A stroke cognitive degeneration risk language prediction method, device and medium
By integrating the ReAct agent with a large language model, and combining multimodal data analysis and lesion-driven strategies, the weights are dynamically adjusted, solving the problems of long time consumption and insufficient accuracy in traditional assessment methods, and realizing early warning and individualized intervention for the risk of cognitive degeneration in stroke.
Patent Information
- Application Number
- CN202611475405.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-21
- Publication Date
- 2026-11-13
AI Technical Summary
Existing technologies struggle to identify subtle language abnormalities in post-stroke cognitive decline in the early stages. Traditional assessment methods are time-consuming and highly subjective, and multimodal fusion prediction accuracy and individualized adaptability are insufficient.
By deeply integrating ReAct intelligent agents with large language models, and through multimodal data analysis, combined with lesion location and clinical status to drive task decomposition and priority ranking, and dynamically adjusting weight fusion, personalized analysis of language, clinical and imaging indicators can be achieved.
It enables early warning and individualized intervention for the risk of cognitive decline in stroke, improves prediction accuracy and efficiency, comprehensively captures language abnormalities, and adapts to individual patient conditions and clinical realities.
Smart Images

Figure CN123095404A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the interdisciplinary fields of artificial intelligence, medical big data and clinical diagnosis and treatment of stroke, and in particular to a method, device and medium for predicting the language risk of cognitive degeneration in stroke. Background Technology
[0002] Cognitive decline after ischemic stroke is an important complication that affects the long-term quality of rehabilitation for patients. Early identification and dynamic assessment of cognitive decline are of great clinical significance for individualized intervention.
[0003] Currently, clinical assessment mainly relies on scales such as the MMSE and cranial imaging examinations. However, scale assessments are time-consuming, highly influenced by physicians' subjective experience, and difficult to achieve high-frequency continuous monitoring. Furthermore, subtle abnormalities in the language expression of patients with mild cognitive decline often appear before the decline in scale scores, making it difficult to capture them in a timely manner using traditional methods.
[0004] Existing AI prediction methods are mostly based on random forests or deep neural networks, classifying structured data such as age, NIHSS score, blood lipids, and lesion volume, which do not make sufficient use of language behaviors that reflect cognitive function. Furthermore, natural language processing often focuses on shallow features such as word count, average sentence length, and pause frequency, making it difficult to identify language abnormalities such as semantic jumps, lack of abstract expression, and confusion of reference.
[0005] Multimodal fusion often uses static modeling with fixed weights, which cannot dynamically adjust the focus of analysis based on lesion location, recovery stage, and clinical feedback, resulting in insufficient individualized prediction accuracy and clinical adaptability. Summary of the Invention
[0006] To address the technical problems described in the background, this application provides a method, device, and medium for predicting the language risk of cognitive decline in stroke. The method includes: acquiring multimodal data from stroke patients; the multimodal data includes language expression data, clinical structured data, and lesion imaging data; performing structured processing on the language expression data to obtain language feature vectors, and determining the language abnormality intensity of each language feature in the language feature vectors; inputting the clinical structured data into a preset ReAct agent, calculating the priority of each language feature analysis task according to a preset task scoring function, and obtaining a priority score for each analysis task; executing a differentiated lesion-driven task based on the lesion location in the clinical structured data, and obtaining a task-driven result; scheduling specialized sub-agents to perform parallel abnormality analysis of each language feature based on the priority score and the task-driven result, and outputting the abnormality intensity result of each language feature; and weighting and fusing the abnormality intensity result, clinical structured data, and lesion imaging data to obtain a stroke cognitive decline risk level.
[0007] In one example, determining the language abnormality intensity of each language feature in the language feature vector specifically includes: obtaining the normal reference range for each language feature in healthy individuals; comparing the actual feature value of each language feature corresponding to a stroke patient with the normal reference range; if the actual feature value is within the normal reference range, the language abnormality intensity of the corresponding language feature is no abnormality; if the actual feature value exceeds the normal reference range, normalization calculation is performed based on the deviation amplitude between the actual feature value and the mean of the reference range to obtain the language abnormality intensity; the larger the deviation amplitude, the higher the language abnormality intensity value.
