 2.0); if the current time window is the first time window, then use the initial value. If an invalid time window exists at the end of the calculation, it is skipped, and the nearest valid time window is selected from the previous time window.

[0139] Continuous stable window ratio Used to reflect whether recent brainwave activity is in a relatively stable state. The most recent time window before the current time window. Within each effective time window (in this embodiment) Set it to 5; if the number of previous effective time windows is insufficient. Then, using the actual quantity as the denominator, we calculate the proportion of windows that simultaneously meet the following three conditions: the normalized window-level homogeneous artifact features of the window. Less than 0.3; the total number of candidate anomalous events within this window is 0; the high-confidence anomalous load of this window. Less than 0.1. The value range is [0,1], when k=1 or the number of historical valid windows is 0. It is 0.

[0140] Generate by combining the above factors :

[0141] ;

[0142] in This is the sigmoid function, whose range is the open interval (0,1), therefore Strictly fall into Within the scope, in terms of engineering implementation, it can be regarded as . , , , These are weighting coefficients for effective compatibility, the degree of anomalous clustering, the upward trend of high-confidence loads, and the proportion of continuous stable windows, respectively. Higher effective compatibility, more pronounced anomalous clustering, and a more significant increase in high-confidence anomalous loads are associated with these factors. The closer This indicates slower residual decay; a higher proportion of a continuous stable window indicates that... The closer This indicates that the residual decay is faster. In this embodiment, Set it to 2.0. Let it be 1.5. Let it be 1.5. Set it to 2.0. Set to 10 seconds. Set to 1200 seconds, of which ;when or When it is 0, = .

[0143] In this embodiment, the high-confidence anomalous load of the current time window This represents the accumulated high-confidence abnormal load within the k-th time window, rather than the load per unit time. After artifact suppression and propagation enhancement processing in the above-mentioned high-confidence anomalous load generation process, its dimensions are consistent with the cumulative candidate anomalous intensity of the lead-level window. Same (i.e., dimensionless normalized intensity value). Due to abnormal residual amount From each They are accumulated after exponential decay, and have the same dimensions, so they can be added directly.

[0144] The core update formula is:

[0145] ;

[0146] in This represents the amount of abnormal residual data after the k-th time window ends. This indicates the amount of abnormal residue after the previous time window ended. Indicates the length of the time window. This represents the abnormal residual decay time constant of the current time window. This represents the preset upper limit of the residual amount (10 in this embodiment). The formula indicates that the anomalous residue left over from the previous time window will decay over time, while the high-confidence anomalous load generated in the current time window will be added to the residual amount. Subsequently, the anomalous residue is normalized according to the normalization scale. Will Converted to normalized abnormal residual amount (Pick The abnormal residual quantity is a brain physiological state quantity generated by the candidate abnormal EEG events, high confidence abnormal load, propagation morphology compatibility and artifact coupling inhibition results measured before the cessation of monitoring. It is used to characterize the continued impact of abnormal EEG activity before the cessation of monitoring on subsequent unmonitored periods.

[0147] In this embodiment, the propagation pattern compatibility within the current time window Coupling index Normalized window-level homogeneous artifact features and high-confidence abnormal load Only after all calculations are complete is it determined whether to update the reference propagation parameters. This delayed update mechanism ensures that the reference parameters for the current time window come only from past information and are not negatively affected by the current calculation results, thus avoiding circular self-verification and preventing invalid time windows from being generated. However, the residual amount still decays according to the actual time interval.

[0148] First, it is assumed that the reference arrival delay, reference arrival delay standard deviation, reference shape change mean vector, and reference shape covariance matrix of all directed edges remain in the state of the previous time window. Then, it is determined whether the current time window is a high-confidence reference update window. A preset upper limit for reference update artifact coupling is set. Reference update propagation compatibility lower bound Reference update upper limit of artifact features and reference update high-reliability load lower limit In this embodiment, Set it to 0.3. Set it to 0.5. Set it to 0.6. Set to 0.1. A time window is allowed to be used to update reference propagation parameters only if it simultaneously meets the following four conditions: window-level coupling index. Not higher than Window-level propagation morphology compatibility Not less than Window-level normalized homology artifact features Not higher than Highly reliable anomaly load in the current time window Not less than .

