Automatic evaluation method and system based on Grasgow coma score

Abstract

The invention provides an automatic evaluation method and system based on a Grasgow coma score. The method comprises the following steps: collecting behavior response, speech response and eye-opening response data of a patient, and generating an initial evaluation data set; based on the initial evaluation data set, performing advanced sequence feature extraction processing by applying a gating loop unit algorithm, performing deep emotion tendency analysis by adopting a deep emotion analysis technology, and generating a behavior speech reaction score; based on the behavior speech reaction score, extracting an eye opening reaction score from the eye opening reaction data by using a multi-modal adaptive fusion algorithm, performing dynamic weighted fusion processing through a preset rule, and generating a comprehensive evaluation index by using a deep reinforcement learning technology; and based on the comprehensive evaluation index, comparing the comprehensive evaluation index with a preset threshold range, automatically judging the current coma level of the patient, and generating a personalized medical suggestion report. According to the technical scheme provided by the invention, the automation and refinement degree of patient coma grade evaluation is remarkably improved.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of medical diagnosis technology, and in particular to an automatic evaluation method and system based on the Glasgow Coma Scale. Background Art

[0002] With the rapid development of medical information technology and artificial intelligence technology, the field of medical diagnosis and evaluation is undergoing profound changes. In modern clinical environments, accurate assessment of patients' coma status is crucial, especially in high-risk environments such as intensive care units and emergency departments. Traditional assessment methods rely on the subjective judgment of doctors, which has large individual differences and time delays. Therefore, there is an urgent need for a method that can automatically and finely assess the patient's coma level to improve the accuracy and efficiency of diagnosis.

[0003] At present, common coma assessment methods mainly include the traditional Glasgow Coma Scale and some rule-based semi-automated assessment tools. Although these methods can reflect the patient's coma state to a certain extent, they still rely on the doctor's experience and subjective judgment, and fail to make full use of multimodal data (such as behavioral response, verbal response, and eye opening response), resulting in incomplete information and affecting the accuracy of the assessment. In addition, most of the existing methods are based on fixed rules and lack a dynamic adjustment mechanism, making it difficult to adapt to the complex and changing clinical environment, especially in emergency situations, which may delay the best treatment time.

[0004] Although existing coma assessment methods are widely used in clinical practice, they still have significant defects, especially the low degree of automation and refinement of the assessment, which makes it difficult to capture subtle changes in the patient's condition, resulting in limited assessment accuracy. The manual assessment process is time-consuming and inefficient, and it is difficult to generate personalized medical advice based on the specific situation of each patient, which affects the optimization of treatment effects. Summary of the invention

[0005] The embodiments of the present application provide an automatic assessment method and system based on the Glasgow Coma Scale, so as to solve the problem of limited assessment accuracy in the prior art.

[0006] In a first aspect, an embodiment of the present application provides an automatic assessment method based on the Glasgow Coma Scale, comprising:

[0007] Collect patient behavioral response data, verbal response data, and eye opening response data to generate an initial evaluation data set;

[0008] Based on the initial evaluation data set, a gated recurrent unit algorithm is used to perform high-level sequence feature extraction processing on the behavioral response data, and a deep sentiment analysis technology is used to perform a deep-level sentiment tendency analysis on the verbal response data to generate a behavioral verbal response score;

[0009] Based on the behavioral and verbal response score, a multimodal adaptive fusion algorithm is used to extract the eye-opening response score from the eye-opening response data, and a dynamic weighted fusion process is performed with the behavioral and verbal response score through preset rules, and a comprehensive evaluation index is generated using deep reinforcement learning technology;

[0010] Based on the comprehensive evaluation index, a comparative analysis is performed with the preset coma level threshold range, the patient's current coma level is automatically determined, and a personalized medical advice report is generated.

[0011] Optionally, based on the initial evaluation data set, the gated recurrent unit algorithm is used to perform high-level sequence feature extraction processing on the behavioral response data, and the deep sentiment analysis technology is used to perform deep-level sentiment tendency analysis on the verbal response data to generate a behavioral verbal response score, including:

[0012] Based on the initial evaluation data set, preprocessing the behavior response data to ensure data format consistency and remove noise, and generating preprocessed behavior response data;

[0013] Based on the pre-processed behavior response data, a gated recurrent unit algorithm is used to control the information flow through a gating mechanism, capture the long-term dependency in the time series, and generate a behavior response feature vector through multi-layer gated recurrent unit processing and fully connected layer mapping;

[0014] Based on the behavior response feature vector, the speech response data is encoded using deep sentiment analysis technology to obtain context-related word vector representation, and the behavior response feature vector is combined with the input into the sentiment classifier to identify key sentiment information and generate a behavior speech feature vector;

[0015] Based on the behavioral speech feature vector, weighted fusion is performed through a preset weighting rule to generate a behavioral speech reaction score.

[0016] Optionally, based on the pre-processed behavior response data, a gated recurrent unit algorithm is used to control the information flow through a gating mechanism, capture long-term dependencies in the time series, and generate a behavior response feature vector through multi-layer gated recurrent unit processing and fully connected layer mapping, including:

[0017] Based on the preprocessed behavioral response data, the preprocessed behavioral response data is input into a gated recurrent unit network for initialization processing to generate an initial state;

[0018] Based on the initial state, a gated recurrent unit algorithm is used to control the flow of information by updating gates and resetting gates, so as to capture long-term dependencies in the time series and generate intermediate state representations;

[0019] Based on the intermediate state representation, high-level features are further extracted through multi-layer gated recurrent units, the internal state is updated, and the current moment behavior feature representation is output through an output gate to generate a deep feature representation;

[0020] Based on the deep feature representation, mapping is performed through multiple fully connected layers to convert high-dimensional feature vectors into fixed-length feature vectors to generate behavioral response feature vectors.

[0021] Optionally, based on the behavior response feature vector, the speech response data is encoded using deep sentiment analysis technology to obtain context-related word vector representation, combined with the behavior response feature vector input into a sentiment classifier, key sentiment information is identified, and a behavior speech feature vector is generated, including:

[0022] Based on the behavior response feature vector, adaptation processing is performed through feature mapping to adapt the speech response data to generate an adapted behavior response feature vector;

[0023] Based on the adaptive behavior response feature vector, the speech response data is encoded and processed using deep sentiment analysis technology, the adaptive behavior response feature vector is introduced, and a context-related word vector representation is generated;

[0024] Based on the context-related word vector representation, combined with the adaptation behavior response feature vector, weighted fusion processing is performed to combine multiple effective information to generate a combined feature vector;

[0025] Based on the combined feature vector, it is input into a dedicated emotion classifier, and deep feature extraction and pattern recognition are performed through a multi-layer neural network to identify key emotion information and generate a behavioral speech feature vector.

[0026] Optionally, based on the behavioral verbal response score, a multimodal adaptive fusion algorithm is used to extract the eye-opening response score from the eye-opening response data, and a dynamic weighted fusion process is performed with the behavioral verbal response score through preset rules, and a comprehensive evaluation index is generated using deep reinforcement learning technology, including:

[0027] Based on the behavioral verbal response score, standardization processing is performed to ensure consistency in scale and distribution, thereby generating a standardized behavioral verbal response score;

[0028] Based on the standardized behavioral speech response score, a multimodal adaptive fusion algorithm is used to extract the eye opening response score from the eye opening response data, perform joint feature extraction processing, and generate a joint feature vector;

[0029] Based on the joint feature vector, a dynamic weighted fusion process is performed through a preset weighting rule to generate a preliminary fusion score;

[0030] Based on the preliminary fusion score, deep reinforcement learning technology is used to continuously learn and optimize the weight allocation strategy, further refine the score, and generate a comprehensive evaluation index.

