Multi-mode recognition and grading intervention system for cognitive impairment of old people
By using a multimodal cognitive impairment identification and grading intervention system for the elderly, and combining speech and scale features for cross-modal in-depth analysis, the system solves the problems of dependence on traditional methods and low accuracy, and achieves efficient and interpretable cognitive impairment assessment and personalized intervention, thereby improving the intelligence and accuracy of the assessment.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-24
- Publication Date
- 2026-03-24
AI Technical Summary
Existing methods for screening cognitive impairment rely on manual assessment, are affected by the assessor's experience and the elderly's education level, and are difficult to achieve large-scale, standardized screening. Traditional machine learning methods cannot capture higher cognitive function problems, and general large models have low diagnostic accuracy in the medical field and cannot provide interpretable diagnostic evidence.
The system employs a multimodal cognitive impairment identification and grading intervention system for the elderly. Through modules for voice acquisition, preprocessing, voice coding, semantic extraction, and scale retrieval, it combines a large language model to conduct cross-modal in-depth analysis, integrates voice, semantic, and scale features, outputs evidence-based cognitive impairment grading and early warning results, and provides personalized intervention plans.
It improves the intelligence and interpretability of cognitive impairment assessment, enhances the accuracy and reliability of assessment, realizes standardized assessment and personalized intervention, solves the dependence and accuracy problems of traditional methods, and ensures data privacy and security.
Smart Images
Figure CN121725829A_ABST
Abstract
Description
Technical Field
[0001] This invention specifically relates to a multimodal cognitive impairment identification and graded intervention system for the elderly. Background Technology
[0002] Cognition is the intelligent processing of knowledge acquisition by an organism, involving a series of voluntary, psychological, and social behaviors such as learning, memory, language, thinking, mental state, and emotion. Cognitive impairment refers to abnormalities in the higher-level intelligent processing of the brain related to learning, memory, and judgment, resulting in severe learning and memory disorders, accompanied by pathological changes such as aphasia, apraxia, agnosia, or apraxia.
[0003] Conventional screening methods primarily involve face-to-face interviews with clinicians or psychologists, combined with structured paper-and-pencil scales. Commonly used scales include the Mini-Mental State Examination (MMSE) and the Montreal Cognitive Assessment (MoCA). In addition, subjective clinical observation and patient history review are relied upon, combining scales with in-person observation to assess disease progression. However, assessment results are highly dependent on the assessor's experience and expertise, and scale results are easily influenced by the elderly person's education level and cultural background. Furthermore, in-person consultations are time-consuming, and accessibility is extremely low for elderly people with limited mobility or living in remote areas. Primary community healthcare institutions generally lack specialized neuropsychiatrists, making it difficult to conduct large-scale, standardized screenings.
[0004] With the explosive growth of medical data and the emergence of high-performance computers, some researchers have used traditional machine learning algorithms to extract features and perform classification predictions on single-modal data, such as speech, image, or behavioral data. For example, they predict the risk of cognitive impairment by analyzing acoustic features such as speech rate and pauses, or by analyzing the results of drawing tests using image processing techniques. However, traditional machine learning methods focus on extracting shallow or pre-defined features and cannot capture the complex and deep-seated higher cognitive function problems caused by cognitive impairment, such as language and semantic regression and logical thinking disorders. At the same time, these systems usually only provide "yes" or "no" classification results and cannot provide interpretable diagnostic evidence that matches clinical scales and guidelines, making them difficult for clinicians to trust and adopt. Later, the introduction of the Transformer architecture sparked a wave of general large language model applications. Currently, it mainly uses general text models for intelligent question answering, simulating doctor-patient dialogues and initial consultations. However, when answering professional questions in the medical field, general models often exhibit "illusions" or lag in knowledge updates, thus affecting the accuracy of diagnosis. At the same time, general large language models (LMMs) cannot understand and follow the scoring criteria, follow-up questioning logic, and clinical interpretation of complex scales such as MMSE or MoCA, and cannot achieve structured and standardized clinical assessment. Summary of the Invention
[0005] This invention provides a multimodal cognitive impairment identification and graded intervention system for the elderly to address the technical problem of inaccurate traditional early warning methods mentioned above. The specific technical solution is as follows:
[0006] A multimodal cognitive impairment identification and tiered intervention system for the elderly includes the following steps:
[0007] The voice acquisition module is used to collect the raw voice data generated by users when they are conducting cognitive assessment questions and answers;
[0008] The speech preprocessing module is used to sequentially perform noise reduction and silence removal on the original speech to obtain a clean speech sequence;
[0009] The speech encoding module is used to encode the clean speech sequence with a Wav2Vec2.0 speech encoder to obtain a high-dimensional speech feature vector aligned with the length of the text token;
[0010] The semantic extraction module is used to transcribe the clean speech sequence into text using the ASR model, and then extract the language semantic feature vector using the BERT model.