[0149] If the current time window meets the preset high-reliability reference update conditions, based on the propagation morphology compatibility exceeding... The degree of generation of reference parameter update rate Specifically, the update rate is equal to the preset maximum update rate. (In this embodiment, it is taken as 0.3) multiplied by and The ratio; due to The update rate does not exceed .

[0150] For the current set of valid lead edges For each directed edge in the array, the reference propagation parameters are updated using an exponentially weighted moving average method, with the weights determined by the aforementioned update rate. Decision. The reference arrival delay update formula is:

[0151] ;

[0152] The reference arrival delay variance, reference morphological change mean vector, and reference morphological covariance matrix are all updated using the same form of exponentially weighted moving average. That is, the updated parameters are the weighted average of the parameters from the previous time window and the current observations (where the covariance update uses the outer product of the current morphological change vector and the updated mean vector as the current observation). A minimum delay variance lower bound is set when updating the reference arrival delay variance. (In this embodiment, it is taken as) To avoid excessively tight matching due to excessively small variance, after updating the reference morphological covariance matrix, minimum eigenvalue protection is performed, i.e., eigenvalue decomposition is performed on the covariance matrix, and if the minimum eigenvalue is less than a preset lower limit... (In this embodiment, it is taken as 0.01). The covariance matrix is ​​then reconstructed after replacing the eigenvalue with this lower bound, ensuring it remains positive definite and avoids degradation. For directed edges not included in the current set of valid leads, no reference parameter update is performed, and the state of the previous time window is maintained. The updated parameters are only used for subsequent time windows.

[0153] In this embodiment, after the k-th time window ends, the normalized abnormal residual amount is... Feedback is sent to the (k+1)th time window. When processing the next time window, the normalized window-level homogeneous artifact features of the next time window are also used. Normalized anomaly residual amount in the k-th time window and the adaptive threshold for the k-th time window A new initial adaptive threshold is generated according to the above adaptive threshold generation rules. This process completes a closed-loop feedback mechanism that affects the sensitivity of candidate anomaly screening in the next time window, thereby improving sensitivity under continuous anomaly conditions and reducing the risk of false alarms under artifact enhancement conditions.

[0154] The risk decision module 400 is used to output the risk warning level of EEG residue before in-hospital transfer based on the final abnormal residual amount through an adaptive decision tree.

[0155] In this embodiment, after all time windows have been processed, the final abnormal residual amount after the last time window ends is obtained. The system outputs a warning level of residual EEG risk before in-hospital transfer using an adaptive decision tree.

[0156] Specifically, the risk decision module 400 outputs three fields: EEG residual risk level, EEG warning confidence level, and abnormal residual trend indication. The EEG residual risk level can be set to low risk, medium risk, high risk, or indeterminate; the EEG warning confidence level can be set to normal confidence, downgraded confidence, or low confidence; and the abnormal residual trend indication can be set to "residual increase indication" or "no residual increase indication". When the EEG warning confidence level is low, the EEG residual risk level is set to indeterminate, and no definitive risk level (low, medium, or high) is output.

[0157] Specifically, the abnormal residual decay time constant of the final time window is obtained, denoted as... The estimated duration of abnormal residue is calculated based on the final abnormal residue amount, the abnormal residue decay time constant of the final time window, and a preset residue safety threshold. The preset residue safety threshold is denoted as... (In this embodiment, the value is 0.5), and the expected duration of the abnormal residue is denoted as... .

[0158] If the final abnormal residue level is not greater than the preset residue safety threshold, the estimated duration of abnormal residue is set to 0. If the final abnormal residue level is greater than the preset residue safety threshold, the estimated duration of abnormal residue is calculated according to the following formula:

[0159] ;in, The natural logarithm is represented. The estimated duration of the anomaly residue represents the estimated time required for the anomaly residue to decay from the final anomaly residue to the preset residue safety threshold without the addition of new high-confidence anomaly loads. The anomaly residue decay time constant of the final time window is taken as the weighted average of the anomaly residue decay time constants of several (3 in this embodiment) final effective time windows close to the time when monitoring stops.

[0160] Furthermore, the reference time for the interruption of target monitoring is also obtained, denoted as... The target monitoring interruption reference time indicates the expected duration during this intra-hospital transport when effective EEG signals cannot be continuously acquired. The target monitoring interruption reference time is primarily based on the expected duration of the monitoring interruption during this intra-hospital transport. The expected duration of the monitoring interruption during this intra-hospital transport refers to the time from the moment monitoring stops or the last reliable EEG sampling moment until the expected moment when the EEG monitoring device is reconnected and effective EEG sampling resumes.