[0031] Optionally, based on the standardized behavioral speech response score, using a multimodal adaptive fusion algorithm, extracting the eye opening response score from the eye opening response data, performing joint feature extraction processing, and generating a joint feature vector, comprises:

[0032] Based on the standardized behavioral verbal response score, dimension expansion processing is performed to ensure the consistency of the dimension, and a dimension consistent score is generated;

[0033] Based on the consistent scores of the dimensions and combined with the eye-opening response data, a multimodal adaptive fusion algorithm is used to perform joint feature extraction processing, dynamically adjust the weight of each modality data, maximize information complementarity and correlation, and generate a preliminary fusion feature vector;

[0034] Based on the preliminary fused feature vector, further optimizing the feature representation by feature enhancement technology to generate an optimized fused feature vector;

[0035] Based on the optimized fusion feature vector, a multimodal adaptive fusion algorithm feature integration module is used to generate a joint feature vector.

[0036] Optionally, the comprehensive assessment index is compared and analyzed with a preset coma level threshold range to automatically determine the patient's current coma level and generate a personalized medical advice report, including:

[0037] Based on the comprehensive evaluation index, a comprehensive evaluation of the patient's current state is performed to generate a current state score;

[0038] Based on the current state score, a comparison analysis is performed with a preset coma level threshold range to generate a coma level interval for the patient;

[0039] Based on the patient's coma level range, a judgment is made through an automatic determination mechanism to generate the patient's current coma level;

[0040] Based on the patient's current coma level, combined with the patient's medical history and other information, a personalized medical advice report is generated.

[0041] In a second aspect, an embodiment of the present application provides an automatic assessment system based on the Glasgow Coma Scale, comprising:

[0042] A collection module, used to collect patient behavioral response data, verbal response data, and eye-opening response data to generate an initial evaluation data set;

[0043] An analysis module, for performing high-level sequence feature extraction processing on the behavioral response data using a gated recurrent unit algorithm based on the initial evaluation data set, and performing deep-level emotional tendency analysis on the verbal response data using a deep sentiment analysis technique to generate a behavioral verbal response score;

[0044] A processing module, for extracting an eye-opening reaction score from the eye-opening reaction data using a multimodal adaptive fusion algorithm based on the behavioral and verbal reaction score, performing dynamic weighted fusion processing with the behavioral and verbal reaction score according to preset rules, and generating a comprehensive evaluation index using deep reinforcement learning technology;

[0045] A generation module is used to compare and analyze the comprehensive evaluation index with a preset coma level threshold range, automatically determine the patient's current coma level, and generate a personalized medical advice report.

[0046] In a third aspect, an embodiment of the present application provides a computing device, comprising a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement an automatic assessment method based on the Glasgow Coma Scale as described in the first aspect.

[0047] In a fourth aspect, an embodiment of the present application provides a computer storage medium storing a computer program. When the computer program is executed by a computer, the automatic assessment method based on the Glasgow Coma Scale as described in the first aspect is implemented.

[0048] In an embodiment of the present application, the patient's behavioral response data, verbal response data and eye opening response data are collected to generate an initial evaluation data set; based on the initial evaluation data set, a gated recurrent unit algorithm is used to perform high-level sequence feature extraction processing on the behavioral response data, and a deep sentiment analysis technology is used to perform a deep-level sentiment tendency analysis on the verbal response data to generate a behavioral verbal response score; based on the behavioral verbal response score, a multimodal adaptive fusion algorithm is used to extract the eye opening response score from the eye opening response data, and a dynamic weighted fusion processing is performed with the behavioral verbal response score through preset rules, and a comprehensive evaluation index is generated using deep reinforcement learning technology; based on the comprehensive evaluation index, a comparative analysis is performed with a preset coma level threshold range to automatically determine the patient's current coma level, and generate a personalized medical advice report. By combining behavioral response data, verbal response data, and eye-opening response data, and utilizing gated recurrent unit algorithms and deep sentiment analysis technology, more comprehensive and in-depth patient information can be captured, thereby significantly improving the accuracy and reliability of the assessment. The entire assessment process is automated, reducing the impact of human intervention and subjective judgment, and improving assessment efficiency and consistency. Based on the comprehensive assessment index, personalized medical advice reports are generated to provide a scientific basis for clinical decision-making and help develop more personalized treatment plans. The use of multimodal adaptive fusion algorithms and deep reinforcement learning technology can dynamically adjust the weight allocation strategy according to actual conditions, ensuring the real-time and adaptability of the assessment results.

[0049] Furthermore, by preprocessing the behavioral response data, the consistency of the data format and noise removal were ensured, thereby improving the accuracy of subsequent feature extraction; the gated recurrent unit algorithm was used to effectively capture the long-term dependencies in the time series, thereby enhancing the model's ability to understand the behavioral response data; combined with deep sentiment analysis technology, not only the context-related word vector representations in the verbal response were extracted, but also the key sentiment information was identified, making the behavioral verbal feature vector richer and more comprehensive; the behavioral verbal feature vectors were fused by using preset weighting rules, and the generated behavioral verbal response score can better reflect the patient's true state, thereby improving the comprehensive expressiveness of the evaluation.

[0050] Furthermore, the behavioral and verbal response scores were standardized to ensure the consistency of scale and distribution of data from different modalities, providing a reliable basis for subsequent fusion; the eye opening response score was extracted from the eye opening response data through a multimodal adaptive fusion algorithm, and joint feature extraction was performed to generate a joint feature vector, thereby enhancing the diversity and expression ability of the features; dynamic weighted fusion processing was performed using preset weighting rules to generate a preliminary fusion score, thereby ensuring that the importance of each modality of data was reasonably reflected and improving the comprehensiveness and accuracy of the evaluation; deep reinforcement learning technology was used to continuously learn and optimize the weight allocation strategy, further refine the score, and generate a more accurate comprehensive evaluation index that can better reflect the patient's actual condition and support more scientific clinical decision-making.

[0051] These and other aspects of the present application will become more clearly understood in the description of the following embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0053] Figure 1 A flowchart of an automatic assessment method based on the Glasgow Coma Scale provided in an embodiment of the present application;

[0054] Figure 2 A schematic diagram of the structure of an automatic assessment system based on the Glasgow Coma Scale provided in an embodiment of the present application;

[0055] Figure 3 A schematic diagram of the structure of a computing device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0056] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0057] In some of the processes described in the specification and claims of this application and the above-mentioned figures, multiple operations that appear in a specific order are included, but it should be clearly understood that these operations may not be executed in the order in which they appear in this article or executed in parallel. The serial numbers of the operations, such as 101, 102, etc., are only used to distinguish between different operations, and the serial numbers themselves do not represent any execution order. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions of "first", "second", etc. in this article are used to distinguish different messages, devices, modules, etc., do not represent the order of precedence, and do not limit the "first" and "second" to be different types.

[0058] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of this application.

[0059] Figure 1 A flowchart of an automatic assessment method based on the Glasgow Coma Scale is provided for an embodiment of the present application, such as Figure 1 As shown, the method includes:

[0060] 101. Collect patient behavioral response data, verbal response data, and eye-opening response data to generate an initial evaluation data set;

[0061] In this step, the behavioral response data includes information such as the patient's motor response and limb activity, which is used to evaluate the patient's motor ability and reaction speed to external stimuli. For example, the patient's limb movement frequency, strength, and reaction time are all part of the behavioral response data.

[0062] Verbal response data covers information such as the patient's voice, intonation, vocabulary selection, etc., reflecting their cognitive and language functions. By analyzing the verbal response data, we can understand the patient's level of consciousness, clarity of thinking, and emotional state.

[0063] The eye opening response data records the patient's eye activities under specific stimulation, such as eye opening frequency, pupil changes, etc., which is used to assess the level of consciousness and can help judge the patient's alertness and cognitive ability to the external environment.

[0064] The initial evaluation dataset is a data set formed by integrating the data of the above three modalities after preprocessing. It provides a basis for subsequent advanced feature extraction and sentiment analysis, and ensures the consistency and integrity of multimodal data.

[0065] In an embodiment of the present application, assuming that in an intensive care unit environment, first, the patient's behavioral response data is monitored in real time through smart sensors and wearable devices to ensure the continuity and integrity of the data; secondly, the patient's verbal response data is collected and processed using speech recognition technology and natural language processing tools to remove background noise and irrelevant information; thirdly, the patient's eye opening response data is captured in combination with a video surveillance system, and time synchronization is performed to ensure the consistency of data in different modalities; finally, all collected data is standardized and normalized to generate a high-quality initial evaluation data set to prepare for subsequent analysis.