[0011] The scale retrieval module is used to input MMSE and / or MoCA scale items into the built-in RAG engine, retrieve the corresponding scoring details in real time, and generate supplementary questions that conform to the clinical questioning logic to form a scale feature vector.
[0012] The feature fusion and evaluation module is used to fuse three types of feature vectors—voice, semantic, and scale—and input them into the built-in large language model, outputting evidence-based evaluation parameters for the cognitive impairment grading and early warning function.
[0013] Furthermore, in the feature fusion and evaluation module, before fusing the three types of feature vectors—speech, semantic, and scale—the long speech sequence is first compressed and mapped to a high-dimensional hidden state sequence of the same length as the text token using the Connectivity Temporal Classification (CTC) algorithm, thereby aligning the speech and text lengths.
[0014] Furthermore, after CTC mapping, a similarity-aware token fusion algorithm is executed to calculate the correlation coefficient with other modalities using text features as anchors, and to weightedly fuse speech and scale features to obtain enhanced features after noise suppression.
[0015] Furthermore, the text features obtained by the similarity-aware token fusion algorithm are input into the reconstruction decoder, and the reconstruction error between the original text features and the fusion features is calculated by a multilayer perceptron. During the model training phase, minimization is used as an auxiliary objective to force the fusion process to retain the original semantic information and suppress the introduction of noise, thereby improving the robustness and generalization ability of subsequent cognitive impairment grading tasks.
[0016] Furthermore, the large language model in the feature fusion and evaluation module adopts a bidirectional Transformer structure. The shared parameters force the distribution of non-textual modal features to converge towards the textual modality. After implicit semantic alignment, cross-attention is used to retrieve the most relevant sentiment cues in speech and scales using text as the query vector.
[0017] Furthermore, the RAG engine in the scale retrieval module searches the knowledge base in real time for the scoring details, questioning logic, and clinical guideline chapters corresponding to the MMSE or MoCA items before each follow-up question, and inputs the search results as prompt words into the large language model to reduce model illusion and ensure the standardization of questioning.
[0018] Furthermore, the multimodal elderly cognitive impairment identification and tiered intervention system also includes:
[0019] The intervention plan output module is used to dynamically match personalized rehabilitation intervention plans from the medical knowledge base based on the evaluation parameters of the graded early warning function, and encrypt the plans and send them back to the terminal to realize a closed loop of evaluation, early warning, intervention and feedback.
[0020] Furthermore, the intervention program output module retrieves the most suitable cognitive training tasks, emotion regulation guidance, and lifestyle suggestions from the knowledge base based on the graded early warning function assessment parameters, user profiles, and evidence-based chain keywords, and pushes them to end users and caregivers in encrypted form.
[0021] Furthermore, all voice data and feature vectors are uploaded to the cloud GPU cluster via HTTPS / SSL encryption. The cloud only receives the de-identified features and does not store the original voice that can identify individuals. Intermediate data is deleted immediately after inference is completed.
[0022] Furthermore, the feature fusion and assessment module outputs tiered early warning function assessment parameters that include at least three categories: normal, mild cognitive impairment (MCI), and mild Alzheimer's disease (AD). The report automatically includes the scale item numbers, key speech indicator values, and cited clinical guidelines, forming a complete chain of evidence-based data.
[0023] The multimodal cognitive impairment identification and graded intervention system for the elderly provided by this invention improves the intelligence and interpretability of the assessment by introducing large models and retrieval enhancement generation technology to perform cross-modal in-depth analysis of speech information. Attached Figure Description
[0024] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0025] Figure 1 This is a schematic diagram of the multimodal cognitive impairment identification and graded intervention system for the elderly in this application;
[0026] Figure 2 This is a schematic diagram of feature extraction in this application;
[0027] Figure 3 This is a schematic diagram of feature alignment and enhancement in this application;
[0028] Figure 4 This is a schematic diagram illustrating the feature fusion and classification of this application. Detailed Implementation
[0029] The embodiments of this application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.