[0161] If the estimated duration of the intra-hospital transport monitoring interruption cannot be obtained, a target monitoring interruption reference time will be selected from the preset standard monitoring interruption reference time set based on the intra-hospital transport project. The preset standard monitoring interruption reference time set includes 5 minutes, 15 minutes, 30 minutes, and 60 minutes. Specifically, 5 minutes represents short-term equipment disconnection or short-distance bedside transport scenarios; 15 minutes represents routine short-distance examinations or short-distance transport scenarios; 30 minutes represents CT scans, routine imaging examinations, or moderately time-consuming transport scenarios; and 60 minutes represents the complete round-trip cycle of MRI or other longer-time-consuming examinations. If the intra-hospital transport project cannot be determined, the target monitoring interruption reference time will be set to 15 minutes by default.

[0162] Before comparing coverage relationships, the estimated duration of abnormal residue and the target monitoring interruption reference time are converted to the same time unit. In this embodiment, the comparison is conducted using seconds as the unit, that is, the standard monitoring interruption reference time expressed in minutes is multiplied by 60 before comparison. When the estimated duration of abnormal residue is not less than the target monitoring interruption reference time, it is determined that the estimated duration of abnormal residue covers the monitoring interruption time expected to be formed during this intra-hospital transfer; when the estimated duration of abnormal residue is less than the target monitoring interruption reference time, it is determined that the estimated duration of abnormal residue does not cover the monitoring interruption time expected to be formed during this intra-hospital transfer.

[0163] To screen for highly reliable propagation contribution evidence that supports high-risk discrimination, this embodiment first generates an adaptive decision tree discrimination threshold. The adaptive decision tree discrimination threshold includes an adaptive propagation compatibility threshold, an adaptive artifact coupling threshold, and a highly reliable anomaly contribution threshold.

[0164] Specifically, valid evaluation windows are selected from the withdrawal window. A valid evaluation window is a time window that simultaneously meets the following two conditions: first, the time window is valid; second, at least one candidate abnormal EEG event exists within the time window. If the number of valid evaluation windows is less than three, the default discrimination thresholds are used, where the default propagation compatibility threshold is 0.4, the default artifact coupling threshold is 0.5, and the default high-confidence abnormality contribution threshold is 0.1.

[0165] If the number of effective evaluation windows reaches three or more, an adaptive discrimination threshold is generated based on the statistical distribution within the effective evaluation windows. Specifically, the adaptive propagation compatibility threshold is generated based on the 75th quantile, median, and difference between the propagation morphology compatibility within the effective evaluation windows, and is limited to the range of 0.3 to 0.8. The adaptive artifact coupling threshold is generated based on the median and median absolute deviation of the window-level coupling index within the effective evaluation windows, and is limited to the range of 0.2 to 0.7. The high-confidence anomaly contribution threshold is generated based on the median and median absolute deviation of the high-confidence anomaly load within the effective evaluation windows, and is limited to the range of 0.05 to 5.0.

[0166] Subsequently, high-credibility contribution windows are selected from all valid time windows. A valid time window is included in the set of high-credibility contribution windows only if it simultaneously meets the following three conditions:

[0167] First, the propagation morphology compatibility is not less than the adaptive propagation compatibility threshold;

[0168] Second, the window-level coupling index is not higher than the adaptive artifact coupling threshold;

[0169] Third, the high-confidence anomaly load is not less than the high-confidence anomaly contribution threshold.

[0170] The set of highly reliable contribution windows is denoted as . The minimum number of high-confidence contribution windows is preset, denoted as... In this embodiment, the number is preferably 3. If the number of high-confidence contribution windows is not less than the preset minimum number of high-confidence contribution windows, it is considered that there is a sufficient number of high-confidence propagation contribution evidence within the withdrawal window; if the number of high-confidence contribution windows is less than the preset minimum number of high-confidence contribution windows, it is considered that the propagation pattern evidence is insufficient to support high-risk confirmation.

