A method and system for intelligent assessment of HCV risk

Abstract

This application discloses an intelligent HCV risk assessment method and system. In the high-risk HCV assessment step, the method acquires electronic medical record text and a domain knowledge graph containing entities with HCV high-risk factors, negation words, and time-limited words. It dynamically generates a context bias matrix and integrates it with a pre-trained language model to accurately output high-risk identification results. In the emotion computing step, it extracts the underlying acoustic features of the patient's speech, processes the emotion and drug side effect-related physiological state features in parallel via a two-stream network, and corrects the emotion features through attention gating to achieve accurate judgment of emotions such as depression and anxiety. This application can assess HCV high-risk risk and patient outpatient compliance, reducing the missed screening rate and the rate of emotion misjudgment.

Description

Technical Field

[0001] This invention belongs to the field of medical cyber-physical systems technology, specifically relating to a system for automated screening, intelligent decision-making, and closed-loop management of outpatient compliance for hepatitis C (HCV). Background Technology

[0002] Hepatitis C, as a global public health challenge, faces management bottlenecks due to multiple breaks in the complete chain of "screening-assessment-intervention-follow-up." While many studies suggest that optimizing management processes can significantly improve health management efficiency, effectively implementing these management concepts from a strategic level into busy clinical practice still faces significant technical challenges: First, the "cognitive dependence" bottleneck in the screening stage: screening of high-risk patients heavily relies on the personal experience and responsibility of clinicians, lacking an automated tool that can proactively and accurately identify risks, leading to a large number of potential patients being missed. Second, the "process disconnect" bottleneck: between antibody positivity and confirmation through RNA testing and initiation of small molecule drug DAA intervention, patients often drop out due to issues such as poor referrals, multiple appointments, and information barriers, resulting in unsatisfactory treatment rates. Third, the "passive follow-up" bottleneck in the outpatient rehabilitation stage: traditional telephone or questionnaire follow-up methods cannot objectively assess patients' treatment adherence and psychological state in real time. In particular, during the treatment cycle of DAA, negative emotions caused by drug side effects, economic pressure, or lack of understanding are key factors leading to treatment interruption, and current technology lacks effective proactive intervention methods.

[0003] Therefore, there is an urgent need for a brand-new technological solution that can not only solidify and automate the execution of optimal management processes, but also achieve accurate screening of high-risk patients and effective perception of their true emotional state through technological innovation, thereby enabling proactive intervention and building a truly meaningful closed-loop management system. Summary of the Invention

[0004] The purpose of this application is to provide an intelligent HCV risk assessment method and system that can realize intelligent HCV risk assessment.

[0005] Firstly, this application provides an intelligent HCV risk assessment method, including HCV high-risk assessment steps, specifically including:

[0006] Obtain the electronic medical record text to be processed;

[0007] A domain knowledge graph is obtained, which includes entities representing high-risk factors for hepatitis C (HCV) and entities representing negation words; wherein, the entities representing negation words are used to identify exclusion scenarios that negate the high-risk factors or are not associated with the patient subject.

[0008] Based on the electronic medical record text and the domain knowledge graph, a context bias matrix is ​​dynamically generated, specifically including:

[0009] Positive enhancement operation: If both patient-related words and high-risk factor entities are detected in the same semantic unit in the electronic medical record text, the attention score of the corresponding element in the context bias matrix is ​​increased.

[0010] Reverse inhibition operation: If a negative word entity and a high-risk factor entity are detected to appear simultaneously in the same semantic unit in the electronic medical record text, or if a high-risk factor entity is a word that is not related to the patient subject, then the attention score of the corresponding word element in the context bias matrix is ​​reduced.

[0011] If the same set of related words simultaneously meets the conditions for both positive enhancement and negative suppression, the negative suppression operation will be performed first.

[0012] The context bias matrix is ​​fused with the original attention score matrix of the pre-trained language model to generate a biased attention score matrix.

[0013] The electronic medical record text is input into a pre-trained language model. The pre-trained language model performs attention calculation based on the biased attention score matrix to achieve semantic parsing and outputs the HCV high-risk assessment result.

[0014] In one possible implementation, the domain knowledge graph also includes time-limited word entities, which are used to define the high-risk time range of HCV high-risk factors;

[0015] If both patient-related words and high-risk factor entities are detected within the same semantic unit in the electronic medical record text, the attention score of the corresponding element in the context bias matrix is ​​increased, including:

[0016] For cases containing only patient-related terms and high-risk factor entities, a basic weighting increase is applied.

[0017] For cases where patient-related words, high-risk factor entities, and time-limited words appear simultaneously within the same semantic unit in electronic medical record text, a higher weighting is applied to reflect the priority of high-risk time associations.