[0030] like Figure 1 The image shows a multimodal cognitive impairment identification and graded intervention system for the elderly according to this application, including: a voice acquisition module, a voice preprocessing module, a voice coding module, a semantic extraction module, a scale retrieval module, and a feature fusion and evaluation module.
[0031] The speech acquisition module collects raw speech data generated by users during cognitive assessment question-and-answer sessions via a smart terminal microphone. The speech preprocessing module performs noise reduction and silence removal on the raw speech to obtain a clean speech sequence. The speech encoding module encodes the clean speech sequence using a Wav2Vec2.0 speech encoder to obtain a high-dimensional speech feature vector aligned with the text token length. The semantic extraction module transcribes the clean speech sequence into text using an ASR model and then extracts the language semantic feature vector using a BERT model. The scale retrieval module inputs MMSE and / or MoCA scale items into the built-in RAG engine, retrieves corresponding scoring details in real time, and generates supplementary questions that conform to clinical questioning logic, forming a scale feature vector. The feature fusion and evaluation module concatenates the speech, semantic, and scale feature vectors and inputs them into a large language model, outputting evidence-based cognitive impairment grading and early warning results. This invention improves the intelligence and interpretability of the assessment by introducing a large model and a retrieval-enhanced generation system for cross-modal deep analysis of speech information. The following is a detailed description of the above modules.
[0032] The voice acquisition module collects raw voice data generated by users during cognitive assessment question-and-answer sessions via the microphone of a smart terminal. DementiaBank is a public dataset specifically for studying cognitive impairment, created by institutions such as the University of Pittsburgh. Its core data includes spoken recordings and transcribed texts of subjects, typically supplemented by cognitive assessment scale scores (such as MMSE, MoCA, etc.) and the final cognitive impairment category diagnosed by the doctor as model training labels to support research on the correlation between language features and cognitive decline. To fully acquire effective voice data from subjects, this invention simulates a doctor-patient dialogue process by incorporating cognitive assessment scale content. A pre-designed cognitive assessment question-and-answer process guides users to interact via voice through a smart terminal (such as a mobile phone / tablet) app. During this process, voice data is collected and uploaded through the smart terminal's microphone. Text input from the user is also supported.
[0033] The speech preprocessing module sequentially performs denoising and silence removal on the raw speech to obtain a clean speech sequence. Specifically, during the data preprocessing stage, all audio data undergoes standardization processes such as denoising and silence removal to ensure data quality.
[0034] The speech encoding module encodes the clean speech sequence using a Wav2Vec2.0 speech encoder to obtain a high-dimensional speech feature vector X aligned with the length of the text token. A .
[0035] like Figure 2 As shown, in the feature extraction process, this invention extracts information from three modalities, including audio, text, and scales, through interactive voice dialogue.
[0036] For the original speech signal, this invention adopts the Wav2Vec 2.0 architecture as the core speech feature encoder to extract context-dependent, high-dimensional feature vectors. These feature vectors serve as the main speech modal inputs and can simultaneously capture acoustic information (such as speech rate, pauses, and pitch).
[0037] The semantic extraction module transcribes clean speech sequences into text using an ASR model, and then extracts the language semantic feature vector X using a BERT model. T Specifically, the ASR model is used to transcribe clean speech sequences into text, and the BERT model is used to further extract high-level language features such as semantic consistency and logical coherence.
[0038] The scale retrieval module inputs MMSE and / or MoCA scale items into the built-in RAG engine, retrieves the corresponding scoring details in real time, and generates supplementary questions that conform to clinical inquiry logic, forming the scale feature vector X. M .
[0039] This invention combines a large model as an interaction port, supplements medical scale information with RAG, and combines the scale to score cognitive function in multiple dimensions, forming scale features. Specifically, before each follow-up question, the RAG engine searches the knowledge base in real time for the scoring details, follow-up question logic, and clinical guideline chapters corresponding to the MMSE or MoCA item, and inputs the search results as prompt words into the large language model to reduce model illusion and ensure the standardization of questioning.
[0040] The feature fusion and evaluation module fuses three types of feature vectors—voice, semantic, and scale—and inputs them into the built-in large language model, outputting evidence-based cognitive impairment classification and early warning results.
