An alzheimer disease early screening method based on AI companion voice interaction

Abstract

The application discloses an early screening method for Alzheimer's disease based on voice interaction of AI accompaniment, and relates to the technical field of computer-aided diagnosis. The method comprises the following steps: S1, end-side perception and preliminary processing; S2, cloud core analysis; S3, multi-device physiological data fusion; and S4, service arrangement and feedback. The application realizes early screening and early warning with no perception and high precision.

Description

Technical Field

[0001] This invention relates to the field of computer-aided diagnostic technology, and more specifically to an AI-assisted voice-interactive method for early screening of Alzheimer's disease. Background Technology

[0002] Current early screening technologies for Alzheimer's disease (AD) are mainly divided into three categories: 1. Biomarker detection: These rely on serum, cerebrospinal fluid, and other samples to detect relevant indicators. They require specialized medical equipment and invasive sampling, making routine home monitoring impossible. Furthermore, they can only reflect physiological abnormalities and are difficult to capture dynamic changes in cognitive function. 2. Traditional cognitive assessment: represented by MoCA and MMSE scales, these rely on manual questioning by medical staff, are highly subjective, are easily affected by the test subject's emotions and immediate state, and are conducted in a serious setting, which may deviate from daily cognitive performance, making it difficult to detect early mild cognitive decline. 3. Intelligent interactive screening: This type is divided into text and image task type and single voice type. Text and image task type has an operational threshold and insufficient data authenticity; single voice type only analyzes voice and text features, does not consider the heterogeneous features of early AD cognition, lacks integration with daily behavior data, and has limited screening accuracy.

[0003] Existing technologies generally have three major limitations: First, they lack the ability to dynamically quantify cognitive asymmetry degradation, with most solutions focusing only on a single cognitive dimension; second, the interaction scenarios are poorly adapted to the behavioral habits of the elderly, and text-based operations pose a usage barrier for older users with low educational backgrounds; and third, the integration of multi-dimensional data is low, and the collaborative analysis of voice cognitive data and physiological activity data has not been achieved, limiting the accuracy of screening.

[0004] Therefore, proposing an AI-based voice interaction-based early screening method for Alzheimer's disease to address the difficulties of existing technologies is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of this, the present invention provides an AI-assisted voice interaction-based early screening method for Alzheimer's disease, which solves the technical problems of difficulty in capturing cognitive degeneration signals, poor scene adaptability, low data integration and insufficient screening accuracy in existing early screening methods for Alzheimer's disease. It provides a seamless, highly adaptable, and multi-dimensional fusion-based AI-assisted voice interaction-based early screening method, which can accurately quantify asymmetric cognitive degeneration, and at the same time construct a closed loop of screening-intervention-monitoring process to adapt to the usage needs of different elderly user groups.

[0006] To achieve the above objectives, the present invention provides the following technical solution: An AI-assisted voice-interactive method for early screening of Alzheimer's disease includes the following steps: S1. End-side perception and preliminary processing: The user's original voice stream is collected through the terminal device, and the voice signal is converted into text data using the ASR automatic speech recognition engine. Simultaneously, local feature extraction is performed on the terminal to capture acoustic features. The original voice data is deleted within 3 seconds after feature extraction is completed. The terminal only uploads the desensitized text data and quantized acoustic feature data to the cloud. S2, Cloud Core Analysis: After receiving data uploaded by the terminal, the cloud server filters out invalid noise and communication interference through the feature cleaning unit, and then performs vector construction to generate semantic vector, temporal vector and behavioral deviation vector. The three are then concatenated into a comprehensive feature vector. The user-specific baseline file is loaded, and the comprehensive feature vector is substituted into the cognitive deviation formula to calculate the cognitive deviation. The trend slope in the sliding window is updated in real time to dynamically determine the cognitive risk level. S3. Multi-device physiological data fusion: The cognitive deviation obtained in S2 is fused with the physiological data indicators collected from multiple devices. The weights of the voice interaction feature indicators and the physiological data indicators are adjusted through a dynamic weight allocation algorithm. Combined with the cognitive deviation evolution curve analysis, a comprehensive cognitive deviation is obtained. S4. Service Orchestration and Feedback: Calculate the comprehensive risk score based on the comprehensive cognitive deviation and trend slope, match the three-level early warning mechanism to execute the corresponding early warning and intervention strategies, and realize the closed loop of the entire process of screening-intervention-monitoring.