[0018] In one possible implementation, the pre-trained language model is a BERT model or a variant thereof.

[0019] In one possible implementation, the HCV high-risk factor entities include a history of blood transfusion, surgical history, intravenous drug addiction, or unclean tattoos; the patient-related terms include “patient,” “person,” or “self-reported”; and the non-patient-related terms include “family member,” “mother,” or “father.”

[0020] In one possible implementation, fusing the context bias matrix with the original attention score matrix of the pre-trained language model to generate a biased attention score matrix includes:

[0021] Pre-trained language model initialization and text encoding: Using the BERT model or its variant as the basic pre-trained language model, the electronic medical record text to be processed is segmented into words, special tags are added and positional encoding is performed according to the requirements of the pre-trained language model, and text embedding vectors are generated.

[0022] Original attention score matrix calculation: Self-attention is calculated on the text embedding vectors using the multi-head attention mechanism of the pre-trained language model to obtain the original attention score matrix. ;

[0023] Context bias matrix normalization: normalization of dynamically generated context bias matrices Min-Max normalization is performed to map the element values ​​to the [0,1] interval, resulting in the normalized bias matrix. ;

[0024] Weighted fusion generates a biased attention score matrix: An adaptive weighted fusion strategy is adopted, and the fusion formula is as follows: ,in This is the attention score matrix after biasing. To integrate the weighting coefficients, To be The matrix expanded by the number of multi-head attention heads has dimensions equal to... Maintain consistency; Each submatrix corresponding to each attention head undergoes Softmax normalization to ensure that the sum of the attention weights for each word is 1, resulting in the final biased attention score matrix. .

[0025] In one possible implementation, the electronic medical record text is input into a pre-trained language model. The pre-trained language model performs attention calculation based on the biased attention score matrix to achieve semantic parsing and outputs an HCV high-risk assessment result, including:

[0026] Pre-trained language models are based on The context-aware feature vector is obtained by weighted summation of the text embedding vectors of the electronic medical record text, using the following formula:

[0027] ;

[0028] in For text embedding vectors, Context-aware feature vectors;

[0029] Will The input is the subsequent fully connected layer and classifier. The confidence level of each HCV high-risk factor is output through the Sigmoid activation function. If the confidence level of a certain HCV high-risk factor is greater than a preset threshold, the corresponding sample is determined to have the high-risk factor. All high-risk factor determination results are integrated to output the HCV high-risk risk assessment result.

[0030] In one possible implementation, the method further includes a step of calculating the emotional state of HCV patients during outpatient follow-up, specifically including:

[0031] Acquire the patient's speech signal and extract the underlying acoustic feature sequence from it;

[0032] The underlying acoustic feature sequence is input in parallel into a dual-stream network, which includes an emotion recognition stream and a physiological state recognition stream. The emotion recognition stream processes the underlying acoustic feature sequence using a convolutional recurrent neural network (CRNN) to generate emotion-related acoustic features. The physiological state recognition stream processes the underlying acoustic feature sequence using a lightweight convolutional network (S-CNN) to identify physiological state acoustic patterns related to drug side effects and generate physiological state features. Using the physiological state features, the emotion-related acoustic features are adaptively adjusted through an attention gate to reduce the interference of physiological fatigue on emotion judgment, thereby obtaining a corrected emotion feature.

[0033] Based on the corrected emotional characteristics, the patient's psychological and emotional state is classified and judged.

[0034] In one possible implementation, the type of the underlying acoustic features includes one or more of the following: Mel frequency cepstral coefficients (MFCC), pitch, speech rate, and speech energy.

[0035] In one possible implementation, the acoustic patterns of the physiological states associated with the drug side effects include characteristic patterns of hoarseness, abnormally slowed speech rate, and reduced energy.

[0036] The method of utilizing the physiological state characteristics and adaptively adjusting the emotion-related acoustic features through attention gating to reduce the interference of physiological fatigue on emotion judgment, thereby obtaining a corrected emotional feature, includes:

[0037] Feature dimension unification: integrating emotion-related acoustic features Physiological characteristics Perform dimensional unification processing;

[0038] Attention gate weight generation: and Element-level addition is performed to obtain the fusion feature; The input consists of an attention-gated network with two fully connected layers and an activation function, and the output is an attention weight matrix. ;

[0039] Adaptive correction of emotion features: A weighted adjustment strategy is used to correct emotion-related acoustic features. The correction formula is as follows: ,in These are the emotional characteristics after preliminary correction. Weight matrix The average value over the length dimension of the acoustic feature sequence This is the compensation coefficient;

[0040] Feature smoothing and enhancement: for corrected features Layer normalization is performed; then, the original emotion-related acoustic features are processed through a residual connection module. Low-frequency components and layer normalization The fusion process yields the final corrected emotional characteristics. .