[0008] In one example, the priority of each language feature analysis task is calculated according to a preset task scoring function to obtain a priority score for each analysis task. Specifically, this includes: obtaining the language abnormality intensity, clinical correlation between the language feature and stroke-related cognitive decline, and historical predictive contribution for each language feature; inputting the language abnormality intensity, clinical correlation, and historical predictive contribution into the preset task scoring function, and calculating the priority score for each language feature analysis task through weighted summation; the task scoring function is: ; Let i be the priority score for the i-th language feature analysis task. Let be the language anomaly intensity of the i-th language feature, with a value range of [0, 1]. denoted as the clinical association degree between the i-th language feature and cognitive decline in stroke, with a value range of [0, 1]. The historical prediction contribution of the i-th language feature, with a value range of [0, 1]; , , These are the weighting coefficients, and .
[0009] In one example, a differentiated lesion-driven task is performed based on the lesion location in the clinical structured data to obtain task-driven results. Specifically, this includes: extracting the lesion location of stroke lesions from the clinical structured data; if the lesion location is in the left frontal lobe, generating a task-driven result that focuses on analyzing verb omissions and syntactic breaks; if the lesion location is in the right frontal lobe, generating a task-driven result that focuses on analyzing pronoun confusion and contextual drift.
[0010] In one example, based on the priority score and the task-driven result, specialized sub-agents are scheduled to perform anomaly analysis of each language feature in parallel, and the anomaly intensity result of each language feature is output. Specifically, this includes: comparing the priority score of each language feature analysis task with a preset priority threshold to determine a first-level priority task and a second-level priority task; the priority of the first-level priority task is greater than the priority of the second-level priority task; combining the lesion identified by the task-driven result with the priority analysis direction, the analysis weight of the corresponding language feature analysis task is preset and targeted; and scheduling specialized sub-agents corresponding to the dimensions of vocabulary, fluency, semantic logic, abstract expression, and referential accuracy to perform language feature anomaly analysis of the corresponding dimensions in parallel according to the task priority order, and outputting the quantitative anomaly intensity result corresponding to each language feature.
[0011] In one example, the abnormal intensity results, clinical structured data, and lesion imaging data are weighted and fused to obtain the stroke cognitive decline risk level. Specifically, this includes: integrating the abnormal intensity results of each language feature, various indicators in the clinical structured data, and various indicators in the lesion imaging data to construct a fusion feature vector; adjusting the fusion weight ratios of the three types of features (language features, clinical data, and imaging data) according to the overall degree of language abnormality and the severity of clinical condition of stroke patients; performing a weighted fusion calculation on the fusion feature vector using the adjusted fusion weights to obtain a comprehensive cognitive decline risk score; and comparing the comprehensive cognitive decline risk score with a preset grading threshold to output the corresponding stroke cognitive decline risk level.
[0012] In one example, after obtaining the risk level of cognitive decline due to stroke, the method further includes: acquiring clinical feedback data on the risk level of cognitive decline due to stroke; the clinical feedback data includes physician scores and suggestion adoption rates; calculating the feedback reward value of the clinical feedback data based on a preset reward function, and updating the weight coefficient of the task priority score and the fusion weight ratio of the fusion feature vector when the feedback reward value is lower than a preset reward threshold.
[0013] In one example, the preset reward function is: ; Rate the doctor To determine the adoption rate of the suggestion, For error penalty items, , , These are the weighting coefficients, and .
[0014] On the other hand, embodiments of this application provide a language prediction device for stroke cognitive degeneration risk, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform any of the above-mentioned language prediction methods for stroke cognitive degeneration risk.
[0015] On the other hand, embodiments of this application provide a non-volatile computer storage medium for predicting the language risk of stroke cognitive decline, which stores computer-executable instructions capable of executing any of the above-mentioned methods for predicting the language risk of stroke cognitive decline.
[0016] The above-described technical solutions adopted in the embodiments of this application can achieve the following beneficial effects: This application deeply integrates the ReAct agent reasoning mechanism with a large language model, breaking through the traditional direct classification and fixed analysis mode. It drives task decomposition and priority ranking based on clinical states such as lesion location and onset time, making language feature analysis more targeted, avoiding ineffective analysis and improving efficiency. Through parallel analysis of multi-dimensional features such as vocabulary, fluency, semantic logic, abstract expression, and reference accuracy by multiple sub-agents, it overcomes the one-sidedness of single features and comprehensively captures early subtle language abnormalities. Combining lesion-driven strategies with dynamic weight fusion, it achieves dynamic adaptation of language, clinical, and imaging indicators, making predictions more consistent with individual patient conditions and clinical reality, and can provide reliable decision support for early warning and individualized intervention of post-stroke cognitive impairment. Attached Figure Description
[0017] To more clearly illustrate the technical solution of this application, some embodiments of this application will be described in detail below with reference to the accompanying drawings, in which: Figure 1 A flowchart illustrating a method for predicting language-related cognitive decline risk in stroke, provided in an embodiment of this application; Figure 2 A flowchart illustrating the steps of a method for predicting the language risk of cognitive decline in stroke, provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a language prediction device for the risk of cognitive decline in stroke, provided in an embodiment of this application. Detailed Implementation
[0018] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0019] Some embodiments of this application will now be described in detail with reference to the accompanying drawings.