[0171] Furthermore, adaptive decision tree discrimination is performed in the following node order:

[0172] Node N1, EEG validity node: Determines the data validity indicator, the number of valid time windows, and the baseline sufficiency indicator. If the data validity indicator is 0, or the number of valid time windows is less than the minimum number of valid time windows, output low confidence, set the EEG residual risk level to undetermined, and output "No residual increase prompt," terminating the judgment. The minimum number of valid time windows is preferably 3. If the data validity indicator is 1, and the number of valid time windows is not less than the minimum number of valid time windows, the judgment continues. In this case, if the baseline sufficiency indicator is 0, output downgraded confidence; if the baseline sufficiency indicator is 1, output normal confidence.

[0173] Node N2, Candidate Abnormal Event Node: Determines whether there are candidate abnormal EEG events in the valid time windows within the evacuation window; if there are no candidate abnormal EEG events in all valid time windows, outputs low risk and "no residual rise prompt", and the judgment terminates; if there are candidate abnormal EEG events in at least one valid time window, proceeds to node N3.

[0174] Node N3, Abnormal Residual Node: Determines whether the final abnormal residual amount is greater than the preset residual safety threshold; if the final abnormal residual amount is not greater than the preset residual safety threshold, outputs low risk and "no residual increase prompt", and terminates the judgment; if the final abnormal residual amount is greater than the preset residual safety threshold, proceeds to node N4.

[0175] Node N4, target coverage node: Determines whether the estimated duration of the abnormal residue covers the target monitoring interruption reference time; if the estimated duration of the abnormal residue is less than the target monitoring interruption reference time, outputs medium risk. If the estimated duration of the abnormal residue is not less than the target monitoring interruption reference time, proceeds to node N5;

[0176] Node N5, High-Confidence Propagation Contribution Node: Determines whether the number of valid leads is not less than 2, and whether the number of high-confidence contribution windows is not less than the preset minimum number of high-confidence contribution windows; if the number of valid leads is not less than 2, and the number of high-confidence contribution windows is not less than the preset minimum number of high-confidence contribution windows, then outputs high risk; if the number of valid leads is less than 2, or the number of high-confidence contribution windows is less than the preset minimum number of high-confidence contribution windows, then outputs medium risk.

[0177] High risk indicates that the final abnormal residual amount exceeds the preset residual safety threshold, the expected duration of the abnormal residual covers the target monitoring interruption reference time, and there are a sufficient number of high-confidence propagation contribution windows within the withdrawal window; medium risk indicates that there are candidate abnormal EEG events within the withdrawal window and the final abnormal residual amount exceeds the preset residual safety threshold, but the expected duration of the abnormal residual does not cover the target monitoring interruption reference time, or the number of high-confidence propagation contribution windows is insufficient to support high-risk confirmation; low risk indicates that there are no candidate abnormal EEG events within the withdrawal window, or the final abnormal residual amount does not exceed the preset residual safety threshold; low confidence indicates that the input EEG data does not meet the minimum validity requirements for deterministic risk level determination.

[0178] For abnormal residual trend indications, when the EEG residual risk level is low or the EEG warning confidence state is low confidence, "No residual increase indication" is directly output; when the EEG residual risk level is medium or high, it is determined whether the abnormal residual amount in the last continuous effective time window near the end of the monitoring stop time is continuously increasing; if the abnormal residual amount in the last continuous effective time window is continuously increasing, and the abnormal residual amount increment of adjacent time windows is not less than the preset increment threshold, then "Residual increase indication" is output, otherwise "No residual increase indication" is output; in this embodiment, the number of last continuous effective time windows is preferably 3, and the preset increment threshold is preferably 0.05;

[0179] In this embodiment, the preset parameters involved in the system can be determined by at least one of the following methods: factory default parameters, departmental configuration parameters, offline calibration parameters of historically labeled EEG data, or patient-specific parameters. The preset parameters are not limited to a fixed, single value; without changing the core processing flow of this embodiment, they can be adjusted according to the EEG monitoring device model, number of leads, sampling rate, type of transport project, and patient risk stratification.

[0180] To verify the stability of the early warning system in this embodiment under strong artifact interference scenarios before monitoring is stopped, an offline test set of 120 multi-lead EEG segments was selected. Each EEG segment was 10 minutes long, with a sampling rate of 256 Hz, using 8 effective leads, and divided into 60 time windows of 10 seconds each. The test set included 60 segments labeled with real abnormal propagation or focal repetitive abnormalities, and 60 segments mainly exhibiting motion artifacts, electrode contact artifacts, or electromyographic artifacts without the risk of leaving real abnormalities. The specific results are shown in the table below.