[0007] Optionally, the terminal device in S1 is a smart speaker or a mobile APP. The acoustic features include speech rate, pause time and pitch jitter. The anonymized text data and the quantized acoustic feature data are transmitted to the cloud in JSON format.

[0008] Optionally, the comprehensive feature vector in S2 can be expressed as follows:

[0009] in, For semantic vectors, For time series vectors, This is the behavior deviation vector. txt Features of the voice interaction side.

[0010] Optional, semantic vector The expression is: ,in, The text content for the current round. This is a pre-trained sentence vector encoder; Pool is the pooling operation. Time vector It consists of response latency, number of edits / corrections, and round interval; Behavior Deviation Vector The score is composed of entity referential consistency score, inconsistency rate, and repeated question rate. It is scored by the Large Language Model (LLM) of LangChain through the tool call and normalized to [0,1].

[0011] Optionally, the formula for cognitive deviation in S2 is:

[0012] in, for Always Cognitive deviation score under the task; For the current moment, the first Class feature vectors; The mean vector of the user's historical baseline; The covariance matrix is ​​the user's historical baseline, used to measure the range of fluctuations in the user's daily behavior; These are the weighting coefficients for semantic, temporal, and behavioral features. T Task type; L For lifestyle categories; P For users' professional-related tasks.

[0013] Optionally, the specific content of the physiological data fusion from multiple devices in S3 includes: fusion of heart rate and step count from smartwatches, medication execution data from smart pillboxes, and activity duration data from smart desk lamps.

[0014] Optionally, the specific details of the triggering conditions and execution strategies for the three-level early warning mechanism in S4 are as follows: Level 1 warning: If the cognitive deviation deviates from the baseline by 10%-30% and does not show a continuous upward trend, the system will increase the frequency of proactive dialogue and collect more interaction samples in a non-intrusive manner. Level 2 warning: If the cognitive deviation exceeds 30% and shows a monotonous upward trend within the sliding window, the system will push an abnormal report to the family's APP and activate the cognitive reinforcement training mode. Level 3 warning: If the comprehensive risk score exceeds 0.8, or if severe disorientation is detected in the user, the system generates a preliminary screening report that meets clinical standards, recommends medical intervention, and sends the data to the community doctor interface.

[0015] Optionally, S1-S4 employ a privacy protection mode that combines local feature extraction with cloud encryption.

[0016] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a voice interaction-based early screening method for Alzheimer's disease with AI companionship, the beneficial effects of which are: 1) Scene adaptability: Seamlessly integrates into daily interactions, improves data authenticity by 30% compared to traditional solutions, and is suitable for the elderly and those with low educational backgrounds; 2) Screening accuracy: Multi-dimensional data fusion improves the early identification accuracy of AD by ≥85%, which is 23% higher than the single voice screening solution, and the consistency with offline scales is ≥85%; 3) Closed-loop process: Integrating screening, intervention, and monitoring to provide users with early intervention time and slow down the cognitive decline process; 4) Privacy and security: Complies with medical data security standards; original voice recordings are deleted within 3 seconds; data transmission and storage are encrypted throughout the entire process. 5) Social and economic effects: No professional medical equipment or medical personnel are required for operation. It can be widely used in community-based home care, elderly care institutions, primary medical institutions and other scenarios. It is suitable for large-scale early screening of Alzheimer's disease and cognitive health management, which can alleviate the problem of uneven distribution of medical resources and reduce the social burden of elderly care and medical care. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention 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 embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0018] Figure 1 A flowchart of an AI-based voice-interactive early screening method for Alzheimer's disease, provided by this invention; Figure 2 The present invention provides an overall flowchart of a voice-interactive Alzheimer's disease early screening method based on AI companionship. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] To address the shortcomings of existing early screening technologies for Alzheimer's disease, this invention features targeted improvements and innovations. Specifically, compared to biomarker detection methods that rely on serum sample colorimetric assays and ultraviolet light detection devices (such as CN114624202A), this invention employs a non-invasive voice interaction method, avoiding invasive sampling, supporting routine monitoring in home environments, and requiring no specialized equipment. Compared to task-based screening schemes that use text-based task interaction and vector matrix analysis (such as CN120766941A), this invention achieves full voice interaction, lowering the operational threshold for elderly users, and improving the applicability and synergy of the solution by integrating physiological data from multiple devices. Compared to traditional scale assessments that rely on human question-and-answer and subjective ratings, this invention achieves objective quantitative analysis and time-series dynamic tracking of cognitive states through non-intrusive data collection, providing a more realistic reflection of daily cognitive performance.