[0041] Understandably, in the original data, the lengths of different modal features are usually inconsistent. Text and scales are usually discrete word or token sequences, while speech is a continuous time series signal. Due to inconsistent sampling rates, the sequence length of speech is usually much longer than that of text and scale modal features. Therefore, to perform feature fusion, the extracted features need to be processed first.
[0042] like Figure 3As shown, this invention employs the Connection Temporal Classification (CTC) algorithm for feature alignment. In the feature fusion and evaluation module, before fusing the three types of feature vectors (speech, semantic, and scale), the CTC algorithm first compresses and maps the long speech sequence to a high-dimensional hidden state sequence of the same length as the text token, achieving alignment between speech and text lengths. Specifically, the original long speech sequence is input into an encoder, and the CTC algorithm calculates the probability of outputting the corresponding text token at each time step. By finding the path with the highest probability and folding duplicates and blanks, CTC effectively learns how to "compress" and map the long speech sequence to a short text sequence, thereby achieving alignment in length and time steps.
[0043] In the embodiments of this application, the feature fusion and evaluation module further executes a similarity-aware token fusion algorithm after CTC mapping. Using text features as anchors, it calculates the correlation coefficient with other modalities and weights the fusion of speech and scale features to obtain enhanced features after noise suppression. It can be noted that after alignment, although speech and text features have the same length, their inherent emotional semantics may be less clear than the text, or even contain more noise. To address this issue, a similarity-aware token fusion algorithm is used to fuse the strong semantics of the text with non-text modalities. Based on similarity calculation, the correlation between the text and other modalities can be obtained. The correlation coefficients are then used to weight the fusion of different modalities to obtain the enhanced speech feature E. A and scale characteristics E M .
[0044] In the embodiments of this application, the feature fusion and evaluation module uses the text features H obtained by the similarity-aware token fusion algorithm after feature fusion. T Input the reconstruction decoder R(·), and calculate the reconstruction error L between the original text features and the original text features using a multilayer perceptron. rec During the model training phase, minimization is used as an auxiliary objective to force the fusion process to retain the original semantic information and suppress the introduction of noise, thereby improving the robustness and generalization ability of subsequent cognitive impairment grading tasks.
[0045] The large language model in the feature fusion and evaluation module adopts a bidirectional Transformer structure. The shared parameters force the distribution of non-textual modal features to converge towards the textual modal. After implicit semantic alignment, cross-attention is used to retrieve the most relevant sentiment cues in speech and scales using text as the query vector.
[0046] Specifically, such as Figure 4 The audio features E after initial fusion are described above. A Text features HT Scale characteristics E M The audio feature C is further updated using a bidirectional Transfoemer structure. A Text features C T Scale characteristics C M Each modality undergoes in-depth feature mining to learn contextual information over a long time span.
[0047] By sharing parameters, the feature distribution of non-textual modalities is forced to converge towards that of textual modalities, achieving implicit semantic alignment. The superior context processing capabilities learned by the textual modality can be "transmitted" to speech and vision through shared parameters, ultimately resulting in three modal features that have learned from each other. Then, audio and scale features are concatenated to form a "database," while text features act as the "questioner," actively retrieving and extracting the most relevant emotional cues from non-textual modalities using textual information, completing the final fusion. Text features search through long sequences of speech and scales, calculating attention weights. High weights indicate that the current speech tone is particularly relevant to the current text or scale information. Based on these weights, weighted information is extracted from the audio and scale feature databases, forming a traceable chain of evidence.
[0048] After the aforementioned cross-attention processing, the fused features are input into the pooling layer and the fully connected layer, and then further fed into the classifier to calculate the three-class classification results, including normal, MCI, and AD, as well as an error L. task During training, this error parameter needs to be minimized to achieve the best model performance. After the results are calculated, they are sent back to the large model's interactive interface, and a feedback report with reasons is generated based on the aforementioned evidence chain. Specifically, the graded warning results output by the feature fusion and evaluation module include at least three categories: normal, mild cognitive impairment (MCI), and mild Alzheimer's disease (AD). The report automatically marks the item number of the scale on which it is based, the value of key speech indicators, and the source of the cited clinical guidelines, forming a complete evidence chain.
[0049] The multimodal cognitive impairment identification and graded intervention system for the elderly in this application also includes: an intervention program output module.
[0050] The intervention plan output module is used to dynamically match personalized rehabilitation intervention plans from the medical knowledge base based on the evaluation parameters of the graded early warning function, and encrypt the plans and send them back to the terminal to realize a closed loop of evaluation, early warning, intervention and feedback.