[0041] In one possible implementation, the categories of the psychological and emotional states include depression, anxiety risk, and normal mood.

[0042] Secondly, this application provides an HCV risk intelligent assessment system, including: a memory and a processor;

[0043] The memory is used to store computer programs;

[0044] The processor is used to invoke the computer program to execute the method described above.

[0045] The aforementioned system can be deployed on a central intelligent management server. The computer program can be divided into a semantic parsing module and a voiceprint emotion calculation module, which are used to implement the step of calculating the emotional state of HCV patients during outpatient follow-up in the HCV high-risk assessment process.

[0046] In one possible implementation, the system further includes a physician workstation; after identifying a high-risk condition, the semantic parsing module automatically sends a prompt message to the physician workstation.

[0047] In one possible implementation, the system further includes a patient follow-up interaction terminal; when the voiceprint emotion calculation module determines that the patient has negative emotions such as depression or anxiety, it automatically generates early warning information and triggers an active intervention process through the physician workstation or the patient follow-up interaction terminal.

[0048] Thirdly, this application provides a computer-readable storage medium storing a computer program that, when run on an electronic device, causes the electronic device to perform the method described above.

[0049] Fourthly, this application provides a computer program product, including a computer program that, when run on an electronic device, causes the electronic device to perform the method described above.

[0050] The specific implementation methods of the second to fourth aspects of this application can refer to the implementation methods of the first aspect, and will not be elaborated here.

[0051] Beneficial effects:

[0052] This application overcomes the bottlenecks of "cognitive dependence" and "insufficient accuracy" in the screening stage, achieving automated and accurate screening of high-risk patients. By constructing a domain knowledge graph containing entities related to HCV high-risk factors, negation words, and time-limited words, this application dynamically generates a context bias matrix and optimizes the attention score matrix of the pre-trained language model. It prioritizes high-risk factors relevant to the patient, effectively eliminating interference from negative statements and non-patient-related information, while emphasizing the priority of high-risk time ranges and their association with high-risk factors. In practical testing on a dataset containing 10,000 real medical records, the accuracy of high-risk factor identification increased from 85.2% in existing general models to 98.7%, with a false negative rate reduced by over 70%. This significantly reduces misjudgments and false negatives caused by "semantic drift," eliminating reliance on physicians' personal experience and addressing the core pain point of large-scale accurate clinical screening.

[0053] This application can accurately capture patients' true emotional state: through a dual-stream network architecture, it processes the emotion-related features of patients' speech and the physiological state features related to drug side effects in parallel. Utilizing an attention gating mechanism, it adaptively adjusts emotional features, effectively reducing the interference of physiological fatigue on emotion judgment, and achieving accurate identification of true negative emotions such as depression and anxiety. Compared to traditional follow-up models and general emotion recognition models, this solution improves the accuracy of emotion recognition by 15.6 percentage points, and improves the F1 score in distinguishing between "fatigue but normal mood" and "depression" by 0.23. It can capture patients' out-of-hospital psychological state in real time and objectively, thereby improving out-of-hospital follow-up management and providing a reliable basis for early psychological intervention.

[0054] Based on this application, a closed-loop management system can be constructed to improve the overall health management effect: The technical solution of this application can link automated and accurate screening, dynamic monitoring of out-of-hospital emotional state, physician workstations, and patient follow-up interactive terminals to form a complete closed loop of "screening-early warning-intervention-follow-up". During the screening stage, it proactively alerts high-risk patients, assisting physicians in timely intervention; during the out-of-hospital stage, it monitors patients' emotional state in real time, and once negative emotions are identified, it immediately triggers the proactive intervention process, effectively reducing the treatment interruption rate caused by emotional problems. At the same time, it solidifies the optimal management process, reducing treatment dropout problems caused by poor referrals and information barriers, and comprehensively improving the efficiency and quality of hepatitis C health management.

[0055] This application boasts strong technical adaptability and outstanding practicality: the pre-trained language model can adopt the BERT model or its variants, the underlying acoustic feature extraction supports multiple common types, and the convolutional recurrent neural network and lightweight convolutional network in the dual-stream network balance recognition accuracy and computational efficiency, making it easy to deploy on common hardware devices such as central intelligent management servers. It can also be conveniently applied through physician workstations and patient follow-up interaction terminals, enabling rapid adaptation to actual clinical scenarios and facilitating widespread adoption. Attached Figure Description

[0056] Figure 1 This is a schematic diagram of the HCV high-risk assessment steps in one embodiment of this application;

[0057] Figure 2 This is a flowchart illustrating the steps for calculating the emotional state of HCV patients during outpatient follow-up in one embodiment of this application.

