Generative content recommendation method and device for cognitive intervention of elderly people with cognitive impairment

Abstract

This application discloses a generative content recommendation method and apparatus for cognitive intervention in elderly people with cognitive impairment. Applied to generative artificial intelligence, it can determine the current cognitive state of the target object after acquiring multimodal data of the target object performing a cognitive assessment task. When a content recommendation request is received, the method determines the target recommendation task corresponding to the target object based on the request. When the target recommendation task is an interactive recommendation scenario, it acquires the target object's historical object information and a set of cognitive intervention content. Then, based on the historical object information, the current cognitive state, and the set of cognitive intervention content, it generates the current task interaction content for the target recommendation task. When feedback content is received regarding the current task interaction content, the method recommends the target content to the target object based on the feedback content and the set of cognitive intervention content. This scheme can improve the adaptability and accuracy of the recommended content.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a generative content recommendation method and apparatus for cognitive intervention in elderly people with cognitive impairment. Background Technology

[0002] With the increasing aging of the global population and the growing awareness of individual health management, the demand for cognitive function maintenance and precise, personalized intervention for elderly people with cognitive impairment is increasing. Cognitive impairment encompasses a range of neurodegenerative diseases, from mild cognitive impairment to Alzheimer's disease, and its core characteristic is the progressive decline in cognitive functions such as memory, attention, and language.

[0003] In related approaches, interventions for individual cognitive impairment mainly rely on standardized training programs and universal content delivery mechanisms. For example, for elderly people with cognitive impairment, a pre-set question bank or fixed cognitive training tasks are usually used. Cognitive assessments are conducted through single-round question-and-answer or multiple-choice questions, and pre-set recommended content is matched based on the cognitive assessment results to enable elderly people with cognitive impairment to maintain their cognition.

[0004] In the process of researching and practicing current technologies, the inventors of this application have found that related technologies are usually based on matching general cognitive training programs with the results of a single cognitive assessment. They cannot provide effective targeted interventions for individual specific situations (such as individual cognitive decline trajectory, cognitive reserve baseline and dynamic change characteristics), resulting in insufficient adaptability and accuracy of recommended content, and making it difficult to effectively support the long-term maintenance and improvement of cognitive function in elderly people with cognitive impairment. Summary of the Invention

[0005] This application provides a generative content recommendation method and apparatus for cognitive intervention in elderly people with cognitive impairment, which can improve the adaptability and accuracy of recommended content, thereby facilitating the long-term maintenance and intervention of cognitive abilities in elderly people with cognitive impairment.

[0006] A generative content recommendation method for cognitive intervention in older adults with cognitive impairment includes: Acquire multimodal data of the target object while performing at least one cognitive assessment task, and determine the current cognitive state of the target object based on the multimodal data; When a content recommendation request for the target object is received, the target recommendation task corresponding to the target object is determined based on the content recommendation request; When the target recommendation task is a recommendation task in an interactive recommendation scenario, obtain the historical object information and cognitive intervention content set of the target object under the target recommendation task; Based on the historical object information, the current cognitive state, and the set of cognitive intervention content, the current task interaction content of the target recommendation task is generated; When feedback is received from the target object regarding the current task interaction content, target recommended content is recommended to the target object based on the feedback content and the cognitive intervention content set.

[0007] Accordingly, embodiments of this application provide a generative content recommendation device for cognitive intervention in elderly people with cognitive impairment, comprising: An acquisition unit is used to acquire multimodal data of a target object performing at least one cognitive assessment task, and to determine the current cognitive state of the target object based on the multimodal data. The receiving unit is configured to, upon receiving a content recommendation request for the target object, determine the target recommendation task corresponding to the target object based on the content recommendation request; The query unit is used to obtain the historical object information and cognitive intervention content set of the target object under the target recommendation task when the target recommendation task is a recommendation task in an interactive recommendation scenario. The generation unit is used to generate the current task interaction content of the target recommendation task based on the historical object information, the current cognitive state, and the cognitive intervention content set. The recommendation unit is used to recommend target content to the target object based on the feedback content and the cognitive intervention content set when it receives feedback content from the target object regarding the current task interaction content.

[0008] In some embodiments, the acquisition unit may be specifically used to extract features from the modal data using a content recommendation model to obtain modal features of the modality; to perform temporal alignment on the modal features and to perform attention weighting on the aligned modal features to obtain fused modal features of the multimodal data; and to perform cognitive evaluation on the target object based on the fused modal features to obtain the current cognitive state of the target object.

[0009] In some embodiments, the acquisition unit may be specifically used to: when the modal data is interactive behavior data, use the content recommendation model to extract features from the interactive behavior data to obtain object behavior features of the target object; when the modal data is speech data, use the content recommendation model to repair the speech data and extract features from the repaired speech data to obtain object speech features of the target object; when the modal data is visual data, use the content recommendation model to perform object recognition on the visual data and extract features from the recognized object information to obtain object visual features of the target object; when the modal data is physiological data, use the content recommendation model to extract features from the physiological data to obtain heart rate variability features and perform time-series modeling on the heart rate variability features to obtain object physiological features; and use any one of the object behavior features, the object speech features, the object visual features, and the object physiological features as the modal features.

[0010] In some embodiments, the acquisition unit may be specifically used to extract multi-dimensional features from the interaction behavior records, aggregate the extracted behavioral features to obtain aggregated behavioral features of the session; construct an object behavior feature sequence of the target object based on the temporal information of the session and the aggregated behavioral features; perform feature transformation on the behavioral features in the object behavior feature sequence to obtain a transformed object behavior feature sequence; perform dimensionality reduction on the transformed object behavior feature sequence, and select at least one object behavior feature from the dimensionality-reduced object behavior feature sequence.

[0011] In some embodiments, the acquisition unit may be specifically used to evaluate at least one cognitive dimension of the target object based on the fused modal features to obtain an initial evaluation result under the cognitive dimension; fuse the initial evaluation results to obtain the initial cognitive state of the target object under the cognitive evaluation task; acquire the historical static cognitive state of the target object, and fuse the initial cognitive state and the historical cognitive state to obtain the current cognitive state of the target object, wherein the historical static cognitive state is obtained by aggregating historical multimodal data from multiple historical sessions.

[0012] In some embodiments, the query unit may be specifically used to obtain the current interaction content of the target object in relation to the target recommendation task, and determine the cognitive training target of the target object in the target recommendation task based on the current interaction content and the current cognitive state; use the content recommendation model to perform thought chain reasoning on the current cognitive state, the current interaction content and the cognitive training target to obtain the initial task framework information of the target recommendation task; and based on the initial task framework information, filter out the historical object information and cognitive intervention content set of the target object under the target recommendation task from at least one knowledge base.

[0013] In some embodiments, the query unit may be specifically used to identify the content scenario of the current interactive content using the content recommendation model, and determine at least one query condition of the knowledge base based on the content scenario; infer the interaction configuration information of the target recommendation task based on the current cognitive state and the cognitive training objective; and construct the task framework of the target recommendation task based on the query conditions and the interaction configuration information to obtain the initial task framework information.

[0014] In some embodiments, the query unit may be specifically used to identify query conditions corresponding to the knowledge base in the initial task framework information, the query conditions including object information query conditions and medical content query conditions; based on the object information query conditions, filter out the object preference information of the target object and the associated content corresponding to the current interaction content in the object knowledge base to obtain the historical object information; based on the medical content query conditions, filter out at least one dimension of task constraint conditions in the medical knowledge base to obtain the cognitive intervention content set.

[0015] In some embodiments, the generation unit may be specifically used to generate initial interactive prompt information corresponding to the target recommendation task based on the initial task framework information and the current cognitive state; add the historical object information and the cognitive intervention content set to the initial interactive prompt information to obtain the interactive prompt information of the target recommendation task; and generate the current task interactive content of the target recommendation task using the content recommendation model based on the interactive prompt information.

[0016] In some embodiments, the recommendation unit may be specifically used to identify the interaction rounds of the target recommendation task in the initial task framework information; when the interaction rounds are multiple rounds, the interaction prompt information is updated based on the feedback content to recommend target recommendation content to the target object; when the interaction rounds are single rounds, the target recommendation content is generated using the content recommendation model based on the feedback content and the current task interaction content, and the target recommendation content is recommended to the target object.

[0017] In some embodiments, the recommendation unit may be specifically used to update the interaction prompt information based on the feedback content, and use the updated interaction prompt information as the interaction prompt information; return to the step of generating the current task interaction content of the target recommendation task based on the interaction prompt information using the content recommendation model, until the interaction round is reached, to obtain the target feedback content; generate the target recommendation content based on the target feedback content and the current task interaction content using the content recommendation model, and recommend the target recommendation content to the target object.

[0018] In some embodiments, the recommendation unit may be specifically configured to generate recommendation prompts for the target recommendation task based on the target feedback content and the current task interaction content; generate at least one candidate recommendation content using the content recommendation model based on preset content generation parameters and the recommendation prompts; filter the candidate recommendation content to obtain filtered content, and perform modal transformation on the filtered content to obtain at least one transformed content; fuse the filtered content and the transformed content to obtain the target recommendation content, and recommend the target recommendation content to the target object.

[0019] In some embodiments, the recommendation unit may also be used to extract object attribute information of the target object from the current task interaction content and the feedback content; perform feature encoding on the object attribute information to obtain incremental object features of the target object; incrementally update the object knowledge base according to the object attribute information and the incremental object features, and use the updated object knowledge base as the object knowledge base.