[0051] This invention adopts an edge-cloud collaborative deployment architecture to achieve resource optimization and privacy and security protection. The core large language model and multimodal features are jointly encoded and deployed on a cloud GPU server cluster for professional inference, while the front-end mobile application is only responsible for lightweight tasks such as data collection, preprocessing and result visualization, thus solving the problem of insufficient computing power of mobile devices.
[0052] All data transmission processes employ HTTPS / SSL encryption protocols combined with medical data anonymization technology to ensure user privacy complies with medical data security standards. All voice data and feature vectors are uploaded to the cloud GPU cluster via HTTPS / SSL encryption. The cloud only receives the anonymized features and does not store the original voice data that can identify individuals. Intermediate data is deleted immediately after inference is completed.
[0053] The system leverages the traceability capabilities of the RAG knowledge base to automatically annotate the scale items, key speech feature indicators, and cited clinical guidelines in the evaluation results, forming a complete chain of evidence-based data and effectively solving the "black box" problem of medical AI.
[0054] The intervention program output module, based on a dynamic matching algorithm, retrieves the most suitable cognitive training tasks, emotion regulation guidance, and lifestyle recommendations from a knowledge base using tiered early warning function assessment parameters, user profiles, and evidence-based keywords. These recommendations are then encrypted and pushed to end-users and caregivers. Specifically, based on the risk level and symptom characteristics determined by the assessment, the system uses a dynamic matching algorithm to retrieve the most suitable personalized rehabilitation intervention strategies from the knowledge base, including customized cognitive training tasks, emotion regulation guidance, and lifestyle recommendations. These are then precisely pushed to users and caregivers via a mobile application, constructing a complete closed loop of "assessment-early warning-intervention-feedback," thereby improving the accuracy and effectiveness of early identification and intervention for cognitive impairment.
[0055] The system described in this invention only provides quantitative assessment parameters of cognitive function and rehabilitation suggestions, and does not directly make disease diagnosis conclusions. All output results are only for reference by qualified clinicians in combination with the patient's overall condition. The final diagnosis must be made independently by the physician in accordance with the "Regulations on the Administration of Medical Institutions" and relevant diagnosis and treatment guidelines.
[0056] This application's system creatively applies large language models (LLMs) with retrieval-enhanced generation techniques and joint speech feature encoding techniques to the identification and intervention of cognitive impairment in the elderly. Specifically, it achieves highly human-like simulated doctor-patient dialogue in the core interaction, leveraging the powerful generation capabilities of LLMs and the voice interaction format to greatly improve assessment efficiency and user experience, addressing the temporal, spatial, and mobility barriers in elderly assessments. Simultaneously, the RAG mechanism enables real-time retrieval of clinical scales and guidelines, ensuring the clinical standardization of intelligent questioning and clarification during dialogues. Secondly, this invention innovatively employs the Wav2Vec2.0 joint encoder to transform the raw speech signal into high-dimensional speech-semantic joint features, which are directly input into the LLM, achieving deep fusion of speech features. This allows the model to capture subtle degradation features in the non-textual dimension, such as speech rate, pauses, and pitch, that traditional scales cannot detect, greatly improving the model's multimodal fusion capability and the accuracy and robustness of recognition. Furthermore, the assessment results possess extremely high clinical interpretability and evidence-based professionalism. This is attributed to the deep integration of the RAG mechanism with the scoring details of scales such as MMSE / MoCA and authoritative treatment guidelines. The system's output reports clearly indicate the scale items, voice indicators, and cited clinical guideline chapters used in the judgments, overcoming the shortcomings of traditional AI black-box models. Ultimately, this method can dynamically generate and push personalized and dynamically adjusted precision intervention plans based on deep analysis results and the RAG knowledge base, supporting tiered and graded clinical intervention goals. Simultaneously, the system employs an edge-cloud collaborative deployment architecture and encrypted transmission technology to ensure the privacy and security of medical data, guaranteeing efficient and reliable operation in real-world scenarios.
[0057] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the above embodiments do not limit the present invention in any way, and all technical solutions obtained by equivalent substitution or equivalent transformation fall within the protection scope of the present invention.