[0058] Figure 3 This is a schematic diagram of the system hardware architecture in one embodiment of this application;

[0059] Figure 4 This is a schematic diagram of the overall system workflow in one embodiment of this application. Detailed Implementation

[0060] To enable those skilled in the art to better understand the present application, the technical solution of the present application will be further described in detail below with reference to the embodiments and accompanying drawings.

[0061] This invention discloses an intelligent HCV risk assessment method and system, aiming to address the technical pain points in existing HCV health management processes, such as missed screening, dropout, and passive follow-up. Based on electronic medical record text and the aforementioned domain knowledge graph, this invention dynamically generates a context bias matrix. This matrix is ​​then used to improve the attention score matrix of a pre-trained language model, accurately identifying high-risk HCV factors related to the patient in the electronic medical record. This achieves automated and high-precision screening of high-risk individuals, overcoming the "cognitive dependence" bottleneck in the screening stage. Furthermore, this application utilizes a two-stream network for outpatient emotional computation. By processing the acoustic features of emotions and physiological states in parallel and employing an attention gating mechanism to suppress the interference of physiological fatigue caused by drug side effects on emotional judgment, it achieves an objective and real-time assessment of the patient's true psychological state. The system integrates the above methods into a central intelligent management server, linking physician workstations and patient follow-up interaction terminals. This not only solidifies the optimal process from screening and diagnosis to outpatient rehabilitation, but also enables proactive intervention through effective perception of the true emotional state of outpatient patients. This significantly reduces the treatment interruption rate caused by psychological problems, constructing a truly automated, intelligent, and closed-loop management system for hepatitis C, encompassing automated screening, intelligent decision-making, and outpatient compliance management. This application addresses the technical problem of insufficient performance of general models in specific medical scenarios. This invention improves existing language models and speech emotion recognition models to better adapt to and address the specific challenges in HCV health management.

[0062] Specific embodiments according to this application will now be described with reference to the accompanying drawings.

[0063] Example 1:

[0064] This application provides an intelligent HCV risk assessment method, including HCV high-risk assessment steps, such as... Figure 1 As shown, it specifically includes:

[0065] S11. Obtain the electronic medical record text to be processed.

[0066] For example, unstructured electronic medical record texts to be processed, such as admission records and progress notes, are obtained from the hospital's HIS / EMR system.

[0067] The electronic medical record text can be preprocessed before being used in subsequent steps.

[0068] S12. Obtain a domain knowledge graph, which includes entities representing high-risk factors for hepatitis C (HCV), time-limited words, and negation words. The time-limited words are used to define the high-risk time range of HCV high-risk factors, and the negation words are used to identify negative statements about high-risk factors or exclusion scenarios related to non-patient subjects.

[0069] The domain knowledge graph can be pre-built before model training.

[0070] In some embodiments, the HCV high-risk factor entities include a history of blood transfusion, surgical history, intravenous drug addiction, or unclean tattoos; the patient-related terms include “patient,” “person,” or “self-reported”; and the non-patient-related terms include “family member,” “mother,” or “father.”

[0071] In some embodiments, time-limiting terms may include “before 1997”, “early 1990s”, etc., because transfusion history prior to a specific time carries a higher risk.

[0072] In some embodiments, the negative word (exclusion word) entity may include “none”, “deny”, “not seen”, “its mother”, “family”, etc.

[0073] S13. Dynamically generate a context bias matrix based on the electronic medical record text and the domain knowledge graph, specifically including:

[0074] Positive enhancement operation: If the same semantic unit in the electronic medical record text is detected to contain both patient-related words and high-risk factor entities (such as "patient + surgical history"), then the attention score of the corresponding element in the context bias matrix is ​​increased.

[0075] Reverse inhibition operation: If the electronic medical record text detects that the same semantic unit contains both a negative word entity and a high-risk factor entity (such as "denial + intravenous drug addiction"), or a high-risk factor entity and words unrelated to the patient subject (such as "his mother + blood transfusion history"), then the attention score of the corresponding word element in the context bias matrix is ​​reduced.

[0076] If the same set of related words simultaneously meets the conditions for both positive enhancement and negative suppression, the negative suppression operation will be performed first.

[0077] In some embodiments, the domain knowledge graph further includes time-limited word entities, which are used to define the high-risk time range of HCV high-risk factors;

[0078] If both patient-related words and high-risk factor entities are detected within the same semantic unit in the electronic medical record text, the attention score of the corresponding element in the context bias matrix is ​​increased, including:

[0079] For cases containing only patient-related words and high-risk factor entities, a base weight increase (e.g., weight coefficient 2.0~2.5) is applied.

