Method and system for detecting abnormal brain function of postoperative delirium patient

Abstract

The invention relates to the field of target detection, and provides a postoperative delirium patient brain function abnormity detection method and system. The method comprises the following steps: acquiring resting state original electroencephalogram data of a postoperative patient and preprocessing the resting state original electroencephalogram data to ensure data quality; feature time is extracted through micro-state analysis, electroencephalogram data under a target micro-state time window are obtained, and key brain dynamic states related to delirium are focused; channel configuration is carried out on electroencephalogram data based on a common spatial pattern algorithm, a plurality of representative electrodes in a characteristic time period are screened out, data redundancy is reduced, and analysis specificity is improved; taking the selected electrode as a brain network node, calculating a phase locking value between the electrodes in a preset frequency band, and extracting a significant difference connection edge under a target frequency band; and extracting a feature vector based on the significant difference edge, inputting the feature vector into the delirium prediction model, and outputting a delirium identification result. According to the method, the abnormal brain function of the postoperative delirium patient is accurately detected, and the method has high application value.

Description

Technical Field

[0001] This invention relates to the field of target detection technology, and more specifically, to a method and system for detecting abnormal brain function in patients with postoperative delirium. Background Technology

[0002] Postoperative delirium is an acute brain syndrome triggered by physiological factors, characterized by inattention, cognitive impairment, and fluctuating states of consciousness. Symptoms include altered consciousness, disorganized behavior, loss of purpose, perceptual abnormalities (such as hallucinations), and mood swings. The severity of symptoms can change rapidly from minutes to hours. This condition is common in postoperative or critically ill patients, especially the elderly; studies show that its incidence in intensive care unit patients can reach 31%.

[0003] Postoperative delirium has a serious negative impact on patient prognosis, closely associated with cognitive decline, increased risk of death, elevated risk of dementia, and increased medical burden. However, current clinical diagnosis and prevention of delirium still rely on traditional methods, primarily through assessment tools such as CAM-ICU for critically ill patients and 3D-CAM for rapid diagnosis. These methods have significant limitations: firstly, they are difficult to identify and diagnose promptly in groups who cannot clearly express their feelings (such as infants and young children, and those with severe cognitive impairment); secondly, diagnostic results depend on subjective judgment, lacking objective laboratory data, electrophysiological or imaging examinations, leading to misdiagnosis or missed diagnosis, resulting in delayed targeted treatment and affecting patient recovery.

[0004] With the development of neuroscience and computer-aided medical technology, electroencephalography (EEG), as a non-invasive, convenient, and high-temporal-resolution neural monitoring technique, has provided a new approach for the objective assessment of delirium. Resting-state EEG data, due to its ease of acquisition and applicability to various patient groups, has become an important data source for studying the neural mechanisms of delirium. However, current EEG-based delirium research still has many shortcomings: Inappropriate time window selection: Traditional analyses often use fixed-length sliding time windows or subjectively selected time points, which cannot reflect the high temporal resolution characteristics of dynamic changes in the brain and may miss key neural activity patterns.

[0005] Channel analysis has its limitations: it relies heavily on empirical full-channel analysis, leading to data redundancy, noise interference, and the influence of irrelevant signals. It may also introduce subjective errors due to human selection of channels, reducing the accuracy of the analysis.

[0006] Inaccurate functional connectivity studies: The analysis of brain network functional connectivity lacks specificity, fails to combine the dynamic state of the brain to select key time periods and channels, and makes it difficult to capture specific neural connectivity abnormalities related to delirium.

[0007] Limited predictive model effectiveness: Existing models are mostly based on single frequency bands or simple features, failing to fully utilize the nonlinear features and multi-frequency functional connectivity information of EEG data. Their classification accuracy and stability are insufficient, making it difficult to meet clinical needs.

[0008] Furthermore, existing research in this field suffers from limitations such as a single sample population, diverse data analysis methods, and low consistency of results, which restricts the credibility of research conclusions. Therefore, there is an urgent need to develop a technical solution that combines dynamic brain state analysis, precise channel selection, and efficient machine learning algorithms to achieve objective and early prediction of postoperative delirium, providing a scientific basis for clinical research. Summary of the Invention

[0009] This invention addresses the technical problems existing in the prior art by providing a method and system for detecting abnormal brain function in patients with postoperative delirium, aiming to provide a scientific basis for research on the neural mechanisms of delirium based on EEG.

[0010] According to a first aspect of the present invention, a method for detecting abnormal brain function in patients with postoperative delirium is provided, comprising: S1, Acquire raw EEG data of the patient at rest after surgery and perform preprocessing; S2, based on microstate analysis, extracts feature time from the preprocessed raw EEG data to obtain EEG data under the target microstate time window; S3, based on the common spatial pattern (CSP), performs channel configuration on the EEG data obtained in S2 to obtain multiple representative electrodes within the characteristic time period; S4. Using the electrodes obtained in S3 as nodes for brain network connectivity analysis, calculate the phase lock value (PLV) between electrodes within a preset frequency band to extract significant difference edges in the target frequency band. S5 extracts feature vectors based on the significant differences extracted in S4 and inputs them into a preset delirium prediction model to output delirium identification results.

[0011] Based on the above technical solution, the present invention can also be improved as follows.

[0012] Optionally, step S1 includes: Based on clinical requirements, the raw EEG data of the postoperative patient at rest and its corresponding electrode coordinates are obtained. The raw EEG data includes EEG data collected from all electrode channels. The raw EEG data is filtered, and artifacts in the raw EEG data are removed based on independent component analysis (ICA). The Kurtosis algorithm is used to automatically detect bad channels, mark and remove the electrodes and EEG data corresponding to the bad channels, and the spherical algorithm is used to interpolate and complete the removed data. The EEG data corresponding to each electrode were averaged and referenced.

[0013] Optionally, step S2 includes: S201, calculate the global field power (GFP) for each electrode in the EEG data of the postoperative patient in the resting state. The global field power (GFP) represents the intensity of the brain electric field at each instant. S202 uses global field power (GFP) as a clustering index and employs a clustering algorithm to cluster the EEG data corresponding to all electrodes into EEG data in multiple microstates. Each microstate represents a quasi-static period corresponding to a certain region in the brain map. The cluster consists of multiple microstates: microstate A, microstate B, microstate C, and microstate D, which correspond one-to-one with the right frontolateral posterior region, left frontolateral posterior region, midline fronto-occipital region, and midline frontal region in the brain map. S203, filter out the EEG data corresponding to microstate C as the EEG data within the characteristic time period.

