Cognitive disease intelligent risk management method based on multi-mode voiceprint data analysis

Abstract

The invention relates to the technical field of cognitive disease risk assessment, and discloses a cognitive disease intelligent risk management method based on multi-modal voiceprint data analysis, the method comprises multi-modal feature fusion, dynamic risk assessment and real-time intervention linkage, by setting a multi-modal feature fusion module, when cognitive disease risk assessment is carried out, a multi-modal voiceprint data analysis module is set, the risk of the cognitive disease is evaluated, and the risk of the cognitive disease is evaluated. The recognition specificity of different cognitive impairment types is guaranteed by fusing acoustic characteristics, rhythm rules and multi-dimensional voiceprint characteristics of semantic logic, meanwhile, a comprehensive feature vector is constructed based on depth feature extraction, neural function degeneration signals implied in voiceprint data can be captured in real time, the misjudgment risk caused by traditional single-mode analysis is reduced, and the recognition accuracy of the cognitive impairment types is improved. According to the method, the reliability of early cognitive impairment screening is improved, a dynamic risk assessment mechanism is constructed, a cognitive state score and an adaptive threshold are compared in real time during individualized cognitive tracking, and abnormal fluctuation of a cognitive function is recognized in time.

Description

Technical Field

[0001] The present invention relates to the technical field of cognitive disease risk assessment, and specifically to an intelligent risk management method for cognitive diseases based on multimodal voiceprint data analysis. Background Art

[0002] Risk assessment refers to the work of quantitatively evaluating the impact and possibility of losses caused by a risk event to people's lives, lives and property before and after the occurrence of the risk event.

[0003] The intelligent risk management method for cognitive diseases based on multimodal voiceprint data analysis is a complex system that integrates multiple advanced technologies. It aims to improve the prediction efficiency and accuracy in the process of cognitive disease risk identification and intervention. As the core module of the system, the multimodal voiceprint analysis engine can integrate voice, voiceprint features and other physiological data to ensure that subtle cognitive decline signals can be accurately captured during the risk monitoring process. In order to enhance the robustness and reliability of the analysis, the system is equipped with an adaptive data fusion algorithm to optimize the overall effect of intelligent risk management for cognitive diseases.

[0004] At present, since cognitive disease risk assessment involves multiple physiological and behavioral dimensions, when conducting real-time monitoring of cognitive impairment, the voiceprint sensor captured by the user's voice data cannot identify in real time whether there is environmental noise interference at the analysis location. When the collected data contains background noise changes, it may cause a large deviation in feature extraction and the accuracy of risk prediction cannot be guaranteed. At the same time, when conducting voiceprint data analysis, it is impossible to dynamically detect whether the analysis angle is abnormal, which may cause a deviation in cognitive feature assessment, and the risk signal cannot be adjusted in real time when it is abnormal. When monitoring cognitive disease risks, due to the high variability and individual differences of user voice patterns, hierarchical regional processing cannot be achieved when performing multimodal data fusion, resulting in possible pattern matching errors in the risk assessment process, further affecting the accuracy of cognitive disease risk management and the effectiveness of early intervention. Summary of the Invention

[0005] (1) Technical problems solved In response to the shortcomings of the existing technology, the present invention provides an intelligent risk management method for cognitive diseases based on multimodal voiceprint data analysis, which solves the problems raised in the above background technology.

[0006] (2) Technical solution To achieve the above objectives, the present invention provides the following technical solutions: a method for intelligent risk management of cognitive diseases based on multimodal voiceprint data analysis, comprising the following steps: S1. Collection and processing: Real-time acquisition of user voice interaction data through terminal devices, simultaneous recording of environmental parameters, extraction of acoustic spectrum features, prosodic temporal features, and semantic association features, and construction of the original voiceprint dataset; S2. Feature standardization: Perform noise reduction segmentation and time-frequency conversion on the original data to generate a voiceprint feature matrix with unified dimensions; S3, multimodal fusion: Fusion of acoustic, prosodic, and semantic features to generate a 128-dimensional deep feature vector; S4. Cognitive assessment: Analyze the correlation between feature vectors and cognitive impairment through the attention LSTM model and output a cognitive status score; S5. Dynamic warning: When the score falls below the preset threshold, a risk warning signal is triggered; S6. Risk profiling: Generate personalized cognitive disease risk reports based on health records; S7, Intelligent Intervention: Matches dynamic intervention plans based on risk reports and pushes them to the user end; S8, Closed-loop optimization: Iteratively update the evaluation model using feedback data; S9. Data storage: Store the analysis results and intervention records in the blockchain database; S10. Decision support: Output risk trend maps and intervention effect reports through a visualization platform.