[0181] Comparison Plan False alarms of artifact samples Abnormal sample underreporting Sensitivity Specificity accuracy AUC Fixed amplitude threshold 25 14 76.7% 58.3% 67.5% 0.72 Adaptive threshold 16 10 83.3% 73.3% 78.3% 0.81 Adaptive threshold + propagation compatibility 8 7 88.3% 86.7% 87.5% 0.89 Adaptive thresholding + propagation compatibility + artifact coupling suppression 4 5 91.7% 93.3% 92.5% 0.94 Complete solution 3 3 95.0% 95.0% 95.0% 0.96

[0182] The results above show that the fixed amplitude threshold scheme is prone to misclassifying multi-lead synchronous jumps as abnormal EEG events in scenarios involving lead removal, body repositioning, and unstable electrode contact, resulting in 25 false positives for artifact samples. The complete scheme in this embodiment, through a combination of homologous artifact features, adaptive abnormality thresholds, propagation morphology compatibility, artifact-abnormality coupling suppression, and abnormal residual attenuation prediction, reduces the number of false positives for artifact samples to 3 and the number of missed abnormal samples to 3, improving the accuracy from 67.5% to 95.0%.

[0183] In one application scenario, this embodiment is used for EEG monitoring and early warning before the in-hospital transfer of critically ill patients in the intensive care unit. A patient with severe traumatic brain injury is receiving multi-lead EEG monitoring at the bedside. Medical staff plan to transfer the patient to the imaging room for a head examination. The estimated time from stopping the bedside EEG monitoring to returning to the ward and reconnecting the EEG monitoring device is 30 minutes.

[0184] Prior to transport, the EEG monitoring device acquired multi-lead EEG signals from the scalp at a sampling rate of 256Hz, and used 8 leads from the international 10-20 system for simplified bedside monitoring. The impedance of each lead electrode was controlled within 5000 ohms. After detecting that the monitoring was about to end, the EEG measurement module automatically locked the continuously acquireable EEG signals within 10 minutes before stopping monitoring as the withdrawal window, and divided this withdrawal window into 10-second time windows. The effective EEG signal duration acquired in this study was 600 seconds, the number of effective time windows was 60, and the number of effective leads was 8, which met the requirements for subsequent EEG residual risk analysis before in-hospital transport.

[0185] Within the withdrawal window, the EEG measurement module extracts indicators from each valid lead and generates homologous artifact features by combining the synchronous jump ratio of multiple leads. It detects motion artifacts and electrode artifacts in some time windows, but most time windows remain valid. It also performs preliminary screening for abnormal events in each valid lead. Within several time windows close to the time of cessation of monitoring, it identifies spikes, spike waves, and spike-slow-wave complexes with durations of approximately 20 to 200 milliseconds and peak amplitudes exceeding 50 to 70 microvolts, and calculates the event-level abnormal intensity.

[0186] The EEG calculation module generates an adaptive abnormality threshold based on the homologous artifact characteristics of the current time window and the abnormal residual amount of the previous time window, and filters the initial screening abnormal events to obtain candidate abnormal EEG events. In the last 5 minutes before monitoring stops, multiple candidate abnormal EEG events show a sequential arrival order that conforms to the lead adjacency relationship between adjacent leads within 50 milliseconds, and the waveform amplitude, duration, frequency band energy, and polarity phase characteristics show a consistent trend. Based on this, the system determines that the relevant time windows have high propagation morphology compatibility. For candidate abnormal EEG events that are synchronously enhanced with motion artifacts or electrode artifacts at near-zero delay, the EEG calculation module reduces their contribution to the high-confidence abnormality load by using the artifact abnormality near-zero delay coupling index.

[0187] The status update module updates the abnormal residual amount window by window. In the last 5 minutes before monitoring stops, high-confidence abnormal loads appear in 4 consecutive valid time windows, and the terminal abnormal residual amount is higher than the residual safety threshold. The system estimates the expected duration of abnormal residual based on the abnormal residual decay time constant of the terminal time window, and the expected duration is 38 minutes. This duration has covered the 30-minute monitoring interruption time expected to occur during this intra-hospital transfer.