[0021] See Figure 1 As shown, this invention discloses an AI-assisted voice interaction-based early screening method for Alzheimer's disease, comprising the following steps: S1. End-side perception and preliminary processing: The user's original voice stream is collected through the terminal device, and the voice signal is converted into text data using the ASR automatic speech recognition engine. Simultaneously, local feature extraction is performed on the terminal to capture acoustic features. The original voice data is deleted within 3 seconds after feature extraction is completed. The terminal only uploads the desensitized text data and quantized acoustic feature data to the cloud. Specifically, the system engages in daily voice interaction with users through AI-assisted conversations via terminal devices. In natural conversations, a "probe embedding" strategy is employed, seamlessly embedding cognitive assessment probe questions as topic extensions into the conversation in no more than 30% of the total dialogue rounds. The topic scheduling module dynamically determines the embedding position and type of the probes based on the cognitive dimension coverage gaps. During the conversation, the terminal device simultaneously collects the user's original voice stream, converts the voice signal into text data using an ASR (Automatic Speech Recognition) engine, and performs local feature extraction on the terminal to capture acoustic features. The original voice data is deleted within 3 seconds of feature extraction completion, and the terminal only uploads the anonymized text data and quantified acoustic feature data to the cloud. S2, Cloud Core Analysis: After receiving data uploaded by the terminal, the cloud server filters out invalid noise and communication interference through the feature cleaning unit, and then performs vector construction to generate semantic vector, temporal vector and behavioral deviation vector. The three are then concatenated into a comprehensive feature vector. The user-specific baseline file is loaded, and the comprehensive feature vector is substituted into the cognitive deviation formula to calculate the cognitive deviation. The trend slope in the sliding window is updated in real time to dynamically determine the cognitive risk level. S3. Multi-device physiological data fusion: The cognitive deviation obtained in S2 is fused with the physiological data indicators collected from multiple devices. The weights of the voice interaction feature indicators and the physiological data indicators are adjusted through a dynamic weight allocation algorithm. Combined with the cognitive deviation evolution curve analysis, a comprehensive cognitive deviation is obtained. S4. Service Orchestration and Feedback: Calculate the comprehensive risk score based on the comprehensive cognitive deviation and trend slope, match the three-level early warning mechanism to execute the corresponding early warning and intervention strategies, and realize the closed loop of the entire process of screening-intervention-monitoring.

[0022] Furthermore, the terminal device in S1 is a smart speaker or a mobile APP. The acoustic features include speech rate, pause time and pitch jitter. The anonymized text data and the quantized acoustic feature data are transmitted to the cloud in JSON format.

[0023] Furthermore, the comprehensive feature vector in S2 is expressed as follows:

[0024] in, For semantic vectors, For time series vectors, This is the behavior deviation vector. txt Features of the voice interaction side.

[0025] Furthermore, semantic vectors The expression is: ,in, The text content for the current round. This is a pre-trained sentence vector encoder; Pool is the pooling operation. Time vector It consists of response latency, number of edits / corrections, and round interval; Behavior Deviation Vector The score is composed of entity referential consistency score, inconsistency rate, and repeated question rate. It is scored by the Large Language Model (LLM) of LangChain through the tool call and normalized to [0,1].

[0026] Furthermore, the formula for cognitive deviation in S2 is:

[0027] in, for Always Cognitive deviation score under the task; For the current moment, the first Class feature vectors; This is the mean vector of the user's historical baseline, which represents the normal state of the user learned by the system during the baseline building period (such as the first 4 months). The covariance matrix is ​​the user's historical baseline, used to measure the range of fluctuations in the user's daily behavior; These are the weighting coefficients for the three types of features: semantic, temporal, and behavioral. For example, the temporal weight can be set higher because a slower speech rate is often the earliest signal. T Task type; L For lifestyle categories; P For users' professional-related tasks.

[0028] Furthermore, the specific content of the S3's integration of physiological data from multiple devices includes: integrating heart rate and step count from smartwatches, medication execution data from smart pillboxes, and activity duration data from smart desk lamps.