[0066] 102. Based on the initial evaluation data set, use the gated recurrent unit algorithm to perform advanced sequence feature extraction processing on the behavioral response data, use deep sentiment analysis technology to perform deep sentiment tendency analysis on the verbal response data, and generate a behavioral verbal response score;

[0067] In this step, the gated recurrent unit is an improved recurrent neural network that can effectively capture dependencies in long time series and control the flow of information through a gating mechanism, thereby enhancing the model's ability to understand time series data. It is particularly suitable for complex dynamic patterns in behavioral response data.

[0068] Advanced sequence feature extraction is to extract representative feature vectors from behavioral response data. These features can reflect the patient's behavioral patterns and their changing trends. Through multi-level time series feature extraction, long-term dependencies can be captured and the accuracy of the assessment can be improved.

[0069] Deep sentiment analysis technology uses deep learning models to encode verbal response data, obtain context-related sentiment word vector representations, and identify key emotional information such as joy, anger, fear, etc. through sentiment classifiers to assess the patient's psychological state.

[0070] The behavioral and verbal response score comprehensively reflects the patient's behavioral and emotional expressions. By weighted fusion of the behavioral response feature vector and the emotional word vector, a comprehensive score is generated. This score not only takes into account the patient's behavioral pattern, but also includes the emotional component, thereby more comprehensively reflecting the patient's status.

[0071] In an embodiment of the present application, assuming that in an emergency department scenario, a gated recurrent unit algorithm is first applied to perform multi-level time series feature extraction on the behavioral response data to capture long-term dependencies and generate a behavioral response feature vector; secondly, a deep learning model is used to encode the verbal response data to obtain context-related sentiment word vector representations, and the behavioral response feature vectors are combined with the behavioral response feature vectors to input into a sentiment classifier to identify key sentiment information; thirdly, the behavioral response feature vectors and sentiment word vectors are fused through preset weighting rules to generate a behavioral speech feature vector; finally, the behavioral speech response score is calculated based on this feature vector, providing an important basis for subsequent comprehensive evaluation.

[0072] Optionally, in step 102, based on the initial evaluation data set, a gated recurrent unit algorithm is used to perform high-level sequence feature extraction processing on the behavior response data, and a deep sentiment analysis technology is used to perform a deep-level sentiment tendency analysis on the speech response data to generate a behavior speech response score, including: based on the initial evaluation data set, preprocessing the behavior response data to ensure that the data format is consistent and remove noise, and generate preprocessed behavior response data; based on the preprocessed behavior response data, a gated recurrent unit algorithm is used to control the flow of information through a gating mechanism, capture long-term dependencies in the time series, and generate a behavior response feature vector through multi-layer gated recurrent unit processing and fully connected layer mapping; based on the behavior response feature vector, the deep sentiment analysis technology is used to encode the speech response data, and a context-related word vector representation is obtained, which is combined with the behavior response feature vector and input into a sentiment classifier to identify key sentiment information and generate a behavior speech feature vector; based on the behavior speech feature vector, weighted fusion is performed through preset weighting rules to generate a behavior speech response score.

[0073] In this step, preprocessing the behavioral response data includes standardizing and denoising the behavioral response data in the initial evaluation data set to ensure consistent data format and remove noise. This process improves the accuracy of subsequent feature extraction and reduces interference from irrelevant information.

[0074] Behavioral response feature vectors are generated by applying multi-layer gated recurrent unit processing and fully connected layer mapping to the preprocessed behavioral response data. These feature vectors capture the time series characteristics of the patient's behavioral pattern and reflect their reaction speed and motor ability to external stimuli.

[0075] Behavior-speech feature vector: The behavior-speech feature vector is a comprehensive feature vector generated by combining the behavior response feature vector with the emotion word vector and inputting it into the emotion classifier to identify key emotion information. This vector not only includes the behavior pattern, but also covers the emotion component, reflecting the patient's condition more comprehensively.

[0076] First, the behavioral response data is preprocessed to ensure the consistency of the data format and remove noise to generate preprocessed behavioral response data; then, based on the preprocessed behavioral response data, the gated recurrent unit algorithm is used to capture the long-term dependencies in the time series, and the behavioral response feature vector is generated through multi-layer gated recurrent unit processing and fully connected layer mapping; then, the speech response data is encoded using deep sentiment analysis technology to obtain context-related word vector representations, and combined with the behavioral response feature vector, it is input into the sentiment classifier to identify key sentiment information and generate a behavioral speech feature vector; finally, based on the behavioral speech feature vector, weighted fusion is performed through preset weighting rules to generate a behavioral speech response score. This process not only improves the accuracy of feature extraction, but also enhances the depth of sentiment analysis, providing richer information support for subsequent evaluation.

[0077] Optionally, based on the pre-processed behavior response data, a gated recurrent unit algorithm is used to control the flow of information through a gating mechanism, capture medium- and long-term dependencies in the time series, and generate a behavior response feature vector through multi-layer gated recurrent unit processing and fully connected layer mapping, including: based on the pre-processed behavior response data, input into a gated recurrent unit network, perform initialization processing, and generate an initial state; based on the initial state, a gated recurrent unit algorithm is used to control the flow of information through an update gate and a reset gate, capture medium- and long-term dependencies in the time series, and generate an intermediate state representation; based on the intermediate state representation, high-level features are further extracted through multi-layer gated recurrent units, internal states are updated, and the current moment behavior feature representation is output through an output gate to generate a deep feature representation; based on the deep feature representation, mapping is performed through multiple fully connected layers to convert high-dimensional feature vectors into fixed-length feature vectors to generate a behavior response feature vector.

[0078] The method is based on the behavior response feature vector, uses deep sentiment analysis technology to encode the speech response data, obtains context-related word vector representation, combines the behavior response feature vector with the input into a sentiment classifier, identifies key sentiment information, and generates a behavior speech feature vector, including: based on the behavior response feature vector, performs adaptation processing through feature mapping to adapt the speech response data and generate an adapted behavior response feature vector; based on the adapted behavior response feature vector, uses deep sentiment analysis technology to encode the speech response data, introduces the adapted behavior response feature vector, and generates a context-related word vector representation; based on the context-related word vector representation, combines the adapted behavior response feature vector, performs weighted fusion processing to combine multiple effective information and generate a combined feature vector; based on the combined feature vector, inputs into a dedicated sentiment classifier, performs deep feature extraction and pattern recognition through a multi-layer neural network, identifies key sentiment information, and generates a behavior speech feature vector.

[0079] In this step, the gated recurrent unit network initialization refers to inputting the preprocessed behavioral response data into the gated recurrent unit network to generate an initial state.

[0080] The update gate and reset gate are the core mechanisms of the gated recurrent unit algorithm, which are used to control the flow of information. The update gate determines which information needs to be updated, while the reset gate determines whether to ignore the state of the previous moment.

[0081] Intermediate state representation refers to the temporary state vectors generated during the operation of the gated recurrent unit algorithm. These vectors reflect the behavioral characteristics of the current moment and serve as the basis for further feature extraction.

[0082] Deep feature representation refers to further extracting high-level features through multi-layer gated recurrent units, updating the internal state, and outputting the behavioral feature representation of the current moment through the output gate. These features capture deeper time series patterns and enhance the expressiveness of the model.

[0083] Fully connected layer mapping refers to converting high-dimensional feature vectors into fixed-length feature vectors through multiple fully connected layers. This process simplifies the feature representation and makes it more suitable for subsequent classification and regression tasks.

[0084] Adapting the behavioral response feature vector refers to adapting the speech response data by performing feature mapping processing on the behavioral response feature vector. This adaptation process ensures the compatibility between different modal data and provides a unified basis for subsequent sentiment analysis.

[0085] Context-related word vector representation refers to the use of deep sentiment analysis technology to encode and process speech response data, combined with word vectors generated by adapting behavioral response feature vectors. These word vectors not only contain the semantic information of the speech content, but also incorporate behavioral characteristics, enhancing the accuracy of sentiment analysis.