[0188] The risk decision module uses an adaptive decision tree to determine the final abnormal residual amount, the expected duration of the abnormal residual, and the number of high-confidence contribution windows. Since there was a continuous high-confidence abnormal load before this transfer, the expected duration of the abnormal residual covers the 30-minute monitoring interruption time, and the data validity and baseline reference both meet the analysis requirements, the risk decision module finally outputs high risk, normal confidence, and gives a residual increase warning.

[0189] In another similar application scenario, a 10-minute evacuation window analysis was also performed before patient transfer. The system obtained 60 valid time windows and 8 valid leads. The EEG calculation module detected candidate abnormal EEG events in the last 5 minutes before monitoring stopped, but only 2 valid time windows met the high-confidence abnormality load condition. The estimated duration of abnormal residual events was 12 minutes, which did not cover the estimated 30-minute monitoring interruption time. The risk decision module ultimately output medium risk with normal confidence.

[0190] Example 2

[0191] like Figure 4 As shown, this embodiment introduces a risk warning method for intra-hospital transfer of critically ill patients based on adaptive decision trees. It is applied to the intra-hospital transfer risk warning system for critically ill patients based on adaptive decision trees in Embodiment 1, and includes the following steps:

[0192] S1. Acquire multi-lead EEG signals from the EEG monitoring device before monitoring is stopped and divide them into multiple sequentially arranged time windows. Extract homologous artifact features and preliminary screening of abnormal events from the effective leads within each time window.

[0193] S2. Based on the homologous artifact features of the current time window and the abnormal residual amount of the previous time window, an adaptive abnormal threshold for the current time window is generated, and then the initial screening of abnormal events is filtered to obtain candidate abnormal EEG events; a brain region lead topology map is generated based on the effective leads, and then the propagation morphology compatibility of the candidate abnormal EEG events in the leads is calculated; based on the propagation morphology compatibility, a protection correction is performed to obtain the corrected homologous artifact features, and the coupling relationship between the candidate abnormal EEG events and the corrected homologous artifact features is calculated, thereby generating a high-confidence abnormal load for the current time window;

[0194] S3. The residual abnormality in the previous time window is attenuated and the high-confidence abnormality load of the current time window is superimposed until all time windows are traversed, and the final residual abnormality representing the electroencephalological state is obtained.

[0195] S4. Based on the final abnormal residual amount, output the risk warning level of EEG residual before in-hospital transfer through an adaptive decision tree.

[0196] Example 3

[0197] This embodiment introduces a computer device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the steps of the adaptive decision tree-based in-hospital transfer risk warning method for critically ill patients as described in Embodiment 2.

[0198] The method for risk warning of intra-hospital transfer of critically ill patients based on adaptive decision trees in Example 2 can be applied in software form, such as by designing it as a standalone program and installing it on a computer device, which can be a processor, embedded medical terminal, bedside monitoring workstation, or server, etc. Alternatively, it can be designed as an embedded program and installed on a computer terminal, such as on a microcontroller.

[0199] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. A risk warning system for intra-hospital transport of critically ill patients based on adaptive decision trees, characterized in that, include: The EEG measurement module is used to acquire multi-lead EEG signals from the EEG monitoring device before monitoring is stopped and divide them into multiple sequentially arranged time windows. It extracts homologous artifact features and preliminary screening of abnormal events from the effective leads within each time window. The EEG calculation module is used to generate an adaptive abnormality threshold for the current time window based on the homology artifact features of the current time window and the abnormal residual amount of the previous time window, and then screen the initial abnormal events to obtain candidate abnormal EEG events; generate a brain region lead topology map based on the effective leads, and then calculate the propagation morphology compatibility of the candidate abnormal EEG events between leads; perform protection correction based on the propagation morphology compatibility to obtain corrected homology artifact features, and calculate the coupling relationship between the candidate abnormal EEG events and the corrected homology artifact features, and then generate a high-confidence abnormality load for the current time window; The state update module is used to control the decay of the abnormal residual amount in the previous time window and to add the high-confidence abnormal load of the current time window until all time windows are traversed to obtain the final abnormal residual amount that represents the electroencephalographic state. The risk decision module is used to output the risk warning level of EEG residue before in-hospital transfer based on the final abnormal residual amount through an adaptive decision tree.