[0007] Preferably, the S1 includes: S11. Use portable terminal devices to collect user daily conversations, reading tasks, and free narrative voice data, and simultaneously record the ambient noise decibel level and speaking time; S12, extracting MFCC coefficients, fundamental frequency trajectory and formant energy of acoustic mode, speech rate fluctuation, pause frequency and intonation entropy of prosodic mode, and sentiment polarity word frequency and syntactic complexity index of semantic mode; S13. Align the multimodal features by timestamps to construct a time-series-associated original voiceprint feature dataset.

[0008] Preferably, the S2 includes: S21, using wavelet transform to perform noise reduction on the speech data, and segmenting the effective speech segments using endpoint detection algorithm; S22. Convert the segmented speech segments into Mel-spectrograms and generate a standardized voiceprint feature matrix with uniform dimension through Z-score normalization.

[0009] Preferably, the S3 includes: S31. Construct a multimodal feature fusion module: The acoustic submodule uses 1D-CNN to extract local spectral features, the prosodic submodule uses Bi-GRU to capture temporal dependencies, and the semantic submodule generates contextual embedding vectors based on the BERT model. S32: The output features of each sub-module are concatenated and compressed through a fully connected layer to generate a 128-dimensional multi-dimensional voiceprint feature vector.

[0010] Preferably, the S4 includes: S41, the cognitive state assessment model includes a three-layer attention LSTM network, and the input layer receives the multi-dimensional voiceprint feature vector; S42, the hidden layer calculates the attention weight formula as follows: , ; in is the attention weight at time step t, is the attention energy value at time step t, is an exponential function, is the time step index, is the total time step of the sequence, is the hidden state at time t, is the memory unit of the previous moment, is a trainable parameter, Hyperbolic tangent activation function; S43. The output layer maps the attention weighted features to the cognitive score interval [0-100] to generate cognitive status score data.

[0011] Preferably, the S6 includes: S61. Integrate user age, medication records, and genetic testing data to construct a health profile vector; S62. Calculate using the risk profile generation algorithm: Risk Index ; in is the cognitive score deviation, is the difference in medication compliance, is the environmental pressure coefficient, is the dynamic weight parameter, is a logarithmic function; S63. Generate a cognitive disease risk report based on risk index grading, including high-risk labels, degeneration rate predictions, and vulnerable brain areas.

[0012] Preferably, the S7 includes: S71. Build a dynamic intervention program library, including cognitive training tasks, medication adjustment suggestions, and social activity programs; S72. Use the multi-armed bandit algorithm to match the optimal strategy: ; in is the historical efficiency of plan a, is the total number of executions, is the number of executions of plan a, To explore factors, For intervention programs, is the natural logarithm function, To select the solution that maximizes the expression; S73. Push personalized intervention instructions to the user terminal through an encrypted channel.

[0013] Preferably, the S8 includes: S81. Collect post-intervention voiceprint data and cognitive scale results as feedback data; S82. Update model parameters through incremental learning: ; in is the cross entropy loss function, is the learning rate, is the regularization coefficient, 、 are model parameters, The loss function parameter The gradient, is the real cognitive state label, is the cognitive state predicted by the model, For parameters The L2 norm of S83. Generate model optimization reports every quarter, including the improvement in area under the ROC curve and changes in the confusion matrix.