[0014] Optionally, step S3 includes: S301, the Common Spatial Pattern (CSP) algorithm is applied to the EEG data extracted by microstate analysis to separate the spatial feature components of each channel in the multi-channel brain-computer interface data; S302, each channel is scored separately based on the spatial feature component weights; S303, based on the scoring results of each channel and the preset score threshold, selects electrodes corresponding to multiple representative channels and extracts the EEG data corresponding to the selected electrodes.

[0015] Optionally, step S4 includes: S401, For the multiple electrodes and their EEG data selected in S3, calculate the phase lock value (PLV) between the electrodes by frequency band to construct the functional connectivity matrix. S402, perform statistical difference analysis on the functional connectivity matrix of the target frequency band to obtain multiple significantly different edges under the target frequency band.

[0016] Optionally, step S4 also includes: S403 plots a cortical phase-locked network (CPLN) based on multiple significantly different edges to visualize anomalous connection patterns.

[0017] Optionally, step S5 includes: S501, extract the phase-locked value (PLV) corresponding to multiple significantly different edges under the target frequency band to form a nonlinear feature vector; S502, the extracted nonlinear feature vector is input into the preset delirium prediction model to obtain the category label of delirium identification and the corresponding risk probability.

[0018] Optionally, before step S5, a delirium prediction model is constructed, specifically including: Resting-state EEG data from multiple known delirium patients and multiple non-delirium patients were used as a sample set and preprocessed. The sample set is processed sequentially through steps S2 to S4 to extract the number of significant differences in the EEG data of the two types of patients in each frequency band. The phase-locked value (PLV) corresponding to the significantly different edge is used as a nonlinear feature, and multiple feature subsets are divided according to frequency band; Based on the CatBoost model architecture, delirium prediction binary classification models were constructed and trained for feature subsets under each frequency band. The delirium prediction binary classification model with better classification performance was selected as the final delirium prediction model.

[0019] Optionally, when processing the sample set in step S2, the category of the target microstate can be determined by whether there is a significant difference in time coverage between the EEG data of the two types of patients.

[0020] According to a second aspect of the present invention, a system for detecting abnormal brain function in patients with postoperative delirium is provided, comprising: The acquisition and preprocessing module is used to acquire and preprocess the raw EEG data of the postoperative patient at rest. The time window filtering module is used to extract feature time from the preprocessed raw EEG data based on microstate analysis to obtain EEG data under the target microstate time window. The channel configuration module is used to configure the EEG data obtained from the time window screening module based on the common spatial pattern (CSP) to obtain multiple representative electrodes within a characteristic time period. The brain network connectivity analysis module is used as the electrodes obtained from the channel configuration module as brain network connectivity analysis nodes, and calculates the phase lock value (PLV) between electrodes within a preset frequency band to extract significant difference edges in the target frequency band. The classification and recognition module is used to extract feature vectors based on the significant difference edges extracted from the brain network connectivity analysis module and input them into a preset delirium prediction model to output delirium recognition results.

[0021] According to a third aspect of the present invention, an electronic device is provided, including a memory and a processor, wherein the processor is configured to execute a computer management program stored in the memory to implement a method for detecting abnormal brain function in patients with postoperative delirium.

[0022] According to a fourth aspect of the present invention, a computer-readable storage medium is provided having a computer management program stored thereon, wherein the computer management program, when executed by a processor, implements the steps of a method for detecting abnormal brain function in patients with postoperative delirium.

[0023] This invention provides a method, system, electronic device, and storage medium for detecting abnormal brain function in patients with postoperative delirium. First, the postoperative resting-state EEG is preprocessed. Then, microstate analysis is used to identify time windows of significantly different target microstates. Subsequently, the most representative electrodes are selected from these windows using a cospatial pattern, and phase-locked values ​​for the target frequency band are calculated using these electrodes as nodes. Significantly different edges are extracted to construct feature vectors. Finally, the features are input into a pre-trained delirium prediction model, achieving rapid and automatic delirium risk identification. This invention achieves high temporal resolution brain state segmentation based on data-driven methods. By utilizing the nonlinear characteristics of brain networks in characteristic microstates, it automatically identifies delirium patients. This solves the problems of performance degradation, data redundancy, and interference from irrelevant channel signals in current EEG-based delirium research, which arise from full-channel analysis. It also eliminates the subjective influence of manually selecting channels to reduce research complexity, resulting in high detection accuracy and high application value. Attached Figure Description

[0024] Figure 1 A flowchart of a method for detecting abnormal brain function in patients with postoperative delirium provided by the present invention; Figure 2 A flowchart of EEG data preprocessing provided for one embodiment; Figure 3 A clustering topology diagram of four microstates corresponding to non-delirious patients and delirious patients, provided for a certain embodiment; Figure 4 A statistical graph showing the time coverage of four microstates for non-delirious patients and delirious patients in a certain embodiment; Figure 5 A schematic diagram of the distribution of channel-selected electrodes on the cerebral cortex provided for one embodiment; Figure 6 A statistical chart of average PLV for each frequency band provided for a certain embodiment; Figure 7 Cortical phase-locked networks for delirium and non-delirium patients in various frequency bands provided in one embodiment; Figure 8 This is a block diagram of a brain function abnormality detection system for postoperative delirium patients provided in an embodiment of the present invention. Detailed Implementation

[0025] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0026] With advancements in neuroscience and computing technology, neuroscientists are able to delve deeper into the working mechanisms of the brain under normal conditions and the extent of functional impairment in disease states. Technological advancements, such as functional magnetic resonance imaging (fMRI), have revealed the interconnectedness of functional brain regions in a network pattern. The dynamic changes in these brain networks are closely related to various brain functions, and therefore, damage to them may be associated with a variety of neurological and psychiatric disorders. On the other hand, electroencephalography (EEG), a widely used technique, can monitor the activity of cortical networks in both healthy and diseased states. EEG is cost-effective and can non-invasively assess synchronous neural activity across local and whole-brain areas. EEG primarily measures the electric potential field on the scalp surface, similar to how magnetoencephalography (MEG) measures magnetic fields. By properly sampling and correctly analyzing this electric field, EEG possesses extremely high temporal resolution, providing reliable information on neuronal activity and its temporal dynamics within milliseconds, a significant advantage compared to other neuroimaging techniques such as fMRI.