[0020] In some embodiments, the recommendation unit can also be used to, when the target recommendation task is a recommendation task in an active recommendation scenario, acquire the target object's historical physiological data within a preset time range and its medical needs information under its current cognitive state; based on the target object's current cognitive state, the historical physiological data, and the medical needs information, use the content recommendation model to generate at least one modality of medical education content corresponding to the medical needs information; use the medical education content as the target recommendation content, and recommend the target recommendation content to the target object.

[0021] Furthermore, embodiments of this application also provide an electronic device, including a processor and a memory, wherein the memory stores an application program, and the processor is used to run the application program in the memory to execute the generative content recommendation method for cognitive intervention in elderly people with cognitive impairment provided in embodiments of this application.

[0022] Furthermore, embodiments of this application also provide a computer-readable storage medium storing a plurality of instructions adapted for loading by a processor to execute steps in any of the generative content recommendation methods for cognitive intervention in elderly people with cognitive impairment provided in embodiments of this application.

[0023] Furthermore, embodiments of this application also provide a computer program product, including a computer program or instructions, which, when executed by a processor, implement the steps in the generative content recommendation method for cognitive intervention in elderly people with cognitive impairment provided in embodiments of this application.

[0024] In this embodiment, after acquiring multimodal data of the target object performing at least one cognitive assessment task, the current cognitive state of the target object can be determined based on the multimodal data. When a content recommendation request for the target object is received, the target recommendation task corresponding to the target object is determined according to the content recommendation request. When the target recommendation task is a recommendation task in an interactive recommendation scenario, the historical object information and cognitive intervention content set of the target object under the target recommendation task are acquired. Then, based on the historical object information, the current cognitive state, and the cognitive intervention content set, the current task interaction content of the target recommendation task can be generated. When feedback content from the target object regarding the current task interaction content is received, the current cognitive state and cognitive state are determined according to the feedback content and the cognitive state. The intervention content set recommends target content to the target group. Because this solution applies generative artificial intelligence, the current cognitive state determined by multimodal data can indicate the target group's real-time cognitive ability level, emotional valence, and cognitive load. This allows the generation of current task interaction content for the target recommendation task based on the current cognitive state and the target group's historical object information. This can accurately match the individual characteristics and immediate needs of the target group (such as elderly people with cognitive impairment), avoiding the serious problem of homogenized recommendation content. It ensures that the recommended content is always adapted to the target group's current cognitive state, thereby improving the adaptability and accuracy of the recommended content. This, in turn, is beneficial to the long-term maintenance and intervention of the cognitive abilities of elderly people with cognitive impairment. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 This is a schematic diagram of a scenario for a generative content recommendation method for cognitive intervention in elderly people with cognitive impairment provided in an embodiment of this application; Figure 2 This is a flowchart illustrating the generative content recommendation method for cognitive intervention in elderly people with cognitive impairment provided in this application embodiment; Figure 3 This is a schematic diagram of the recommended scenario selection interface provided in the embodiments of this application; Figure 4 This is a schematic diagram of the recommended task interface provided in an embodiment of this application; Figure 5 This is a flowchart illustrating the generative content recommendation method for cognitive intervention in elderly people with cognitive impairment provided in this application embodiment; Figure 6This is a schematic diagram of the generative content recommendation device for cognitive intervention in elderly people with cognitive impairment provided in an embodiment of this application; Figure 7 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0027] The technical solutions of the embodiments of this application 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.

[0028] This application provides a generative content recommendation method and related equipment for cognitive intervention in elderly people with cognitive impairment. The related equipment may include a generative content recommendation device for cognitive intervention in elderly people with cognitive impairment, an electronic device, a computer program product, and a computer-readable storage medium. The generative content recommendation device for cognitive intervention in elderly people with cognitive impairment can be integrated into an electronic device, which may be a server or a terminal, etc.

[0029] The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery network (CDN), and big data and artificial intelligence platforms. The terminal can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, etc., but is not limited to these. The terminal and server can be directly or indirectly connected via wired or wireless communication, which is not limited herein.

[0030] For example, see Figure 1Taking the integration of a generative content recommendation device for cognitive intervention in elderly people with cognitive impairment into an electronic device as an example, the electronic device can determine the current cognitive state of the target object based on the multimodal data obtained after the target object performs at least one cognitive assessment task. When a content recommendation request is received for the target object, the device can determine the target recommendation task corresponding to the target object based on the content recommendation request. When the target recommendation task is a recommendation task in an interactive recommendation scenario, the device can obtain the target object's historical object information and cognitive intervention content set under the target recommendation task. Then, based on the historical object information, the current cognitive state, and the cognitive intervention content set, the device can generate the current task interaction content of the target recommendation task. When the device receives feedback from the target object on the current task interaction content, it can recommend the target recommendation content to the target object based on the feedback content and the cognitive intervention content set. Therefore, the adaptability and accuracy of the recommended content can be improved.

[0031] It is understood that, in the specific implementation of this application, multimodal data of the target object, historical object information and other related data are involved. When the following embodiments of this application are applied to specific products or technologies, permission or consent is required, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0032] The following sections provide detailed descriptions of each example. It should be noted that the order in which the embodiments are described is not intended to limit the preferred order of the embodiments.

[0033] This embodiment will be described from the perspective of a generative content recommendation device for cognitive intervention in elderly people with cognitive impairment. Specifically, this generative content recommendation device for cognitive intervention in elderly people with cognitive impairment can be integrated into an electronic device, which can be a server or a terminal, etc. The terminal can include tablet computers, laptops, personal computers (PCs), wearable devices, virtual reality devices, or other smart devices that can perform generative content recommendation for cognitive intervention in elderly people with cognitive impairment.

[0034] like Figure 2 As shown, this generative content recommendation method for cognitive intervention in elderly people with cognitive impairment is applied to generative artificial intelligence. The specific process of this method is as follows: 101. Obtain multimodal data of the target object when performing at least one cognitive assessment task, and determine the current cognitive state of the target object based on the multimodal data.

[0035] The target group can be understood as the individual or target group participating in cognitive intervention, such as elderly users, especially those at high risk of cognitive decline (such as mild cognitive impairment MCI, subjective cognitive decline) or those in the cognitive health maintenance stage. Cognitive intervention can be understood as a series of intervention measures to purposefully promote, maintain or improve an individual's cognitive function through training tasks and educational activities.

[0036] Cognitive assessment tasks can be understood as lightweight interactive tasks used to quickly assess the cognitive level of a target. These tasks can take the form of short games or micro-tasks (e.g., 1-2 minutes), such as number recitation games, graphic matching games, short-term color memory games, simple instruction following games, and quick number spot-the-difference games.

[0037] The multimodal data can include modal data from at least two modalities of the conversation generated by the target object during the cognitive assessment task. A conversation can be understood as the complete interaction process generated by the target object during the completion of the cognitive assessment task, including a full-process interaction record of task start, task in progress, and task end (interruption or completion). Modal data can include any one of the following: interactive behavior data, speech data, visual data, and physiological data of the target object during the execution of at least one cognitive assessment task. Interactive behavior data can be understood as operation-related data generated by the target object during the execution of the cognitive assessment task. Interactive behavior data can include interactive behavior records of at least one conversation, such as click coordinates, operation duration, response time, operation interval, operation accuracy, correctness, error type, number of retries, etc. Speech data can be understood as audio-related data generated by the target object during the cognitive assessment process, including speech content and acoustic features, such as response speech content, speech rate, pause intervals, speech energy, fundamental frequency changes, etc. Visual data can be understood as image-related data generated by the target object during the cognitive assessment process, such as the target object's facial expressions, eye contact, body movements, and environmental images. Physiological data can be understood as physiological data generated by the target subject during the cognitive assessment process (which can be collected using wearable devices), such as heart rate, heart rate variability, respiratory rate, etc.

[0038] The current cognitive state can be understood as a set of quantitative indicators reflecting the target object's current and historical cognitive abilities, obtained through comprehensive analysis of multimodal data. This current cognitive state can include the target object's cognitive ability level (such as memory level, attention level, reaction ability level, logical ability level, executive function, etc.), emotional valence (positive / neutral / negative), cognitive load intensity (strong, medium, low), distribution of interests and hobbies, baseline ability levels for each cognitive dimension, baseline physiological indicators in daily activities, and characteristics of interactive behavioral habits.

[0039] As an example, the target's current cognitive state may include cognitive ability level: medium (score 65 / 100), emotional valence: slightly positive (+0.3), and cognitive load intensity: low (35%).

[0040] There are several ways to acquire multimodal data of the target object when performing at least one cognitive assessment task. For example, multimodal data can be collected in real time through sensor modules integrated in the terminal device (such as recording interaction behavior data through a touch screen, collecting voice data through a built-in microphone, collecting visual data through a front-facing camera, or synchronizing physiological data through a wearable device connected via Bluetooth). Alternatively, multimodal data streams from external dedicated acquisition devices can be received through a preset data interface. Or, system logs and event callback interaction behavior data during the execution of the cognitive assessment task can be used, etc.

[0041] After obtaining multimodal data on the target object performing at least one cognitive assessment task, the target object's current cognitive state can be determined based on the multimodal data. The specific steps for determining the target object's current cognitive state based on multimodal data are as follows (S11-S13): S11. Use a content recommendation model to extract features from the modal data to obtain modal features.