[0020] Figure 1 This is a flowchart illustrating a method for predicting language-based cognitive decline risk in patients with stroke, provided as an embodiment of this application. This method can be applied to various business domains. Certain input parameters or intermediate results in this process can be manually adjusted to help improve accuracy.
[0021] The analysis method involved in the embodiments of this application can be implemented by a terminal device or a server, and this application does not impose any special limitations on it. For ease of understanding and description, the following embodiments are all described in detail using a server as an example.
[0022] Based on this Figure 1 The process may include the following steps: S101: Acquire multimodal data of stroke patients; the multimodal data includes language expression data, clinical structured data, and lesion imaging data.
[0023] In some embodiments of this application, by connecting the hospital's clinical data acquisition system with a standardized acquisition terminal, the synchronous and complete acquisition of three types of multimodal data is achieved. The acquisition process follows the principles of standardization, repeatability, and non-interference to ensure the authenticity and validity of the data.
[0024] The language expression data specifically encompasses three complementary categories, comprehensively covering three dimensions: oral expression, written expression, and language comprehension. First, oral expression data involves guiding patients through a 3-minute free-expression task using a standardized voice acquisition terminal. External interference is avoided during the acquisition process to ensure clear speech. Acquisition parameters are specified as a sampling frequency of 16kHz, a sampling precision of 16bit, and a mono channel. After acquisition, the speech is transcribed to obtain the spoken text. Second, written text data uses a standardized writing task, requiring patients to complete a designated narrative task within 10 minutes. An example task involves describing three events completed that morning, including the time, place, process, and feelings associated with each event. After acquisition, the written text undergoes standardized preprocessing to remove corrections and standardize font formatting, resulting in standardized written text data. Third, language comprehension feedback data uses standardized logical reasoning questions to assess patients' language comprehension abilities. The questions are designed to fit everyday scenarios with a reasonable difficulty gradient. Example questions include whether it's reasonable to eat before brushing teeth and whether the statement "apples and bananas are both fruits" is correct. Patients provide their judgments orally or in writing, and the judgment results and reaction times are recorded simultaneously as supplementary data for language comprehension characteristics.
[0025] Structured clinical data is automatically and synchronously collected through the hospital information system, eliminating the need for manual data entry. The core clinical data collected includes basic patient information, stroke-related information, assessment indicators, and physiological indicators. Lesion imaging data is obtained by connecting to the medical imaging system, specifically including imaging characteristics such as the size, shape, and location of stroke lesions.
[0026] After initial verification of completeness and format standardization, all collected multimodal data are uniformly stored in the clinical data platform to provide a data foundation for subsequent feature processing and analysis.
[0027] S102: Perform structured processing on the language expression data to obtain language feature vectors, and determine the language anomaly intensity of each language feature in the language feature vectors.
[0028] In some embodiments of this application, the collected unstructured language data is converted into standardized structured feature vectors, and the language feature vectors are defined as follows: It contains five feature components, each of which accurately represents the patient's language function status through quantitative calculation.
[0029] in, The average sentence length is used to quantify the completeness of a patient's language expression. The calculation method is to divide the total number of words by the total number of sentences. Spoken language is counted according to the vocabulary after speech transcription and sentences are divided according to pause intervals. Written text is counted according to the actual written vocabulary and sentences are divided according to punctuation marks. The pause frequency is used to quantify the fluency of a patient's spoken expression. Pauses exceeding 800ms are defined as abnormal pauses. The calculation method is to divide the number of abnormal pauses by the total duration of the spoken expression. The number of semantic jumps is used to quantify the logical coherence of patients' language expression. The topic relevance of two consecutive sentences is analyzed through a large language model, and the number of topic deviations is counted. The deviation criterion is that the content of the second sentence is not related to the core topic of the first sentence and cannot be related by inference from the context. The pronoun error rate is used to quantify the accuracy of patients' semantic reference. It counts the number of incorrect references of pronouns such as he, she, it, this, that, etc. The calculation method is to divide the number of incorrect pronoun references by the total number of times pronouns are used. Incorrect references include situations such as unclear references and incorrect references. The abstract vocabulary ratio is used to quantify a patient's abstract thinking ability. It is a statistical measure of the proportion of abstract nouns such as emotion, memory, thinking, and hope that do not have specific physical counterparts. The calculation method is to divide the number of abstract nouns by the total number of words.
