ICU patient delirium risk early screening system based on AI electrocardiogram analysis

Abstract

The invention relates to an ICU patient delirium risk early screening system based on AI electrocardiogram analysis. The system comprises a neural signal extraction module, a component mode determination module, a delirium type identification module, a risk trajectory prediction module and a risk alarm generation module, wherein the neural signal extraction module extracts neuromodulation related signal fragments based on an electrocardiogram signal sequence and a preset neuromodulation specificity index; a component mode determination module extracts time sequence features from the fragments and determines an abnormal component distribution mode; the delirium type identification module is used for matching potential delirium types through the feature database; the risk trajectory prediction module predicts a risk trajectory in combination with a long-short-term memory network; a risk alert generation module evaluates a risk level and generates an alert. With the adoption of the system, early and dynamic screening of delirium risks can be realized, the screening accuracy and timeliness are improved through multi-module collaborative analysis, and accurate monitoring support is provided for ICU (Intensive Care Unit) patients.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and healthcare, and in particular relates to an early screening system for delirium risk in ICU patients based on AI electrocardiogram analysis. Background Technology

[0002] With the deep integration of artificial intelligence and healthcare technologies, intelligent monitoring technology for critically ill patients based on multimodal physiological signal analysis has emerged. This technology can dynamically assess disease risk by continuously collecting vital sign data, and its advantage lies in overcoming the limitations of traditional single-time-point assessments. Current ICU delirium screening methods are mostly based on clinical scales and manual observation: medical staff mainly rely on the CAM-ICU scale for regular checks, combined with recording abnormal manifestations by observing patient behavior. Current screening methods have the following problems: scale assessment depends on patient cooperation and has poor applicability to sedated patients; manual observation struggles to capture nocturnal symptom fluctuations, easily missing subclinical cases; traditional methods lack the ability to quantitatively analyze abnormal neuroregulation in electrocardiogram signals, making it difficult to establish a mapping relationship from physiological parameters to delirium types, resulting in short warning windows and high false alarm rates. Summary of the Invention

[0003] Therefore, it is necessary to provide an early screening system for delirium risk in ICU patients based on AI electrocardiogram analysis that can solve the above problems.

[0004] Firstly, this application provides an early screening system for delirium risk in ICU patients based on AI electrocardiogram analysis, comprising:

[0005] The neural signal extraction module is used to extract neural regulation-related signal fragments based on the electrocardiogram signal sequences of ICU patients and according to preset neural regulation-specific indicators.

[0006] The component pattern determination module is used to extract the temporal features of neural modulation-related signal segments and determine the distribution pattern of abnormal components based on the temporal features.

[0007] The delirium type identification module is used to identify potential delirium types based on the distribution patterns of abnormal components and through a preset delirium type feature database.

[0008] The risk trajectory prediction module, based on potential delirium types and combined with abnormal component distribution patterns, uses a preset long short-term memory network to predict delirium risk trajectories.

[0009] The risk alert generation module determines the delirium risk level based on the delirium risk trajectory and generates a delirium risk alert by combining the delirium risk trajectory and the delirium risk level.

[0010] In one embodiment, the preset neural regulation-specific indicators of the neural signal extraction module include heart rate variability indicators, QT interval variability indicators, and electrocardiogram complexity indicators.

[0011] Based on the electrocardiogram signal sequences of ICU patients, neural regulation-related signal fragments are extracted according to preset neural regulation-specific indicators, including:

[0012] The electrocardiogram signal sequence is divided into multiple time window sequences using a preset sliding window.

[0013] For each time window series, perform the following calculations:

[0014] SDNN values ​​and LF / HF ratios were calculated as indicators of heart rate variability.

[0015] The QT interval detection method was used to calculate the dynamic fluctuation parameter value of the QT interval as an indicator of QT interval variability.

[0016] The sample entropy calculation method is used to calculate the quantified value of the complexity of the electrocardiogram signal as an index of electrocardiogram complexity;

[0017] For each time window sequence, the heart rate variability index, QT interval variability index, and electrocardiogram complexity index are compared with preset specific index thresholds. Time window sequences that meet the specific index thresholds are selected as neural regulation-related signal segments.

[0018] In one embodiment, the component pattern determination module is further configured to:

[0019] Based on neural modulation-related signal fragments, time-domain features, frequency-domain features, and time-frequency-domain features are extracted as temporal features;

[0020] By combining specific indicators of neural regulation, the temporal features are weighted and fused to obtain fused temporal features;

[0021] Clustering algorithms are used to perform clustering on the fused temporal features to obtain feature clusters;

[0022] Calculate the feature similarity between each feature cluster and the preset normal baseline features;

[0023] Clusters of features with similarity below a preset similarity threshold are selected as abnormal clusters.

[0024] Extract the feature distribution parameters of the temporal features of abnormal clusters;

[0025] A distribution feature matrix is ​​constructed based on the feature distribution parameters to serve as the distribution pattern of abnormal components.

[0026] In one embodiment, the delirium type identification module's preset delirium type feature database contains type feature matrices corresponding to various potential delirium types:

[0027] Based on the distribution patterns of abnormal components, potential delirium types are identified through a pre-defined delirium type feature database, including:

[0028] Determine the percentage overlap of feature dimensions between the distribution feature matrix and each type of feature matrix to obtain the feature dimension overlap rate;

[0029] Select type feature matrices with feature dimension overlap rates higher than preset feature dimension overlap rate thresholds to form candidate type feature matrices;

[0030] Extract the eigenvalues ​​of each feature dimension of the distribution feature matrix and the candidate type feature matrix;

[0031] Based on the distribution feature matrix and the feature matrix of any candidate type, the weighted deviation value is calculated by combining the eigenvalues ​​and feature weights.

[0032] Based on the weighted bias value, the delirium types corresponding to the top N candidate type feature matrices with the smallest weighted bias value are selected and collectively used as potential delirium types.

[0033] In one embodiment, the preset long short-term memory network of the risk trajectory prediction module includes an attention fusion layer, a double hidden layer, a dropout layer and an output layer. The attention fusion layer adopts a multi-head attention mechanism. The first and second hidden layers in the double hidden layer each contain 80 long short-term memory units and adopt the ELU activation function. The dropout layer has a dropout rate of 0.35. The output layer adopts a linear activation function with feature weights.