[0029] Furthermore, the formula for calculating the comprehensive risk score in S4 is as follows:

[0030] in, ∈[0,1] represents the comprehensive risk score at time t; The normalized comprehensive cognitive deviation is the result after integrating the deviation of voice interaction and the deviation of physiological data, and is normalized to [0,1]. This represents the trend slope (linear regression slope) of cognitive deviation within the sliding window. The sigmoid function maps the slope to [0,1]. A positive slope and a larger value indicate a more pronounced cognitive decline trend. The multi-device physiological abnormality index is obtained by weighted normalization of the deviations of each physiological indicator from the baseline. ∈[0,1]; , , Let be the weighting coefficient, satisfying + + =1, which can be adjusted based on clinical validation (e.g.) =0.5, =0.3, =0.2).

[0031] Among them, normalized comprehensive cognitive deviation The calculation formula is:

[0032] in, Weighting coefficients for life tasks and professional tasks ( ), for The value after min-max normalization for The value after min-max normalization.

[0033] Trend slope The calculation formula is:

[0034] in, The size of the sliding window (e.g., 5 scrolls). The time interval between adjacent rounds. For the first ti The normalized comprehensive cognitive deviation of the wheel For the first t - i -1 round of normalized comprehensive cognitive deviation.

[0035] Furthermore, the specific details of the triggering conditions and execution strategies for the three-level early warning mechanism in S4 are as follows: Level 1 warning: If the cognitive deviation deviates from the baseline by 10%-30% and does not show a continuous upward trend, the system will increase the frequency of proactive dialogue and collect more interaction samples in a non-intrusive manner. Level 2 warning: If the cognitive deviation exceeds 30% and shows a monotonous upward trend within the sliding window, the system will push an abnormal report to the family's APP and activate the cognitive reinforcement training mode. Level 3 warning: If the comprehensive risk score exceeds 0.8, or if severe disorientation is detected in the user, the system generates a preliminary screening report that meets clinical standards, recommends medical intervention, and sends the data to the community doctor interface.

[0036] Specifically, Level 1 warning (low risk / observation period): Triggering condition: Cognitive deviation Slightly above the baseline (e.g., deviating by 10%-30%), but without forming a continuous upward trend.

[0037] System Behavior: Enhanced data handling without intrusion. The system automatically increases the frequency of proactive dialogue (e.g., proactively asking "What did you eat for lunch today?") to collect more samples for secondary confirmation, without disturbing family members for the time being.

[0038] Level 2 Warning (Medium Risk / Intervention Expectation): Triggering conditions: The deviation increases significantly (e.g., >30%) and shows a monotonically increasing trend within the sliding window (e.g., 5 rounds).

[0039] Systemic behavior: Proactive intervention in sync with family members.

[0040] Send an anomaly report (including specific evidence, such as "increased rate of repeated questions") to the family members' APP.

[0041] Activate the "Cognitive Reinforcement Training" mode and add specific memory-awakening dialogues (such as playing old operas and asking for the title of the opera).

[0042] Level 3 Warning (High Risk / Medical Treatment Period): Triggering condition: Comprehensive risk score Exceeding a high threshold (e.g., >0.8), or detecting severe loss of orientation (e.g., inability to identify time / location).

[0043] System behavior: Medical intervention recommendations, generating preliminary screening reports that meet clinical standards (connected to MoCA scale dimensions), directly recommending medical treatment, and sending data to community doctors via interface.

[0044] Furthermore, S1-S4 employ a privacy protection mode that combines local feature extraction with cloud encryption.

[0045] See Figure 2 As shown, in this embodiment, the system adopts a layered architecture of edge perception-cloud analysis-service closed loop. The first layer is the edge perception and preliminary processing layer, deployed on the user's smart speaker or mobile app. Its core function is to collect the user's raw speech stream in real time and convert the speech signal into text data using the built-in ASR (Automatic Speech Recognition) engine. During this process, the terminal processor simultaneously performs local feature extraction, specifically capturing acoustic features that cannot be obtained from text alone, including non-linguistic indicators such as speech rate, pause time, and pitch jitter. To strictly protect user privacy, this system implements a "calculate and destroy" strategy. After feature extraction is completed, the raw speech data must be completely deleted within 3 seconds. The terminal only uploads anonymized text data and quantized acoustic feature data (transmitted in JSON format) to the cloud, ensuring no leakage of biometric data.