[0030] After feature extraction, the language abnormality intensity of each language feature is further determined: First, the normal reference range for language features in healthy individuals of the same age range is obtained, which is obtained through multi-center clinical data statistics; then, the actual feature values of the language features corresponding to stroke patients are compared with the normal reference range. If the actual feature value is within the normal reference range, the language abnormality intensity of the corresponding language feature is 0, indicating that the feature is not abnormal; if the actual feature value exceeds the normal reference range, normalization is performed based on the deviation between the actual feature value and the mean of the reference range to obtain the language abnormality intensity with a value range of [0,1], as shown in the formula: , These are the actual eigenvalues. The mean of the reference interval is used as the reference value. The greater the deviation between the actual feature value and the mean of the reference interval, the higher the intensity of the language anomaly. The closer it is to 1, the more obvious the abnormality of the feature.
[0031] S103: Input the clinical structured data into a preset ReAct agent, calculate the priority of each language feature analysis task according to the preset task scoring function, and obtain the priority score of each analysis task.
[0032] In some embodiments of this application, relying on the dynamic reasoning mechanism of the ReAct agent, the priority ranking of the prediction task is driven by the patient's clinical status, breaking the traditional fixed language feature analysis mode.
[0033] In the specific calculation process, the patient's core clinical features, such as lesion location, onset time, and NIHSS score, are first extracted from the clinical structured data. At the same time, the intensity of language abnormality corresponding to each language feature is obtained. The parameters include three categories: the clinical association between the language feature and cognitive decline in stroke, and the historical predictive contribution of the language feature.
[0034] The three parameters mentioned above are then input into a preset task scoring function, and a priority score for each type of language feature analysis task is calculated through weighted summation. The task scoring function is as follows: ;in, Let i be the priority score for the i-th language feature analysis task. Let be the language anomaly intensity of the i-th language feature, with a value range of [0, 1]. The clinical correlation between the i-th language feature and cognitive decline in stroke is determined by clinical data statistics and ranges from [0, 1]. The larger the value, the higher the correlation. The historical prediction contribution of the i-th language feature is the degree to which the language feature contributes to the prediction of cognitive degradation risk in past prediction cases. It is obtained by training with historical data and has a value range of [0, 1]. , , These are the weighting coefficients, and It can be dynamically adjusted based on clinical practice and model training results.
[0035] This task scoring function allows for the dynamic quantification of the importance of each analytical task in conjunction with the individual clinical characteristics of the patient, providing a quantitative basis for subsequent task breakdown and scheduling.
[0036] S104: Execute a differentiated lesion-driven task based on the lesion location in the clinical structured data to obtain the task-driven result.
[0037] In some embodiments of this application, this step combines the functional differences of brain regions in stroke lesions to formulate targeted lesion-driven task strategies, further improving the personalization and accuracy of task breakdown and conforming to the actual clinical pathological characteristics.
[0038] In practice, the location information of stroke lesions is first extracted from clinical structured data to identify the specific brain region to which the lesion belongs. Then, based on the different impacts of different brain regions on language function, a differentiated analysis focus strategy is implemented: if the lesion is located in the left frontal lobe, since the left frontal lobe is mainly responsible for language expression and syntactic construction, the task-driven results that focus on verb omissions and syntactic breaks are generated, and the analysis weight of the corresponding language features is strengthened in subsequent analyses; if the lesion is located in the right frontal lobe, since the right frontal lobe is mainly responsible for semantic understanding and context perception, the task-driven results that focus on referential confusion and context drift are generated, and the in-depth mining of the corresponding semantic features is emphasized in subsequent analyses.
[0039] The results of this lesion-driven task, together with the priority score, will serve as the basis for subsequent sub-agent scheduling, ensuring that the language feature analysis aligns with the patient's lesion pathology characteristics.