[0086] A combined feature vector refers to a comprehensive feature vector generated by weighted fusion processing, combining context-related word vector representation and adaptive behavior response feature vector. This combination method makes full use of a variety of effective information and provides richer input for the sentiment classifier.

[0087] A dedicated emotion classifier refers to a multi-layer neural network designed specifically for identifying key emotional information. Through deep feature extraction and pattern recognition, this classifier can accurately identify the patient's emotional tendencies and generate behavioral speech feature vectors.

[0088] In the embodiment of the present application, the pre-processed behavior response data is first input into the gated recurrent unit network for initialization processing, and the gated recurrent unit algorithm is used to control the flow of information through the update gate and the reset gate to capture the long-term dependency in the time series and generate the intermediate state representation; secondly, based on the intermediate state representation, high-level features are further extracted through multi-layer gated recurrent units, the internal state is updated, and the behavior feature representation at the current moment is output through the output gate, and the mapping is performed through multiple fully connected layers to convert the high-dimensional feature vector to a fixed-length feature vector to generate a behavior response feature vector. Thirdly, adaptation processing is performed through feature mapping to adapt the speech reaction data to generate an adapted behavior response feature vector, encoding processing is performed to generate a context-related word vector representation, and weighted fusion processing is performed to generate a combined feature vector; finally, it is input into a dedicated emotion classifier, and deep-level feature extraction and pattern recognition are performed through a multi-layer neural network to identify key emotional information and generate a behavior speech feature vector.

[0089] Assuming that in an environment of an Alzheimer's care center, the collected behavioral response data is first preprocessed to ensure that the data format is consistent and noise is removed, and high-quality preprocessed behavioral response data is generated and input into the gated recurrent unit network for initialization processing to generate an initial state; secondly, the gated recurrent unit algorithm is used to control the flow of information through update gates and reset gates, capture long-term dependencies in time series, generate intermediate state representations, further extract high-level features through multi-layer gated recurrent units, update internal states, and output the current moment behavioral feature representation through output gates, and map through multiple fully connected layers to convert high-dimensional feature vectors to fixed lengths. feature vector to generate a behavioral response feature vector; again, based on the behavioral response feature vector, adaptation processing is performed through feature mapping to adapt the verbal response data, encoding processing is performed, and the adapted behavioral response feature vector is introduced to generate a context-related word vector representation, and the context-related word vector representation and the adapted behavioral response feature vector are combined to perform weighted fusion processing to generate a combined feature vector; finally, it is input into a dedicated emotion classifier, and deep feature extraction and pattern recognition are performed through a multi-layer neural network to identify key emotional information and generate a behavioral speech feature vector, which provides doctors with a more comprehensive basis for patient status assessment and supports the formulation of personalized treatment plans.

[0090] The present application takes into account the problem that the prior art is not accurate enough in capturing the long-term dependencies of time series, so the embodiment of the invention proposes this optional solution, which enhances the ability to understand time series data by introducing the update gate and reset gate mechanism in the gated recurrent unit algorithm, so as to solve the technical problem that traditional methods are difficult to process long time interval information; in addition, the existing methods are relatively simple in expressing nonlinear characteristics, resulting in insufficient expressiveness and flexibility of the model; therefore, this solution introduces a variety of nonlinear activation functions and transformations, such as sine functions, exponential functions, etc., to capture potential time series dependencies and improve the expressiveness of the model.

[0091] Optionally, based on the initial state, a gated recurrent unit algorithm is used to control the flow of information by updating gates and resetting gates, capturing long-term dependencies in the time series, and generating an intermediate state representation, including:

[0092] Based on the initial state, concatenate the input data with the hidden state vector of the previous time step, and map them to a suitable feature space through linear transformation;

[0093] Apply nonlinear activation functions to introduce nonlinear characteristics, enhance the model's expressiveness, and introduce sine and exponential functions to further capture potential time series dependencies to generate update gate outputs;

[0094] The update gate output is calculated using the following formula:

[0095]

[0096] Among them, z t is the update gate output at time step t; W z is the weight matrix of the update gate; h t-1 is the hidden state vector of the previous time step; x t is the input vector of the current time step; b z is the bias term of the update gate; σ is the nonlinear activation function Sigmoid function; α is the coefficient for adjusting the exponential function; β is the coefficient for adjusting the amplitude of the sine function; γ is the coefficient for adjusting the frequency of the sine function;

[0097] Based on the update gate output, perform an element-by-element multiplication operation on the hidden state vector of the previous time step to control the flow of information, transform it through a specific weight matrix, and introduce a bias term to generate a reset gate output;

[0098] The reset gate output is calculated using the following formula:

[0099]

[0100] Among them, r t is the reset gate output at time step t; Wr is the weight matrix of the reset gate; z t is the update gate output at time step t; h t-1 is the hidden state vector of the previous time step; x t is the input vector of the current time step; b r is the bias term of the reset gate; σ is the nonlinear activation function Sigmoid function; ⊙ is the element-by-element multiplication operation; δ is the coefficient for adjusting the exponential function; η is the coefficient for adjusting the amplitude of the cosine function; θ is the coefficient for adjusting the frequency of the cosine function; λ is the coefficient for adjusting the logarithmic function; μ is the coefficient for adjusting the exponential function;

[0101] Based on the reset gate output, element-by-element selective transmission is performed, and rich feature representation is introduced through high-order polynomial transformation and dynamic threshold adjustment. Combined with the update gate output, weighted merging is performed through an adaptive combination mechanism to generate an intermediate state representation.

[0102] This method aims to better capture the long-term dependencies in time series and enhance the expressiveness of the model. Based on the gated recurrent unit algorithm, it controls the flow of information through update gates and reset gates, combines multiple nonlinear transformations, and generates intermediate state representations, thereby achieving effective extraction of complex time series features; maps input data to a suitable feature space through linear transformation, and introduces nonlinear characteristics by applying nonlinear activation functions; calculates the update gate and reset gate outputs through specific formulas to control the information flow of the hidden state vector of the previous time step; finally, generates intermediate state representations through element-by-element selective transmission and adaptive combination mechanisms.

[0103] In the update gate output, the input mapping term W z ·[h t-1 ,x t ]: used to concatenate the input data and the hidden state vector of the previous time step and map them to the appropriate feature space to ensure the consistency of data dimensions; bias compensation term b z : Provides an offset for linear transformation, making the model more flexible and better able to adapt to different data distributions; nonlinear activation term The Sigmoid function is introduced as a nonlinear activation function to enhance the model's expressiveness and enable the model to capture more complex patterns; the exponential adjustment term exp(-α(W z ·[h t-1 ,x t ])):The exponential function is used to capture long-term dependencies, and the adjustment coefficient α controls the exponential decay speed to ensure that the model can effectively process long time interval information; the sinusoidal modulation term β·sin(γ·(W z ·[h t-1 ,x t])):Capture periodic characteristics through sine function, adjust amplitude β and frequency γ, and further enrich the time series analysis ability of the model;

[0104] Among them, the weight matrix W z and the bias term b z It is obtained by optimizing the back propagation algorithm during training; the nonlinear activation function σ is a predefined Sigmoid function; the coefficients α, β, and γ are obtained through cross-validation or hyperparameter tuning; the hidden state vector h of the previous time step t-1 and the input vector x at the current time step t Directly retrieved from previous network layers or data sources;

[0105] At the reset gate output, the information selection item z t ⊙h t-1 :Control the information flow through element-by-element multiplication operation, retain important information, and avoid interference from irrelevant information; input mapping item W r ·[z t ⊙h t-1 ,x t ]: used to concatenate the update gate output with the hidden state vector of the previous time step and map it to the appropriate feature space to ensure the consistency of data dimensions; bias compensation term b r : Provides an offset for linear transformation, making the model more flexible and better able to adapt to different data distributions; nonlinear activation term The Sigmoid function is introduced as a nonlinear activation function to enhance the model's expressiveness and enable the model to capture more complex patterns; the exponential adjustment term exp(-δ(W r ·[z t ⊙h t-1 ,x t ])):The exponential function is used to capture long-term dependencies, and the adjustment coefficient δ controls the exponential decay speed to ensure that the model can effectively process long time interval information; the cosine modulation term η·cos(θ·(W r ·[z t ⊙h t-1 ,x t ])):The cosine function is used to capture the periodic characteristics, adjust the amplitude η and frequency θ, and further enrich the time series analysis ability of the model; the logarithmic enhancement term log(1+exp(μ·(W r ·[z t ⊙h t-1 ,x t ]))):By introducing the logarithmic function to enhance the model's expressiveness and adjusting the coefficients λ and μ, the model can better handle extreme values;