[0034] Based on potential delirium types and anomalous component distribution patterns, a pre-defined long short-term memory network is used to predict delirium risk trajectories, including:

[0035] The distribution feature matrix is ​​concatenated with the type feature matrix corresponding to the potential delirium type to form the model input matrix;

[0036] Based on the model input matrix and feature weights, an attention fusion layer is used to perform attention weighting to form an attention weighting matrix;

[0037] Based on the attention weighting matrix, the first hidden layer is used to extract temporal correlation features to obtain the primary correlation vector;

[0038] Based on the primary correlation vector, a dropout layer is used to perform random deactivation denoising to obtain the denoised correlation vector;

[0039] Based on the denoised correlation vector, a second hidden layer is used to extract deep temporal evolution features to obtain a high-level evolution vector;

[0040] The output layer is used to perform temporal trajectory mapping on the high-level evolution vector to form delirium risk trajectory.

[0041] In one embodiment, the risk alert generation module is further configured to:

[0042] Based on the delirium risk trajectory and combined with the preset risk level classification standards, the threshold range of each risk level is determined;

[0043] Based on the delirium risk trajectory, a sliding window is used to statistically quantify the risk at preset time points;

[0044] By comparing the quantified risk value with the threshold range, the risk level corresponding to each time point can be obtained.

[0045] By combining the risk level with the delirium risk trajectory, a delirium risk alert is generated.

[0046] In one embodiment, the delirium type identification module is further configured to use the following formula to filter out the delirium types corresponding to the top N candidate type feature matrices with the smallest weighted deviation values ​​based on the weighted deviation values, and collectively use them as potential delirium types:

[0047]

[0048] in, This is a type of potential delirium. Let k be the delirium type corresponding to the k-th candidate type feature matrix, where k is the index identifier of the candidate type feature matrix. The set of indices for the candidate type feature matrix. This represents the weighted bias value corresponding to the feature matrix of the k-th candidate type. The feature dimension overlap rate between the feature matrix of the k-th candidate type and the distribution feature matrix. is the feature dimension weight of the feature matrix of the k-th candidate type, N is the preset screening threshold, and D is the total number of feature dimensions.

[0049] Secondly, this application also provides a method for early screening of delirium risk in ICU patients based on AI electrocardiogram analysis, including:

[0050] Based on the electrocardiogram signal sequences of ICU patients, neural regulation-related signal fragments were extracted according to preset neural regulation-specific indicators;

[0051] Temporal features of neural modulation-related signal fragments are extracted, and the distribution patterns of abnormal components are determined based on these temporal features.

[0052] Based on the distribution patterns of abnormal components, potential delirium types are identified through a pre-set delirium type feature database;

[0053] Based on potential delirium types and combined with abnormal component distribution patterns, a pre-set long short-term memory network is used to predict delirium risk trajectories.

[0054] Based on the delirium risk trajectory, the delirium risk level is determined, and a delirium risk alert is generated by combining the delirium risk trajectory and the delirium risk level.

[0055] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to realize the functions of the above-mentioned early screening system for delirium risk in ICU patients based on AI electrocardiogram analysis.

[0056] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the functions of the aforementioned AI-based electrocardiogram analysis-based early screening system for delirium risk in ICU patients.

[0057] The aforementioned AI-based electrocardiogram (ECG) analysis-based early screening system for delirium risk in ICU patients utilizes a neural signal extraction module to extract neural modulation-related signal fragments based on the ECG signal sequences of ICU patients and preset neural modulation-specific indicators, achieving preliminary screening of raw ECG data. A component pattern determination module extracts the temporal features of these signal fragments and determines the distribution patterns of abnormal components, revealing potential abnormalities through feature analysis and providing a foundation for subsequent identification. A delirium type identification module uses a preset delirium type feature database to match abnormal distribution patterns and identify potential delirium types, improving the targeted nature of the diagnosis. A risk trajectory prediction module, combined with a long short-term memory network, predicts risk trajectories based on abnormal component distribution patterns and potential delirium types, achieving dynamic risk monitoring. A risk alarm generation module determines the risk level based on the risk trajectory and generates alarms to ensure timely intervention, achieving early and dynamic screening for delirium risk. Attached Figure Description

[0058] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0059] Figure 1 This is a structural diagram of an early screening system for delirium risk in ICU patients based on AI electrocardiogram analysis, according to the present invention.

[0060] Figure 2 This is a flowchart of an early screening method for delirium risk in ICU patients based on AI electrocardiogram analysis, according to the present invention. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0062] In one embodiment, such as Figure 1 As shown, an early screening system for delirium risk in ICU patients based on AI electrocardiogram analysis is provided. This embodiment illustrates the system by deploying it on a terminal. It is understood that the system can also be deployed on a server, or in an architecture that includes both a terminal and a server, and is implemented through the interaction between the terminal and the server.

[0063] The implementation environment includes: terminal devices such as embedded smart monitors connected to electrocardiogram (ECG) sensors to collect patients' ECG signal sequences in real time, and transmitting the pre-processed data to the server via a wireless network; the server integrates the system's neural signal extraction module, component pattern determination module, delirium type identification module, risk trajectory prediction module, and risk alarm generation module, supporting high-concurrency processing. Application scenarios include: when ICU patients exhibit autonomic nervous system regulation abnormalities, the terminal collects raw ECG data, and the server initiates multi-module collaborative analysis: the neural signal extraction module filters related signal fragments based on preset neural regulation-specific indicators (such as heart rate variability and QT interval variability) and outputs them to the component pattern determination module; the component pattern determination module extracts temporal features and generates abnormal component distribution patterns, which are then passed to the delirium type identification module; the delirium type identification module matches potential delirium types through a feature database and inputs them into the risk trajectory prediction module; the risk trajectory prediction module uses a long short-term memory network to predict the risk trajectory; and the risk alarm generation module assesses the risk level and generates an alarm, which is then returned to the terminal for display, achieving a closed-loop interaction from data collection to early warning, improving screening timeliness.

[0064] In this embodiment, the system includes:

[0065] The neural signal extraction module 101 is used to extract neural regulation-related signal fragments based on the electrocardiogram signal sequence of ICU patients and according to preset neural regulation-specific indicators.