[0027] Generally, the frequency range of electroencephalogram (EEG) signals is considered to be 0.5-40Hz, which can be divided into five frequency bands: Delta band: The frequency range is 0.5-4Hz. The Delta band signal is a major feature of EEG recorded during deep sleep. Within this frequency range, Delta waves typically have large amplitudes (75-200μV) and show strong uniformity across the entire scalp.

[0028] Theta band: This band ranges from 4 to 8 Hz. Theta wave signals are typically detected in the temporal and parietal lobes of the brain. EEG activity in this band is generally associated with episodic and working memory functions. Furthermore, certain abnormalities in the central nervous system can trigger increased theta wave activity, a phenomenon often more pronounced in patients with mental illnesses. Theta wave activity is also commonly seen in cognitive processes involving memory.

[0029] Alpha band: The frequency range is (8-12Hz). Alpha signals dominate during wakefulness and are most prominent in the occipital lobe. Alpha frequency increases from early childhood to adulthood. Alpha frequency is positively correlated with cognitive performance and therefore decreases with age or age-related neurological diseases.

[0030] Beta band: The frequency range is (12-30Hz). The signal of the Beta band is characterized by increased alertness and concentration. This wave is often found in the frontal lobe, temporal lobe and central area of ​​the brain. It is more easily detected when a person is engaged in logical reasoning and mathematical calculation.

[0031] Gamma band: This band, ranging from 30 Hz upwards, is associated with information processing (e.g., recognition of sensory stimuli) and the initiation of voluntary movement. In general, the slowest cortical rhythms are associated with idle brain activity, while the fastest rhythms are associated with information processing.

[0032] The Broadman partitioning system is based on observations of the cellular structure and function of different brain regions, with each number representing a unique functional area of ​​the cerebral cortex. These regions encompass various cognitive and perceptual functions, including perception, motor control, emotion regulation, and language processing. Through Broadman partitioning, researchers can better understand the functions and interrelationships of different brain regions. This system is widely used in neuroscience research, helping scientists understand the connection between brain structure and function and promoting in-depth exploration of brain function. This partitioning system provides a convenient framework for studying specific areas of the brain, helping to reveal the complex mechanisms of cognition, emotion, and behavior.

[0033] By understanding the functions of the five main frequency bands of EEG signals and the functions of the Brodman partitioning, it can be seen that the brain has close intrinsic connections with different frequency bands during different cognitions, emotions, tasks, and diseases. Furthermore, the coordinated functioning of various brain regions is closely related to physiological activities such as consciousness, cognition, emotion, and behavior. Therefore, the research on delirium detection using resting-state EEG in this invention has a profound physiological basis.

[0034] Figure 1 A flowchart of a method for detecting abnormal brain function in patients with postoperative delirium provided by the present invention is shown below. Figure 1 As shown, the method includes steps S1 to S5: S1, Acquire raw EEG data of the patient at rest after surgery and perform preprocessing; S2, based on microstate analysis, extracts feature time from the preprocessed raw EEG data to obtain EEG data under the target microstate time window; S3, based on the common spatial pattern (CSP), performs channel configuration on the EEG data obtained in S2 to obtain multiple representative electrodes within the characteristic time period; S4. Using the electrodes obtained in S3 as nodes for brain network connectivity analysis, calculate the phase lock value (PLV) between electrodes within a preset frequency band to extract significant difference edges in the target frequency band. S5 extracts feature vectors based on the significant differences extracted in S4 and inputs them into a preset delirium prediction model to output delirium identification results.

[0035] Understandably, given the deficiencies in the background technology, this invention proposes a method for detecting abnormal brain function in patients with postoperative delirium. This method first preprocesses the postoperative resting-state EEG, then uses microstate analysis to identify target microstate time windows with significant differences. Subsequently, it uses a cospatial pattern to select the most representative electrodes from this window, and uses these electrodes as nodes to calculate the phase-locked value of the target frequency band. Significantly different edges are extracted to construct feature vectors, and finally, the features are input into a pre-trained delirium prediction model to achieve rapid and automatic delirium risk identification.

[0036] This embodiment realizes brain state segmentation based on data-driven high temporal resolution. It utilizes the nonlinear characteristics of brain networks in characteristic microstates to automatically identify delirium patients. It solves the problems of performance degradation, data redundancy, and interference of study results by channel signals irrelevant to the study in current EEG-based delirium research. It also eliminates the influence of subjective factors such as manually selecting channels to reduce research complexity.

[0037] In one possible embodiment, such as Figure 2 As shown, step S1 includes sub-steps S101 to S104: S101, based on clinical requirements, acquires the raw EEG data of the postoperative patient in a resting state and their corresponding electrode coordinates, and imports the data. The raw EEG data includes EEG data acquired from all electrode channels.

[0038] For example, in one implementation scenario, a 32-channel saline cap (Greentek, Wuhan, China) conforming to the international 10-10 system standard and an EEG acquisition device manufactured by Gilching BrainProduce, Germany, were used to record the EEG signals of the research subjects. The complete EEG acquisition equipment consists of a signal amplifier, battery, EEG signal recording software, and EEG signal analysis software. Before wearing the EEG cap, the subject's scalp needs to be cleaned to remove oil and dirt to ensure good contact between the electrodes and the scalp, and the impedance between the electrodes and the scalp should be kept below 10kΩ. When wearing the saline cap for EEG acquisition, it is necessary to ensure that the cap is correctly positioned front-to-back and side-to-side, and that the electrode cap can properly cover the subject's head. The position of the electrodes should be adjusted to align with the international 10-10 system markers to ensure accurate lead placement. The subject's EEG signals were acquired at a sampling rate of 500Hz. Due to the low signal intensity, the signals were first amplified by an amplifier, then processed by an analog-to-digital converter, and finally presented using Recorder software for subsequent processing.

[0039] EEG signals are affected by various factors during acquisition, such as noise generated by vibrations caused by operator contact during surgery; 50Hz or 60Hz power frequency interference from electrical equipment in the acquisition environment; physiological artifacts generated by other electrical signals such as electrooculogram and electrocardiogram; as well as slip artifacts and motor malfunctions. To reduce the impact of these factors on the accuracy and reliability of EEG data, the acquired raw EEG data needs to be preprocessed according to steps S102 to S104.

[0040] S102, the raw EEG data is filtered and artifacts in the raw EEG data are removed based on independent component analysis (ICA).