[0080] For cases where patient-related words, high-risk factor entities, and time-limited words (such as "myself + 1990 + blood transfusion history") appear simultaneously in the same semantic unit of electronic medical record text, a higher weighting (such as a weighting coefficient of 2.5 to 3.0) is adopted to reflect the priority of high-risk time association.

[0081] S14. The context bias matrix is ​​fused with the original attention score matrix of the pre-trained language model (e.g., element-wise addition) to generate the biased attention score matrix.

[0082] In some embodiments, the pre-trained language model is a BERT model or a variant thereof.

[0083] In some embodiments, the attention score matrix is ​​obtained by fusion followed by Softmax normalization.

[0084] In some embodiments, the specific implementation process is as follows:

[0085] S14.1, Pre-trained Language Model Initialization and Text Encoding: Using the BERT model or its variants as the basic pre-trained language model, the electronic medical record text to be processed is segmented according to the requirements of the pre-trained language model, special tags are added ([CLS] at the beginning of the sentence and [SEP] at the end of the sentence), and positional encoding is performed, generating a dimension of... The text embedding vector, where This represents the number of electronic medical record text samples processed in a single batch by the pre-trained language model. The length of the electronic medical record text sequence. The hidden layer dimension of the pre-trained language model;

[0086] S14.2 Calculation of the original attention score matrix: The original attention score matrix is ​​obtained by calculating self-attention on the text embedding vectors through the multi-head attention mechanism of the pre-trained language model. Its dimensions are ,in The number of attention heads in a multi-head attention mechanism; where Indicates the first In the nth sample, the nth The first thing to pay attention to. The word for the first The original attention weights of each word;

[0087] Context bias matrix normalization: normalization of dynamically generated context bias matrices Min-Max normalization is performed to map the element values ​​to the [0,1] interval, resulting in the normalized bias matrix. This ensures that the values ​​match the range of the original attention score matrix.

[0088] S14.3, Weighted fusion to generate the biased attention score matrix: An adaptive weighted fusion strategy is adopted, and the fusion formula is as follows: ,in This is the attention score matrix after biasing. The fusion weight coefficients (ranging from 0.6 to 0.8, with the optimal value determined through cross-validation) To be The matrix expanded by the number of multi-head attention heads has dimensions equal to... Maintain consistency (will) along Dimensional copying extension, generation Dimensions and Consistent); To The submatrix corresponding to each attention head (dimension 1) Perform Softmax normalization operations on each word to ensure that the sum of the attention weights for each word is 1, thus obtaining the final biased attention score matrix. The formula is: ;in For sample index, For the attention head index, Softmax along Dimension calculation;

[0089] S15. Input the electronic medical record text into the pre-trained language model. The pre-trained language model performs attention calculation based on the biased attention score matrix to achieve semantic parsing and output the HCV high-risk assessment result.

[0090] In some embodiments, the pre-trained language model is based on We obtain the context-aware feature vector by performing a weighted summation on the text embedding vectors, using the following formula:

[0091] ;

[0092] in For text embedding vectors, Context-aware feature vectors;

[0093] The feature vector The input is then fully connected layer and classifier. The confidence level of each HCV high-risk factor is output through the Sigmoid activation function. When the confidence level of a certain HCV high-risk factor is greater than a preset threshold (such as 0.75 by default, which can be adjusted according to the actual scenario), the sample is determined to have the high-risk factor. All high-risk factor determination results are integrated to output the HCV high-risk risk assessment result (high risk / medium risk / low risk).

[0094] For example, all high-risk factor determination results are integrated, and combined with risk level rules (e.g., the presence of 2 or more HCV high-risk factors indicates high risk, the presence of 1 HCV high-risk factor indicates medium risk, and the presence of 0 HCV high-risk factors indicates low risk), to output the HCV high-risk risk assessment result.

[0095] This solution enables intelligent screening of high-risk HCV patients based on a context-biased attention mechanism. Based on the biased attention score matrix, the model is "forced" to focus on HCV high-risk factors directly related to the patient, thereby accurately identifying genuine risk signals and effectively eliminating interference from irrelevant information. For example, the model can accurately identify high risk from "the patient had a blood transfusion history in the 1990s" and exclude interference from irrelevant information such as "the patient's family had a blood transfusion history," ultimately outputting accurate HCV high-risk assessment results and confidence levels.

[0096] Compared to directly using the general BERT model, this method significantly reduces false positives and false negatives caused by semantic drift. Tested on a dataset containing 10,000 real medical records, the accuracy of high-risk factor identification increased from 85.2% to 98.7%, and the false negative rate decreased by more than 70%. This addresses the core pain point of general language models in parsing HCV medical records, which suffers from semantic drift due to a lack of domain knowledge and inaccurate identification of key high-risk factors, leading to missed screenings due to the model's "medical acumen," thus solving the bottleneck problem of large-scale clinical screening.