[0014] Preferably, between S1 and S2, the following is also included: S1a, data enhancement processing: Perform sample balancing on the original voiceprint dataset and use a generative adversarial network to generate simulated voiceprint data for specific cognitive impairment types; The training samples are expanded through time domain masking and frequency domain noise addition technology to generate enhanced voiceprint datasets; S1b, Clinical Correlation Verification: The enhanced voiceprint data was correlated with the MMSE cognitive scale results, and a subset of voiceprint features with a Pearson correlation coefficient greater than 0.85 was selected; The original voiceprint dataset is updated based on the screening results.

[0015] Preferably, the S7 further includes: S7a, Dynamic Interaction Optimization: Collect voiceprint feedback data in real time when the user executes the intervention plan, and calculate task response delay and speech coherence index; When response latency exceeds 30% of baseline values and the coherence index decreases, a three-level intervention strategy is triggered: Level 1 adjustment: reduce the complexity of cognitive training tasks; Secondary adjustment: insert relaxation guidance voice clips; Level 3 adjustment: Activate emergency medical assistance protocol; S7b, Gamification Incentives: Virtual reward points are generated based on changes in voiceprint characteristics. The points redemption rules are as follows: ; in is the weekly change in cognitive score, is the absolute value of the change in cognitive score, The rule for calculating points when the cognitive score increases is: This is the rule for calculating points when the cognitive score drops. Points are used to unlock higher-level training levels.

[0016] (3) Beneficial effects Compared with the existing technology, the present invention provides an intelligent risk management method for cognitive diseases based on multimodal voiceprint data analysis, which has the following beneficial effects: 1. In the present invention, by setting up a multimodal feature fusion module, when conducting cognitive disease risk assessment, the multi-dimensional voiceprint features of acoustic characteristics, rhythmic laws and semantic logic are integrated to ensure the identification specificity of different types of cognitive impairment. At the same time, based on deep feature extraction, a comprehensive feature vector is constructed, which can capture the neural function degeneration signals implicit in the voiceprint data in real time, reduce the risk of misjudgment caused by traditional single-modal analysis, and improve the reliability of early cognitive impairment screening.

[0017] 2. In the present invention, by constructing a dynamic risk assessment mechanism, when conducting individualized cognitive tracking, the cognitive status score and the adaptive threshold are compared in real time to promptly identify abnormal fluctuations in cognitive function, so that the system can trigger a graded warning signal at the early stage of cognitive degradation. When it is found that the assessment indicators deviate from the normal trajectory, it can automatically generate a personalized risk profile by combining environmental parameters and health records, ensuring that the risk assessment results evolve dynamically with the user's physiological state, reducing the lag effect of traditional static assessment.

[0018] 3. In the present invention, by establishing a closed-loop intervention system, when executing cognitive disease risk management, the optimal intervention plan is screened in the strategy library through an intelligent matching algorithm, and the intervention intensity is dynamically adjusted based on real-time voiceprint feedback data. When it is detected that the user's execution effect deviates from expectations, it can automatically switch to a multi-level intervention strategy to ensure that the intervention measures are always synchronized with changes in cognitive status, thereby improving the initiative and sustainability of health management and reducing the risk of cognitive disease deterioration.

[0019] 4. In this invention, by integrating blockchain evidence storage technology, when managing cognitive health data, credible evidence of assessment results and intervention records is achieved through distributed storage, preventing key medical data from being tampered with and lost. At the same time, smart contracts are used to automatically verify data integrity, ensuring traceability of the entire process from data collection to medical decision-making, and providing an irrefutable original chain of evidence for long-term follow-up research on cognitive diseases. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a schematic diagram of the overall system architecture of the present invention. DETAILED DESCRIPTION