2. The intra-hospital transfer risk early warning system for critically ill patients based on adaptive decision tree as described in claim 1, characterized in that, The electroencephalogram (EEG) measurement module specifically includes: Starting from the moment when monitoring stops, trace back forward and extract continuous multi-lead EEG signals that meet the preset validity conditions to form a withdrawal window, and divide the withdrawal window into multiple time windows of equal length; Within the evacuation window, a preset validity determination is performed on each lead to obtain the valid leads. Local artifact features of the valid leads within each time window are extracted, and the synchronization jump ratio between valid leads is calculated. Based on the local artifact features and the synchronization jump ratio, homologous artifact features for each time window are generated.

3. The intra-hospital transfer risk early warning system for critically ill patients based on adaptive decision tree as described in claim 1, characterized in that, In the EEG calculation module, an adaptive abnormality threshold for the current time window is generated based on the homologous artifact features of the current time window and the abnormal residual amount of the previous time window. Specifically, this includes: The initial adaptive threshold for the current time window is calculated using the following formula: in, This is the initial adaptive threshold for the current time window. The preset basic anomaly threshold, The features of homologous artifacts after normalization within the current time window. This represents the anomalous residual amount after normalization from the previous time window. and These are the preset artifact adjustment coefficient and residual amount adjustment coefficient, respectively. and The lower and upper limits are truncated to preset thresholds; The initial adaptive threshold is constrained by a rate of change so that the rate of change of the initial adaptive threshold relative to the adaptive anomaly threshold of the previous time window does not exceed a preset safety ratio, and the constrained value is then used as the adaptive anomaly threshold of the current time window.

4. The intra-hospital transfer risk early warning system for critically ill patients based on adaptive decision tree according to claim 3, characterized in that, The EEG calculation module calculates the propagation morphology compatibility of candidate abnormal EEG events across leads, specifically including: The brain region lead topology map includes multiple directed edges representing the EEG propagation relationship between leads. Each directed edge represents the propagation direction of abnormal events from the starting lead to the ending lead. For each directed edge, in the corresponding starting lead and ending lead, candidate abnormal EEG events with a peak time difference not less than a preset synchronization judgment threshold and not exceeding a preset maximum allowable delay are selected to form a matchable event pair. The peak time difference between the matching event pair and the starting lead and the ending lead is obtained as the current waveform propagation arrival delay. Based on the current waveform propagation arrival delay and the reference arrival delay of the previous time window, the time delay soft matching value of the matching event pair is calculated. Extract the waveform morphology change features of the matching event pair in the endpoint lead relative to the starting lead, and calculate the morphological evolution compatibility based on the reference morphological change distribution of the previous time window. Based on the time delay soft matching value and the morphological evolution compatibility, the edge-level propagation morphological compatibility is generated, and the weighted average of the edge-level propagation morphological compatibility of each directed edge is calculated to obtain the propagation morphological compatibility.

5. The intra-hospital transfer risk early warning system for critically ill patients based on adaptive decision tree according to claim 4, characterized in that, The step of correcting the homologous artifact features based on the propagation morphology compatibility and calculating the coupling relationship specifically includes: The number of windows in which the candidate abnormal EEG events occur consecutively in each of the effective leads and the average abnormal intensity are obtained. A lead-level focal repetition abnormality protection value is generated for each effective lead, and the maximum value among the lead-level focal repetition abnormality protection values ​​is taken as the window-level focal repetition abnormality protection value. The larger value between the propagation morphology compatibility and the window-level focal repetition abnormality protection value is taken as the effective compatibility. Based on the homologous artifact features of the current time window, obtain the lead-level normalized homologous artifact features; using the effective compatibility, perform an attenuation operation on the lead-level normalized homologous artifact features to obtain the corrected homologous artifact features. The current time window is divided into multiple sub-windows, and the sub-window-level candidate anomaly intensity and the corrected sub-window-level homologous artifact feature are generated for each valid lead in each sub-window. Within a preset maximum correlation delay range, the delay correlation is calculated between the corrected sub-window-level homologous artifact feature sequence and the sub-window-level candidate anomaly intensity sequence for each valid lead, and the maximum positive correlation value corresponding to each valid lead is taken as the lead-level coupling index of that valid lead.