[0042] Content recommendation models can be understood as a large language model that takes multimodal data as input and generates recommended content for a target audience. A modality can be understood as the specific form or carrier in which data is generated and exists during the target audience's cognitive assessment task. Each modality corresponds to a data type with unique attributes, such as interactive behavior modality, voice modality, visual modality, physiological modality, etc. Modal features can be understood as semantically expressive vector representations extracted from modal data.

[0043] For example, when the modal data is interactive behavior data, the modal feature corresponding to the modal data can be the object's behavioral features; or, when the modal data is speech data, the modal feature corresponding to the modal data can be the object's speech features; or, when the modal data is visual data, the modal feature corresponding to the modal data can be the object's visual features; or, when the modal data is physiological data, the modal feature corresponding to the modal data can be the object's physiological features, and so on.

[0044] Before using the content recommendation model to extract features from the modal data, the collected modal data can be preprocessed to ensure the accuracy of subsequent feature extraction. For example, outliers in each modal data can be repaired, missing values ​​can be removed, and noise can be removed.

[0045] After preprocessing the collected modal data, a content recommendation model can be used to extract features from the modal data to obtain the modal features, as follows: (1) When the modal data is interactive behavior data, the content recommendation model is used to extract features from the interactive behavior data to obtain the object behavior features of the target object.

[0046] Among them, object behavior features can be understood as vector representations extracted from interaction behavior data that characterize the cognitive performance and operational habits of the target object.

[0047] When the modal data is interactive behavior data, the method of using a content recommendation model to extract features from the interactive behavior data to obtain the object behavior features of the target object can specifically include: extracting multi-dimensional features from the interactive behavior records, aggregating the extracted behavioral features to obtain the aggregated behavioral features of the session, constructing the object behavior feature sequence of the target object based on the temporal information of the session and the aggregated behavioral features, transforming the behavioral features in the object behavior feature sequence to obtain the transformed object behavior feature sequence, reducing the dimensionality of the transformed object behavior feature sequence, and selecting at least one object behavior feature from the dimensionality-reduced object behavior feature sequence.

[0048] A single complete session can include multiple interaction processes, each of which can have its own corresponding interaction behavior data. Interaction behavior data can be understood as the operation-related data generated by the target object when performing a cognitive assessment task. Interaction behavior data can include the interaction behavior records of at least one session.

[0049] Among them, behavioral features can be understood as vector representations extracted from the interaction behavior records of a single interaction process, based on multiple dimensions such as task performance, interactive behavior, and emotional investment. Specifically, these can include task performance features, interactive behavior features, emotional investment features, etc. Aggregated behavioral features can be understood as vector representations obtained by statistically aggregating behavioral features from multiple dimensions. Temporal information can be understood as the chronological order of the aggregated behavioral features of each interaction process within a single complete session. The object behavioral feature sequence can be understood as a sequence of aggregated behavioral features of the target object corresponding to each interaction process in that session, ordered according to temporal information. The transformed object behavioral feature sequence can be understood as a set of vectors with a unified format and consistent dimensions after normalization and encoding. The dimensionality-reduced object behavioral feature sequence can be understood as the object behavioral feature sequence after dimensionality reduction.

[0050] Optionally, in some implementations, after obtaining the interaction behavior records corresponding to each of the multiple interaction processes in a single complete session, features can be extracted from the interaction behavior records of a single interaction process from multiple dimensions such as task performance (e.g., accuracy, error type, reaction time, etc.), interaction behavior (e.g., click accuracy, interaction flow, etc.), and emotional investment (completion rate, abandonment rate and abandonment point, interaction duration and interaction pattern, content preference, etc.). This allows for the acquisition of multiple dimensions of behavioral features corresponding to each interaction process. Subsequently, for each interaction process, its multiple behavioral features can be statistically aggregated (e.g., mean, annotation difference, trend, etc.) to obtain aggregated behavioral features that comprehensively reflect the overall performance of the interaction process in the session.

[0051] Optionally, in some implementations, after obtaining the aggregated behavioral features that comprehensively reflect the overall performance of the interaction process in the session, the precise timestamps corresponding to each interaction process and their order of occurrence in the session can be extracted as time sequence information. Then, the aggregated behavioral features of all interaction processes are arranged in order according to the time sequence information corresponding to each interaction process to construct an object behavior feature sequence that can fully present the changes in the cognitive state of the target object throughout the session (the object behavior feature sequence is consistent with the actual occurrence logic of the interaction process).

[0052] Optionally, in some implementations, after obtaining the object behavior feature sequence, feature transformation can be performed on the behavior features to obtain a transformed object behavior feature sequence. Specifically, feature transformation of the behavior features may include operations such as normalization and encoding of the behavior features. For example, the behavior features may be Z-score normalized or quantile normalized to eliminate the influence of different feature dimensions, or the behavior features may be one-hot encoded to convert them into numerical features, and so on.

[0053] Optionally, in some implementations, after obtaining the transformed object behavior feature sequence, since its feature dimension may be very high, the transformed object behavior feature sequence can be dimensionality reduced to remove redundancy and noise while retaining the main information, thereby obtaining a dimensionality-reduced object behavior feature sequence. For example, principal component analysis can be used to reduce the dimensionality of the transformed object behavior feature sequence to focus on the main variation direction of the transformed object behavior feature sequence; alternatively, linear discriminant analysis can be used to reduce the dimensionality of the transformed object behavior feature sequence; or, an autoencoder-based nonlinear transformation can be used to reduce the dimensionality of the transformed object behavior feature sequence, and so on.

[0054] Optionally, in some implementations, after obtaining the dimensionality-reduced object behavior feature sequence, at least one object behavior feature can be selected from the dimensionality-reduced object behavior feature sequence. For example, selection can be based on the variance contribution rate of the principal components, retaining the top N object behavior features whose cumulative variance contribution rate reaches a preset threshold (e.g., 90%); or, importance scores of each feature can be calculated based on a tree model (e.g., random forest), selecting the top N object behavior features by score; or, mutual information can be used to select object behavior features that are strongly correlated with the cognitive state assessment target; or, domain knowledge can be combined to specifically select object behavior features that have strong representational power for cognitive functions (e.g., memory, attention) (e.g., reaction time variability, task accuracy), and so on.

[0055] (2) When the modal data is speech data, the content recommendation model is used to repair the speech data, and the repaired speech data is used to extract the object speech features of the target object.

[0056] The restored speech data can be understood as the speech signal after restoration processes such as speech activity detection, environmental noise reduction, and accent adaptation. The target speech features can be understood as vector representations extracted from the restored speech data that can characterize the speech content and cognitive state of the target object.

[0057] Optionally, in some implementations, after acquiring the speech data, the speech data can be repaired to obtain repaired speech data. For example, the speech activity detection module can distinguish speech segments from non-speech segments (such as silence or environmental noise) in the speech data, and apply a preset noise reduction algorithm (such as spectral subtraction or a deep learning-based noise reduction model) to suppress background noise and improve the signal-to-noise ratio. Alternatively, the accent deviation of the target object can be corrected based on a preset corpus adaptation module. Or, the speech rate normalization process can be appropriately performed on excessively slow or irregular speech rates to improve recognition efficiency, and so on.

[0058] Optionally, in some implementations, after obtaining the repaired speech data, feature extraction can be performed on the repaired speech data to obtain the target object's speech features. For example, the speech content in the repaired speech data can be converted into text in real time, and the semantic features of the text can be extracted to embed the target object's speech features. Alternatively, open-source toolkits (such as eGeMAPS) can be used to extract acoustic features in parallel from the repaired speech data, such as fundamental frequency, speech rate, pauses, energy, and other paralinguistic features, to embed the target object's speech features. Or, text semantic features and acoustic features can be combined to construct multi-dimensional fused target object speech features, comprehensively characterizing the target object's language expression ability, emotional valence, and cognitive state, and so on.

[0059] (3) When the modal data is visual data, the content recommendation model is used to identify objects in the visual data and extract features from the identified object information to obtain the visual features of the target object.

[0060] Among these, object information can be understood as various types of information identified from visual data, such as facial information (the number of faces appearing in the visual data, human characteristics, facial expressions, etc.), scene information (the scene in which the visual data is located, such as a family living room, park, hospital, school, etc.), and object information (specific items appearing in the visual data, such as teacups, photo albums, etc.). Object visual features can be understood as vector representations extracted from the identified object information that can characterize the semantics of visual content and the context related to the target object.

[0061] Optionally, in some implementations, after acquiring the visual data, the visual data can be preprocessed to obtain preprocessed visual data. For example, the visual data can be denoised and filtered to improve the signal-to-noise ratio; or, the visual data can be color corrected and white balance adjusted to improve color consistency; or, the visual data can be image enhanced and contrast stretched to highlight key features; or, geometric correction and distortion repair can be performed to eliminate lens distortion; or, missing area repair or invalid data filling can be performed to ensure the integrity of the visual impairment data, and so on.

[0062] Optionally, in some implementations, after obtaining the preprocessed visual data, object recognition can be performed on the visual data to obtain object information. For example, object detection technology (such as YOLO or Faster R-CNN) in a preset recognition algorithm can be used to identify object information in the visual image; or scene classification technology (such as Place365) in a preset recognition algorithm can be used to determine the scene information of the image; or face recognition technology (such as MTCNN or RetinaFace) in a preset recognition algorithm can be used to locate and identify face information in the image; or emotion recognition technology in a preset recognition algorithm can be used to identify the expression state of the detected face, and so on.