[0066] Among them, the preset specific indicators of neural regulation are quantifiable physiological parameters such as heart rate variability, QT interval variability, or electrocardiogram complexity. In practice, based on these indicators, the electrocardiogram signal sequences of ICU patients—continuously acquired electrophysiological data streams—can be used to filter out signal segments related to neural regulation by using signal processing techniques (such as sliding window segmentation of sequences) and calculating indicator thresholds for comparison, thus providing basic data input for subsequent analysis.

[0067] The component pattern determination module 102 is used to extract the temporal features of neural regulation-related signal segments and determine the distribution pattern of abnormal components based on the temporal features.

[0068] Among them, the temporal features of neural regulation-related signal segments can cover multiple feature types such as time domain features (e.g., mean, variance), frequency domain features (e.g., power spectral density), or time-frequency domain features (e.g., wavelet transform results). Based on these temporal features, the distribution patterns of anomalous components can be determined through feature fusion techniques (e.g., weighted fusion) and pattern recognition methods (e.g., clustering algorithms). For example, by calculating feature similarity and screening anomalous clusters to construct a distribution feature matrix as the distribution pattern of anomalous components, a quantitative characterization of potential anomalies can be achieved.

[0069] The delirium type identification module 103 is used to identify potential delirium types based on the distribution pattern of abnormal components and through a preset delirium type feature database.

[0070] Among them, based on the distribution pattern of abnormal components, potential delirium types can be identified through a preset delirium type feature database (containing type feature matrices corresponding to various potential delirium types). Through pattern matching techniques such as feature dimension overlap rate calculation and weighted deviation value comparison (e.g., by screening candidate feature matrices and determining multiple types of output based on the principle of minimizing deviation), accurate differentiation of delirium subtypes can be achieved.

[0071] The risk trajectory prediction module 104, based on potential delirium types and combined with abnormal component distribution patterns, uses a preset long short-term memory network to predict delirium risk trajectories.

[0072] Among them, based on the potential delirium type (i.e., multiple delirium subtypes matched from the feature database) and the distribution pattern of abnormal components (such as the distribution feature matrix), the delirium risk trajectory can be predicted by a preset long short-term memory network (including components such as attention fusion layer, hidden layer and output layer to achieve temporal modeling). In practice, the dynamic prediction of the evolution of delirium risk can be achieved through steps such as feature matrix splicing, attention weighting, temporal correlation feature extraction and trajectory mapping.

[0073] The risk alert generation module 105 determines the delirium risk level based on the delirium risk trajectory and generates a delirium risk alert by combining the delirium risk trajectory and the delirium risk level.

[0074] Among them, the delirium risk level is determined based on the delirium risk trajectory (i.e., the time-series risk quantification data obtained from the prediction module) (such as by using preset risk level classification standards such as threshold range comparison and sliding window statistical technology), and delirium risk alarm is generated by combining the delirium risk trajectory and delirium risk level. In practice, timely intervention on delirium risk can be achieved through methods such as dynamic calculation of risk quantification value, adaptive adjustment of level threshold and multi-level alarm triggering mechanism, supporting closed-loop monitoring.

[0075] The aforementioned AI-based electrocardiogram (ECG) analysis-based early screening system for delirium risk in ICU patients automatically analyzes ECG signals through a neural signal extraction module to avoid reliance on patient cooperation; a component pattern determination module enables quantitative analysis to overcome the limitations of manual observation; a delirium type identification module matches a feature database; a risk trajectory prediction module provides dynamic early warnings to extend the window period; and a risk alarm generation module assesses the risk level in real time. Through the collaboration of multiple modules, the system improves the accuracy and timeliness of early delirium screening and reduces the false alarm rate.

[0076] In one embodiment, the preset neural regulation-specific indicators of the neural signal extraction module 101 include heart rate variability indicators, QT interval variability indicators, and electrocardiogram complexity indicators.

[0077] Based on the electrocardiogram signal sequences of ICU patients, neural regulation-related signal fragments are extracted according to preset neural regulation-specific indicators, including:

[0078] The electrocardiogram signal sequence is divided into multiple time window sequences using a preset sliding window.

[0079] For each time window series, perform the following calculations:

[0080] SDNN values ​​and LF / HF ratios were calculated as indicators of heart rate variability.

[0081] The QT interval detection method was used to calculate the dynamic fluctuation parameter value of the QT interval as an indicator of QT interval variability.

[0082] The sample entropy calculation method is used to calculate the quantified value of the complexity of the electrocardiogram signal as an index of electrocardiogram complexity;

[0083] For each time window sequence, the heart rate variability index, QT interval variability index, and electrocardiogram complexity index are compared with preset specific index thresholds. Time window sequences that meet the specific index thresholds are selected as neural regulation-related signal segments.

[0084] For example, preset neural regulation-specific indicators include heart rate variability, QT interval variability, and electrocardiogram complexity. In implementation, a preset sliding window (window size can be set to 5 minutes, step size 1 minute, adjusted according to the frequency of clinical data acquisition) can be used to divide the continuous electrocardiogram signal sequence into multiple non-overlapping and continuous time window sequences. Indicator calculations are performed separately for each time window sequence. The heart rate variability indicator is calculated by measuring the standard deviation of the sinus beat RR interval (i.e., the SDNN value) and the low-frequency power versus high-frequency power. The power ratio (i.e., the LF / HF ratio) is obtained. When calculating the SDNN value, the RR intervals corresponding to all sinus beats within the identified time window are used. The dispersion of this group of RR intervals is calculated using the standard deviation formula. When calculating the LF / HF ratio, a Fast Fourier Transform is performed on the RR interval sequence to separate the power values ​​of the 0.04-0.15Hz low-frequency band and the 0.15-0.4Hz high-frequency band, and then the ratio of the two is taken. The QT interval variability index adopts a QT interval detection method based on waveform feature recognition (combining thresholding and slope detection to locate Q). The Q point and T wave endpoint of the RS complex are used to calculate the dynamic fluctuation parameters such as the difference and standard deviation of adjacent QT intervals within the time window. The ECG complexity index is calculated using the sample entropy method. The embedding dimension m=2 and the similarity tolerance r=0.2 times the standard deviation of the ECG signal within the time window are set. The sample entropy value is calculated by statistically analyzing the probability distribution of similar patterns in the signal sequence, and this is used as the quantitative value of the ECG signal complexity. For each time window sequence, the calculated heart rate variability index, QT interval variability index, and ECG complexity index are compared with preset specific index thresholds (determined based on clinical big data statistics, such as SDNN value threshold set to 100ms, LF / HF ratio threshold set to 2.5, QT interval dynamic fluctuation parameter value threshold set to 30ms, and sample entropy value threshold set to 0.8, which can be dynamically adjusted according to clinical validation results). Time window sequences that meet the corresponding threshold requirements for all three indicators (or the judgment rule that at least one indicator meets the threshold is set according to clinical needs) are selected. The selected time window sequences are the neural regulation-related signal segments.