[0041] Regarding filtering, this invention employs the FIR filter from the EEGLAB toolkit, using a bandpass filter with a frequency range of 0-40Hz for the raw EEG signal. This frequency range encompasses commonly used ranges for EEG signal analysis while avoiding interference from 50 or 60Hz power line interference and high-frequency signals such as electromyography (EMG) on the EEG signal.

[0042] Then, artifact removal is performed: the raw EEG data acquired contains artifacts caused by various factors. EEGLAB has an important function—Independent Component Analysis (ICA)—which can remove artifacts from the raw signal. ICA decomposes EEG data into statistically independent components. These independent components can represent the activity of different regions or neural networks within the brain, or they may represent external interference or physiological artifacts. Artifact removal is a crucial step. ICA is an effective method for artifact identification and removal. By analyzing the components decomposed by ICA, researchers can identify those components representing artifacts and remove them from the raw data, thereby purifying the EEG signal.

[0043] S103 uses the Kurtosis algorithm to automatically detect bad channels, marks and removes the electrodes and EEG data corresponding to the bad channels, and uses the spherical algorithm to interpolate and complete the removed data.

[0044] It is worth noting that during EEG signal acquisition, signal quality may degrade or become unusable due to factors such as prolonged use, physical damage leading to aging, poor contact with the scalp, electromagnetic interference in the environment, and equipment malfunction. Therefore, bad channels need to be detected and interpolated during the preprocessing stage. Kurtosis is used to automatically detect bad channels, and the electrode names corresponding to the bad channels are marked with red lines. Then, noise signals are removed using a thresholding method. To ensure the continuity of EEG data, a spherical algorithm is also used to perform interpolation.

[0045] S104, average reference processing is performed on the EEG data corresponding to each electrode.

[0046] Understandably, rereferencing is used to improve the quality of EEG data, thereby providing more accurate information for subsequent analysis. Rereferencing is used in this embodiment of the invention. Common rereferencing methods include average reference, zero reference, and specific electrode reference. The average reference method uses the average potential of all electrodes as a common reference value, assuming that the average of all signals measured on the scalp surface is zero. Therefore, the activity of any electrode is measured relative to the average activity of the entire scalp potential. Because average reference is more suitable for the research scheme of this invention, it is preferentially used in the preprocessing stage.

[0047] After cleaning the data using the preprocessing methods described above, relatively clean EEG data can be obtained for subsequent research. Based on the data information, this embodiment extracts 90 seconds of preprocessed data for each sample.

[0048] Microstates are a method for studying the dynamic characteristics of brain activity. They involve a more detailed division and analysis of brain activity over a time scale, aiming to reveal the functional network organization and information processing methods of the brain in different time periods. Specifically, in the field of neuroscience, microstates refer to relatively stable spatial patterns of brain activity within a certain time range, although these patterns may change over different time periods. Microstates can be obtained through complex signal processing and analysis of electroencephalogram (EEG) data, typically exhibiting a relatively stable spatial distribution of EEG activity patterns that correspond to the activity characteristics of the brain in different functional states.

[0049] The core idea of ​​microstate analysis is to divide continuous EEG signals or brain imaging data into different microstates, and to quantify and compare these microstates in order to reveal the functional network organization and information processing methods of the brain in different states.

[0050] Based on the above-mentioned microstate analysis approach, in one possible embodiment, step S2 uses Cartool software to perform microstate calculations on the patients' resting-state EEG data (90s of EEG data were extracted from each patient) after preprocessing in step S1. For example, firstly, the K-means algorithm is used for individual-level microstate clustering; secondly, the K-means algorithm is used for group-level clustering to obtain a group-level template topographic map; finally, the template topographic map is fitted to the original EEG data. If the EEG data of non-delirious patients and delirious patients are each divided into two groups for microstate clustering analysis, a microstate transition sequence is obtained for each subject group after microstate clustering. In this embodiment, the original EEG is segmented based on this sequence, and then the EEG signals corresponding to the characteristic microstates are extracted, laying the foundation for studying the brain spatial characteristics of patients with resting-state delirium.

[0051] More specifically, step S2 includes sub-steps S201 to S203.

[0052] S201, calculate the global field power (GFP) for each electrode in the postoperative resting-state EEG data of the patient. Global field power (GFP) is an important EEG signal analysis indicator used to describe the intensity and pattern of overall brain electrical activity. For example, in this embodiment, global field power (GFP) characterizes the intensity of the brain's electric field at each instant.

[0053] Understandably, in microstate analysis, GFP can quickly reveal the overall activity level of EEG signals, facilitating comparisons of EEG activity across different time periods or experimental conditions. Using GFP transforms multi-channel EEG data into a single indicator, reducing data complexity and facilitating subsequent cluster analysis. This dimensionality reduction helps researchers better understand the overall characteristics of EEG signals, promoting the discovery and classification of microstates. Furthermore, GFP provides a comprehensive description of overall EEG activity, representing the neural activity state of the entire brain. Therefore, using GFP in microstate analysis allows for a more comprehensive examination of the overall patterns of brain activity, rather than being limited to the activity of specific brain regions or channels. Choosing GFP as a clustering indicator in EEG signal microstate analysis simplifies the data analysis process while providing a comprehensive description of overall EEG activity, helping to reveal the overall characteristics and patterns of brain activity.

[0054] GFP is calculated as follows: (1)

[0055] In equation (1), i represents the electrode index and n represents the number of electrodes. At time point t, Ui(t) represents the voltage measurement of electrode i relative to the reference point, while ui(t) represents the average instantaneous potential of all electrodes at that time point.

[0056] The peak of the GFP curve indicates the moment when the electric field strength is strongest and the signal-to-noise ratio of the topographic map is highest. In microstate analysis, the electric field topographic map at the peak of the GFP curve is considered as the discrete state of the EEG, and the changes in the signal are regarded as a series of state transitions.

[0057] S202 uses global field power (GFP) as a clustering index and employs a clustering algorithm to cluster the EEG data corresponding to all electrodes into EEG data in multiple microstates. Each microstate represents a quasi-static period corresponding to a certain region in the brain map. The cluster consists of multiple microstates: microstate A, microstate B, microstate C, and microstate D, which correspond one-to-one with the right frontolateral posterior region, left frontolateral posterior region, midline fronto-occipital region, and midline frontal region in the brain map.