[0021] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0022] See also Figure 1 The intelligent risk management method for cognitive diseases based on multimodal voiceprint data analysis includes the following steps: S1. Collection and processing: Real-time acquisition of user voice interaction data through terminal devices, simultaneous recording of environmental parameters, extraction of acoustic spectrum features, prosodic temporal features, and semantic association features, and construction of the original voiceprint dataset; S2. Feature standardization: Perform noise reduction segmentation and time-frequency conversion on the original data to generate a voiceprint feature matrix with unified dimensions; S3, multimodal fusion: Fusion of acoustic, prosodic, and semantic features to generate a 128-dimensional deep feature vector; S4. Cognitive assessment: Analyze the correlation between feature vectors and cognitive impairment through the attention LSTM model and output a cognitive status score; S5. Dynamic warning: When the score falls below the preset threshold, a risk warning signal is triggered; S6. Risk profiling: Generate personalized cognitive disease risk reports based on health records; S7, Intelligent Intervention: Matches dynamic intervention plans based on risk reports and pushes them to the user end; S8, Closed-loop optimization: Iteratively update the evaluation model using feedback data; S9. Data storage: Store the analysis results and intervention records in the blockchain database; S10, Decision support: Output risk trend maps and intervention effect reports through the visualization platform; S1 includes: S11. Use portable terminal devices to collect user daily conversations, reading tasks, and free narrative voice data, and simultaneously record the ambient noise decibel level and speaking time; S12, extracting MFCC coefficients, fundamental frequency trajectory and formant energy of acoustic mode, speech rate fluctuation, pause frequency and intonation entropy of prosodic mode, and sentiment polarity word frequency and syntactic complexity index of semantic mode; S13, aligning the multimodal features by timestamp to construct a time-series-associated original voiceprint feature dataset; S2 includes: S21, using wavelet transform to perform noise reduction on the speech data, and segmenting the effective speech segments using endpoint detection algorithm; S22, converting the segmented speech segments into mel-spectrograms, and generating a standardized voiceprint feature matrix with uniform dimensions through Z-score normalization; S3 includes: S31. Construct a multimodal feature fusion module: The acoustic submodule uses 1D-CNN to extract local spectral features, the prosodic submodule uses Bi-GRU to capture temporal dependencies, and the semantic submodule generates contextual embedding vectors based on the BERT model. S32, concatenating the output features of each submodule and compressing them through a fully connected layer to generate a 128-dimensional multi-dimensional voiceprint feature vector; S4 includes: S41, the cognitive state assessment model includes a three-layer attention LSTM network, and the input layer receives the multi-dimensional voiceprint feature vector; S42, the hidden layer calculates the attention weight formula as follows: , ; in is the attention weight at time step t, is the attention energy value at time step t, is an exponential function, is the time step index, is the total time step of the sequence, is the hidden state at time t, is the memory unit of the previous moment, is a trainable parameter, Hyperbolic tangent activation function; S43, the output layer maps the attention weighted features to the cognitive score interval [0-100] to generate cognitive status score data; S6 includes: S61. Integrate user age, medication records, and genetic testing data to construct a health profile vector; S62. Calculate using the risk profile generation algorithm: Risk Index ; in is the cognitive score deviation, is the difference in medication compliance, is the environmental pressure coefficient, is the dynamic weight parameter, is a logarithmic function; S63. Generate a cognitive disease risk report based on risk index grading, including high-risk labels, degeneration rate predictions, and vulnerable brain areas; S7 includes: S71. Build a dynamic intervention program library, including cognitive training tasks, medication adjustment suggestions, and social activity programs; S72. Use the multi-armed bandit algorithm to match the optimal strategy: ; in is the historical efficiency of plan a, is the total number of executions, is the number of executions of plan a, To explore factors, For intervention programs, is the natural logarithm function, To select the solution that maximizes the expression; S73. Pushing personalized intervention instructions to the user terminal through an encrypted channel; The S8 includes: S81. Collect post-intervention voiceprint data and cognitive scale results as feedback data; S82. Update model parameters through incremental learning: ; in is the cross entropy loss function, is the learning rate, is the regularization coefficient, 、 are model parameters, The loss function parameter The gradient, is the real cognitive state label, is the cognitive state predicted by the model, For parameters The L2 norm of S83. Generate model optimization reports every quarter, including the improvement in area under the ROC curve and changes in confusion matrix; Between S1 and S2 also include: S1a, data enhancement processing: Perform sample balancing on the original voiceprint dataset and use a generative adversarial network to generate simulated voiceprint data for specific cognitive impairment types; The training samples are expanded through time domain masking and frequency domain noise addition technology to generate enhanced voiceprint datasets; S1b, Clinical Correlation Verification: The enhanced voiceprint data was correlated with the MMSE cognitive scale results, and a subset of voiceprint features with a Pearson correlation coefficient greater than 0.85 was selected; Updating the original voiceprint dataset based on the screening result; The S7 also includes: S7a, Dynamic Interaction Optimization: Collect voiceprint feedback data in real time when the user executes the intervention plan, and calculate task response delay and speech coherence index; When response latency exceeds 30% of baseline values and the coherence index decreases, a three-level intervention strategy is triggered: Level 1 adjustment: reduce the complexity of cognitive training tasks; Secondary adjustment: insert relaxation guidance voice clips; Level 3 adjustment: Activate emergency medical assistance protocol; S7b, Gamification Incentives: Virtual reward points are generated based on changes in voiceprint characteristics. The points redemption rules are as follows: ; in is the weekly change in cognitive score, is the absolute value of the change in cognitive score, The rule for calculating points when the cognitive score increases is: This is the rule for calculating points when the cognitive score drops. Points are used to unlock higher-level training levels.