[0046] Secondly, there is the Cloud Analysis Layer, which is the heart of the algorithm in this invention. After receiving the data uploaded by the terminal, the cloud server first filters out invalid noise and communication interference through the feature cleaning unit. Then, it enters the crucial Vector Construction stage: the system calls the pre-trained SimCSE model to encode the text data and generate semantic vectors. Simultaneously, a time-series vector is generated by combining the acoustic quantization data uploaded from the edge. Furthermore, it utilizes a large language model (LLM) to perform in-depth analysis of the logical coherence of the dialogue content, generating behavioral deviation vectors. After vector construction is complete, the risk assessment unit loads the user's proprietary baseline file (containing historical averages). Covariance Matrix Substituting the above comprehensive feature vectors into the cognitive deviation formula The system performs calculations and updates the trend slope within the sliding window in real time to dynamically determine the current level of cognitive risk.

[0047] Finally, the Service Layer is responsible for executing specific interaction strategies based on the analysis results. The early warning strategy unit matches the calculated risk score with the preset three-level early warning rules. When the system determines that the user is at level two risk (medium risk), the intervention push module sends an abnormality report to the family's APP, while dynamically adjusting the prompt word strategy for the next round of dialogue. Without interrupting the user's normal communication, it naturally guides the user to perform cognitive reinforcement training such as memory recall or logical repetition. When the user is determined to be at level three risk, the system generates a preliminary screening report that meets clinical standards and recommends medical intervention. The dynamic adjustment of the above dialogue strategies is achieved through a large language model orchestration framework (such as LangChain), thereby completing an automated closed loop from screening to intervention.

[0048] In another specific embodiment: Equipment deployment and baseline building: The user is a 65-year-old retired male with a high school education. He routinely uses a smart speaker for weather inquiries and plays traditional opera, and engages in casual text conversations with the AI ​​via a family app. System deployment: The smart speaker is equipped with an ARM Cortex-A53 edge SoC (end-to-end inference latency ≤500ms), linked to a smartwatch and smart lock, and connected to a community healthcare platform. The backend uses LangChain as the dialogue orchestration framework, responsible for dialogue flow scheduling, probe question embedding, and structured extraction of dialogue metadata and behavioral tags. Baseline construction period: Semantic, temporal, and behavioral feature distributions were statistically analyzed for both daily life and professional tasks. K-means was used to cluster historical conversations to obtain individualized baseline intervals and thresholds (equivalent to setting...). , (and risk threshold). Baseline characteristics: 3-5 proactive conversations per day, ≤1 repeated question per week, speaking speed of 120-140 words per minute, ≤2 pauses per day (each pause ≤3 seconds).

[0049] Screening process and anomaly identification: In the second month after baseline, the system simultaneously detected anomalies in three dimensions of evidence: speech, text, and behavioral physiology. On the speech side, orientation-related questions increased to 5-6 times per day, with repeated questions like "What time is it now / What day of the week is it today" within 10 minutes, and lexical generalization (e.g., referring to "Peking Opera" as "that old opera"), with lexical diversity TTR dropping to 0.45 (baseline 0.65). On the text side, response latency increased to 7-9 seconds, with repeated messages like "How's the weather tomorrow?" sent 3-4 times per day, and contextual reference backtracking failures (e.g., asking "Where did I say I was going?" after mentioning "going to the community to buy groceries"). The backend extracted information from each round of dialogue. (Semantic / temporal / behavioral) deviations were calculated, with a textual cognitive deviation of 52%. Behavioral and physiological aspects: door lock opening frequency decreased to 1-2 times / day, step count decreased to 3000-4000 (60% deviation from baseline), and heart rate decreased to 80-90 beats / minute with increased fluctuations. The system used a dual-task asymmetric quantification model to calculate the deviation from daily life tasks. ≈60%, deviation from professional tasks ≈12% (the recognition rate of opera terminology remained basically unchanged); the trend was monotonously increasing within 5 sliding windows, with a change rate of 12%. ≈28%, overall risk score It was determined to be of medium risk and met the heterogeneous characteristic of "life declines first while professional skills are maintained".

[0050] Result verification: The system triggered a medium-risk warning and pushed medical advice and interpretable evidence (repeated questioning rate, response delay, rhythm deviation, and activity decline) to the family's APP. Clinically, the MoCA scale score was 21 (range of mild cognitive impairment), and combined with the medical history, a diagnosis of mild cognitive impairment (MCI) was made, consistent with the clinical characteristics of the prodromal stage of Alzheimer's disease (AD), which is consistent with the system screening conclusion, verifying the consistency between the screening results and clinical assessment in a home setting.