[0058] Understandably, interpreting EEG recordings as a series of topographical states, using the electric field topography at the peak of the GFP curve, presents two significant characteristics. First, although many potential maps exist in multichannel recordings, most of the signal is represented by only a few topographical features. Interestingly, most studies examining resting-state EEGs report four prototype microstates that explain a major portion of global topographical changes. These four maps are labeled A, B, C, and D: right fronto-posterior, left fronto-right posterior, midline fronto-occipital, and midline fronto-topographical maps. The quasi-static period formed by such single maps is called a microstate. Therefore, interpreting EEG as a time-varying potential topography, the entire signal can be simplified and represented by a set of discrete, alternating topographical maps.

[0059] In cluster analysis, the potential topography at all GFP peaks is first extracted and then fed into a clustering algorithm. This algorithm categorizes these topographic features into several classes based on their similarity, regardless of their order of occurrence. Next, the topographic map at each GFP peak is assigned to one of these classes, thus re-presenting the EEG signal as a sequence of microstate classes. The frequency of occurrence of each microstate is the average number of times the microstate dominates per second during recording. The coverage of a microstate is a portion of the total recording time during which the microstate dominates. Temporal coverage refers to the total duration of each microstate class as a percentage of the total resting-state EEG time. The topographic shapes of the four microstate maps (A, B, C, and D) are frequently compared between different groups and behavioral states. Figure 3The diagram shows the topological clusters of four microstates for patients without delirium and patients with delirium. In microstate analysis, changes in brain state are described based on changes in these parameters.

[0060] For example, in this embodiment, the Modified K-Means Clustering algorithm is used for cluster analysis.

[0061] The improved K-means clustering algorithm is a stochastic clustering method based on a linear model of EEG data. Each EEG data vector (topographic map) Xi is simulated as a linear combination of M representative microstate maps, with microstate indices l = 0…M-1 and n... ch The residual vector consists of three identical and independently distributed Gaussian random variables. i. The general expression for this data model is: (2)

[0062] Furthermore, the solution can be found by minimizing the cost functional using equation (3).

[0063] (3)

[0064] As mentioned above, competitive backfit selection That is, having index L i A single microstate diagram to represent the instantaneous EEG topology. Instead of using the coefficient α il A linear combination of M microstate diagrams. For the linear model, competitive backfitting is equivalent to selecting αiL. i = l, and for l≠L i α il = 0.

[0065] The algorithm uses M randomly selected input vectors. , i = 0, ..., M 1. Initialization is performed, using EEG data vectors at M randomly selected GFP maxima in this case. The modified K-means clustering algorithm iteratively finds the local minimum of the cost function, performing two computational steps in each iteration. First, in each iteration, A is assigned according to the current cluster. l• Microstate sequence L i The calculation principle is the same as that of AAHC, so I won't go into too much detail. Next, using the previously calculated microstate label sequence L... iThe microstate graph is then updated. The updated microstate graph Al• is defined as the normalized eigenvector of matrix Sl with the largest eigenvalue, and Sl is defined as shown in equations (4) to (5): (4)

[0066] (5)

[0067] Using the eigenvector method is equivalent to solving for all n ch Calculate the maximum value in the column vector U such that |U| = 1. Understand L i Choosing the correct index is necessary for the above summation. Sequence L i The cluster assignment is essentially defined and changes during optimization. Convergence is evaluated by the change in relative residual variance. Given a microstate assignment L... i EEG data at the GFP maximum ( The relationship between the residual variance and the variance is shown in equation (6), and they are directly proportional.

[0068] (6)

[0069] For optimization purposes, the normalization constant for obtaining the correct variance value can be omitted, as it does not change during iteration. Regarding EEG topology, the eigenvector method implicitly ignores the polarity of the map because the eigenvector only defines n. ch One direction in the dimensional EEG sensor space. Since the initialization steps are randomized, different algorithm runs will produce different microstate diagrams. To approach the global maximum of the optimization problem, it is recommended to run the K-means clustering algorithm multiple times.

[0070] The best result from ten runs is selected to define the microstate. For each run of the algorithm, the convergence criterion is set to a maximum of 500 iterations. The relative error is 10 -6 The cross-validation (CV) selection algorithm based on previous research is shown in equation (7): (7)

[0071] in, It is calculated based on equation (6), including the entire EEG dataset X. ij The time index i = 0, ..., nt 1. The optimal K-means clustering algorithm is the one with the smallest cross-validation (CV) value. The optimal number of clusters was determined to be 4 using the cross-validation (CV) criterion, so the range for the number of clusters was chosen to be 2-6. Then, the optimal clustering coefficient was determined to be 4 based on the maximum globally explained variance (GEV). The patient's brain states were clustered into four classes based on GFP: microstate A, microstate B, microstate C, and microstate D, corresponding one-to-one with the right frontolateral posterior region, left frontolateral posterior region, midline fronto-occipital region, and midline frontal region on the brain map.

[0072] S203, filter out the EEG data corresponding to microstate C as the EEG data within the characteristic time period.

[0073] Figure 4 This study presents statistical results on the time coverage of four microstates in two groups: patients without delirium and patients with delirium. Statistical analysis of microstate indicators revealed a statistical difference in the time coverage of microstate C between patients with and without delirium. This may reflect that the EEG signals corresponding to microstate C contain relevant information distinguishing the two patient groups. Specifically, the study found that the time coverage of microstate C was significantly lower in patients with delirium than in those without. Furthermore, research on fMRI scans of the brain during microstate C revealed activation associated with the dorsal anterior cingulate cortex, bilateral inferior frontal gyrus cortex, and right insula. These brain regions are related to consciousness, cognitive function, and pain regulation. A decrease in the time coverage of microstate C may lead to a decline in the activation levels of these brain regions, thereby affecting the patient's consciousness, cognition, and pain regulation functions. Postoperatively, patients may develop delirium as a clinical manifestation due to the decline in these functional levels. In addition, some researchers have proposed that the functional connectivity of the brain in microstate C is significantly related to diseases such as consciousness and cognition. Therefore, this invention extracts the EEG signals corresponding to microstate C of each patient based on the state sequence, and studies the level of brain functional connectivity in this state in the subsequent step S4.