[0097] In some embodiments, the method further includes a step of calculating the emotional state of HCV patients during outpatient follow-up, such as... Figure 2 As shown, it specifically includes:

[0098] S21. Acquire the patient's speech signal and extract the underlying acoustic feature sequence from it.

[0099] In some embodiments, the types of the underlying acoustic features include one or more of the following: Mel frequency cepstral coefficients (MFCC), pitch, speech rate, and speech energy.

[0100] In some embodiments, the patient's voice signal is generated by the patient recording a voice diary through a follow-up interactive terminal.

[0101] S22. The underlying acoustic feature sequence is input in parallel into a dual-stream network, which includes an emotion recognition stream and a physiological state recognition stream. The emotion recognition stream processes the underlying acoustic feature sequence through a convolutional recurrent neural network to learn and extract general acoustic patterns related to emotional states such as depression and anxiety, so as to generate emotion-related acoustic features. The physiological state recognition stream processes the underlying acoustic feature sequence through a lightweight convolutional network to identify physiological state acoustic patterns related to drug side effects and generate physiological state features.

[0102] The lightweight convolutional network is specifically trained using labeled speech data from patients suffering from fatigue and weakness caused by HCV drug side effects (such as ribavirin). It focuses on recognizing physiological acoustic patterns, such as hoarseness, abnormally slowed speech rate, and reduced energy, and generates physiological state features.

[0103] In some embodiments, the acoustic patterns of physiological states related to drug side effects include characteristic patterns of hoarseness, abnormally slowed speech rate, and reduced energy.

[0104] S23. Utilizing the physiological state characteristics, the emotion-related acoustic features are adaptively adjusted through an attention gate to reduce the interference of physiological fatigue on emotion judgment, thereby obtaining a corrected emotion feature.

[0105] The specific implementation process is as follows:

[0106] S23.1, Feature Dimension Unification: Unifying the emotion-related acoustic features output from the emotion recognition stream. (dimension is) Physiological state characteristics output by physiological state recognition stream (dimension is) ) Perform dimensionality unification processing, using a 1×1 convolutional layer to... Dimensions transformed into Consistent, resulting in dimension 1 Post-adaptation physiological characteristics ;in, The number of speech samples processed in a single batch. The length of the acoustic feature sequence extracted from a single speech signal. The dimension of the output features for emotion recognition stream (the dimensional scale representing the emotion features). The dimension of the output features for identifying physiological states (the dimensional scale representing the features of physiological states).

[0107] S23.2, Attention gate weight generation: and Element-level addition is performed to obtain the fusion feature. ;Will The input consists of an attention gate network with two fully connected layers and an activation function. The first fully connected layer... Compress the feature dimension to The second fully connected layer Restore the feature dimensions to Dimensions, ultimately outputting the attention weight matrix (dimension is) The weight values ​​range from [0,1] and are used to characterize the degree to which the emotional features of each position and each dimension are affected by physiological fatigue. The calculation formula is: ,in The first fully connected layer Activation function, This is the activation function for the second fully connected layer;

[0108] S23.3 Adaptive Correction of Emotional Features: A weighted adjustment strategy is used to correct emotion-related acoustic features. The correction formula is as follows: ,in The emotional characteristics after preliminary correction have suppressed the interference component of physiological fatigue. Weight matrix exist The average value across dimensions This is a compensation coefficient (ranging from 0.1 to 0.3, used to preserve reasonable emotional characteristic intensity); the formula uses weights... It suppresses feature components severely affected by physiological fatigue, while using a compensation coefficient to avoid the loss of effective emotional information due to over-suppression;

[0109] S23.4 Feature Smoothing and Enhancement: For the corrected features... Layer normalization is performed to eliminate feature distribution differences between different samples, resulting in the following: The original emotion-related acoustic features are then processed through a residual connection module. low-frequency components and The fusion process yields the final corrected emotional characteristics. This ensures the stability and effectiveness of the features; the formula is:

[0110] ;

[0111] in, Indicates the corrected features The result of layer normalization express The low-frequency components, through the analysis of The result is obtained by performing 1D average pooling, with a pooling kernel size of 3, a stride of 1, and a fill value of 1.