[0085] In one embodiment, the component pattern determination module 102 is further configured to:

[0086] Based on neural modulation-related signal fragments, time-domain features, frequency-domain features, and time-frequency-domain features are extracted as temporal features;

[0087] By combining specific indicators of neural regulation, the temporal features are weighted and fused to obtain fused temporal features;

[0088] Clustering algorithms are used to perform clustering on the fused temporal features to obtain feature clusters;

[0089] Calculate the feature similarity between each feature cluster and the preset normal baseline features;

[0090] Clusters of features with similarity below a preset similarity threshold are selected as abnormal clusters.

[0091] Extract the feature distribution parameters of the temporal features of abnormal clusters;

[0092] A distribution feature matrix is ​​constructed based on the feature distribution parameters to serve as the distribution pattern of abnormal components.

[0093] Specifically, time-domain features, frequency-domain features, and time-frequency-domain features can be extracted as time-series features. Time-domain features include the signal's mean, variance, peak value, trough value, kurtosis, skewness, and standard deviation of the interval between adjacent peaks. Frequency-domain features can be obtained through Fast Fourier Transform (FFT) to acquire the signal's power spectral density, center frequency, frequency band energy proportion, and bandwidth. Time-frequency-domain features can be obtained by performing a three-level wavelet transform using the db4 wavelet basis function to extract the mean and variance of wavelet coefficients at each scale. Combining these with three neural regulation-specific indicators—heart rate variability, QT interval variability, and electrocardiogram complexity—the analytic hierarchy process (AHP) is used to determine... Assign weights to each indicator (e.g., heart rate variability 0.4, QT interval variability 0.3, ECG complexity 0.3), and assign corresponding weights to each time-series feature based on its correlation with specific neural regulation indicators. Perform weighted fusion of all time-series features to obtain fused time-series features. K-means clustering can be used to cluster the fused time-series features. The number of clusters K is determined by the elbow rule (range 5-8). Randomly initialize K cluster centers, calculate the Euclidean distance between each fused time-series feature and each cluster center, and assign the features to... The cluster containing the nearest cluster center is iteratively updated until the change in the center between two consecutive iterations is less than a preset threshold (e.g., 0.001), resulting in a feature cluster. The feature similarity between each feature cluster and a preset normal baseline feature is calculated. This preset normal baseline feature is a standard cluster feature obtained from neuromodulation-related signal fragments of 1000 delirium-free ICU patients, obtained through identical feature extraction, weighted fusion, and clustering. The feature similarity is calculated using a cosine similarity algorithm. Feature clusters with feature similarity below a preset similarity threshold (clinically validated at 0.7) are selected. The abnormal clusters are identified as anomalous clusters. Feature distribution parameters of the temporal features of these anomalous clusters are extracted. These feature distribution parameters include the mean, variance, standard deviation, median, upper quartile, lower quartile, and peak value of the probability density function for each temporal feature. A distribution feature matrix is ​​constructed based on these feature distribution parameters as the anomalous component distribution pattern. During matrix construction, each temporal feature of the anomalous cluster is used as a row, and the corresponding feature distribution parameters are used as columns. The distribution parameters of each temporal feature are sequentially filled into the corresponding positions in the matrix according to a preset order. The resulting two-dimensional numerical matrix is ​​the anomalous component distribution pattern.

[0094] In one embodiment, the delirium type identification module 103's preset delirium type feature database contains type feature matrices corresponding to various potential delirium types:

[0095] Based on the distribution patterns of abnormal components, potential delirium types are identified through a pre-defined delirium type feature database, including:

[0096] Determine the percentage overlap of feature dimensions between the distribution feature matrix and each type of feature matrix to obtain the feature dimension overlap rate;

[0097] Select type feature matrices with feature dimension overlap rates higher than preset feature dimension overlap rate thresholds to form candidate type feature matrices;

[0098] Extract the eigenvalues ​​of each feature dimension of the distribution feature matrix and the candidate type feature matrix;

[0099] Based on the distribution feature matrix and the feature matrix of any candidate type, the weighted deviation value is calculated by combining the eigenvalues ​​and feature weights.

[0100] Based on the weighted bias value, the delirium types corresponding to the top N candidate type feature matrices with the smallest weighted bias value are selected and collectively used as potential delirium types.

[0101] For example, a pre-defined delirium type feature database is established, which is based on clinically diagnosed potential delirium types (such as manic delirium, depressive delirium, mixed delirium, etc.). By collecting neuromodulation-related signal fragments from corresponding patients, and through the same feature extraction, weighted fusion, cluster analysis, and matrix construction process as the component pattern determination module, a set of type feature matrices specific to each delirium type is formed. The feature dimensions of each type feature matrix are consistent with the distribution feature matrix corresponding to the abnormal component distribution. In implementation, the overlap ratio of feature dimensions between the distribution feature matrix and each type feature matrix in the database is determined to obtain the feature dimension overlap rate. Type feature matrices with a feature dimension overlap rate higher than a pre-defined feature dimension overlap rate threshold (set to 70% based on clinical data validation) are selected and integrated to form candidate type feature matrices. The feature matrix set is used to extract the specific values ​​of each feature dimension in the distribution feature matrix and each candidate type feature matrix as feature values. The correspondence of feature dimensions is matched based on feature name, calculation method and physiological meaning. Based on the distribution feature matrix and any candidate type feature matrix, the feature weights corresponding to each feature dimension (consistent with the weights used in the weighted fusion of time-series features in the component pattern determination module, determined by the analytic hierarchy process) are used to calculate the weighted deviation value between the two. Based on the weighted deviation values ​​corresponding to all candidate type feature matrices, they are sorted in ascending order, and the top N candidate type feature matrices with the smallest weighted deviation values ​​(N is the preset screening quantity threshold, which can be set to 3 according to clinical diagnostic needs) are selected. The delirium types corresponding to these matrixes are used as the potential delirium types of the patient.