[0074] EEG signals contain a wealth of information and can be used to analyze interactive activities in different brain regions. However, when analyzing a single brain disease or problem, whole-electrode analysis can lead to information redundancy or interference from irrelevant signals. Therefore, electrode selection is particularly important when analyzing a class of problems. Step S3 utilizes the Common Spatial Pattern (CSP) algorithm for electrode selection. CSP is a powerful signal processing technique specifically designed to extract spatial features that distinguish different states from multi-channel signals. It not only enhances data quality but also optimizes channel configuration by identifying the most important channels in the analysis. CSP is an algorithm for extracting features through spatial filtering in binary classification tasks. The core of CSP lies in finding a suitable spatial filter for signal projection through matrix diagonalization. This process highlights the variance differences between the two classes of signals, achieving the goal of extracting feature vectors with significant classification capabilities.

[0075] In this embodiment of the invention, characteristic electrodes for delirium patients are selected by using CSP analysis on the time series of differential states. Specifically, step S3 uses the CSP algorithm to select the most representative electrodes based on the EEG signals under the characteristic states (microstate C) extracted in step S2, and then conducts further research. Figure 5 This demonstrates the distribution of channel-selected electrodes on the cerebral cortex. Figure 5 The yellow electrodes are the selected characteristic electrodes. The results of this invention show that the location distribution of these characteristic electrodes on the cerebral cortex basically matches the activation status of brain regions in each microstate.

[0076] In one possible embodiment, step S3 includes sub-steps S301 to S303.

[0077] S301, the Common Spatial Pattern (CSP) algorithm is applied to the EEG data extracted by microstate analysis to separate the spatial feature components of each channel in the multi-channel brain-computer interface data; S302: Each channel is scored separately based on the spatial feature component weights to find the most effective spatial distribution features for the classification task. S303, based on the characteristics of the spatial distribution data, a suitable threshold is preset; Based on the scoring results of each channel and the preset score threshold, the electrodes with scores above the score threshold in the analysis are selected as the electrodes corresponding to the most representative channels, and the EEG data corresponding to the selected electrodes are extracted.

[0078] Understandably, step S3, through the CSP algorithm, can find the optimal channel configuration scheme while considering multiple constraints, thereby improving the accuracy and reliability of EEG signal recording. It can effectively handle various constraints in the EEG channel selection process, such as signal interference, scalp location, and electrode number limitations. CSP channel selection provides a more accurate and reliable EEG channel configuration scheme for neuroscience research and clinical medicine, offering a more precise research scope for subsequent brain functional connectivity analysis.

[0079] Electroencephalography (EEG) signals capture the electrical activity generated by neurons in the brain. They contain a wealth of information, encompassing not only the brain's physiological state but also higher-order functions such as memory and consciousness. Therefore, EEG signal analysis is a convenient and effective tool for exploring the mysteries of the brain. However, EEG analysis is a very challenging task because brain signals are highly complex and non-stationary biological signals with highly complex dynamic behavior.

[0080] To better extract the rich information from EEG signals in the time and frequency domains and address their inherent limitations, numerous evaluation metrics describing EEG connectivity have been developed. Based on the statistical correlation between signals, functional connectivity of the brain can be measured using metrics such as coherence, phase synchronization, generalized synchronization, and Granger causality. Phase lock value (PLV) is a quantitative measure of the degree of phase synchronization. PLV separates the phase and amplitude components, whose influence is negligible. Furthermore, due to the influence of eye movements and other factors, the instantaneous amplitude of EEG signals typically varies widely, making it suitable for EEG analysis.

[0081] The calculation principle of PLV will now be illustrated with an example.

[0082] Given signal (Y) i and Y j Y and the target frequency (f) can be calculated in three steps. i and Y j PLV at point f: (1) Based on the target frequency, filter Y using a bandpass filter. i and Y j Perform passband filtering to obtain X i and X j .

[0083] (2) The instantaneous phase Φ is calculated by Hilbert transform.

[0084] (3) After obtaining the instantaneous phases Φi and Φj of the two signals, the PLV value between the two signals is calculated based on the phase difference θ(t) between the two signals.

[0085] The composition of the signal X(t) obtained after the original signal Y(t) passes through the bandpass filter is shown in equation (8): (8)

[0086] in The sequence obtained by performing the Hilbert transform on X(t) is shown in equation (9): (9)

[0087] Where PV represents the Cauchy median.

[0088] Equation (10) is the method for calculating the instantaneous phase. The instantaneous phases of the two passing signals are calculated to obtain Φi and Φj respectively. The signal difference θ(t) between the two signals is calculated by equation (11), and then the PLV between the two signals is calculated by equation (12).

[0089] (10)

[0090] (11)

[0091] (12)

[0092] In equation (12), N represents x i All points within the range. PLV is an average value that varies between 0 and 1. If PLV = 1, it means x i and x j Fully synchronized.

[0093] In step S4 of this embodiment of the invention, based on the above principle, the phase-locked value (PLV) is used to quantitatively analyze the brain functional connectivity between the most representative electrodes in a specific state.

[0094] In one possible embodiment, step S4 includes sub-steps S401 to S403.

[0095] S401, For the multiple electrodes and their EEG data selected in S3, calculate the phase lock value (PLV) between the electrodes by frequency band to construct the functional connectivity matrix. S402, perform statistical difference analysis on the functional connectivity matrix of the target frequency band to obtain multiple significantly different edges under the target frequency band; S403 uses multiple significantly different edges to plot the Cortical Phase Locked Network (CPLN) to visualize anomalous connection patterns.

[0096] This embodiment uses the HERMES toolkit in the Matlab platform to calculate the PLV value between the selected electrodes, with a time window length of 4s and a sliding overlap rate of 50%.

[0097] Understandably, steps S2 and S3 determine the EEG signal time periods and electrode ranges for brain functional connectivity analysis, followed by step S4. The data analysis in step S4 comprises three parts: First, to describe the average connection strength between electrodes in each frequency band, the average PLV value in the brain network constructed from the selected electrodes is calculated and statistically analyzed. Second, to analyze the causal relationships between connections, a PLV connectivity matrix for each frequency band is constructed, and statistical analysis and discussion are performed to identify edges with statistical differences. Then, based on these statistically different edges, a Cortical Phase Locked Network (CPLN) is constructed to visually demonstrate the differences in the brain networks of delirium patients.

[0098] In a specific implementation scenario, after channel selection in step S3, 11 EEG channels corresponding to the electrodes are obtained. The average PLV between the 11 EEG channels in each frequency band is calculated for patients who developed delirium post-surgery and those who did not. The average connectivity strength and statistical differences within the brain network formed by these electrodes are also analyzed. Figure 6 The average PLV statistics for each frequency band are shown. The average PLV levels of the Alpha and Beta bands are significantly enhanced, with the enhancement of brain connectivity being more pronounced in the Alpha band. Therefore, the Alpha band was selected as the target frequency band.