[0106] Among them, the weight matrix W r and the bias term b rIt is obtained by optimizing the back propagation algorithm during training; the nonlinear activation function σ is a predefined Sigmoid function; the coefficients δ, η, θ, λ, μ are obtained through cross-validation or hyperparameter tuning; the update gate output z t , the hidden state vector h of the previous time step t-1 and the input vector x at the current time step t Directly retrieved from previous network layers or data sources;

[0107] Assume that in an EEG analysis system, it is necessary to automatically evaluate the coma score; Assume that W z =[0.5,-0.3]; h t-1 =[0.8,0.6]; x t =[0.7,0.4]; b z =0.2; α=0.1; β=0.5; γ=0.2;

[0108]

[0109] Assume W r =[0.4,-0.2]; z t =0.72; h t-1 =[0.8,0.6]; x t =[0.7,0.4]; b r =0.1; δ=0.1; eta=0.4; θ=0.3; λ=0.6; μ=0.2;

[0110]

[0111] Assuming that the threshold is set to 0.6, since the final result 0.68 output by the reset gate is greater than the corresponding set threshold, it indicates that the input data at the current time step performs well in terms of information flow control and feature capture, and is suitable for further analysis or as the final state representation. Through the above steps, time series data can be effectively processed, and high-quality behavioral and emotional feature representations can be extracted, thereby improving the accuracy and reliability of patient status assessment.

[0112] 103. Based on the behavioral and verbal response score, a multimodal adaptive fusion algorithm is used to extract the eye-opening response score from the eye-opening response data, and a dynamic weighted fusion process is performed with the behavioral and verbal response score through preset rules, and a comprehensive evaluation index is generated using deep reinforcement learning technology;

[0113] In this step, the multimodal adaptive fusion algorithm aims to integrate data from different modalities and maximize the information complementarity of each modality data. By adaptively adjusting the weight distribution strategy, it can better handle multi-source heterogeneous data and improve the accuracy and reliability of the evaluation.

[0114] The eye opening response score reflects the patient's level of consciousness and is extracted from the eye opening response data through a multimodal adaptive fusion algorithm. The score can capture the patient's eye activity characteristics, such as eye opening frequency, pupil changes, etc., and further supplement the behavioral and verbal response score to enhance the comprehensiveness of the assessment.

[0115] Dynamic weighted fusion processing refers to weighted fusion of data of different modalities according to preset rules to generate a preliminary fusion score. This processing method can dynamically adjust the weight allocation strategy according to actual conditions to ensure that the importance of each modality data is reasonably reflected.

[0116] The comprehensive assessment index is the final score generated by further refining the preliminary fusion score. The weight allocation strategy is continuously optimized through deep reinforcement learning technology, so that the comprehensive assessment index can more accurately reflect the patient's coma state and support more scientific clinical decision-making.

[0117] In an embodiment of the present application, assuming that in a recovery room after a neurosurgery operation, first, standardized processing is performed based on the behavioral verbal response score to ensure the consistency of scale and distribution of data of different modalities; secondly, a multimodal adaptive fusion algorithm is used to extract the eye opening reaction score from the eye opening reaction data, and a joint feature extraction process is performed to generate a joint feature vector; thirdly, the joint feature vector is dynamically weighted and fused according to a preset weighting rule to generate a preliminary fusion score; finally, deep reinforcement learning technology is used to continuously optimize the weight allocation strategy, further refine the score, and generate a comprehensive evaluation index to comprehensively reflect the patient's current status.

[0118] Optionally, in step 103, based on the behavioral and verbal reaction score, a multimodal adaptive fusion algorithm is used to extract the eye opening reaction score from the eye opening reaction data, and a dynamic weighted fusion processing is performed with the behavioral and verbal reaction score through preset rules, and a comprehensive evaluation index is generated using deep reinforcement learning technology, including: based on the behavioral and verbal reaction score, standardization processing is performed to have scale and distribution consistency to generate a standardized behavioral and verbal reaction score; based on the standardized behavioral and verbal reaction score, a multimodal adaptive fusion algorithm is used to extract the eye opening reaction score from the eye opening reaction data, and a joint feature extraction processing is performed to generate a joint feature vector; based on the joint feature vector, a dynamic weighted fusion processing is performed through preset weighting rules to generate a preliminary fusion score; based on the preliminary fusion score, deep reinforcement learning technology is used to continuously learn and optimize the weight allocation strategy, further refine the score, and generate a comprehensive evaluation index.

[0119] Among them, based on the standardized behavioral verbal response score, a multimodal adaptive fusion algorithm is used to extract the eye opening response score from the eye opening response data, and a joint feature extraction process is performed to generate a joint feature vector, including: based on the standardized behavioral verbal response score, a dimension expansion process is performed to ensure dimension consistency, and a dimension consistency score is generated; based on the dimension consistency score, combined with the eye opening response data, a multimodal adaptive fusion algorithm is used to perform a joint feature extraction process, dynamically adjust the weights of each modality data, maximize information complementarity and correlation, and generate a preliminary fused feature vector; based on the preliminary fused feature vector, the feature representation is further optimized through feature enhancement technology to generate an optimized fused feature vector; based on the optimized fused feature vector, a multimodal adaptive fusion algorithm feature integration module is used to generate a joint feature vector.

[0120] In this step, standardization refers to adjusting the scale and distribution of behavioral verbal response scores to eliminate the dimensional differences between different modal data.

[0121] Joint feature extraction processing refers to extracting representative feature vectors from data of different modalities and combining these feature vectors into a comprehensive feature representation. Through this method, more dimensional information can be captured and the expressiveness of the model can be enhanced.

[0122] The preliminary fusion score is a score generated by applying a preset weighting rule to the joint feature vector

[0123] Deep reinforcement learning technology is a machine learning method that further refines scoring by continuously learning and optimizing weight distribution strategies. This technology can automatically adjust parameters based on feedback to improve the accuracy and stability of evaluation results.

[0124] Dimensionality expansion processing refers to processing the standardized behavioral verbal response scores to ensure their dimensional consistency with the eyes opening response data.

[0125] The dimension consistency score is a behavioral verbal response score generated after dimension expansion processing. It has the same dimensional structure as the eye-opening response data, providing a unified basis for subsequent joint feature extraction.

[0126] The preliminary fused feature vector is a feature vector generated by a multimodal adaptive fusion algorithm based on the dimensional consistency score and the eye opening response data.

[0127] The optimized fused feature vector is generated by applying feature enhancement techniques to the preliminary fused feature vector. This optimization improves the quality of feature representation and makes it more reflective of the patient's true status.

[0128] The joint feature vector is a comprehensive feature vector finally generated by the feature integration module of the multimodal adaptive fusion algorithm, which integrates the effective information of all modal data.

[0129] In the embodiment of the present application, the behavioral and verbal response scores are first standardized to ensure that they have consistency in scale and distribution, and the multimodal adaptive fusion algorithm is used to extract the eye opening reaction scores from the eye opening reaction data, and a joint feature extraction process is performed to generate a joint feature vector; secondly, dynamic weighted fusion processing is performed through preset weighting rules, and deep reinforcement learning technology is used to continuously learn and optimize the weight allocation strategy to further refine the score, and the eye opening reaction data is combined with the multimodal adaptive fusion algorithm to dynamically adjust the weights of each modal data to maximize information complementarity and correlation, and generate a preliminary fused feature vector; finally, based on the preliminary fused feature vector, the feature representation is further optimized through feature enhancement technology, and the feature integration module of the multimodal adaptive fusion algorithm is used to generate a joint feature vector.