[0063] Optionally, in some implementations, after obtaining the identified object information, feature extraction can be performed to obtain the object visual features of the target object. For example, a visual encoder can be used to encode the visual data as a whole to extract the object visual features; or, the identified object information can be feature extracted separately according to its category to obtain scene features, object features, and facial emotion features, and the extracted features can be fused together to form the object visual features of the target object; or, the object information in the image and the relationships between object information can be modeled to generate the object visual features of the target object, and so on.

[0064] (4) When the modal data is physiological data, the content recommendation model is used to extract features from the physiological data to obtain heart rate variability features, and the heart rate variability features are time-series modeled to obtain the physiological features of the object.

[0065] Heart rate variability (HRV) characteristics can be understood as a set of physiological indicators calculated from physiological data to quantify minute differences in continuous heartbeat cycles (i.e., RR intervals or NN intervals). In this application, HRV characteristics may include time-domain features and frequency-domain features. Time-domain features may include, but are not limited to, mean heart rate, standard deviation of heart rate, standard deviation of NN intervals (SDNN), root mean square of the difference between adjacent NN intervals (RMSSD), and the percentage of adjacent NN interval differences greater than 50 ms (pNN50). Frequency-domain features may include, but are not limited to, total power, low-frequency power (LF), high-frequency power (HF), and the LF / HF ratio. The physiological characteristics of the target object can be understood as vector representations extracted from the HRV characteristics that characterize changes in the physiological state of the target object.

[0066] Optionally, in some implementations, after obtaining the physiological data of the target object, feature extraction can be performed on the physiological data to obtain heart rate variability features. For example, the raw physiological data can be preprocessed (e.g., abnormal heartbeat data can be removed, missing values ​​can be filled, and noise can be filtered), and then the SDNN (24-hour NN interval standard deviation) in the time domain features can be calculated to obtain heart rate variability features. Alternatively, the RR interval sequence can be analyzed in the frequency domain using Fast Fourier Transform (FFT) or Wavelet Transform to extract frequency domain features such as total power, LF (0.04-0.15Hz), and HF (0.15-0.4Hz) to obtain heart rate variability features. Or, the sliding window method can be used to calculate heart rate variability features in different time periods, and so on.

[0067] Optionally, in some implementations, after obtaining the heart rate variability (HRV) features, temporal modeling can be performed on the HRV features to obtain the physiological characteristics of the object. For example, the HRV features extracted in different time periods can be arranged in chronological order to construct a temporal feature sequence, which can be directly used as the physiological characteristics of the object; or, a recurrent neural network or a long short-term memory network can be used to perform temporal modeling on the temporal feature sequence to capture the temporal dependence and changing trend of the HRV features to obtain the physiological characteristics of the object; or, a temporal convolutional network can be used to extract local and global temporal patterns in the temporal feature sequence to generate a compact temporal feature representation to obtain the physiological characteristics of the object; or, statistics (such as mean, variance, and trend slope) of the HRV features in different time windows can be calculated and fused with the original temporal features to construct multi-scale physiological characteristics of the object, and so on.

[0068] (5) Take any one of the object’s behavioral features, speech features, visual features and physiological features as modal features.

[0069] Specifically, it can acquire multimodal data of the target object performing at least one cognitive assessment task, and extract features from each modality of the multimodal data to obtain object behavior features, object speech features, object visual features, and object physiological features, so that any one of them can be used as a modal feature.

[0070] S12. Perform temporal alignment on the modal features and apply attention weighting to the aligned modal features to obtain the fused modal features of the multimodal data.

[0071] Aligned modal features can be understood as feature representations that, after timestamp synchronization, ensure consistency across modal features over time. Fusion modal features can be understood as unified feature representations formed after attention-weighted fusion.

[0072] Optionally, in some implementations, after obtaining the modal features corresponding to each modality, dynamic time warping or interpolation algorithms based on key event points can be used to unify all modal feature sequences onto the same time base according to the timestamps of each modal data to perform temporal alignment of the modal features. Then, attention features are extracted from the aligned modal features, where attention features can indicate the correlation between modal features. Subsequently, attention weights of modal features can be determined based on the attention features, and the attention weights are normalized to obtain the fused features of the modal features.

[0073] Here, attention features can be understood as attention scores. Attention weights can be numerical parameters obtained by calculating the correlation between features of different modalities, used to quantify the contribution of different modalities to the final fused features. In this application, attention weights can be obtained through a cross-attention mechanism, for example, by using object behavior features as queries, object speech features as keys, and object visual features and object physiological features as values, and calculating the attention weights through a softmax function.

[0074] S13. Based on the fused modal features, perform cognitive assessment on the target object to obtain the target object's current cognitive state.

[0075] Specifically, the method of performing cognitive evaluation on the target object based on the fused modal features to obtain the target object's current cognitive state may include: evaluating at least one cognitive dimension of the target object based on the fused modal features to obtain an initial evaluation result for the cognitive dimension; then fusing the initial evaluation results to obtain the target object's initial cognitive state under the cognitive evaluation task; then obtaining the target object's historical static cognitive state and fusing the initial cognitive state and the historical cognitive state to obtain the target object's current cognitive state.

[0076] The cognitive dimension can include the target subject's cognitive ability level, emotional valence, and cognitive load intensity. The target subject's cognitive ability level can be understood as a comprehensive ability encompassing major cognitive domains such as memory, attention, reaction time, and logical reasoning. The target subject's emotional valence can be understood as the emotional tendency of the target subject during the cognitive assessment task, such as a positive, negative, or neutral emotional state. The target subject's cognitive load intensity can be understood as the degree of psychological resource consumption borne by the target subject when processing the cognitive assessment task. The initial assessment result under the cognitive dimensions can be understood as the quantitative score corresponding to each cognitive dimension. The initial cognitive state can be understood as the set of quantitative scores integrating the target subject's cognitive ability level, emotional valence, and cognitive load intensity, indicating the real-time assessment result of the ongoing cognitive assessment task.

[0077] Among them, historical static cognitive state can be evaluated by aggregating historical multimodal data from multiple historical conversations. Historical static cognitive state can characterize the relatively stable cognitive and behavioral characteristics of the target object over a long period of time, including but not limited to: the target object's interest distribution, baseline ability levels of each cognitive dimension (such as typical score ranges for memory, attention, and executive function), typical emotional response patterns, baseline physiological indicators in daily activities (such as resting heart rate and normal range of heart rate variability), as well as interactive behavior habits and cognitive reserve levels.

[0078] After obtaining the fused modal features, a regression sub-model of the content recommendation model can be used to decode the fused modal features to perform cognitive evaluation of the target object, obtaining initial evaluation results under the cognitive dimensions. Specifically, this can include using the cognitive ability level output layer of the regression sub-model to output the quantitative scores of each cognitive domain (such as memory, attention, reaction, and logic) of the target object's predicted cognitive ability level; using the emotional valence output layer of the regression sub-model to output the quantitative evaluation of the target object's predicted emotional valence (positive / neutral / negative) and fine-grained emotional categories (such as pleasure and anxiety); and using the cognitive load output layer of the regression sub-model to output the quantitative evaluation of the target object's predicted cognitive load intensity. Based on the quantitative scores of each output layer, initial evaluation results under each cognitive dimension can be obtained. For example, cognitive ability level: medium (score 65 / 100), emotional valence: slightly positive (+0.3), cognitive load intensity: low (35%). After obtaining the initial assessment results for each cognitive dimension, these results can be fused to obtain the target object's initial cognitive state during the cognitive assessment task. This fusion can be done in several ways. For example, the initial assessment results for each cognitive dimension can be simply concatenated into vectors to obtain the target object's initial cognitive state (e.g., moderate cognitive ability, slightly positive emotional valence, low cognitive load). Alternatively, different weights can be assigned to the assessment results of each cognitive dimension based on their importance or reliability (the weights can be set according to different task types or target object characteristics), and then a weighted average can be performed to obtain the target object's initial cognitive state. Another approach is to map the initial cognitive results for each cognitive dimension to a unified cognitive state classification label based on a preset rule mapping table or state transition model to obtain the target object's initial cognitive state, and so on.

[0079] After obtaining the initial cognitive state of the target object, the historical static cognitive state of the target object can be obtained. There are several ways to obtain the historical static cognitive state of the target object. For example, the average vector of the cognitive state of several sessions within a preset time period can be extracted from the object knowledge base of the target object to obtain the historical static cognitive state. Alternatively, the historical cognitive state sequence can be aggregated based on a time decay weighting method, so that the recent state is given a higher weight to obtain the historical static cognitive state. Or, when there is a lack of sufficient historical cognitive states (such as a new target object), a clustering algorithm can be used to find a group of objects similar to the target object, and the average cognitive state of the group can be used as the historical static cognitive state, and so on.

[0080] After obtaining the historical static cognitive state of the target object, the initial cognitive state and the historical cognitive state can be fused to obtain the target object's current cognitive state. This fusion can be done in several ways. For example, linear interpolation can be performed on the two, and the fusion weights can be dynamically adjusted based on the ratio of the duration of the current cognitive assessment task to the average duration of historical cognitive assessment tasks to obtain the target object's current cognitive state. Alternatively, weights can be assigned based on the proximity of the timestamp of the session corresponding to the initial cognitive state to the current time, with more recent sessions receiving higher weights, and then combined with a weighted average based on the historical static cognitive state, and so on.