[0099] like Figure 7 The diagram shows the cortical phase-locked network (CPLN) for delirious and non-delirious patients at each frequency band. Step S4 generates the CPLN based on multiple significantly different edges, which can then be used as input to a pre-defined delirium prediction model. Step S5 yields the delirium prediction result corresponding to the EEG data being tested, indicating whether the tested patient is delirious.

[0100] Before step S5, the process also includes building and training a delirium prediction model, specifically including: 1. Take the resting-state EEG data of several known delirium patients and the resting-state EEG data of several non-delirium patients as two groups of samples in the sample set and preprocess them in step S1; 2. The sample set is processed sequentially through steps S2 to S4 to extract the number of significant differences between the two types of patients' EEG data in each frequency band; wherein, when processing the sample set through step S2, the category of the target microstate is determined by whether there is a significant difference in time coverage between the two types of patients' EEG data. 3. The phase-locked value (PLV) corresponding to the significantly different edge is used as a nonlinear feature, and multiple feature subsets are divided according to frequency band; 4. Based on the CatBoost model architecture, delirium prediction binary classification models are constructed and trained for feature subsets under each frequency band. The delirium prediction binary classification model with better classification performance is selected as the final delirium prediction model.

[0101] For example, the functional connectivity of the brain to 11 EEG electrodes selected based on CSP channels was analyzed. By calculating the PLV between electrodes in each frequency band, 55 undirected EEG nonlinear features were obtained for each frequency band. These feature subsets, corresponding to the frequency bands, are denoted as Fdelta, Ftheta, Falpha, and Fbeta. Feature selection will then be performed on these feature subsets.

[0102] The feature selection method used in this implementation scenario is based on statistical methods. Statistical methods were used to explore which connections showed significant differences between the two patient groups. Based on the analysis results, features that did not show statistical differences between the two groups of data corresponding to the two patient groups were removed from the original feature subset. After the entire feature selection process was completed, four new feature matrices were obtained based on the feature set corresponding to each frequency band. The new feature subsets obtained based on feature selection and the number of features in each feature subset are shown in Table 1.

[0103] Table 1 Number of features in the feature subset

[0104] Based on the CatBoost model architecture, delirium prediction binary classification models were constructed and trained using feature subsets for each frequency band, and their classification accuracy was verified. For example, resting-state EEG data from 46 patients (including 23 delirium patients and 23 non-delirium patients) were used to construct the model during the entire study. In this embodiment, a recognition model was constructed using feature subsets for each frequency band, and the classification metrics of each feature were compared. The classification accuracy of the feature sets across the four frequency bands was analyzed. Specific results are shown in Table 2.

[0105] Table 2 Classification results of features in each frequency band

[0106] The results in Table 2 show that when using the CatBoost model for classification model research, the classifier using the feature subset corresponding to the alpha band as feature input has better accuracy, sensitivity, specificity, and volatility than the classification model using feature subsets of the beta, delta, and theta bands. As can be seen from Table 2, although the feature dimension of the alpha band is slightly higher than other bands, its classification model has better classification performance, which effectively compensates for this. Therefore, this embodiment of the invention selects the feature vector of the alpha band to construct the delirium classification model.

[0107] After constructing and training the delirium prediction model, delirium prediction can be achieved through step S5. In one possible embodiment, step S5 includes sub-steps S501-S502: S501, extract the phase-locked value (PLV) corresponding to multiple significantly different edges under the target frequency band from the output of step S4 to form a nonlinear feature vector. S502, the extracted nonlinear feature vector is input into the preset delirium prediction model to obtain the category label of delirium identification and the corresponding risk probability.

[0108] Figure 8 A structural diagram of a brain function abnormality detection system for postoperative delirium patients provided in an embodiment of the present invention is shown below. Figure 8 As shown, a system for detecting abnormal brain function in patients with postoperative delirium includes an acquisition and preprocessing module, a time window filtering module, a channel configuration module, a brain network connectivity analysis module, and a classification and recognition module, wherein: The acquisition and preprocessing module is used to acquire and preprocess the raw EEG data of the postoperative patient at rest. The time window filtering module is used to extract feature time from the preprocessed raw EEG data based on microstate analysis to obtain EEG data under the target microstate time window. The channel configuration module is used to configure the EEG data obtained from the time window screening module based on the common spatial pattern (CSP) to obtain multiple representative electrodes within a characteristic time period. The brain network connectivity analysis module is used as the electrodes obtained from the channel configuration module as brain network connectivity analysis nodes, and calculates the phase lock value (PLV) between electrodes within a preset frequency band to extract significant difference edges in the target frequency band. The classification and recognition module is used to extract feature vectors based on the significant difference edges extracted from the brain network connectivity analysis module and input them into a preset delirium prediction model to output delirium recognition results.

[0109] It is understood that the postoperative delirium patient brain function abnormality detection system provided by the present invention corresponds to the postoperative delirium patient brain function abnormality detection method provided in the foregoing embodiments. The relevant technical features of the postoperative delirium patient brain function abnormality detection system can be referred to the relevant technical features of the postoperative delirium patient brain function abnormality detection method, and will not be repeated here.

[0110] The present invention provides a method and system for detecting abnormal brain function in patients with postoperative delirium, and its main innovations are as follows: (1) In view of the current shortcomings in the study of the neural mechanism of delirium based on EEG, this invention applies microstate analysis method to the study of time window selection for delirium based on EEG for the first time, which solves the problem of fixed window length limitation of sliding time window or subjective selection of time window. It performs state self-division of resting EEG signals of delirium and non-delirium patients, and realizes brain state division based on data-driven high temporal resolution.

[0111] (2) This invention is the first to apply the co-space pattern algorithm to channel selection in the analysis of brain connectivity in delirium patients, solving the problems of performance degradation, data redundancy, and interference from irrelevant channel signals in current EEG-based delirium research. It also eliminates the subjective influence of manually selecting channels to reduce research complexity. In short, our research, for the first time, enables precise research and characterization of brain connectivity and brain networks in delirium patients at high temporal resolution using engineering methods. This is expected to deepen our understanding of the neural mechanisms of delirium and enrich research findings in this field both domestically and internationally.

[0112] (3) For the first time, it was proposed to use the nonlinear characteristics of brain networks in characteristic microstates to automatically identify delirium patients, and finally establish a stable and efficient delirium disease prediction and classification system.