[0040] S105: Based on the priority score and the task-driven result, schedule the special sub-agents to perform anomaly analysis of each language feature in parallel and output the anomaly intensity result of each language feature.
[0041] In some embodiments of this application, based on the obtained task priorities and task-driven results, multiple specialized sub-agents are scheduled to work in parallel to comprehensively extract the patient's language features, covering different dimensions of language expression, and avoiding the limitations of single feature analysis.
[0042] In practice, the priority score of each language feature analysis task is first compared with a preset priority threshold to determine first-level and second-level priority tasks. The priority of the first-level priority task is higher than that of the second-level priority task. The default priority threshold is 0.7. When the priority score is greater than 0.7, the corresponding task is a first-level priority task; otherwise, it is a second-level priority task. At the same time, based on the priority analysis direction corresponding to the lesions clearly defined by the task-driven results, the analysis weight of the corresponding language feature analysis task is preset and targeted to enhance the analysis depth of lesion-related features.
[0043] After completing task classification and weight adjustment, specialized sub-agents corresponding to the dimensions of vocabulary, fluency, semantic logic, abstract expression, and reference accuracy are scheduled to conduct parallel analysis of language feature anomalies in the corresponding dimensions according to task priority: the vocabulary feature agent focuses on analyzing vocabulary richness, the rationality of verb / noun usage, and vocabulary error rate; the fluency feature agent focuses on analyzing core indicators such as pause patterns and speech rate changes in spoken expression; the semantic logic feature agent focuses on analyzing the logical coherence and thematic consistency of language expression; the abstract expression feature agent focuses on assessing the patient's abstract thinking ability, reflected by the use of abstract vocabulary; and the reference accuracy feature agent focuses on analyzing the correctness of pronoun usage, reflecting the completeness of semantic understanding and expression.
[0044] After each sub-agent completes its analysis, it outputs the quantified anomaly intensity result corresponding to each language feature, which is then uniformly summarized into the subsequent fusion module.
[0045] S106: The abnormal intensity results, clinical structured data, and lesion imaging data are weighted and fused to obtain the risk level of cognitive decline in stroke.
[0046] In some embodiments of this application, this step dynamically integrates language features, clinical indicators, and imaging indicators from multiple dimensions. A dynamic weight update mechanism ensures that the contribution of each indicator matches the patient's actual condition, ultimately outputting an accurate cognitive decline risk level and providing a clear decision-making basis for clinical intervention.
[0047] In practice, the abnormal intensity results of each language feature, various indicators in the clinical structured data, and various indicators in the lesion imaging data are first integrated to construct a multi-dimensional fusion feature vector. , where F is the language feature vector, C is the clinical data vector, and I is the image data vector.
[0048] Subsequently, based on the overall degree of language abnormality and the severity of clinical condition of stroke patients, the fusion weight ratios of the three types of features—language features, clinical data, and imaging data—are dynamically adjusted. The more significant the feature abnormality, the higher the weight ratio of the corresponding category. The update of the dynamic fusion weights follows the constraint that the sum of the weight coefficients is 1.
[0049] The adjusted fusion weights are used to perform a weighted fusion calculation on the fusion feature vector to obtain a comprehensive risk score for cognitive decline. The score ranges from [0, 100], and the formula is as follows: .
[0050] in, To fuse feature vectors, for The dimension; The feature weights learned by the model; For large models, it is a nonlinear feature mapping function; This is the model bias term, obtained from training with clinical data.
[0051] Finally, the comprehensive risk score of cognitive decline is compared and matched with the preset grading thresholds to output the corresponding stroke cognitive decline risk level. The risk level is divided into four levels: low risk, corresponding to the score range [0,20], with basically normal language features, stable clinical indicators, and no obvious cognitive decline trend; low to medium risk, corresponding to the score range (20,40], with a few minor abnormalities in language features, basically stable clinical indicators, and a slight risk of cognitive decline, requiring regular follow-up; medium to high risk, corresponding to the score range (40,70], with multiple abnormalities in language features, abnormal clinical indicators, and a significant cognitive decline trend, requiring targeted intervention; high risk, corresponding to the score range (70,100], with severely abnormal language features, significantly abnormal clinical indicators, and a rapid rate of cognitive decline, requiring emergency intervention and intensive treatment.
[0052] The output includes a comprehensive risk score, abnormal indicators for each dimension, and corresponding clinical intervention suggestions, making it easy for clinicians to refer to and apply.