[0023] Example 1: Multimodal feature acquisition and fusion When elderly users wear smart glasses for daily conversations at home, the system collects continuous voice streams through bone conduction microphones. The front-end processor separates ambient noise from effective human voices in real time and uses spectral subtraction to eliminate interference from the TV sound source in the background. The feature extraction stage simultaneously performs three types of analysis: the acoustic channel calculates the time-varying trajectory of the Mel-frequency cepstral coefficients to capture the characteristics of muscle decline of the vocal organs; the rhythmic channel monitors the abnormal extension of the pause interval between sentences to identify the decline of the prefrontal executive function; the semantic channel analyzes the frequency of word repetitions through a lightweight language model to detect hippocampal memory encoding disorders. After time alignment, the multimodal features are input into the feature fusion module for cross-modal association learning. When it is detected that "the high-frequency attenuation of the acoustic spectrum exceeds the critical value and the semantic repetition rate exceeds the threshold", a fusion feature vector is automatically generated and a clinical annotation label is added.

[0024] Example 2: Dynamic risk assessment and hierarchical warning After a patient with mild cognitive impairment used the system for four consecutive weeks, the risk assessment module showed a typical pathological evolution trajectory. In the initial stage, the system captured three instances of chronological confusion when the patient described his breakfast, but the acoustic characteristics remained stable, and a yellow observation-level warning was generated. In the third week, uncontrolled fundamental frequency jitters continued to appear in the nighttime voice data. Combined with the medication records, it showed that the patient had missed taking cholinesterase inhibitors for two consecutive days. The model calculated that the risk probability of hippocampal degeneration had risen to the dangerous threshold. The system immediately executed a third-order response, activated a red warning signal and pushed it to the children's mobile phones, automatically retrieved the patient's brain MRI data from last year, superimposed the current voiceprint degeneration characteristics, and generated a customized risk report marking the frontal and temporal lobe areas as priority intervention targets. The report was then transmitted to the attending physician's workstation via the medical network.

[0025] Example 3: Closed-loop intervention strategy execution and optimization Cognitive enhancement intervention is initiated for the above-mentioned high-risk patients. The strategy library first matches the dual-task training plan, requiring patients to repeat the number sequence while stepping to train execution functions. Mobile sensors monitor the execution status in real time. When the acoustic analysis finds that the amplitude of voice tremor exceeds 40% of the baseline value, the system immediately inserts mindfulness breathing guidance audio to implement first-level adjustment. After the semantic analysis module detects that the number error rate drops below the threshold, it automatically unlocks the second-level complex task. The intervention adopts a gamification mechanism throughout. Patients who successfully maintain the standard speaking speed for three minutes will receive a virtual medal. Accumulated points can be exchanged for the length of video consultation. After each round of training, the system dynamically adjusts the difficulty parameters of the next round of training and updates the benefit matrix by comparing the changes in the voiceprint spectrum energy distribution before and after the intervention.