[0081] It should be noted that the current cognitive state of the target object is a dynamic balance result that takes into account both immediate performance and long-term trends. Integrating the initial cognitive state of the target object with its historical static cognitive state not only improves the stability and continuity of the cognitive state assessment of the target object, but also provides a richer and more three-dimensional cognitive context for personalized content recommendation. This allows for a better distinction between the short-term fluctuations and long-term changes of the target object, enabling more accurate and stable recommendation decisions.

[0082] 102. When a content recommendation request for a target object is received, determine the target recommendation task corresponding to the target object based on the content recommendation request.

[0083] A content recommendation request can be understood as an instruction triggered proactively by the target object or triggered at a specific time (such as upon completion of a cognitive assessment task), requesting the generation or delivery of the next personalized interactive content for the target object. The content recommendation request carries scene selection interactive content, which can be understood as the interaction content generated between the target object and the interactive interface when selecting a recommended scene.

[0084] The recommendation scenarios can include interactive recommendation scenarios and proactive recommendation scenarios. Interactive recommendation scenarios can be understood as recommendation scenarios that require the target object to participate and interact with the terminal's interactive interface. Interactive recommendation scenarios can include at least one recommendation task, which can be a cognitive training task. A cognitive training task can be understood as a digital intervention task that uses structured and interactive interactive content to specifically train, maintain, or improve the target object's cognitive abilities (such as memory, attention, executive function, etc.). For example, recommendation tasks in interactive recommendation scenarios can include guided memory-related cognitive training tasks (such as a memory box, which focuses on training episodic memory and semantic memory), guided computation-related cognitive training tasks (such as a daily challenge, which focuses on training attention and logical reasoning), guided emotion and relaxation-related cognitive training tasks (such as a music walkthrough, which focuses on emotion regulation and psychological relaxation), and so on.

[0085] In this context, proactive recommendation scenarios can be understood as recommendation scenarios that analyze the target audience's cognitive state, behavioral habits, or health status, without requiring the target audience to participate in the interaction. Targeted recommendation tasks can be understood as recommendation tasks where the target audience makes a selection and the recommendation is ultimately presented to the target audience.

[0086] Specifically, the method of determining the target recommendation task corresponding to the target object based on the content recommendation request may include: selecting interactive content based on the scenario, filtering out the target recommendation scenario corresponding to the content recommendation request from the candidate recommendation scenarios, and then, when the target recommendation scenario includes multiple candidate recommendation tasks, receiving the task selection interactive content returned by the target object, and filtering out the target recommendation task corresponding to the target object from the candidate recommendation tasks based on the task selection interactive content, and when the target recommendation scenario includes one candidate recommendation task, using the candidate recommendation task as the target recommendation task corresponding to the target object.

[0087] The candidate recommendation scenario can include at least one of interactive recommendation scenarios and proactive recommendation scenarios. Each candidate recommendation scenario can include at least one candidate recommendation task (e.g., an interactive recommendation scenario can include multiple candidate recommendation tasks, while a proactive recommendation scenario includes only one). The target recommendation scenario can be understood as a recommendation scenario selected by the target object and ultimately presented to that object. The task selection interaction content can be understood as the feedback information when the target object selects from multiple candidate recommendation tasks. Examples include clicking on a candidate recommendation task, voice confirmation, and gesture selection.

[0088] As an example, such as Figure 4 As shown, the terminal has an interactive interface. After the target subject completes the cognitive assessment task, the interface presents a recommended task selection screen. This screen includes the target subject's daily health report (generated based on their current cognitive state), a personalized greeting, and two candidate recommendation scenarios. These scenarios can be interactive or proactive. The interactive scenario presents multiple candidate tasks, such as a memory box, a daily challenge, or a music walkthrough. The proactive scenario presents one candidate task, such as a health education class. The terminal can then receive the scenario selection interaction content returned by the target subject to determine the corresponding target recommendation scenario. If the target recommendation scenario is interactive, the target subject can further select from multiple candidate tasks within that scenario. The terminal can receive the task selection interaction content returned by the target subject and filter the candidate tasks to select the target recommendation task that best suits them. If the target recommendation scenario is proactive, then the candidate task within that proactive scenario is selected as the target recommendation task for the target subject.

[0089] 103. When the target recommendation task is a recommendation task in an interactive recommendation scenario, obtain the target object's historical object information and cognitive intervention content set in the target recommendation task.

[0090] When the target recommendation task is a recommendation task in an interactive recommendation scenario, the server needs to prepare context information for the upcoming personalized target recommendation task, which may include extracting historical object information from the object knowledge base and extracting a set of cognitive intervention content from the medical knowledge base.

[0091] The method of obtaining the target object's historical object information and cognitive intervention content set under the target recommendation task may specifically include the following steps (S31-S33): S31: Obtain the current interaction content of the target object in relation to the target recommendation task, and determine the cognitive training target of the target object in the target recommendation task based on the current interaction content and the current cognitive state.

[0092] The current interaction content can be understood as the multimodal information input by the target object in real time during the execution of the target recommendation task, such as voice descriptions, uploaded images, and text responses. The current interaction content indicates the target object's current intention and expression. The cognitive training objective can be understood as the cognitive training direction set for the current recommendation task based on the target object's current cognitive state and current interaction content. For example, the cognitive training objective could be to deepen the ability to extract details from contextual memory of a certain event or to improve the target object's language fluency.

[0093] The method for obtaining the current interaction content of the target object for the target recommendation task can specifically include: obtaining the task attribute information of the target recommendation task, identifying candidate interaction items corresponding to the target recommendation task in the task attribute information, and then, when there is only one candidate interaction item, obtaining the content of the target object's interaction under the candidate interaction item to obtain the current interaction content; and, when there are multiple candidate interaction items, determining the target interaction item selected by the target object among the candidate interaction items, and obtaining the content of the target object's interaction under the target interaction item to obtain the current interaction content. The task attribute information can be understood as a predefined data structure describing the basic characteristics and interaction requirements of the task within the target recommendation. The task attribute information can include task type (such as memory-based, calculation-based, relaxation-based, etc.) and pre-input content (such as images, speech, text, etc.). Different recommendation tasks correspond to different task types and different required pre-input content. For example, the task type of the "Memories Box" task can be memory-based, and its corresponding pre-input modality can include images.

[0094] In this context, candidate interaction items can be understood as the types of pre-recommendation content supported by the target object in the target recommendation task. For example, the candidate interaction item for the "Memories Box" task can be one (such as an image), the candidate interaction item for the "Music Tour" task can be one (such as voice), and the candidate interaction items for "Today's Challenge" can be multiple (such as images, voice, text, etc.). The target interaction item can be understood as the interaction item finally selected by the target object from multiple candidate interaction items. When there is only one candidate interaction item, that unique item automatically becomes the target interaction item. When there are multiple candidate interaction items, the final target interaction item is determined through a secondary selection by the target object.

[0095] As an example, when the target recommendation task is determined to be the "Memories Box" task, its task attribute information is parsed. That is, the task type of the "Memories Box" task can be memory-related, and its corresponding pre-input modality can include images, and the number of its candidate interaction items is one. Then, the target object can be guided to upload an image, and the image uploaded by the target object can be used as the current interaction content.

[0096] As another example, when the target recommended task is determined to be a music roaming task, its task attribute information is parsed. That is, the task type of the music roaming task can be relaxation, and its corresponding pre-input modality can include voice, and the number of its candidate interaction items is one. Then, the target object can be guided to upload voice or audio, and the voice or audio uploaded by the target object can be used as the current interaction content.

[0097] As another example, when the target recommended task is determined to be the "Today's Challenge" task, its task attribute information is parsed. That is, the task type of the "Today's Challenge" task can be a calculation type, and its corresponding pre-input modality can include images, voice, text, and the selection content of the receiving target object (the "Today's Challenge" task has pre-set interactive scene construction information). The number of its candidate interactive items can be multiple. The system will first present multiple sub-challenge options (such as Friends Card Game, Number Maze). After the target object selects one of them (such as clicking Friends Card Game), the selected sub-challenge becomes the target interactive item. Then, the target object can be guided to perform specific interactions under the target interactive item (such as uploading relevant photos or entering a friend's name), and these interactive contents will be used as the current interactive content.

[0098] It should be noted that the task attribute information not only defines the task type and the pre-input content, but also implicitly includes the goal orientation, cognitive training dimensions and scene construction rules of the recommended task. The degree of matching between the current interactive content of the target object and the task attributes will directly affect the accuracy and scene adaptability of the subsequent personalized content generation.

[0099] Optionally, in some implementations, after obtaining the target object's current interaction content for the target recommendation task, the cognitive training objective for the target object in the target recommendation task can be determined based on the current interaction content and the current cognitive state. There are various ways to determine the cognitive training objective. For example, it can be determined by matching the semantic structure of the current interaction content with the cognitive state to identify the cognitive dimensions where the target object performs weakly. Alternatively, it can be adjusted based on the target object's recent training records and adaptive rules. If the target object's current cognitive state shows that a certain cognitive dimension is under high load or fatigue, the training objective can be appropriately adjusted. Or, multiple factors such as short-term task completion, immediate experience, and long-term cognitive improvement trends can be considered simultaneously, and the optimal training objective can be iteratively selected through reinforcement learning or evolutionary algorithms, and compared with historical training effect data to ensure the scientific nature and sustainability of the objective setting, etc.