[0130] Assuming that in a brain injury rehabilitation center environment, the behavioral and verbal response scores are first standardized to ensure that they have scale and distribution consistency, and the dimension expansion process is performed to ensure dimensional consistency, and a dimension consistency score is generated; secondly, combined with the eye-opening reaction data, a multimodal adaptive fusion algorithm is used to dynamically adjust the weights of each modality data to maximize information complementarity and correlation, generate a preliminary fusion feature vector, and further optimize the feature representation through feature enhancement technology to generate an optimized fusion feature vector; thirdly, the multimodal adaptive fusion algorithm feature integration module is used to generate a joint feature vector, and dynamic weighted fusion processing is performed through preset weighting rules to generate a preliminary fusion score; finally, based on the preliminary fusion score, deep reinforcement learning technology is used to continuously learn and optimize the weight allocation strategy, further refine the score, and generate a comprehensive evaluation index. In this way, doctors can obtain a more comprehensive and accurate patient status assessment and support the formulation of personalized rehabilitation plans.

[0131] This application takes into account that

[0132] Optionally, based on the dimensional consistency score, combined with the eye-opening response data, a multimodal adaptive fusion algorithm is used to perform joint feature extraction processing, dynamically adjust the weight of each modality data, maximize information complementarity and correlation, and generate a preliminary fusion feature vector, including:

[0133] Based on the consistent scores of the dimensions, the most representative features are extracted through feature selection technology to reduce redundant information;

[0134] Apply multimodal data alignment technology to ensure the temporal synchronization of data of different modalities to generate dimensionally consistent scoring feature representations;

[0135] The dimension consistent scoring feature representation is calculated using the following formula:

[0136]

[0137] Among them, S' is the dimension consistent score feature representation; S is the dimension consistent score; W 1 is the weight matrix for feature extraction; b 1 is the bias term for feature extraction; σ is the nonlinear activation function; α is the coefficient for adjusting the exponential function; β is the coefficient for adjusting the amplitude of the sine function; γ is the coefficient for adjusting the frequency of the sine function; δ is the coefficient for adjusting the logarithmic function; η is the coefficient for adjusting the exponential function;

[0138] Based on the consistent score feature representation of the dimension, it is spliced ​​with the eye-opening response data, and high-order polynomial transformation and exponential logarithmic function modulation are performed. An adaptive weight adjustment mechanism is introduced to dynamically adjust the weight of each modal data to maximize information complementarity and correlation, so as to generate a fusion feature intermediate representation;

[0139] The intermediate representation of the fused features is calculated using the following formula:

[0140]

[0141] Among them, F is the intermediate representation of fusion features; S' is the feature representation based on dimension consistency score; E is the eye opening response data; W 2 is the weight matrix extracted by the final fusion feature; b 2 is the bias term extracted by the final fusion feature; σ is the nonlinear activation function; ζ is the coefficient for adjusting the exponential function; λ is the coefficient for adjusting the amplitude of the cosine function; θ is the coefficient for adjusting the frequency of the cosine function; μ is the coefficient for adjusting the exponential decay function; v is the coefficient for adjusting the square term in the exponential decay function; κ is the coefficient for adjusting the logarithmic function; ω is the coefficient for adjusting the exponential function; [S', E] is the concatenation result of the dimension consistent score feature representation S' and the eye opening response data E;

[0142] Based on the intermediate representation of the fused features, multi-layer perceptron processing is performed to capture deeper feature relationships, combine the importance of each modal data, effectively integrate new and old information, and generate a preliminary fused feature vector.

[0143] In the dimension-consistent score feature representation, the linear transformation mapping W 1 S: used to map the dimension consistency score S to a suitable feature space to ensure the consistency of data dimensions; this step helps to convert the input data into a form suitable for subsequent processing; bias compensation term b 1 : Provides an offset for linear transformation, making the model more flexible and better able to adapt to different data distributions; the bias term allows the model to fit different baseline levels; the nonlinear activation term The Sigmoid function is introduced as a nonlinear activation function to enhance the model's expressiveness and enable the model to capture more complex patterns. The Sigmoid function compresses the output to between 0 and 1, which is suitable for binary classification tasks or situations where probability output is required. The exponential adjustment term exp(-α(W 1 ·S)): The exponential function is used to capture long-term dependencies, and the adjustment coefficient α controls the exponential decay speed to ensure that the model effectively processes information with long time intervals. The exponential function can smoothly reduce the influence of distant data points. The sinusoidal modulation term β·sin(γ·(W 1 ·S)): The sine function is used to capture periodic features, adjust the amplitude β and frequency γ, and further enrich the model's time series analysis capabilities; the sine function can capture periodic changes in the data; the logarithmic enhancement term δ·log(1+exp(η·(W 1 S))): The model’s expressiveness is enhanced by introducing a logarithmic function, and the coefficients δ and η are adjusted so that the model can better handle extreme values. The logarithmic function is particularly useful when dealing with extremely large or small values, and can prevent numerical overflow.

[0144] Among them, the weight matrix W 1 and the bias term b 1 It is obtained through the optimization of the back-propagation algorithm in the training process; the nonlinear activation function σ is a predefined Sigmoid function; the coefficients α, β, γ, δ, η are obtained through cross-validation or hyperparameter tuning; the dimension consistency score S is directly obtained from the previous network layer or data source;

[0145] In the intermediate representation of the fused features, the concatenation term [S', E] is input: the dimension-consistent score feature representation S' is concatenated with the eye-opening response data E to ensure the integrity of the multimodal data; the concatenation operation retains all the original information for subsequent processing; the linear transformation mapping W 2 [S', E]: used to map the concatenated data to a suitable feature space to ensure consistent data dimensions; this step helps convert the input data into a form suitable for subsequent processing; bias compensation term b 2 : Provides an offset for linear transformation, making the model more flexible and better able to adapt to different data distributions; the bias term allows the model to fit different baseline levels; nonlinear activation term: The Sigmoid function is introduced as a nonlinear activation function to enhance the model's expressiveness and enable the model to capture more complex patterns. The Sigmoid function compresses the output to between 0 and 1, which is suitable for binary classification tasks or situations where probability output is required. The exponential adjustment term exp(-ζ(W 2·[S',E])):The exponential function is used to capture long-term dependencies. The adjustment coefficient ζ controls the exponential decay speed to ensure that the model can effectively process information with long time intervals. The exponential function can smoothly reduce the influence of distant data points. The cosine modulation term λ·cos(θ·(W 2 ·[S',E])):The cosine function is used to capture periodic characteristics, adjust the amplitude λ and frequency θ, and further enrich the time series analysis ability of the model; the cosine function can capture the periodic changes in the data; the exponential decay term μ·exp(-ν·(W 2 ·[S',E]) 2 ): By introducing an exponential decay function to enhance the model's expressiveness, the coefficients μ and ν are adjusted so that the model can better handle extreme values; the exponential decay function can suppress the influence of outliers to a certain extent; the logarithmic enhancement term κ·log(1+exp(ω·(W 2 [S', E]))):The model expressiveness is enhanced by introducing logarithmic functions and adjusting coefficients κ and ω so that the model can better handle extreme values. Logarithmic functions are particularly useful when dealing with extremely large or small values ​​and can prevent numerical overflow.

[0146] Among them, the weight matrix W 2 and the bias term b 2 It is obtained by optimizing the back-propagation algorithm during the training process; the nonlinear activation function σ is a predefined Sigmoid function; the coefficients ζ, λ, θ, μ, ν, κ, ω are obtained by cross-validation or hyperparameter tuning; the dimensionally consistent score feature representation S' and the eye-opening response data E are directly obtained from the previous network layer or data source;

[0147] Assume that in an intelligent health monitoring system, it is used to evaluate the health status of the elderly in real time; Assume that W 1 =[0.6,-0.4]; S=[0.75,0.55]; b 1 =0.15; α=0.12; β=0.45; γ=

[0148] 0.23; δ=0.38; η=0.32;

[0149]

[0150] Assume W 2 =[0.5,-0.3]; S'=[0.71,0.66]; E=[0.68,0.45]; b 2 =0.12; ζ=0.11; λ=0.42; θ=0.31; μ=0.58; v=0.21; κ=0.49; ω=0.37;

[0151]

[0152] Assuming that the threshold is set to θ = [0.6, 0.65], since each element of the final result F = [0.67, 0.73] of the fusion feature intermediate representation is greater than the corresponding set threshold, it shows that the current health information of the elderly performs well in terms of understanding and feature capture, and is suitable for further use or as the final information representation. Through the above steps, multimodal data can be effectively processed, the effective fusion of multimodal data and the capture of deep feature relationships are achieved, and high-quality behavioral and emotional feature representations are extracted, thereby improving the accuracy and reliability of health status assessment.