[0026] Example 4: System deployment and clinical verification In the complete system deployed in the neurology department of a tertiary hospital, edge computing terminals are embedded in hearing aid devices to achieve low-power real-time processing of voice data. The core engine is deployed in a medical private cloud, and a containerized architecture is used to run multimodal analysis models. Blockchain dual-channel verification is enabled in key medical data storage links, and voiceprint feature hash values are stored in a local security chip. Risk reports and intervention plans are uploaded to the chain after asymmetric encryption. Clinical verification data show that in the continuous monitoring of 126 patients with early Alzheimer's disease: the system warning time point is an average of 11.3 weeks earlier than the traditional scale examination, dynamic intervention slows the rate of cognitive decline, and blockchain evidence successfully intercepts two illegal data tampering behaviors. The family-side application of dementia patients supports voice command operation, and the complex function interface is simplified to complete within three steps. The operation success rate for users over 75 years old exceeds 98%.

[0027] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or device comprising the element.

[0028] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent risk management method for cognitive diseases based on multimodal voiceprint data analysis, characterized by: The following steps are involved: S1. Collection and processing: Real-time acquisition of user voice interaction data through terminal devices, simultaneous recording of environmental parameters, extraction of acoustic spectrum features, prosodic temporal features, and semantic association features, and construction of the original voiceprint dataset; S2. Feature standardization: Perform noise reduction segmentation and time-frequency conversion on the original data to generate a voiceprint feature matrix with unified dimensions; S3, multimodal fusion: Fusion of acoustic, prosodic, and semantic features to generate a 128-dimensional deep feature vector; S4. Cognitive assessment: Analyze the correlation between feature vectors and cognitive impairment through the attention LSTM model and output a cognitive status score; S5. Dynamic warning: When the score falls below the preset threshold, a risk warning signal is triggered; S6. Risk profiling: Generate personalized cognitive disease risk reports based on health records; S7, Intelligent Intervention: Matches dynamic intervention plans based on risk reports and pushes them to the user end; S8, Closed-loop optimization: Iteratively update the evaluation model using feedback data; S9. Data storage: Store the analysis results and intervention records in the blockchain database; S10. Decision support: Output risk trend maps and intervention effect reports through a visualization platform.

2. The intelligent risk management method for cognitive diseases based on multimodal voiceprint data analysis according to claim 1, characterized in that: Said S1 comprises: S11. Use portable terminal devices to collect user daily conversations, reading tasks, and free narrative voice data, and simultaneously record the ambient noise decibel level and speaking duration; S12, extracting MFCC coefficients, fundamental frequency trajectory and formant energy of acoustic mode, speech rate fluctuation, pause frequency and intonation entropy of prosodic mode, and sentiment polarity word frequency and syntactic complexity index of semantic mode; S13. Align the multimodal features by timestamps to construct a time-series-associated original voiceprint feature dataset.

3. The intelligent risk management method for cognitive diseases based on multimodal voiceprint data analysis according to claim 1 is characterized by: The S2 includes: S21, using wavelet transform to perform noise reduction on the speech data, and segmenting the valid speech segments using endpoint detection algorithm; S22. Convert the segmented speech segments into Mel-spectrograms and generate a standardized voiceprint feature matrix with uniform dimension through Z-score normalization.

4. The intelligent risk management method for cognitive diseases based on multimodal voiceprint data analysis according to claim 1, characterized in that: The S3 includes: S31. Construct a multimodal feature fusion module: The acoustic submodule uses 1D-CNN to extract local spectral features, the prosodic submodule uses Bi-GRU to capture temporal dependencies, and the semantic submodule generates contextual embedding vectors based on the BERT model. S32: The output features of each sub-module are concatenated and compressed through a fully connected layer to generate a 128-dimensional multi-dimensional voiceprint feature vector.