[0100] S32: Use a content recommendation model to perform thought chain reasoning on the current cognitive state, current interactive content, and cognitive training objective to obtain the initial task framework information for the objective recommendation task.

[0101] The initial task framework information can be understood as a structured instruction or prompt template that describes the context configuration, knowledge query requirements and interaction control parameters required for the execution of the target recommendation task. It defines the basic logic of task generation and execution, the required external knowledge sources, and the system roles and interaction rules.

[0102] For example, the initial task framework information can be as follows: {Model Role Setting: You are a patient and friendly cognitive training assistant, skilled at helping target individuals with cognitive training through conversation based on the following given content;} The target object's current cognitive state: [Current Cognitive State] Current interaction content for the target recommendation task: [Current interaction content]; Cognitive training objectives for the target audience: [Cognitive training objectives]; Interaction rounds for the target recommendation task: [Number of interaction rounds]; Interaction type for the target recommendation task: [Interaction type]; Required external knowledge: [Required knowledge base: Object knowledge base; Required data: Historical object information]; [Required knowledge base: Medical knowledge base; Required data: Collection of cognitive intervention content]; ....} The method of using a content recommendation model to perform thought chain reasoning on the current cognitive state, current interactive content, and cognitive training objective to obtain the initial task framework information of the target recommendation task can specifically include: using a content recommendation model to identify the content scenario of the current interactive content, and based on the content scenario, determining at least one query condition for a knowledge base; then, inferring the interaction configuration information of the target recommendation task based on the current cognitive state and cognitive training objective; and finally, constructing the task framework of the target recommendation task based on the query conditions and interaction configuration information to obtain the initial task framework information.

[0103] The content scenario can be understood as a semantic theme or task execution context abstracted from the current interactive content, such as family event recollection, computational challenge tasks, and music-based mood regulation. Interaction configuration information can be understood as the control parameters for the target recommendation task during task execution, such as interaction rounds, conversation style, guidance strategies, and feedback rules.

[0104] The knowledge base can include an object knowledge base for the target object and a medical knowledge base for the current cognitive state. The object knowledge base can be understood as a vector database or graph database storing the target object's personal historical data (such as life events, interests, social relationships, and cognitive reserve level). The medical knowledge base can be understood as a domain knowledge graph or document library storing professional content such as cognitive training methodologies, disease knowledge, and health education materials. Query conditions can be understood as search instructions or semantic vectors used to accurately retrieve relevant information from the knowledge base. Query conditions can include object information query conditions and medical content query conditions. Object information query conditions can be used to query historical object information of the target object in the object knowledge base. For example, for a children's wedding scenario, the object information query condition could be a semantic vector: [wedding, children, 20XX], used to retrieve relevant historical information. Medical content query conditions can be used to query the set of cognitive intervention content corresponding to the current cognitive state in the medical knowledge base. For example, for the goal of deepening episodic memory, the medical content query condition could be a semantic vector: [episodic memory, detail retrieval, guiding question examples]. The object information query conditions and medical query vectors can be semantic vectors, graph vectors, or other suitable vector representations, which are not limited here.

[0105] After obtaining the current interaction content, a content recommendation model can be used to identify the content scenario of the current interaction content. Based on the content scenario, the query conditions for object information in the object knowledge base and the query conditions for medical content in the medical knowledge base are determined respectively, so that the query operation can be performed in the corresponding knowledge base to obtain the external knowledge required for task generation.

[0106] Once the current cognitive state and cognitive training objective are obtained, the interaction configuration information for the target recommendation task can be inferred. For example, if the current cognitive state indicates that the target object's emotional valence is positive but its cognitive ability level is moderate, and the cognitive training objective is to deepen episodic memory, the inferred interaction configuration information might include: interaction rounds: 4 rounds (gradually deepening), first round question type: open-ended (guiding the narration of an event overview), subsequent question strategy: from asking about event details to emotional experience, feedback mechanism: high-frequency encouragement, etc.

[0107] After obtaining the query conditions and interaction configuration information, a task framework for the target recommendation task can be constructed based on these conditions and configuration information, thus obtaining initial task framework information. There are several ways to construct the task framework to obtain initial task framework information. For example, pre-defined templates can be used to fill in the inferred parameters into a structured template to obtain initial task framework information; alternatively, natural language descriptions of the initial task framework information can be generated through prompt engineering; or a graph-based representation method can be used to construct a task flowchart with interaction rounds as nodes and logical flows as edges, which can then be further converted into initial task framework information, and so on.

[0108] For example, in the case of a target recommendation task being a "memory box," upon receiving input from the target object (such as a family photo), the content recommendation model performs thought chain reasoning based on the task type and the target object's input content. This reasoning can include the following steps: Reasoning Step 1: The target object's input content is a group photo with the target object as the main subject and festive decorations in the background, suggesting a family celebration scene; Reasoning Step 2: Based on the content scene, it is necessary to retrieve relevant information about similar events from the object knowledge base (such as weddings, birthdays, etc.) and to retrieve guidance methods for deepening contextual memory from the medical knowledge base; Reasoning Step 3: Infer the interaction configuration information for the target recommendation task, such as using open-ended questions (5W1H) in the first round to guide the target object to narrate the event, and based on the richness of the feedback content, the next round can ask for emotional details or object associations. The above reasoning steps can then be analyzed to obtain the initial task framework information for the target recommendation task.

[0109] The target knowledge base can store personal historical data corresponding to the target, such as life events, hobbies, social relationships, cognitive reserve level, etc. In this application, the cognitive reserve level of the target can be understood as the cognitive resources and abilities accumulated by the target through continuous accumulation, which reflects the potential resilience of the target's brain in the face of age-related changes or pathological damage. In this application, the cognitive reserve level of the target can usually be comprehensively assessed through the target's life history information, mainly covering the following aspects: (1) The target's education and occupational background, which may include the target's formal years of education (or highest academic level), and the complexity of the occupation (such as the requirements for knowledge and skills, decision-making ability and interpersonal communication). (2) Cognitive activities and lifestyle, which may include whether they frequently participate in mental activities such as reading, writing, chess, learning new skills, etc., as well as the degree of social participation (such as community activities, volunteer service, frequency of interaction with relatives and friends) and physical activity level. (3) Early intellectual foundation (such as the intellectual foundation of elderly people with cognitive impairment before the onset of the disease), which can be traced through the target's childhood intelligence test, vocabulary ability or adult standardized intelligence test (such as WAIS), as a stable predictive indicator of cognitive reserve. In this application, a specially designed cognitive reserve scale can also be used to assess the cognitive reserve level of the target subjects, such as the Cognitive Reserve Index Questionnaire (CRIq), Cognitive Reserve Questionnaire (CRQ), Cognitive Reserve Scale (CRS), or Life Experience Questionnaire (LEQ).

[0110] S33: Based on the initial task framework information, filter out the historical object information and cognitive intervention content set of the target object under the target recommendation task from at least one knowledge base.

[0111] Specifically, the method of filtering historical object information and cognitive intervention content set of the target object under the target recommendation task from at least one knowledge base based on the initial task framework information may include: identifying the query conditions corresponding to the knowledge base in the initial task framework information; then, filtering the object preference information of the target object and the associated content corresponding to the current interaction content in the object knowledge base according to the object information query conditions to obtain historical object information; and then, filtering at least one dimension of task constraint conditions in the medical knowledge base based on the medical content query conditions to obtain the cognitive intervention content set.

[0112] Among these, object preference information can be understood as the target object's interests, behavioral habits, selection preferences, and emotional feedback displayed in historical interactions, such as preferences for specific topics (e.g., gardening, history), interaction forms of preferences (e.g., voice priority), and the range of task difficulty preferred. Related content can be understood as historical records in the object's knowledge base that are semantically, visually, or contextually associated with the current interaction content, such as previously uploaded images on similar topics, related people or events mentioned, and historical answers in similar tasks.

[0113] Task constraints can be understood as professional principles and limitations extracted from the medical knowledge base to guide the task generation and interaction process. These include, but are not limited to: training method guidelines applicable to the current cognitive state (e.g., prioritizing error-free learning for MCI target subjects), questioning constraints (e.g., avoiding negative questions to reduce cognitive load), interaction rhythm suggestions (e.g., providing clear positive feedback after each round of interaction), content safety boundaries (e.g., avoiding words that induce anxiety or confusion), and cognitive ability adaptation rules (e.g., limiting single information load to no more than 3 items if cognitive ability is weak). At least one dimension of task constraints can be understood as a set of structured rules extracted from the medical knowledge base to guide or restrict the task generation and interaction process in different aspects, with each dimension corresponding to a specific constraint theme or intervention principle. For example: the methodological dimension specifies the cognitive intervention methods to be used in training, such as error-free learning or progressive prompting; or the interaction dimension limits the questioning methods, feedback formats, and language styles, such as using open-ended questions and avoiding complex sentence structures; or the difficulty dimension controls the complexity, duration, and information load of the task, such as a single task duration of ≤5 minutes and no more than 4 options presented each time; or the safety dimension ensures that the content is ethical and avoids triggering negative emotions, such as prohibiting words that may cause anxiety and prioritizing the use of encouraging language.