[0153] 104. Based on the comprehensive assessment index, a comparative analysis is performed with a preset coma level threshold range to automatically determine the patient's current coma level and generate a personalized medical advice report.

[0154] In this step, the preset coma level threshold range is a numerical interval set according to the Glasgow Coma Scale Standard, which is used to divide different coma levels. By comparing and analyzing the comprehensive assessment index with these threshold ranges, the patient's coma level can be automatically determined.

[0155] Comparative analysis refers to comparing the comprehensive assessment index with the preset coma level threshold range to determine the patient's current coma level. This process not only improves the degree of automation of the assessment, but also ensures the objectivity and consistency of the results.

[0156] The personalized medical advice report is a detailed report generated based on the evaluation results, combined with the patient's medical history and other clinical information. The report contains specific treatment measures and follow-up observation recommendations to help doctors make more accurate decisions and support efficient clinical decision-making processes.

[0157] In an embodiment of the present application, assuming that in a rehabilitation center environment, first, the generated comprehensive assessment index is compared and analyzed with the preset coma level threshold range to determine the patient's current coma level; secondly, based on the assessment results, combined with the patient's medical history and other clinical information, a detailed medical advice report is generated; thirdly, the report contains specific treatment measures and follow-up observation suggestions to help doctors make more accurate decisions; finally, the report is sent to the attending physician and related medical staff through an automated system to ensure the timeliness and accuracy of information transmission and support efficient clinical decision-making processes.

[0158] Optionally, the step 104 includes: based on the comprehensive evaluation index, performing a comparative analysis with a preset coma level threshold range, automatically determining the patient's current coma level, and generating a personalized medical advice report, including: based on the comprehensive evaluation index, comprehensively evaluating the patient's current state and generating a current state score; based on the current state score, performing a comparative analysis with a preset coma level threshold range to generate a patient's coma level interval; based on the patient's coma level interval, determining through an automatic determination mechanism to generate the patient's current coma level; based on the patient's current coma level, combining the patient's medical history and other information, generating a personalized medical advice report.

[0159] In this step, the comprehensive assessment index is the final score generated through multimodal data fusion and deep learning technology, which reflects the patient's current overall state. The index integrates multiple information such as behavioral response, verbal response and eye opening response, providing a key basis for the subsequent coma level determination.

[0160] The current status score is a specific numerical value generated after a comprehensive assessment of the patient's current status. It is based on a comprehensive assessment index and combines other relevant factors, such as time series characteristics and emotional tendencies, to more accurately reflect the patient's immediate condition.

[0161] The automatic determination mechanism refers to the process of determining the patient's coma level range through preset rules and algorithms to generate the patient's current coma level. This mechanism can complete the determination quickly and accurately, reducing the impact of human intervention and subjective judgment.

[0162] In the embodiment of the present application, the current state of the patient is firstly comprehensively evaluated based on the comprehensive evaluation index to generate the current state score; then, based on the current state score, a comparative analysis is performed with the preset coma level threshold range to generate the patient's coma level interval; then, based on the patient's coma level interval, a judgment is made through an automatic judgment mechanism to generate the patient's current coma level; finally, based on the patient's current coma level, combined with the patient's medical history and other information, a personalized medical advice report is generated. This process not only realizes the automation and refinement of the evaluation, but also provides a scientific basis for clinical decision-making, and enhances the personalization and effectiveness of the treatment plan.

[0163] Assuming that in a neurointensive care unit environment, first, the automated system is used to update the comprehensive assessment index in real time to ensure the timeliness of the assessment; secondly, through the comparative analysis module, the current status score is automatically compared with the preset threshold range to determine the coma level interval; thirdly, the automatic judgment mechanism quickly generates the patient's current coma level according to the preset rules; finally, the system automatically generates a personalized medical advice report containing treatment recommendations and observation points, and sends it to the attending physician and related medical staff through the electronic medical record system to ensure the timeliness and accuracy of information transmission and support efficient clinical decision-making processes.

[0164] In summary, steps 101 to 104 cover the complete process from preliminary processing of multimodal data to deep fusion feature extraction, aiming to provide an efficient and accurate patient status assessment framework that meets the strict requirements of real-time and accuracy in clinical applications.

[0165] Figure 2 A schematic diagram of the structure of an automatic assessment system based on the Glasgow Coma Scale is provided for an embodiment of the present application. Figure 2 As shown, the device comprises:

[0166] A collection module 21 is used to collect patient behavior response data, speech response data and eye opening response data to generate an initial evaluation data set;

[0167] An analysis module 22 is used to perform high-level sequence feature extraction processing on the behavioral response data using a gated recurrent unit algorithm based on the initial evaluation data set, and to perform deep-level emotional tendency analysis on the verbal response data using a deep sentiment analysis technique to generate a behavioral verbal response score;

[0168] A processing module 23 is used to extract an eye-opening reaction score from the eye-opening reaction data based on the behavior and speech reaction score by using a multimodal adaptive fusion algorithm, and dynamically weighted and fused the score with the behavior and speech reaction score according to preset rules, and generate a comprehensive evaluation index by using deep reinforcement learning technology;

[0169] The generating module 24 is used to compare and analyze the comprehensive evaluation index with the preset coma level threshold range, automatically determine the patient's current coma level, and generate a personalized medical advice report.

[0170] Figure 2 The automatic assessment system based on the Glasgow Coma Scale can be performed Figure 1A method for automatic evaluation based on the Glasgow Coma Scale described in the illustrated embodiment, the implementation principle and technical effects of which will not be elaborated further. For the automatic evaluation system based on the Glasgow Coma Scale in the above embodiment, the specific manners in which each module and unit perform operations have been described in detail in the embodiment related to the method, and will not be elaborated here.

[0171] In a possible design, Figure 2 The automatic evaluation system based on the Glasgow Coma Scale in the illustrated embodiment can be implemented as a computing device, such as Figure 3 as shown, the computing device may include a storage component 31 and a processing component 32;

[0172] The storage component 31 stores one or more computer instructions, wherein the one or more computer instructions are called and executed by the processing component 32.

[0173] The processing component 32 is configured to: collect patient behavior response data, speech response data, and eye-opening response data to generate an initial evaluation data set; based on the initial evaluation data set, use the gated recurrent unit algorithm to perform high-level sequence feature extraction processing on the behavior response data, and use deep sentiment analysis technology to perform in-depth sentiment tendency analysis on the speech response data to generate a behavior speech response score; based on the behavior speech response score, use the multi-modal adaptive fusion algorithm to extract an eye-opening response score from the eye-opening response data, and perform dynamic weighted fusion processing with the behavior speech response score through a preset rule, and use deep reinforcement learning technology to generate a comprehensive evaluation index; based on the comprehensive evaluation index, perform comparative analysis with a preset coma level threshold range to automatically determine the current coma level of the patient and generate a personalized medical advice report.

[0174] Among them, the processing component 32 may include one or more processors to execute computer instructions to complete all or part of the steps in the above method. Of course, the processing component may also be implemented by one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components for executing the above method.

[0175] The storage component 31 is configured to store various types of data to support operations at the terminal. The storage component can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0176] Of course, the computing device may also include other components, such as input / output interfaces, display components, communication components, etc.

[0177] The input / output interface provides an interface between the processing component and the peripheral interface module, which may be an output device, an input device, etc.