[0112] This step is crucial for achieving "physiological-psychological" decoupling. The physiological state features output by the physiological state recognition stream are not directly used for classification; instead, they are fed into an attention gate. This gate generates a set of weights based on the intensity of the physiological state features, which are used to adaptively reweight the emotion-related acoustic features output by the emotion recognition stream. For example, when the physiological state recognition stream detects a strong "fatigue" pattern, the attention gate adaptively adjusts (enhances or suppresses) the attention given to specific frequency or temporal features in the emotion recognition stream. For instance, it "tells" the emotion recognition stream: "The currently detected low energy and slow speech rate are likely due to physiological fatigue. Please reduce the weight of these features in judging 'depressive' emotions and pay more attention to other cues such as tone changes," because it judges these acoustic manifestations to be more likely to originate from physiological rather than psychological factors. Simultaneously, it may increase the weight of other depression-related features such as "flat tone changes." Through this adjustment, the original "emotion-related acoustic features" are corrected into a corrected emotional feature that better reflects the true psychological state.

[0113] S24. Based on the corrected emotional characteristics, classify and judge the patient's psychological and emotional state.

[0114] In some embodiments, the corrected emotional features are fed into the final classifier (such as a fully connected layer) to output the patient's psychological and emotional state assessment results.

[0115] In some embodiments, the categories of psychological and emotional states include depression, anxiety risk, and normal mood.

[0116] This method solves the problem of general emotion recognition models not being applicable to patient populations. It can effectively distinguish between physical fatigue and psychological depression, avoiding false alarms caused by drug side effects, thus obtaining a purer and more accurate judgment of emotional state that excludes interference from physiological state. It enables the accurate identification of the true psychological emotions (such as depression and anxiety) of HCV patients, even when they commonly experience drug side effects such as fatigue and weakness (which can contaminate speech signals), thereby assessing the risk of treatment interruption.

[0117] In tests on HCV patients, compared to the baseline CRNN model, this invention improved emotion recognition accuracy by 15.6 percentage points. Particularly in distinguishing between the easily confused states of "fatigue but normal mood" and "depression," the harmonic mean of F1 score (precision and recall) improved by 0.23, demonstrating superior performance. This precise, non-contact negative emotion early warning enables early psychological intervention, significantly reducing treatment discontinuation rates due to emotional problems and producing unexpected positive effects in clinical practice.

[0118] It should be understood that the numbers S11 to S15 and S21 to S24 are only used to distinguish and facilitate the expression of different steps, and do not necessarily constitute a restriction on the execution order between the steps.

[0119] Example 2:

[0120] This application provides an HCV risk intelligent assessment system, including: a memory and a processor;

[0121] The memory is used to store computer programs;

[0122] The processor is configured to invoke the computer program to execute the method as described in Embodiment 1.

[0123] In some embodiments, the system described above can be deployed on a central intelligent management server. The computer program can be divided into a semantic parsing module and a voiceprint emotion calculation module, which are used to implement the step of calculating the emotional state of HCV patients during outpatient follow-up in the HCV high-risk assessment process.

[0124] In some embodiments, the system further includes a physician workstation; after identifying a high risk of HCV, the semantic parsing module automatically sends a prompt message to the physician workstation.

[0125] In some embodiments, the system further includes a patient follow-up interaction terminal; when the voiceprint emotion calculation module determines that the patient has negative emotions such as depression or anxiety, it automatically generates early warning information and triggers an active intervention process through the physician workstation or the patient follow-up interaction terminal.

[0126] like Figure 3 and Figure 4As shown, in some embodiments, the HCV risk intelligent assessment system provided in this application includes a central intelligent management server 100, one or more physician workstations 200, and one or more patient follow-up interaction terminals 300 in hardware. The central intelligent management server 100, as the core of the system, deploys key algorithm modules based on Embodiment 1 of this application, including a semantic parsing module 110 and a voiceprint emotion computing module 120. It is responsible for data processing, model calculation, task scheduling, and risk warning. The physician workstation 200 is the operating interface for medical staff, typically a PC or tablet. It is used to receive screening prompts (such as HCV high-risk assessment results) and patient emotion warnings from the central intelligent management server 100, and to make subsequent health management decisions. The patient follow-up interaction terminal 300 is the device used by patients, typically an APP or dedicated mini-program on a smartphone. It is used to collect patients' voice diaries, receive medication reminders, and health education materials. The blockchain node 400 is used to store key health management data and follow-up records, ensuring the immutability and traceability of the data. This system can realize HCV health management and compliance control.

[0127] Optionally, the system may also include blockchain nodes.

[0128] Example 3:

[0129] This embodiment provides a computer-readable storage medium storing a computer program. When the computer program is run on an electronic device, it causes the electronic device to perform the method described in Embodiment 1.

[0130] Example 4:

[0131] This embodiment provides a computer program product, including a computer program that, when run on an electronic device, causes the electronic device to perform the method described in Embodiment 1.