Claims
1. A multimodal cognitive impairment identification and tiered intervention system for the elderly, characterized in that, Includes the following steps: The voice acquisition module is used to collect the raw voice data generated by users when they are conducting cognitive assessment questions and answers; The speech preprocessing module is used to sequentially perform noise reduction and silence removal on the original speech to obtain a clean speech sequence; The speech encoding module is used to encode the clean speech sequence with a Wav2Vec2.0 speech encoder to obtain a high-dimensional speech feature vector aligned with the length of the text token; The semantic extraction module is used to transcribe the clean speech sequence into text using the ASR model, and then extract the language semantic feature vector using the BERT model. The scale retrieval module is used to input MMSE and / or MoCA scale items into the built-in RAG engine, retrieve the corresponding scoring details in real time, and generate supplementary questions that conform to the clinical questioning logic to form a scale feature vector. The feature fusion and evaluation module is used to fuse three types of feature vectors—voice, semantic, and scale—and input them into the built-in large language model, outputting evidence-based evaluation parameters for the cognitive impairment grading and early warning function.
2. The multimodal cognitive impairment identification and graded intervention system for the elderly according to claim 1, characterized in that, In the feature fusion and evaluation module, before fusing the three types of feature vectors—speech, semantic, and scale—the long speech sequence is first compressed and mapped to a high-dimensional hidden state sequence of the same length as the text token using the Connection Temporal Classification (CTC) algorithm, thus achieving alignment between speech and text lengths.
3. The multimodal cognitive impairment identification and graded intervention system for the elderly according to claim 2, characterized in that, After CTC mapping, a similarity-aware token fusion algorithm is further executed to calculate the correlation coefficient with other modalities using text features as anchors, and to weight and fuse speech and scale features to obtain enhanced features after noise suppression.
4. The multimodal cognitive impairment identification and graded intervention system for the elderly according to claim 3, characterized in that, The text features obtained by the similarity-aware token fusion algorithm are input into the reconstruction decoder, and the reconstruction error between the original text features and the fusion features is calculated by a multilayer perceptron. During the model training phase, minimization is used as an auxiliary objective to force the fusion process to retain the original semantic information and suppress the introduction of noise, thereby improving the robustness and generalization ability of subsequent cognitive impairment grading tasks.
5. The multimodal cognitive impairment identification and graded intervention system for the elderly according to claim 1, characterized in that, The large language model in the feature fusion and evaluation module adopts a bidirectional Transformer structure. The shared parameters force the distribution of non-textual modal features to converge towards the textual modality. After implicit semantic alignment, cross-attention is used to retrieve the most relevant sentiment cues in speech and scales using text as the query vector.
6. The multimodal cognitive impairment identification and graded intervention system for the elderly according to claim 1, characterized in that, The RAG engine in the scale retrieval module searches the knowledge base in real time for the scoring details, questioning logic, and clinical guideline chapters of the corresponding MMSE or MoCA item before each follow-up question, and inputs the search results as prompt words into the large language model to reduce model illusion and ensure the standardization of questioning.
7. The multimodal cognitive impairment identification and graded intervention system for the elderly according to claim 1, characterized in that, The multimodal cognitive impairment identification and tiered intervention system for the elderly also includes: The intervention plan output module is used to dynamically match personalized rehabilitation intervention plans from the medical knowledge base based on the evaluation parameters of the graded early warning function, and encrypt the plans and send them back to the terminal to realize a closed loop of evaluation, early warning, intervention and feedback.
8. The multimodal cognitive impairment identification and graded intervention system for the elderly according to claim 1, characterized in that, The intervention program output module retrieves the most suitable cognitive training tasks, emotion regulation guidance, and lifestyle suggestions from the knowledge base based on the hierarchical early warning function assessment parameters, user profiles, and evidence-based chain keywords, and then pushes them to end users and caregivers in encrypted form.
9. The multimodal cognitive impairment identification and graded intervention system for the elderly according to claim 1, characterized in that, All voice data and feature vectors are uploaded to the cloud GPU cluster via HTTPS / SSL encryption. The cloud only receives the de-identified features and does not store the original voice that can identify individuals. Intermediate data is deleted immediately after inference is completed.
10. The multimodal cognitive impairment identification and graded intervention system for the elderly according to claim 1, characterized in that, The feature fusion and assessment module outputs tiered early warning function assessment parameters that include at least three categories: normal, mild cognitive impairment (MCI), and mild Alzheimer's disease (AD). The report automatically marks the scale item number, key speech index values, and cited clinical guidelines to form a complete chain of evidence.