[0053] It should be noted that, although the embodiments in this application are based on... Figure 1 Steps S101 to S106 will be described sequentially, but this does not mean that steps S101 and S106 must be performed in a strict order. The reason this embodiment follows this order is... Figure 1 The order in which steps S101 to S106 are described is provided to facilitate understanding of the technical solutions of the embodiments of this application by those skilled in the art. In other words, in the embodiments of this application, the order of steps S101 to S106 can be appropriately adjusted according to actual needs.
[0054] Figure 2 This application provides a flowchart illustrating the steps of a method for predicting language-related cognitive decline risk in patients with stroke. Figure 2 The diagram illustrates the complete execution flow of this application, from data input to risk output and then to closed-loop optimization: the starting point is a multimodal data acquisition unit, which is labeled with the acquisition paths and specifications for three types of data: language expression data, clinical structured data, and lesion imaging data; following this is the language feature structuring processing unit, which shows the process of generating standardized language feature vectors and corresponding language abnormality intensities after feature extraction, benchmark comparison, and normalization calculation of unstructured language data; further down is the ReAct intelligent agent task decomposition core unit, which presents the processing logic in two branches: one is the analysis task priority calculation branch based on the task scoring function, and the other is the differentiated lesion-driven task branch based on the lesion location. The two branches jointly output task scheduling instructions; The subsequent arrangement includes a parallel analysis unit with multiple specialized sub-agents, illustrating the parallel working architecture of five types of sub-agents: vocabulary, fluency, semantic logic, abstract expression, and referential accuracy. It also shows the flow of the abnormal intensity results of language features in the corresponding dimensions output by each sub-agent. Next is the risk prediction fusion unit, demonstrating the complete process of dynamically adjusting and weighting the three types of features—language features, clinical data, and imaging data—to obtain a comprehensive risk score, which is then matched with graded thresholds to output a four-level cognitive decline risk level. Finally, a clinical feedback optimization branch is introduced, illustrating the closed-loop iterative path of updating the task score weights and fusion weight parameters in reverse after the clinical feedback data is quantitatively evaluated using a reward function. Overall, the sequence, data flow, and logical connections of each step in this method are clearly presented.
[0055] Figure 3 A schematic diagram of the structure of a language prediction device for the risk of cognitive decline in stroke provided in this application embodiment includes: At least one processor; and, A memory that is communicatively connected to at least one processor; wherein, A method for predicting the language risk of cognitive decline in stroke, wherein the memory stores instructions that can be executed by at least one processor, and the instructions are executed by at least one processor to enable at least one processor to perform any of the above.
[0056] Some embodiments of this application provide a non-volatile computer storage medium for predicting the language risk of stroke cognitive decline, storing computer-executable instructions that can execute any of the above-mentioned methods for predicting the language risk of stroke cognitive decline.
[0057] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the device and medium embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the description of the method embodiments.
[0058] The devices and media provided in this application are one-to-one with the methods. Therefore, the devices and media also have similar beneficial technical effects as their corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the devices and media will not be repeated here.
[0059] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0060] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0061] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0062] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0063] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.
[0064] Memory may include non-persistent storage in computer-readable media, random access memory (RAM), and non-volatile memory such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.
[0065] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.
[0066] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.
[0067] The above are merely embodiments of this application and are not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the technical principles of this application should fall within the protection scope of this application.
Claims
1. A method for predicting the language risk of cognitive decline in stroke, characterized in that, The method includes: Acquire multimodal data of stroke patients; the multimodal data includes language expression data, clinical structured data, and lesion imaging data; The language expression data is structured to obtain language feature vectors, and the language anomaly intensity of each language feature in the language feature vectors is determined. The clinical structured data is input into a preset ReAct agent, and the priority of each language feature analysis task is calculated according to a preset task scoring function to obtain the priority score of each analysis task. A differentiated lesion-driven task is performed based on the lesion location in the clinical structured data to obtain task-driven results; Based on the priority score and the task-driven result, the special sub-agents are scheduled to perform anomaly analysis of each language feature in parallel and output the anomaly intensity result of each language feature. The abnormal intensity results, clinical structured data, and lesion imaging data are weighted and fused to obtain the risk level of cognitive decline in stroke.