[0114] It's important to note that different cognitive states correspond to different medical knowledge sub-bases and task constraints. For example, if the current cognitive state indicates attention deficit, the medical knowledge base can provide task constraints such as using prominent visual cues, reducing irrelevant information interference, and keeping the task duration within 3 minutes. Or, if the current cognitive state indicates decreased episodic memory, it can provide guidance methods such as using personally relevant materials, adopting a time-place-person-event question structure, and encouraging detailed descriptions and emotional connections. Or, if the current cognitive state indicates low mood, it can provide task constraints such as starting with empathetic language, incorporating positive narratives and humor, and guiding through suggestions rather than direct corrections, and so on.

[0115] Among them, based on the object information query conditions, the object preference information of the target object and the associated content corresponding to the current interaction content are filtered out in the object knowledge base. There are multiple ways to obtain historical object information, such as, for example, retrieval based on vector similarity, or, keyword and tag matching, or, graph relationship query, etc.

[0116] There are several ways to obtain a set of cognitive intervention content by filtering out at least one dimension of task constraints from the medical knowledge base based on medical content query conditions. For example, the current cognitive state can be matched with a preset index of the medical knowledge base to obtain the corresponding task constraints. Alternatively, a classifier can be used for condition mapping, inputting the current cognitive state into a trained classifier and outputting the corresponding task constraints of at least one dimension. Or, semantic retrieval and sorting can be used to transform the medical content query conditions into search statements, find relevant task constraints in the medical knowledge base, and return them in a sorted manner based on relevance.

[0117] Based on the initial task framework information, historical object information and cognitive intervention content sets in the medical knowledge base can be retrieved from the object knowledge base. Then, personalized, scientific and effective cognitive training content can be generated for the target object by combining the target object's specific preferences and medical guiding principles.

[0118] Optionally, in some implementations, the cognitive reserve level of the target object can be retrieved from the object knowledge base based on the object information query conditions. This allows the target object's cognitive reserve level to be added to the initial task framework information. Furthermore, by combining the target object's specific preferences, medical guidelines, and cognitive reserve level, the generated personalized cognitive training content (such as adjusting the difficulty, complexity, and focus of the cognitive training tasks) can dynamically adapt to the target object's cognitive resource background. For example, for target objects with a high cognitive reserve level, more complex, abstract, and challenging cognitive training tasks can be generated and recommended; while for target objects with a low cognitive reserve level or signs of cognitive decline, cognitive training tasks focused on consolidating the foundation, maintaining function, and enhancing confidence can be prioritized and recommended.

[0119] 104. Based on historical object information, current cognitive state, and cognitive intervention content set, generate the current task interaction content for the target recommendation task.

[0120] The current task interaction content can be understood as the first or subsequent rounds of guiding and interactive content generated by the target recommendation task. Its form can be text, voice, image or a combination thereof, and it aims to start or promote task execution in a user-friendly way, such as asking questions, showing task instructions, and guiding the target audience to participate.

[0121] The method of generating the current task interaction content of the target recommendation task based on historical object information, current cognitive state, and cognitive intervention content set can specifically include generating initial interaction prompt information corresponding to the target recommendation task based on initial task framework information and current cognitive state; then, adding historical object information and cognitive intervention content set to the initial interaction prompt information to obtain the interaction prompt information of the target recommendation task; and finally, using a content recommendation model to generate the current task interaction content of the target recommendation task based on the interaction prompt information.

[0122] The initial interactive prompts can be understood as a basic template for task guidance built upon the task framework and cognitive state, without incorporating personalized information. The interactive prompts can be understood as complete, structured task generation instructions that integrate historical object information and a set of cognitive intervention content, including key information such as role settings, task objectives, historical object information, a set of cognitive intervention content, and generation format requirements.

[0123] The process involves generating initial interactive prompts for the target recommended task based on the initial task framework information and the current cognitive state. Historical object information and a set of cognitive intervention content are then added to these initial prompts. Specifically, a template-filling method can be used. Predefined prompt templates for different task types (e.g., memory-based, calculation-based) include placeholders such as "[role setting]" and "[cognitive training goal]". The corresponding template is selected based on the task type in the initial task framework information and filled in according to the current cognitive state to generate the initial interactive prompts. Subsequently, retrieved historical object information (e.g., the target object was previously a teacher) and a set of cognitive intervention content (e.g., questioning techniques for deepening contextual memory) are filled into specific positions on the templates to form complete interactive prompts.

[0124] After receiving the interactive prompts, the method of generating the current task interaction content for the target recommendation task using a content recommendation model can take several forms. For example, an end-to-end generation method can be used, where the interactive prompts are taken as input and the content recommendation model directly outputs task guidance text or structured responses that meet the requirements. Alternatively, conditional generation control can be combined, adjusting parameters such as temperature, repetition penalties, and keyword guidance to control the stability and security of the generated content. Or, the tone, sentence complexity, and emotional color of the generated content can be automatically adjusted based on the target audience's historical preferences or current cognitive state to better suit their habits. Or, multiple versions can be generated and the best version can be selected as the final output.

[0125] Among these, methods for controlling the stability and security of generated content by combining conditional generation control and adjusting temperature parameters, repetition penalties, keyword guidance, etc., can include: In terms of security, the medical knowledge base undergoes multiple rounds of review by domain experts during the construction phase to ensure the professional credibility of the knowledge sources used for retrieval enhancement. Furthermore, after content generation, the model output content can be intercepted and cleaned in real time through built-in multi-level content filtering, such as sensitive word matching, negative sentiment recognition, medical fact consistency verification, and cultural compatibility detection, to prevent inappropriate or misleading content from reaching users. In terms of reliability assurance, a confidence assessment can be introduced to output a confidence score for each round of current task interaction content generated by the content recommendation model (such as the judgment result based on generation probability, semantic consistency, or a dedicated evaluation model). When the confidence score is lower than a preset threshold, a fallback strategy can be automatically triggered (such as enabling alternative templates, simplifying the description, or transferring it to manual review). In addition, for high-risk or high-complexity task scenarios (such as health guidance after the first diagnosis, medication reminders, etc.), a hybrid mode of model generation and manual review can be supported. The current task interaction content can be finally confirmed by a neurologist or cognitive rehabilitation therapist before it can be pushed, ensuring zero error output in critical scenarios.

[0126] It should be noted that the current task interaction content is not a fixed output; its generation process is dynamic and adaptable. That is, the subsequent interaction content can be adjusted in real time based on the subsequent feedback from the target object, thereby achieving continuous personalized guidance and support during the task execution process.

[0127] 105. When feedback is received from the target object regarding the current task interaction content, target recommended content is recommended to the target object based on the feedback content and the cognitive intervention content set.

[0128] Feedback content can be understood as the target audience's response to the current task interaction, such as multimodal interaction information like voice replies, text input, selection operations, and emotional expressions. For example, in the "Memories Box" task, if the question is: "This photo is so heartwarming! Can you tell me what day it was taken and what happy event happened?", the target audience's feedback could be: "This was taken at my son's wedding last National Day, at a restaurant in my hometown." This feedback content can include specific information provided by the target audience, their emotional inclination, and their cognitive engagement.

[0129] The target recommendation content can be understood as the summary, guidance, or encouragement content generated and recommended to the target user after the target recommendation task, based on the context of the entire conversation (especially the target user's feedback). This could include a summary of the target user's shared story, health advice extracted from the entire target recommendation content, an encouraging evaluation, or guidance for a natural transition to the next task. Its core function is to complete the closed loop of the current training task and enhance the training effect.

[0130] Specifically, the method of recommending target content to the target object based on the feedback content and the cognitive intervention content set can include: identifying the interaction rounds of the target recommendation task in the initial task framework information; when there are multiple interaction rounds, updating the interaction prompts based on the feedback content to recommend target content to the target object; when there is a single interaction round, generating target recommendation content using a content recommendation model based on the feedback content and the current task interaction content, and recommending the target recommendation content to the target object.

[0131] The interaction rounds can be understood as the total number of question-and-answer interactions planned with the target object in this target recommendation task. A single round is a question-and-answer interaction, while multiple rounds include multiple consecutive follow-up questions, guidance, or task advancement.

[0132] Optionally, in some implementations, when the interaction round is single-round, the target recommendation task can be designed as a one-time question and answer. After receiving the target object's feedback on the single question, the target recommendation content (such as a brief summary and encouragement) can be generated directly by combining the cognitive intervention content set (such as a summary template for this task type and a library of encouraging phrases) and the current target recommendation task can be ended.

[0133] Optionally, in some implementations, when feedback content is obtained and the interaction rounds are multiple, the interaction prompt information can be updated based on the feedback content, and the updated interaction prompt information can be used as the interaction prompt information. Then, the process returns to the step of generating the current task interaction content of the target recommendation task based on the interaction prompt information using a content recommendation model, until the interaction rounds are reached and the target feedback content is obtained. Then, based on the target feedback content and the current task interaction content, the target recommendation content is generated using a content recommendation model, and the target recommendation content is recommended to the target object.

[0134] In this context, the target feedback content can be understood as the integration of the target object's feedback content across all interaction rounds in a multi-round interactive task. For example, consider the "Memory Box" task, assuming a preset three-round interaction: Round 1 asks: When was this photo taken? The target object's feedback: Taken last year at my son's wedding; Round 2, based on the previous feedback, asks: What was the specific date? The target object's feedback: October 1st; Round 3, based on the previous feedback, continues to ask: What holiday was it? The target object's feedback: National Day. Therefore, the target object's target feedback content can include a combination of the feedback content from Rounds 1, 2, and 3.