6. The intra-hospital transfer risk early warning system for critically ill patients based on adaptive decision tree according to claim 5, characterized in that, The high-confidence abnormal load generated in the EEG calculation module for the current time window specifically includes: For the effective lead m, its lead-level high-confidence anomalous load The calculation formula is: ； in, The intensity of candidate abnormalities at the lead level is extracted based on the candidate abnormal EEG events. The coupling index of the lead level is... For the compatibility of the aforementioned propagation mode, This is the protection value for the window-level focal recurrence anomaly. This is a preset smoothing mapping function. , , These are preset parameters; Each valid lead is evaluated to obtain a lead weight. Based on the lead weight, the lead-level high-confidence anomalous load of all valid leads is weighted and summed to generate the high-confidence anomalous load of the current time window.

7. The intra-hospital transfer risk early warning system for critically ill patients based on adaptive decision tree according to claim 1, characterized in that, The state update module performs attenuation control on the residual anomalous amount from the previous time window and superimposes the high-confidence anomalous load from the current time window, specifically including: Based on the residual anomaly after the end of the previous time window, the high-confidence anomaly load of the current time window, the time window length, the anomaly residual decay time constant of the current time window, and the upper limit of the constrained residual amount, the residual anomaly after the end of the current time window is generated, and the update formula is as follows: ； in, This is the updated abnormal residual amount. This refers to the abnormal residual amount in the previous time window. The length of the time window; The abnormal residual decay time constant is generated by weighted fusion based on the effective compatibility, high-confidence abnormal load, number of candidate abnormal EEG events, and number of continuous time windows that meet the preset stability conditions. For the current time window, the high-reliability anomalous load, This is the preset upper limit for residual amount; The state update module further includes: when it is determined that the current time window meets the preset high-confidence reference update conditions, based on the propagation parameters extracted in the current time window and in accordance with the weighted smoothing rule, updating the reference arrival delay of the previous time window and the reference shape change distribution of the previous time window on the directed edge.

8. The intra-hospital transfer risk early warning system for critically ill patients based on adaptive decision tree according to claim 1, characterized in that, The risk decision-making module outputs a pre-hospital transfer risk warning level for EEG residual effects via the adaptive decision tree, specifically including: Obtain the final abnormal residual amount, and based on the abnormal residual decay time constant of the end time window, calculate the expected duration of abnormal residual required for the abnormal residual to decay from the final abnormal residual amount to the preset residual safety threshold. Obtain the target monitoring interruption reference time, which is determined based on the expected duration of the intra-hospital transfer monitoring interruption. The time windows that meet the preset conditions within the evacuation window are used as effective evaluation windows. An adaptive decision tree discrimination threshold is generated based on the statistical distribution of the effective evaluation windows, and the effective time windows that meet the conditions of the adaptive decision tree discrimination threshold are included in the high-confidence contribution window set. Based on the coverage relationship between the estimated duration of the abnormal residual and the reference time of the target monitoring interruption, as well as the number of windows in the high-confidence contribution window set, an adaptive decision tree discrimination is performed to output the risk warning level of EEG residual.

9. A method for early warning of intra-hospital transport risks of critically ill patients based on adaptive decision trees, applied to the intra-hospital transport risk early warning system for critically ill patients based on adaptive decision trees as described in any one of claims 1-8, characterized in that, Includes the following steps: The multi-lead EEG signals of the EEG monitoring device before monitoring was stopped were acquired and divided into multiple sequentially arranged time windows. Homologous artifact features and preliminary screening of abnormal events in the effective leads within each time window were extracted. Based on the homology artifact features of the current time window and the abnormal residual amount of the previous time window, an adaptive abnormality threshold for the current time window is generated, and then the initial screening of abnormal events is filtered to obtain candidate abnormal EEG events; a brain region lead topology map is generated based on the effective leads, and then the propagation morphology compatibility of the candidate abnormal EEG events in the leads is calculated; based on the propagation morphology compatibility, a protection correction is performed to obtain the corrected homology artifact features, and the coupling relationship between the candidate abnormal EEG events and the corrected homology artifact features is calculated, thereby generating a high-confidence abnormality load for the current time window; The residual abnormality from the previous time window is attenuated and superimposed with the high-confidence abnormality load of the current time window until all time windows are traversed, and the final residual abnormality representing the electroencephalological state is obtained. Based on the final abnormal residual amount, an adaptive decision tree is used to output the risk warning level of EEG residuals before in-hospital transfer.

10. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method for early warning of risks in the intra-hospital transfer of critically ill patients based on adaptive decision trees as described in claim 9.

Images