[0132] The specific implementation of the system, electronic device, computer-readable storage medium, and computer program product provided in this application can be referred to the specific embodiments of the above methods, and will not be repeated here.

[0133] Obviously, those skilled in the art should understand that the various units or steps of this application described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device, or fabricating them separately as individual integrated circuit modules, or fabricating multiple modules or steps into a single integrated circuit module. Thus, this application is not limited to any particular combination of hardware and software.

[0134] The above description is merely a preferred embodiment of this application and is 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, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for intelligent assessment of HCV risk, characterized in that, This includes the HCV high-risk assessment steps, specifically: Obtain the electronic medical record text to be processed; A domain knowledge graph is obtained, which includes entities representing high-risk factors for hepatitis C (HCV) and entities representing negation words; wherein, the entities representing negation words are used to identify exclusion scenarios that negate the high-risk factors or are not associated with the patient subject. Based on the electronic medical record text and the domain knowledge graph, a context bias matrix is ​​dynamically generated, specifically including: Positive enhancement operation: If both patient-related words and high-risk factor entities are detected in the same semantic unit in the electronic medical record text, the attention score of the corresponding element in the context bias matrix is ​​increased. Reverse inhibition operation: If a negative word entity and a high-risk factor entity are detected to appear simultaneously in the same semantic unit in the electronic medical record text, or if a high-risk factor entity is a word that is not related to the patient subject, then the attention score of the corresponding word element in the context bias matrix is ​​reduced. If the same set of related words simultaneously meets the conditions for both positive enhancement and negative suppression, the negative suppression operation will be performed first. The context bias matrix is ​​fused with the original attention score matrix of the pre-trained language model to generate a biased attention score matrix, including: Pre-trained language model initialization and text encoding: Using the BERT model or its variant as the basic pre-trained language model, the electronic medical record text to be processed is segmented into words, special tags are added and positional encoding is performed according to the requirements of the pre-trained language model, and text embedding vectors are generated. Original attention score matrix calculation: Self-attention is calculated on the text embedding vectors using the multi-head attention mechanism of the pre-trained language model to obtain the original attention score matrix. ; Context bias matrix normalization: normalization of dynamically generated context bias matrices Min-Max normalization is performed to map the element values ​​to the [0,1] interval, resulting in the normalized bias matrix. ; Weighted fusion generates a biased attention score matrix: An adaptive weighted fusion strategy is adopted, and the fusion formula is as follows: ,in This is the attention score matrix after biasing. To integrate the weighting coefficients, To be The matrix expanded by the number of multi-head attention heads has dimensions equal to... Maintain consistency; Each submatrix corresponding to each attention head undergoes Softmax normalization to ensure that the sum of the attention weights for each word is 1, resulting in the final biased attention score matrix. ; The electronic medical record text is input into a pre-trained language model. The pre-trained language model performs attention calculation based on the biased attention score matrix to achieve semantic parsing and outputs the HCV high-risk assessment result.

2. The method according to claim 1, characterized in that, The domain knowledge graph also includes time-limited word entities, which are used to define the high-risk time range of HCV high-risk factors; If both patient-related words and high-risk factor entities are detected within the same semantic unit in the electronic medical record text, the attention score of the corresponding element in the context bias matrix is ​​increased, including: For cases containing only patient-related terms and high-risk factor entities, a basic weighting increase is applied. For cases where patient-related words, high-risk factor entities, and time-limited words appear simultaneously within the same semantic unit in electronic medical record text, a higher weighting is applied to reflect the priority of high-risk time associations.

3. The method according to claim 1, characterized in that, The process involves inputting the electronic medical record text into a pre-trained language model. The pre-trained language model performs attention calculations based on the biased attention score matrix to achieve semantic parsing and outputs HCV high-risk assessment results, including: Pre-trained language models are based on The context-aware feature vector is obtained by weighted summation of the text embedding vectors of the electronic medical record text, using the following formula: ； in For text embedding vectors, Context-aware feature vectors; Will The input is the subsequent fully connected layer and classifier. The confidence level of each HCV high-risk factor is output through the Sigmoid activation function. If the confidence level of a certain HCV high-risk factor is greater than the preset threshold, the corresponding sample is determined to have the high-risk factor. All high-risk factor determination results are integrated to output the HCV high-risk risk assessment result.

4. An intelligent HCV risk assessment system, characterized in that, include: Memory and processor; The memory is used to store computer programs; The processor is configured to invoke the computer program to perform the method as described in any one of claims 1 to 3.

5. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed on an electronic device, causes the electronic device to perform the method as described in any one of claims 1 to 3.

6. A computer program product, comprising a computer program, characterized in that, When the computer program is run on an electronic device, it causes the electronic device to perform the method as described in any one of claims 1 to 3.

Images