[0102] In one embodiment, the preset long short-term memory network of the risk trajectory prediction module 104 includes an attention fusion layer, a double hidden layer, a dropout layer and an output layer. The attention fusion layer adopts a multi-head attention mechanism. The first and second hidden layers in the double hidden layer each contain 80 long short-term memory units and adopt the ELU activation function. The dropout layer has a dropout rate of 0.35. The output layer adopts a linear activation function with feature weights.

[0103] Based on potential delirium types and anomalous component distribution patterns, a pre-defined long short-term memory network is used to predict delirium risk trajectories, including:

[0104] The distribution feature matrix is ​​concatenated with the type feature matrix corresponding to the potential delirium type to form the model input matrix;

[0105] Based on the model input matrix and feature weights, an attention fusion layer is used to perform attention weighting to form an attention weighting matrix;

[0106] Based on the attention weighting matrix, the first hidden layer is used to extract temporal correlation features to obtain the primary correlation vector;

[0107] Based on the primary correlation vector, a dropout layer is used to perform random deactivation denoising to obtain the denoised correlation vector;

[0108] Based on the denoised correlation vector, a second hidden layer is used to extract deep temporal evolution features to obtain a high-level evolution vector;

[0109] The output layer is used to perform temporal trajectory mapping on the high-level evolution vector to form delirium risk trajectory.

[0110] Specifically, the pre-defined Long Short-Term Memory (LSTM) network includes an attention fusion layer, two hidden layers, a dropout layer, and an output layer. The attention fusion layer employs an 8-head multi-head attention mechanism (each head has an attention dimension of 16). Both the first and second hidden layers of the two hidden layers contain 80 LSTM units, and both layers use the ELU (Exponential Linear Unit) activation function to enhance gradient propagation. The random dropout rate of the dropout layer is set to 0.35 (i.e., randomly setting the output of 35% of neurons to zero to avoid overfitting). The output layer uses a linear activation function with feature weights (feature weights and component pattern determination module). The weights used in the time-series feature weighted fusion are kept consistent. In implementation, the distribution feature matrix corresponding to the abnormal component distribution pattern and the type feature matrix corresponding to each potential delirium type can be concatenated column-wise (ensuring alignment of features with the same physiological meaning during concatenation; if dimensional differences exist, zero-padding can be used to complete to a consistent dimension), forming a model input matrix with unified dimensions. Based on the model input matrix and preset feature weights, the attention weights of each head are calculated separately through the 8-head multi-head attention mechanism of the attention fusion layer (attention weights are obtained through matrix operations of the query vector, key vector, and value vector; the query vector is generated by linear transformation of the model input matrix). The key vector and value vector are obtained by multiplying the feature weight matrix and the model input matrix, respectively. The attention outputs of each head are concatenated and fused through a linear transformation to form an attention weighting matrix, which enables the model to focus on key features related to delirium risk. The attention weighting matrix is ​​input into the first hidden layer, and the temporal information in the matrix is ​​modeled through 80 LSTM units to extract the temporal correlation of features under different time windows. After processing by the ELU activation function, a primary correlation vector is obtained. Based on this primary correlation vector, a random deactivation and noise reduction operation is performed through a dropout layer to randomly block the neuron outputs corresponding to some redundant features, resulting in... The model's generalization ability is improved by denoising the correlation vector. The denoised correlation vector is then input into the second hidden layer, where 80 LSTM units are used to further mine the deep temporal evolution patterns of the features (such as the trend of feature changes over time, abrupt change nodes, etc.). After processing by the ELU activation function, a high-level evolution vector is obtained. The high-level evolution vector is then input into the output layer, where a linear activation function with feature weights is used to map its temporal trajectory. This outputs the delirium risk quantification value corresponding to each time node within a preset future time range (such as the next 72 hours, with a time step of 1 hour). These risk quantification values ​​arranged in chronological order together constitute the delirium risk trajectory.

[0111] In one embodiment, the risk alert generation module 105 is further configured to:

[0112] Based on the delirium risk trajectory and combined with the preset risk level classification standards, the threshold range of each risk level is determined;

[0113] Based on the delirium risk trajectory, a sliding window is used to statistically quantify the risk at preset time points;

[0114] By comparing the quantified risk value with the threshold range, the risk level corresponding to each time point can be obtained.

[0115] By combining the risk level with the delirium risk trajectory, a delirium risk alert is generated.

[0116] For example, based on the delirium risk trajectory output by the risk trajectory prediction module, and combined with a preset risk level classification standard, the threshold range for each risk level is determined. This preset risk level classification standard is based on clinical delirium occurrence data and prognostic results of 5,000 ICU patients, and is calibrated using the quartile method, dividing the risk into four levels: low risk threshold range 0-0.3, medium risk threshold range 0.3-0.6, high risk threshold range 0.6-0.8, and very high risk threshold range >0.8. Each threshold range can be dynamically fine-tuned according to the characteristics of the hospital's ICU patient population. Based on the delirium risk trajectory, a preset sliding window (window size set to 3 time steps, i.e., 3 hours, step size set to 1 time step, i.e., 1 hour) is used to statistically analyze the risk quantification value at each preset time node. The arithmetic mean of all risk quantification values ​​within each sliding window is taken as the risk quantification value for the corresponding preset time node (the time node corresponding to the center of the window). The risk quantification value at each time node is compared with the threshold range for each risk level. If the risk quantification value falls within a certain threshold range, the risk level corresponding to that time node is determined to be within that range. The risk level is categorized as high risk (e.g., a risk quantification value of 0.75). A delirium risk alert is generated by combining the risk level at each time point with the complete delirium risk trajectory. The alert content may include the patient's unique identifier ID, the risk level and corresponding risk quantification value at each time point, the trend of the risk trajectory (rising / falling / stable, determined by the slope of the difference in risk quantification values ​​between adjacent time points; a slope > 0.05 indicates rising, < -0.05 indicates falling, and the rest indicates stable), and specific time intervals for high-risk and extremely high-risk points. The alert format can combine visual pop-ups and audio prompts. Visual pop-ups correspond to different colors based on risk level (low risk: green; medium risk: yellow; high risk: orange; extremely high risk: red). Audio prompts are set to different frequencies based on risk level (low risk: single low-pitched alert; medium risk: two medium-pitched alerts; high risk: three high-pitched alerts; extremely high risk: continuous beeping alert with a 0.5-second interval). The alert information can also be synchronized to the ICU monitoring center's backend system and the corresponding medical staff's mobile terminal APP, and supports read / unread status marking and historical alert query functions.