5. The method for intelligent risk management of cognitive diseases based on multimodal voiceprint data analysis according to claim 1, characterized in that: The S4 includes: S41, the cognitive state assessment model includes a three-layer attention LSTM network, and the input layer receives the multi-dimensional voiceprint feature vector; S42, the hidden layer calculates the attention weight formula as follows: ， ； in is the attention weight at time step t, is the attention energy value at time step t, is an exponential function, is the time step index, is the total time step of the sequence, is the hidden state at time t, is the memory unit of the previous moment, is a trainable parameter, Hyperbolic tangent activation function; S43. The output layer maps the attention weighted features to the cognitive score interval [0-100] to generate cognitive status score data.

6. The intelligent risk management method for cognitive diseases based on multimodal voiceprint data analysis according to claim 1, characterized in that: The S6 includes: S61. Integrate user age, medication records and genetic testing data to construct a health record vector; S62. Calculate using the risk profile generation algorithm: Risk Index ; in is the cognitive score deviation, is the difference in medication compliance, is the environmental pressure coefficient, is the dynamic weight parameter, is a logarithmic function; S63. Generate a cognitive disease risk report based on risk index grading, including high-risk labels, degeneration rate predictions, and vulnerable brain areas.

7. The intelligent risk management method for cognitive diseases based on multimodal voiceprint data analysis according to claim 1, characterized in that: The S7 includes: S71. Build a dynamic intervention program library, including cognitive training tasks, medication adjustment suggestions, and social activity programs; S72. Use the multi-armed bandit algorithm to match the optimal strategy: ； in is the historical efficiency of plan a, is the total number of executions, is the number of executions of plan a, To explore factors, For intervention programs, is the natural logarithm function, To select the solution that maximizes the expression; S73. Push personalized intervention instructions to the user terminal through an encrypted channel.

8. The intelligent risk management method for cognitive diseases based on multimodal voiceprint data analysis according to claim 1 is characterized by: The S8 includes: S81. Collect post-intervention voiceprint data and cognitive scale results as feedback data; S82. Update model parameters through incremental learning: ； in is the cross entropy loss function, is the learning rate, is the regularization coefficient, 、 are model parameters, The loss function parameter The gradient, is the real cognitive state label, is the cognitive state predicted by the model, For parameters The L2 norm of S83. Generate model optimization reports every quarter, including the improvement in the area under the ROC curve and the changes in the confusion matrix.

9. The intelligent risk management method for cognitive diseases based on multimodal voiceprint data analysis according to claim 1, characterized in that: Also included between S1 and S2: S1a, data enhancement processing: Perform sample balancing on the original voiceprint dataset and use a generative adversarial network to generate simulated voiceprint data for specific cognitive impairment types; The training samples are expanded through time domain masking and frequency domain noise addition technology to generate enhanced voiceprint datasets; S1b, Clinical Correlation Verification: The enhanced voiceprint data was correlated with the MMSE cognitive scale results, and a subset of voiceprint features with a Pearson correlation coefficient greater than 0.85 was selected; The original voiceprint dataset is updated based on the screening results.

10. The intelligent risk management method for cognitive diseases based on multimodal voiceprint data analysis according to claim 1, characterized in that: The S7 further includes: S7a, Dynamic Interaction Optimization: Collect voiceprint feedback data in real time when the user executes the intervention plan, and calculate task response delay and speech coherence index; When response latency exceeds 30% of baseline values and the coherence index decreases, a three-level intervention strategy is triggered: Level 1 adjustment: reduce the complexity of cognitive training tasks; Secondary adjustment: insert relaxation guidance voice clips; Level 3 adjustment: Activate emergency medical assistance protocol; S7b, Gamification Incentives: Virtual reward points are generated based on changes in voiceprint characteristics. The points redemption rules are as follows: ； in is the weekly change in cognitive score, is the absolute value of the change in cognitive score, The rule for calculating points when the cognitive score increases is: This is the rule for calculating points when the cognitive score drops. Points are used to unlock higher-level training levels.

Images