[0113] In summary, this invention demonstrates that microstate analysis is of significant reference value for time selection during resting-state EEG analysis and effectively solves the problem of low temporal resolution in traditional time selection methods. Furthermore, the application of the co-space pattern algorithm, commonly used in brain-computer interfaces, to channel selection of resting-state EEG signals is feasible, and the selection results are consistent with previous studies. Finally, based on the above two methods, a precise analysis of brain functional connectivity in postoperative delirium patients and non-delirium patients was conducted, revealing significant differences between the two groups. Combining machine learning and other tools, a predictive model for postoperative delirium was constructed, which may provide new insights for the prediction and diagnosis of postoperative delirium in the future.

[0114] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0115] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0116] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0117] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for detecting abnormal brain function in patients with postoperative delirium, characterized in that, include: S1, Acquire raw EEG data of the patient at rest after surgery and perform preprocessing; S2, based on microstate analysis, extracts feature time from the preprocessed raw EEG data to obtain EEG data under the target microstate time window; S3, based on the common spatial pattern (CSP), performs channel configuration on the EEG data obtained in S2 to obtain multiple representative electrodes within the characteristic time period; S4. Using the electrodes obtained in S3 as nodes for brain network connectivity analysis, calculate the phase lock value (PLV) between electrodes within a preset frequency band to extract significant difference edges in the target frequency band. S5 extracts feature vectors based on the significant differences extracted in S4 and inputs them into a preset delirium prediction model to output delirium identification results.

2. The method for detecting abnormal brain function in patients with postoperative delirium according to claim 1, characterized in that, Step S1 includes: Based on clinical requirements, the raw EEG data of the postoperative patient at rest and its corresponding electrode coordinates are obtained. The raw EEG data includes EEG data collected from all electrode channels. The raw EEG data is filtered, and artifacts in the raw EEG data are removed based on independent component analysis (ICA). The Kurtosis algorithm is used to automatically detect bad channels, mark and remove the electrodes and EEG data corresponding to the bad channels, and the spherical algorithm is used to interpolate and complete the removed data. The EEG data corresponding to each electrode were averaged and referenced.

3. The method for detecting abnormal brain function in patients with postoperative delirium according to claim 1, characterized in that, Step S2 includes: S201, calculate the global field power (GFP) for each electrode in the EEG data of the postoperative patient in the resting state. The global field power (GFP) represents the intensity of the brain electric field at each instant. S202 uses global field power (GFP) as a clustering index and employs a clustering algorithm to cluster the EEG data corresponding to all electrodes into EEG data in multiple microstates. Each microstate represents a quasi-static period corresponding to a certain region in the brain map. The cluster consists of multiple microstates: microstate A, microstate B, microstate C, and microstate D, which correspond one-to-one with the right frontolateral posterior region, left frontolateral posterior region, midline fronto-occipital region, and midline frontal region in the brain map. S203, filter out the EEG data corresponding to microstate C as the EEG data within the characteristic time period.

4. A method for detecting abnormal brain function in patients with postoperative delirium according to claim 1 or 3, characterized in that, Step S3 includes: S301, the Common Spatial Pattern (CSP) algorithm is applied to the EEG data extracted by microstate analysis to separate the spatial feature components of each channel in the multi-channel brain-computer interface data; S302, each channel is scored separately based on the spatial feature component weights; S303, based on the scoring results of each channel and the preset score threshold, selects electrodes corresponding to multiple representative channels and extracts the EEG data corresponding to the selected electrodes.

5. The method for detecting abnormal brain function in patients with postoperative delirium according to claim 1, characterized in that, Step S4 includes: S401, For the multiple electrodes and their EEG data selected in S3, calculate the phase lock value (PLV) between the electrodes by frequency band to construct the functional connectivity matrix. S402, perform statistical difference analysis on the functional connectivity matrix of the target frequency band to obtain multiple significantly different edges under the target frequency band.

6. The method for detecting abnormal brain function in patients with postoperative delirium according to claim 5, characterized in that, Step S4 also includes: S403 plots a cortical phase-locked network (CPLN) based on multiple significantly different edges to visualize anomalous connection patterns.

7. A method for detecting abnormal brain function in patients with postoperative delirium according to claim 5 or 6, characterized in that, Step S5 includes: S501, extract the phase-locked value (PLV) corresponding to multiple significantly different edges under the target frequency band to form a nonlinear feature vector; S502, the extracted nonlinear feature vector is input into the preset delirium prediction model to obtain the category label of delirium identification and the corresponding risk probability.

8. The method for detecting abnormal brain function in patients with postoperative delirium according to claim 1, characterized in that, Before step S5, the process also includes constructing a delirium prediction model, specifically including: Resting-state EEG data from multiple known delirium patients and multiple non-delirium patients were used as a sample set and preprocessed. The sample set is processed sequentially through steps S2 to S4 to extract the number of significant differences in the EEG data of the two types of patients in each frequency band. The phase-locked value (PLV) corresponding to the significantly different edge is used as a nonlinear feature, and multiple feature subsets are divided according to frequency band; Based on the CatBoost model architecture, delirium prediction binary classification models were constructed and trained for feature subsets under each frequency band. The delirium prediction binary classification model with better classification performance was selected as the final delirium prediction model.

9. A method for detecting abnormal brain function in patients with postoperative delirium according to claim 8, characterized in that, When the sample set is processed in step S2, the category of the target microstate is determined by whether there is a significant difference in time coverage between the EEG data of the two types of patients.

10. A system for detecting abnormal brain function in patients with postoperative delirium, characterized in that, include: The acquisition and preprocessing module is used to acquire and preprocess the raw EEG data of the postoperative patient in a resting state. The time window filtering module is used to extract feature time from the preprocessed raw EEG data based on microstate analysis to obtain EEG data under the target microstate time window. The channel configuration module is used to configure the EEG data obtained from the time window screening module based on the common spatial pattern (CSP) to obtain multiple representative electrodes within a characteristic time period. The brain network connectivity analysis module is used as the electrodes obtained from the channel configuration module as brain network connectivity analysis nodes, and calculates the phase lock value (PLV) between electrodes within a preset frequency band to extract significant difference edges in the target frequency band. The classification and recognition module is used to extract feature vectors based on the significant difference edges extracted from the brain network connectivity analysis module and input them into a preset delirium prediction model to output delirium recognition results.

Images