[0117] In one embodiment, the delirium type identification module 103 is further configured to use the following formula to filter out the delirium types corresponding to the top N candidate type feature matrices with the smallest weighted deviation values ​​based on the weighted deviation values, and collectively use them as potential delirium types:

[0118]

[0119] in, This is a type of potential delirium. Let k be the delirium type corresponding to the k-th candidate type feature matrix, where k is the index identifier of the candidate type feature matrix. The set of indices for the candidate type feature matrix. This represents the weighted bias value corresponding to the feature matrix of the k-th candidate type. The feature dimension overlap rate between the feature matrix of the k-th candidate type and the distribution feature matrix. is the feature dimension weight of the feature matrix of the k-th candidate type, N is the preset screening threshold, and D is the total number of feature dimensions.

[0120] Specifically, the index set for the candidate type feature matrix. For each index k in the candidate type feature matrix, the weighted bias value corresponding to the k-th candidate type feature matrix is ​​called. The overlap rate of feature dimensions between the candidate matrix and the distribution feature matrix The clinical importance of the feature dimensions in the feature matrix of each candidate type for the diagnosis of delirium is determined in advance (this can be determined by the analytic hierarchy process, with values ​​ranging from 0 to 1, and the sum of the feature dimension weights of the candidate matrices in the same batch being 1, such as candidate matrices with a high proportion of key physiological features). The feature dimension weights are determined by setting the weight to 0.8 (and 0.5 for features with a high proportion of minor features). The comprehensive evaluation value is calculated according to the formula, that is This value comprehensively reflects the matching degree between the candidate type feature matrix and the distribution feature matrix (the smaller the weighted deviation value, the higher the feature dimension overlap rate, and the greater the feature dimension weight, the better the comprehensive evaluation value). The comprehensive evaluation values ​​of all candidate type feature matrices are sorted in ascending order using the rank function (the smaller the ranking after sorting, the better the comprehensive matching degree). Delirium types corresponding to candidate type feature matrices whose ranking after sorting is ≤ a preset screening quantity threshold N (N is set to 3 after clinical validation, and can be dynamically adjusted to 2-5 according to the diversity of delirium types in the ICU patient population) are selected. Finally, all the selected Tk values ​​are grouped into a set, which represents the potential delirium types that need to be identified. This allows for the precise screening of the top N potential delirium types that best match the distribution pattern of abnormal components in patients. The total number of feature dimensions, D, serves only as a baseline reference for relevant parameters (e.g., for verifying feature dimension overlap). (The calculation is reasonable), and it does not participate in the direct calculation process of the formula.

[0121] The aforementioned AI-based electrocardiogram (ECG) analysis-based early screening system for delirium risk in ICU patients utilizes a neural signal extraction module to extract neuromodulation-related signal fragments based on the ECG signal sequences of ICU patients and preset neuromodulation-specific indicators. This eliminates the need for active patient cooperation and addresses the poor applicability of traditional scale assessments to sedated patients. A component pattern determination module extracts the temporal characteristics of these signal fragments and identifies the distribution patterns of abnormal components, enabling quantitative analysis of neuromodulation abnormalities in ECG signals. This overcomes the shortcomings of traditional methods, which lack such quantitative analysis capabilities and struggle to establish a mapping relationship between physiological parameters and delirium types. A delirium type identification module further enhances this capability. Based on the distribution patterns of abnormal components, potential delirium types are identified through a pre-set delirium type feature database, laying the foundation for accurate risk assessment. The risk trajectory prediction module combines potential delirium types with abnormal component distribution patterns and uses a pre-set long short-term memory network to predict delirium risk trajectories. This can dynamically capture symptom fluctuations, avoiding the difficulty of manually observing nighttime abnormalities and missing subclinical cases, and extending the warning window period. The risk alarm generation module judges the risk level based on the risk trajectory and generates alarms. Through systematic analysis of multiple modules, the false alarm rate is reduced, enabling early and dynamic screening of delirium risk in ICU patients and providing precise monitoring support for clinical practice.

[0122] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0123] Based on the same inventive concept, this application also provides a method for early screening of delirium risk in ICU patients based on AI electrocardiogram analysis, which implements the aforementioned system for early screening of delirium risk in ICU patients based on AI electrocardiogram analysis. The solution provided by this method is similar to the implementation described in the above system. Therefore, the specific limitations of one or more embodiments of the method for early screening of delirium risk in ICU patients based on AI electrocardiogram analysis provided below can be found in the limitations of the system for early screening of delirium risk in ICU patients based on AI electrocardiogram analysis described above, and will not be repeated here.

[0124] In one exemplary embodiment, such as Figure 2 As shown, an early screening method for delirium risk in ICU patients based on AI electrocardiogram analysis is provided, including:

[0125] S01, based on the electrocardiogram signal sequence of ICU patients, extracts neural regulation-related signal fragments according to preset neural regulation-specific indicators;

[0126] S02, extract the temporal features of neural regulation-related signal segments, and determine the distribution pattern of abnormal components based on the temporal features;

[0127] S03, Based on the distribution pattern of abnormal components, identify potential delirium types through a pre-set delirium type feature database;

[0128] S04, based on potential delirium types and combined with abnormal component distribution patterns, uses a pre-set long short-term memory network to predict delirium risk trajectories;

[0129] S05. Based on the delirium risk trajectory, determine the delirium risk level, and generate a delirium risk alert by combining the delirium risk trajectory and the delirium risk level.