[0051] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0052] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A voice-interactive method for early screening of Alzheimer's disease based on AI companionship, characterized in that, Includes the following steps: S1. End-side perception and preliminary processing: The user's original voice stream is collected through the terminal device, and the voice signal is converted into text data using the ASR automatic speech recognition engine. Simultaneously, local feature extraction is performed on the terminal to capture acoustic features. The original voice data is deleted within 3 seconds after feature extraction is completed. The terminal only uploads the desensitized text data and quantized acoustic feature data to the cloud. S2, Cloud Core Analysis: After receiving data uploaded by the terminal, the cloud server filters out invalid noise and communication interference through the feature cleaning unit, and then performs vector construction to generate semantic vector, temporal vector and behavioral deviation vector. The three are then concatenated into a comprehensive feature vector. The user-specific baseline file is loaded, and the comprehensive feature vector is substituted into the cognitive deviation formula to calculate the cognitive deviation. The trend slope in the sliding window is updated in real time to dynamically determine the cognitive risk level. S3. Multi-device physiological data fusion: The cognitive deviation obtained in S2 is fused with the physiological data indicators collected from multiple devices. The weights of the voice interaction feature indicators and the physiological data indicators are adjusted through a dynamic weight allocation algorithm. Combined with the cognitive deviation evolution curve analysis, a comprehensive cognitive deviation is obtained. S4. Service Orchestration and Feedback: Calculate the comprehensive risk score based on the comprehensive cognitive deviation and trend slope, match the three-level early warning mechanism to execute the corresponding early warning and intervention strategies, and realize the closed loop of the entire process of screening-intervention-monitoring.

2. The method for early screening of Alzheimer's disease based on AI companionship and voice interaction as described in claim 1, characterized in that, In S1, the terminal device is a smart speaker or a mobile APP. The acoustic features include speech rate, pause time and pitch jitter. The anonymized text data and the quantized acoustic feature data are transmitted to the cloud in JSON format.

3. The method for early screening of Alzheimer's disease based on AI-assisted voice interaction as described in claim 1, characterized in that, The comprehensive feature vector in S2 is expressed as follows: in, For semantic vectors, For time series vectors, This is the behavior deviation vector. txt Features of the voice interaction side.

4. The method for early screening of Alzheimer's disease based on AI companionship and voice interaction according to claim 3, characterized in that, semantic vectors The expression is: ,in, The text content for the current round. This is a pre-trained sentence vector encoder; Pool is the pooling operation. Time vector It consists of response latency, number of edits / corrections, and round interval; Behavior Deviation Vector The score is composed of entity referential consistency score, inconsistency rate, and repeated question rate. It is scored by the Large Language Model (LLM) of LangChain through the tool call and normalized to [0,1].

5. The method for early screening of Alzheimer's disease based on AI-assisted voice interaction as described in claim 1, characterized in that, The formula for cognitive deviation in S2 is: in, for Always Cognitive deviation score under the task; For the current moment, the first Class feature vectors; The mean vector of the user's historical baseline; The covariance matrix is ​​the user's historical baseline, used to measure the range of fluctuations in the user's daily behavior; These are the weighting coefficients for semantic, temporal, and behavioral features. T Task type; L For lifestyle categories; P For users' professional-related tasks.

6. The method for early screening of Alzheimer's disease based on AI-assisted voice interaction as described in claim 1, characterized in that, The specific content of the S3's integration of physiological data from multiple devices includes: the integration of heart rate and step count from a smartwatch, medication administration data from a smart pillbox, and activity duration data from a smart desk lamp.

7. The method for early screening of Alzheimer's disease based on AI-assisted voice interaction as described in claim 1, characterized in that, The specific details of the triggering conditions and execution strategies for the three-level early warning mechanism in S4 are as follows: Level 1 warning: If the cognitive deviation deviates from the baseline by 10%-30% and does not show a continuous upward trend, the system will increase the frequency of proactive dialogue and collect more interaction samples in a non-intrusive manner. Level 2 warning: If the cognitive deviation exceeds 30% and shows a monotonous upward trend within the sliding window, the system will push an abnormal report to the family's APP and activate the cognitive reinforcement training mode. Level 3 warning: If the comprehensive risk score exceeds 0.8, or if severe disorientation is detected in the user, the system generates a preliminary screening report that meets clinical standards, recommends medical intervention, and sends the data to the community doctor interface.

8. The method for early screening of Alzheimer's disease based on AI-assisted voice interaction according to claim 1, characterized in that, S1-S4 employ a privacy protection mode that combines local feature extraction with cloud encryption.

Images