2. The method according to claim 1, characterized in that, Determining the language anomaly intensity of each language feature in the language feature vector specifically includes: Obtain the normal reference range for each language feature in healthy individuals; The actual feature value of each language feature corresponding to the stroke patient is compared with the normal reference interval; If the actual feature value is within the normal reference range, the language anomaly intensity of the corresponding language feature is no anomaly. If the actual feature value exceeds the normal reference range, the language abnormality intensity is calculated by normalization based on the deviation between the actual feature value and the mean of the reference range; the larger the deviation, the higher the language abnormality intensity value.
3. The method according to claim 1, characterized in that, The priority of each language feature analysis task is calculated based on a pre-defined task scoring function, resulting in a priority score for each analysis task. Specifically, this includes: Obtain the intensity of language abnormality, the clinical correlation between language features and cognitive decline in stroke, and the historical predictive contribution for each language feature; The language abnormality intensity, clinical relevance, and historical predictive contribution are input into a preset task scoring function, and the priority score corresponding to the analysis task of each language feature is obtained by weighted summation. The task scoring function is: ; Let i be the priority score for the i-th language feature analysis task. Let be the language anomaly intensity of the i-th language feature, with a value range of [0, 1]. denoted as the clinical association degree between the i-th language feature and cognitive decline in stroke, with a value range of [0, 1]. The historical prediction contribution of the i-th language feature, with a value range of [0, 1]; , , These are the weighting coefficients, and .
4. The method according to claim 1, characterized in that, Based on the lesion locations in the aforementioned clinical structured data, a differentiated lesion-driven task is performed to obtain task-driven results, specifically including: Extract the location of stroke lesions from the aforementioned structured clinical data; If the lesion is located in the left frontal lobe, a task-driven result is generated that focuses on analyzing verb loss and syntactic breaks. If the lesion is located in the right frontal lobe, a task-driven result will be generated that focuses on analyzing referential confusion and contextual drift.
5. The method according to claim 1, characterized in that, Based on the priority score and the task-driven result, the specialized sub-agents are scheduled to perform anomaly analysis for each language feature in parallel, and output the anomaly intensity result for each language feature, specifically including: The priority score of each language feature analysis task is compared with a preset priority threshold to determine first-level priority tasks and second-level priority tasks; the priority of the first-level priority task is higher than the priority of the second-level priority task. Based on the lesion-specific analysis direction clearly defined by the task-driven results, the analysis weight of the corresponding language feature analysis task is pre-defined and targeted to increase; Specialized sub-agents corresponding to the dimensions of vocabulary, fluency, semantic logic, abstract expression, and referential accuracy are scheduled to perform language feature anomaly analysis in parallel according to task priority, and output the quantitative anomaly intensity result for each language feature.
6. The method according to claim 1, characterized in that, The abnormal intensity results, clinical structured data, and lesion imaging data are weighted and fused to obtain the stroke cognitive decline risk level, specifically including: The abnormal intensity results of each language feature, various indicators in the clinical structured data, and various indicators in the lesion imaging data are integrated to construct a fusion feature vector; Based on the overall degree of language abnormality and the severity of clinical condition of stroke patients, the fusion weight ratios of the three types of features—language features, clinical data, and imaging data—are adjusted. The fusion feature vector is weighted and fused using the adjusted fusion weights to obtain a comprehensive risk score for cognitive decline. The comprehensive risk score of cognitive decline is compared and matched with a preset grading threshold to output the corresponding risk level of cognitive decline due to stroke.
7. The method according to claim 1, characterized in that, After obtaining the risk level of cognitive decline due to stroke, the method further includes: Obtain clinical feedback data on the stroke cognitive decline risk level; the clinical feedback data includes physician scores and recommendation adoption rate; The feedback reward value of the clinical feedback data is calculated based on a preset reward function, and the weight coefficient of the task priority score and the fusion weight ratio of the fusion feature vector are updated when the feedback reward value is lower than the preset reward threshold.
8. The method according to claim 7, characterized in that, The preset reward function is: ; Rate the doctor To determine the adoption rate of the suggestion, For error penalty items, , , These are the weighting coefficients, and .
9. A language prediction device for the risk of cognitive decline in stroke, characterized in that, include: At least one processor; as well as, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform a language prediction method for stroke cognitive degeneration risk as described in any one of claims 1-8.
10. A language prediction storage medium for stroke cognitive decline risk, storing computer-executable instructions, characterized in that, The computer-executable instructions are capable of executing the language prediction method for stroke cognitive decline risk as described in any one of claims 1-8.