[0130] In one embodiment, the preset neural regulation-specific indicators include heart rate variability, QT interval variability, and electrocardiogram complexity.

[0131] Based on the electrocardiogram signal sequences of ICU patients, neural regulation-related signal fragments are extracted according to preset neural regulation-specific indicators, including:

[0132] The electrocardiogram signal sequence is divided into multiple time window sequences using a preset sliding window.

[0133] For each time window series, perform the following calculations:

[0134] SDNN values ​​and LF / HF ratios were calculated as indicators of heart rate variability.

[0135] The QT interval detection method was used to calculate the dynamic fluctuation parameter value of the QT interval as an indicator of QT interval variability.

[0136] The sample entropy calculation method is used to calculate the quantified value of the complexity of the electrocardiogram signal as an index of electrocardiogram complexity;

[0137] For each time window sequence, the heart rate variability index, QT interval variability index, and electrocardiogram complexity index are compared with preset specific index thresholds. Time window sequences that meet the specific index thresholds are selected as neural regulation-related signal segments.

[0138] In one embodiment, temporal features of neural modulation-related signal segments are extracted, and the distribution pattern of abnormal components is determined based on the temporal features, including:

[0139] Based on neural modulation-related signal fragments, time-domain features, frequency-domain features, and time-frequency-domain features are extracted as temporal features;

[0140] By combining specific indicators of neural regulation, the temporal features are weighted and fused to obtain fused temporal features;

[0141] Clustering algorithms are used to perform clustering on the fused temporal features to obtain feature clusters;

[0142] Calculate the feature similarity between each feature cluster and the preset normal baseline features;

[0143] Clusters of features with similarity below a preset similarity threshold are selected as abnormal clusters.

[0144] Extract the feature distribution parameters of the temporal features of abnormal clusters;

[0145] A distribution feature matrix is ​​constructed based on the feature distribution parameters to serve as the distribution pattern of abnormal components.

[0146] In one embodiment, the preset delirium type feature database contains type feature matrices corresponding to various potential delirium types;

[0147] Based on the distribution patterns of abnormal components, potential delirium types are identified through a pre-defined delirium type feature database, including:

[0148] Determine the percentage overlap of feature dimensions between the distribution feature matrix and each type of feature matrix to obtain the feature dimension overlap rate;

[0149] Select type feature matrices with feature dimension overlap rates higher than preset feature dimension overlap rate thresholds to form candidate type feature matrices;

[0150] Extract the eigenvalues ​​of each feature dimension of the distribution feature matrix and the candidate type feature matrix;

[0151] Based on the distribution feature matrix and the feature matrix of any candidate type, the weighted deviation value is calculated by combining the eigenvalues ​​and feature weights.

[0152] Based on the weighted bias value, the delirium types corresponding to the top N candidate type feature matrices with the smallest weighted bias value are selected and collectively used as potential delirium types.

[0153] In one embodiment, the preset long short-term memory network includes an attention fusion layer, a double hidden layer, a dropout layer, and an output layer. The attention fusion layer adopts a multi-head attention mechanism. In the double hidden layer, the first and second hidden layers each contain 80 long short-term memory units and adopt the ELU activation function. The dropout layer has a dropout rate of 0.35. The output layer adopts a linear activation function with feature weights.

[0154] Based on potential delirium types and anomalous component distribution patterns, a pre-defined long short-term memory network is used to predict delirium risk trajectories, including:

[0155] The distribution feature matrix is ​​concatenated with the type feature matrix corresponding to the potential delirium type to form the model input matrix;

[0156] Based on the model input matrix and feature weights, an attention fusion layer is used to perform attention weighting to form an attention weighting matrix;

[0157] Based on the attention weighting matrix, the first hidden layer is used to extract temporal correlation features to obtain the primary correlation vector;

[0158] Based on the primary correlation vector, a dropout layer is used to perform random deactivation denoising to obtain the denoised correlation vector;

[0159] Based on the denoised correlation vector, a second hidden layer is used to extract deep temporal evolution features to obtain a high-level evolution vector;

[0160] The output layer is used to perform temporal trajectory mapping on the high-level evolution vector to form delirium risk trajectory.

[0161] In one embodiment, based on the delirium risk trajectory, the delirium risk level is determined, and a delirium risk alert is generated by combining the delirium risk trajectory and the delirium risk level, including:

[0162] Based on the delirium risk trajectory and combined with the preset risk level classification standards, the threshold range of each risk level is determined;

[0163] Based on the delirium risk trajectory, a sliding window is used to statistically quantify the risk at preset time points;

[0164] By comparing the quantified risk value with the threshold range, the risk level corresponding to each time point can be obtained.

[0165] By combining the risk level with the delirium risk trajectory, a delirium risk alert is generated.

[0166] In one embodiment, based on the weighted bias value, the delirium types corresponding to the top N candidate type feature matrices with the smallest weighted bias values ​​are selected and collectively used as potential delirium types, achieved through the following formula:

[0167]

[0168] in, This is a type of potential delirium. Let k be the delirium type corresponding to the k-th candidate type feature matrix, where k is the index identifier of the candidate type feature matrix. The set of indices for the candidate type feature matrix. This represents the weighted bias value corresponding to the feature matrix of the k-th candidate type. The feature dimension overlap rate between the feature matrix of the k-th candidate type and the distribution feature matrix. is the feature dimension weight of the feature matrix of the k-th candidate type, N is the preset screening threshold, and D is the total number of feature dimensions.

[0169] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the functions of an AI-based electrocardiogram analysis-based early screening system for delirium risk in ICU patients as described above.

[0170] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, performs the functions described in the system embodiments above.

[0171] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0172] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. An early screening system for delirium risk in ICU patients based on AI electrocardiogram analysis, characterized in that, The system includes: The neural signal extraction module is used to extract neural regulation-related signal fragments based on the electrocardiogram signal sequences of ICU patients and according to preset neural regulation-specific indicators. The component pattern determination module is used to extract the temporal features of the neural modulation-related signal segments and determine the abnormal component distribution pattern based on the temporal features. The delirium type identification module is used to identify potential delirium types based on the distribution pattern of the abnormal components and through a preset delirium type feature database. The risk trajectory prediction module, based on the potential delirium type and combined with the abnormal component distribution pattern, uses a preset long short-term memory network to predict the delirium risk trajectory. The risk alert generation module determines the delirium risk level based on the delirium risk trajectory, and generates a delirium risk alert by combining the delirium risk trajectory and the delirium risk level.