[0178] The communication component is configured to facilitate, among other things, wired or wireless communications between the computing device and other devices.

[0179] Among them, the computing device can be a physical device or an elastic computing host provided by a cloud computing platform, etc. In this case, the computing device can refer to a cloud server, and the above-mentioned processing components, storage components, etc. can be basic server resources rented or purchased from the cloud computing platform.

[0180] The present application also provides a computer storage medium storing a computer program, wherein the computer program can achieve the above-mentioned Figure 1 The illustrated embodiment is an automatic assessment method based on the Glasgow Coma Scale.

[0181] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0182] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0183] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0184] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. An automatic assessment method based on the Glasgow Coma Scale, characterized in that: include: Collect patient behavioral response data, verbal response data, and eye opening response data to generate an initial evaluation data set; Based on the initial evaluation data set, a gated recurrent unit algorithm is used to perform high-level sequence feature extraction processing on the behavioral response data, and a deep sentiment analysis technology is used to perform a deep-level sentiment tendency analysis on the verbal response data to generate a behavioral verbal response score; Based on the behavioral and verbal response score, a multimodal adaptive fusion algorithm is used to extract the eye-opening response score from the eye-opening response data, and a dynamic weighted fusion process is performed with the behavioral and verbal response score through preset rules, and a comprehensive evaluation index is generated using deep reinforcement learning technology; Based on the comprehensive evaluation index, a comparative analysis is performed with the preset coma level threshold range, the patient's current coma level is automatically determined, and a personalized medical advice report is generated.

2. The method according to claim 1, characterized in that Based on the initial evaluation data set, the gated recurrent unit algorithm is used to perform high-level sequence feature extraction processing on the behavioral response data, and the deep sentiment analysis technology is used to perform deep-level sentiment tendency analysis on the verbal response data to generate a behavioral verbal response score, including: Based on the initial evaluation data set, preprocessing the behavior response data to ensure data format consistency and remove noise, and generating preprocessed behavior response data; Based on the pre-processed behavior response data, a gated recurrent unit algorithm is used to control the information flow through a gating mechanism, capture the long-term dependency in the time series, and generate a behavior response feature vector through multi-layer gated recurrent unit processing and fully connected layer mapping; Based on the behavior response feature vector, the speech response data is encoded using deep sentiment analysis technology to obtain context-related word vector representation, and the behavior response feature vector is combined with the input into the sentiment classifier to identify key sentiment information and generate a behavior speech feature vector; Based on the behavioral speech feature vector, weighted fusion is performed through a preset weighting rule to generate a behavioral speech reaction score.

3. The method according to claim 2, characterized in that Based on the pre-processed behavior response data, the gated recurrent unit algorithm is used to control the information flow through the gating mechanism, capture the long-term dependency in the time series, and generate the behavior response feature vector through multi-layer gated recurrent unit processing and full connection layer mapping, including: Based on the preprocessed behavioral response data, the preprocessed behavioral response data is input into a gated recurrent unit network for initialization processing to generate an initial state; Based on the initial state, a gated recurrent unit algorithm is used to control the flow of information by updating gates and resetting gates, so as to capture long-term dependencies in the time series and generate intermediate state representations; Based on the intermediate state representation, high-level features are further extracted through multi-layer gated recurrent units, the internal state is updated, and the current moment behavior feature representation is output through an output gate to generate a deep feature representation; Based on the deep feature representation, mapping is performed through multiple fully connected layers to convert high-dimensional feature vectors into fixed-length feature vectors to generate behavioral response feature vectors.

4. The method according to claim 2, characterized in that: The method of encoding the speech response data based on the behavior response feature vector using deep sentiment analysis technology to obtain context-related word vector representation is combined with the behavior response feature vector and input into a sentiment classifier to identify key sentiment information and generate a behavior speech feature vector, including: Based on the behavior response feature vector, adaptation processing is performed through feature mapping to adapt the speech response data to generate an adapted behavior response feature vector; Based on the adaptive behavior response feature vector, the speech response data is encoded and processed using deep sentiment analysis technology, the adaptive behavior response feature vector is introduced, and a context-related word vector representation is generated; Based on the context-related word vector representation, combined with the adaptation behavior response feature vector, weighted fusion processing is performed to combine multiple effective information to generate a combined feature vector; Based on the combined feature vector, it is input into a dedicated emotion classifier, and deep feature extraction and pattern recognition are performed through a multi-layer neural network to identify key emotion information and generate a behavioral speech feature vector.

5. The method according to claim 1, characterized in that Based on the behavior and speech reaction score, a multimodal adaptive fusion algorithm is used to extract the eye-opening reaction score from the eye-opening reaction data, and a dynamic weighted fusion process is performed with the behavior and speech reaction score through preset rules, and a comprehensive evaluation index is generated using deep reinforcement learning technology, including: Based on the behavioral verbal response score, standardization processing is performed to ensure consistency in scale and distribution, thereby generating a standardized behavioral verbal response score; Based on the standardized behavioral speech response score, a multimodal adaptive fusion algorithm is used to extract the eye opening response score from the eye opening response data, perform joint feature extraction processing, and generate a joint feature vector; Based on the joint feature vector, a dynamic weighted fusion process is performed through a preset weighting rule to generate a preliminary fusion score; Based on the preliminary fusion score, deep reinforcement learning technology is used to continuously learn and optimize the weight allocation strategy, further refine the score, and generate a comprehensive evaluation index.

6. The method according to claim 5, characterized in that The method of extracting the eye-opening reaction score from the eye-opening reaction data based on the standardized behavioral speech reaction score and performing joint feature extraction processing to generate a joint feature vector includes: Based on the standardized behavioral verbal response score, dimension expansion processing is performed to ensure the consistency of the dimension, and a dimension consistent score is generated; Based on the consistent scores of the dimensions and combined with the eye-opening response data, a multimodal adaptive fusion algorithm is used to perform joint feature extraction processing, dynamically adjust the weight of each modality data, maximize information complementarity and correlation, and generate a preliminary fusion feature vector; Based on the preliminary fused feature vector, further optimizing the feature representation by feature enhancement technology to generate an optimized fused feature vector; Based on the optimized fusion feature vector, a multimodal adaptive fusion algorithm feature integration module is used to generate a joint feature vector.

7. The method according to claim 1, characterized in that Based on the comprehensive evaluation index, a comparative analysis is performed with a preset coma level threshold range to automatically determine the patient's current coma level and generate a personalized medical advice report, including: Based on the comprehensive evaluation index, a comprehensive evaluation of the patient's current state is performed to generate a current state score; Based on the current state score, a comparison analysis is performed with a preset coma level threshold range to generate a coma level interval for the patient; Based on the patient's coma level range, a judgment is made through an automatic determination mechanism to generate the patient's current coma level; Based on the patient's current coma level, combined with the patient's medical history and other information, a personalized medical advice report is generated.

8. An automatic assessment system based on the Glasgow Coma Scale, characterized in that: include: A collection module, used to collect patient behavioral response data, verbal response data, and eye-opening response data to generate an initial evaluation data set; An analysis module, for performing high-level sequence feature extraction processing on the behavioral response data using a gated recurrent unit algorithm based on the initial evaluation data set, and performing deep-level emotional tendency analysis on the verbal response data using a deep sentiment analysis technique to generate a behavioral verbal response score; A processing module, for extracting an eye-opening reaction score from the eye-opening reaction data using a multimodal adaptive fusion algorithm based on the behavioral and verbal reaction score, performing dynamic weighted fusion processing with the behavioral and verbal reaction score according to preset rules, and generating a comprehensive evaluation index using deep reinforcement learning technology; A generation module is used to compare and analyze the comprehensive evaluation index with a preset coma level threshold range, automatically determine the patient's current coma level, and generate a personalized medical advice report.

9. A computing device, characterized in that It comprises a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement an automatic assessment method based on the Glasgow Coma Score as described in any one of claims 1 to 7.

10. A computer storage medium, characterized in that: A computer program is stored, and when the computer program is executed by a computer, the automatic assessment method based on the Glasgow Coma Scale as claimed in any one of claims 1 to 7 is implemented.

Images