2. The system according to claim 1, characterized in that, The preset neural regulation-specific indicators of the neural signal extraction module include heart rate variability, QT interval variability, and electrocardiogram complexity. The electrocardiogram signal sequence based on ICU patients, according to preset neural regulation-specific indicators, extracts neural regulation-related signal fragments, including: The electrocardiogram signal sequence is divided into multiple time window sequences using a preset sliding window. For each of the aforementioned time window sequences, perform the following calculations: The SDNN value and LF / HF ratio were calculated as indicators of heart rate variability. The QT interval detection method is used to calculate the dynamic fluctuation parameter value of the QT interval as the QT interval variability index. The sample entropy calculation method is used to calculate the quantified value of the complexity of the electrocardiogram signal as the electrocardiogram complexity index; For each of the aforementioned time window sequences, the heart rate variability index, the QT interval variability index, and the electrocardiogram complexity index are compared with preset specificity index thresholds, and time window sequences that meet the specificity index thresholds are selected as the neural regulation-related signal segments.

3. The system according to claim 1, characterized in that, The component pattern determination module is also used for: Based on the neural modulation-related signal fragments, time-domain features, frequency-domain features, and time-frequency-domain features are extracted as the temporal features; By combining the aforementioned neural regulation-specific indicators, the temporal features are weighted and fused to obtain fused temporal features; A clustering algorithm is used to perform clustering on the fused temporal features to obtain feature clusters; Calculate the feature similarity between each of the aforementioned feature clusters and the preset normal baseline features; Feature clusters with similarity lower than a preset similarity threshold are selected as abnormal clusters. Extract the feature distribution parameters of the temporal features of the abnormal clusters; A distribution feature matrix is ​​constructed based on the aforementioned feature distribution parameters, serving as the distribution pattern of the abnormal components.

4. The system according to claim 3, characterized in that, The delirium type identification module's preset delirium type feature database contains type feature matrices corresponding to various potential delirium types: The step of identifying potential delirium types based on the abnormal component distribution pattern and through a pre-set delirium type feature database includes: The overlap ratio of feature dimensions between the distribution feature matrix and each type feature matrix is ​​determined to obtain the feature dimension overlap rate. Select the type feature matrix whose feature dimension overlap rate is higher than the preset feature dimension overlap rate threshold, and form a candidate type feature matrix; Extract the eigenvalues ​​of each feature dimension of the distribution feature matrix and the candidate type feature matrix; Based on the distribution feature matrix and any of the candidate type feature matrices, and combining the feature values ​​and feature weights, a weighted deviation value is calculated; Based on the weighted deviation value, the delirium types corresponding to the top N candidate type feature matrices with the smallest weighted deviation value are selected and collectively used as the potential delirium types.

5. The system according to claim 4, characterized in that, The preset long short-term memory network of the risk trajectory prediction module includes an attention fusion layer, a double hidden layer, a discard layer, and an output layer. The attention fusion layer adopts a multi-head attention mechanism. The first and second hidden layers of the double hidden layer each contain 80 long short-term memory units and adopt the ELU activation function. The discard layer has a discard rate of 0.

35. The output layer adopts a linear activation function with feature weights. The method of predicting delirium risk trajectories using a preset long short-term memory network based on the potential delirium type and the abnormal component distribution pattern includes: The distribution feature matrix is ​​concatenated with the type feature matrix corresponding to the potential delirium type to form the model input matrix; Based on the model input matrix and the feature weights, the attention fusion layer is used to perform attention weighting to form an attention weighting matrix; Based on the attention weighting matrix, temporal correlation features are extracted using the first hidden layer to obtain a primary correlation vector; Based on the primary correlation vector, random deactivation denoising is performed using the discard layer to obtain a denoised correlation vector; Based on the denoised correlation vector, the deep temporal evolution features are extracted using the second hidden layer to obtain the high-level evolution vector; The output layer is used to perform temporal trajectory mapping on the high-level evolution vector to form the delirium risk trajectory.

6. The system according to claim 1, characterized in that, The risk alert generation module is also used for: Based on the delirium risk trajectory and in conjunction with the preset risk level classification criteria, the threshold range for each risk level is determined. Based on the delirium risk trajectory, a sliding window is used to statistically quantify the risk at preset time points; The risk quantification value is compared with the threshold range to obtain the risk level corresponding to each time point; The delirium risk alert is generated by combining the risk level with the delirium risk trajectory.

7. The system according to claim 4, characterized in that, The delirium type identification module is further configured to use the following formula to filter out the delirium types corresponding to the top N candidate type feature matrices with the smallest weighted deviation values ​​based on the weighted deviation values, and collectively use them as the potential delirium types: in, This is a type of potential delirium. Let k be the delirium type corresponding to the k-th candidate type feature matrix, where k is the index identifier of the candidate type feature matrix. The set of indices for the candidate type feature matrix. This represents the weighted bias value corresponding to the feature matrix of the k-th candidate type. The feature dimension overlap rate between the feature matrix of the k-th candidate type and the distribution feature matrix. is the feature dimension weight of the feature matrix of the k-th candidate type, N is the preset screening threshold, and D is the total number of feature dimensions.

8. A method for early screening of delirium risk in ICU patients based on AI electrocardiogram analysis, characterized in that, The method includes: Based on the electrocardiogram signal sequences of ICU patients, neural regulation-related signal fragments were extracted according to preset neural regulation-specific indicators; Temporal features of the neural regulation-related signal segments are extracted, and the distribution pattern of abnormal components is determined based on the temporal features. Based on the abnormal component distribution pattern, potential delirium types are identified through a pre-set delirium type feature database; Based on the potential delirium types and the abnormal component distribution patterns, a preset long short-term memory network is used to predict delirium risk trajectories. Based on the delirium risk trajectory, the delirium risk level is determined, and a delirium risk alert is generated by combining the delirium risk trajectory and the delirium risk level.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it performs the functions of the system according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it performs the functions of the system according to any one of claims 1 to 7.

Images