[0135] Optionally, in some implementations, after obtaining the target feedback content, recommendation prompts for the target recommendation task can be generated based on the target feedback content and the current task interaction content. Then, based on preset content generation parameters and recommendation prompts, at least one candidate recommendation content is generated using a content recommendation model. The candidate recommendation content is then filtered to obtain filtered content, and modal transformation is performed on the filtered content to obtain at least one transformed content. Finally, the filtered content and the transformed content are fused to obtain the target recommendation content, and the target recommendation content is recommended to the target object.

[0136] The recommended prompts can be understood as structured instructions to guide the generation of summary or extended recommended content. These may include a task review (a brief summary of the core content and interaction process of the training task), a summary of the target audience's performance (highlights and progress in the target audience's cognitive performance extracted from the feedback), assessment of the achievement of cognitive training objectives (evaluating the degree of completion of the preset training objectives), required notes (professional advice and precautions provided by the medical knowledge base), and output format requirements (clearly specifying the tone, length limits, and other formatting specifications of the generated content). Preset content generation parameters can be understood as technical configuration parameters that control the quality and style of the generated content. These parameters may include creative control parameters (such as a temperature value to control the randomness of generation, a top-p value to control the range of vocabulary selection), generation length limit parameters (minimum and maximum generation length constraints), language style preference parameters (configurations for different tone styles such as formal, friendly, and encouraging), and so on.

[0137] Here, candidate recommended content can be understood as multiple alternative recommended content generated by the content recommendation model based on recommendation prompts. Filtered content can be understood as recommended content that has undergone security filtering. Transformed content can be understood as content that transforms the filtered text modality into other modalities to enrich the target audience's experience.

[0138] After obtaining the target recommendation prompts, a content recommendation model can be used to generate content based on the prompts. During the content generation process, the content style, length, creativity, and language features of the recommended content can be controlled based on preset content generation parameters, thereby obtaining multiple candidate recommended content.

[0139] After obtaining multiple candidate recommendations, a content safety filter in the content recommendation model can be used to filter these recommendations across multiple dimensions to obtain the filtered content. There are various ways to filter multiple candidate recommendations across multiple dimensions. For example, sensitive word detection can be performed to block inappropriate or offensive words; negative emotion detection can be performed to filter expressions that may induce anxiety, frustration, or negative emotions; medical accuracy verification can be performed to ensure that the recommendations meet medical standards and are free of misleading information; relevance assessment can be performed to ensure that the content is highly relevant to the feedback from the target audience and the training objectives; or language fluency and cultural suitability verification can be performed, and so on.

[0140] After obtaining the filtered content, modal transformation can be performed on it to obtain at least one transformed content. There are various ways to perform modal transformation on the filtered content. For example, the filtered content can be converted into a speech modal using text-to-speech technology; it can be converted into an image modal in the form of an illustrated card or infographic using image-text generation technology; it can be converted into a video modal in the form of a short animation or explanatory video using dynamic synthesis technology; or it can be converted into a structured interactive component, such as a clickable task card or progress reminder, and so on.

[0141] After obtaining the filtered content and the transformed modality, the filtered content and the transformed content can be merged to obtain the target recommended content. There are various fusion methods. For example, spatiotemporal synchronization fusion can be used to ensure precise synchronization in presentation time for text, voice, images, and video (e.g., when a specific keyword is spoken, the interface highlights the corresponding text and displays related images). Alternatively, semantic enhancement fusion can be used to make the content of different modalities semantically complementary (e.g., when a voice broadcast makes a summary statement, the interface displays a structured text list of key points and related diagrams). Another approach is sentiment consistency fusion to ensure that the emotional tone conveyed by different modalities is consistent (e.g., encouraging text is paired with a gentle voice tone and warm background images). Finally, the most suitable modality combination can be adaptively selected for recommendation based on the target audience's terminal device support, and so on.

[0142] After obtaining the target recommended content, the target recommended content can be recommended to the target object. There are multiple ways to recommend the target recommended content to the target object. For example, when the execution subject in this embodiment is the server, the target recommended content can be sent to the terminal for display through push notifications, interface callbacks, etc.; when the execution subject in this embodiment is the terminal, multimodal content can be directly displayed on the local interface through local GUI rendering.

[0143] Optionally, in some implementations, after obtaining the current task interaction content and feedback content, the object attribute information of the target object can be extracted from the current task interaction content and feedback content. Then, the object attribute information can be feature-encoded to obtain the incremental object features of the target object. Then, the object knowledge base can be incrementally updated according to the object attribute information and the incremental object features, and the updated object knowledge base can be used as the object knowledge base.

[0144] Among them, object attribute information can be understood as information related to newly added target objects identified and extracted from the current interaction process. For example, new life events (such as the target object mentioning for the first time that the daughter brought her grandson to visit last week), supplementary preference information (such as the user explicitly stating that they don't like watching opera and prefer listening to old songs), emotional state markers (such as the user showing a pleasant mood when mentioning a past event in this interaction), cognitive performance snapshots (such as the average reaction time in this interaction being 10% shorter than last week), and so on.

[0145] Incremental object features can be understood as vectorization of object attribute information obtained through feature encoding. Incremental object features facilitate efficient similarity calculation and rapid retrieval within the object knowledge base. In this application, the incremental update of the object knowledge base is a continuous learning process. After each interaction, new facts, sentiments, and behavioral cues related to the target object are extracted from the conversation, transformed into incremental object features, and then fused with or updated with historical object information stored in the object knowledge base, thereby achieving progressive improvement of the object knowledge base.

[0146] Optionally, in some implementations, after updating the object knowledge base, the long-term interest distribution, behavioral pattern baseline, and cognitive ability evolution trajectory of the target object can be recalculated or adjusted based on the historical object information in the updated object knowledge base. Furthermore, the current cognitive state of the target object can be dynamically calibrated based on the updated object knowledge base, or more accurate personalized context can be provided for task generation and scene matching in subsequent sessions. For example, if the update adds a mention of red Tang suits by the target object, this information can be prioritized in subsequent tasks involving themes such as clothing, colors, or family rituals to enhance the relevance and emotional resonance of the generated content.

[0147] As an example, taking the interactive scenario with the target recommended content as the "Memories Box" task as an example, let's set the elderly user as Grandma Li, 76 years old, with mild cognitive impairment. In other words, elderly user Grandma Li can complete cognitive assessment tasks (such as a number recitation game) through an interactive interface, and obtain multimodal data of elderly users with cognitive impairment when performing the cognitive assessment task (such as delayed response, intermittent speech, and slightly elevated heart rate). This allows analysis of Grandma Li's current cognitive state (such as decreased episodic memory, positive mood, and moderate cognitive load). The interactive interface then displays a recommended task selection interface to Grandma Li. After receiving feedback from Grandma Li on the task selection interaction, the system can determine the target recommended task selected by Grandma Li based on the task selection interaction content. When the target recommended task is a recommended task in an interactive recommendation scenario, the system can obtain the content of Grandma Li's interaction under the target interactive item based on the task attribute information of the target recommended task, thus obtaining the current interaction content. For example, if the target recommended task is the "Memories Box" in today's challenge task, then Grandma Li can be guided to upload an image (such as a photo of herself and her husband in front of the peonies in the old courtyard in 1987). Afterward, the system can identify the content scene in the image: peonies, courtyard, and couple's photo.

[0148] Then, based on the current interaction content and current cognitive state, the cognitive training objective for the elderly user, Grandma Li, in this target recommendation task can be determined. A thought chain reasoning process is then performed on the current cognitive state, current interaction content, and cognitive training objective to obtain the initial task framework information for the target recommendation task. This determines the object information query conditions and medical content query conditions to be retrieved from the object knowledge base. Based on these object information query conditions, the following can be retrieved from the object knowledge base: Grandma Li's husband planted peonies in 1987, he passed away in 2015, and the peonies were transplanted to her current residential area (historical object information). Based on these medical content query conditions, the following can be retrieved from the medical knowledge base: using personal materials, five-element questioning, positive feedback in each round, and avoiding grief-oriented approaches (cognitive intervention content set). Then, based on the historical object information, current cognitive state, and cognitive intervention content set, interaction prompts are obtained. A content recommendation model is then used to generate the current task interaction content for the target recommendation task based on these prompts, and this is verified through the interaction interface. The system displays the current task interaction content to elderly user Grandma Li and receives her feedback. For example, the first round of interaction could be, "What spring was this taken? Is the courtyard still there?", and Grandma Li's feedback could be, "It was taken in 1987, the year I retired. The courtyard was demolished, and the peonies were moved to the current community." The second round could be, "How did Grandpa come up with the idea of ​​planting peonies?", and Grandma Li's feedback could be, "He knew I liked peonies, so the year I got married, he asked someone to buy seedlings and secretly planted them, saying it was a wedding gift for me." The third round could be, "What did Grandpa say when the first flower bloomed?", and Grandma Li's feedback could be, "He said, 'This flower looks like you.'" Based on the target feedback content and cognitive intervention content received from Grandma Li, recommendation prompts are obtained. A content recommendation model can be used to generate content, and the generated content is processed to obtain the target recommendation content. For example, the recommended content for the auditory modality could be, "You remember every detail of how your grandfather planted peonies, even the words he said. Memory is like a peony; if you nurture it with care, it will bloom."; the recommended content for the visual modality could be an illustration of a peony.

[0149] Optionally, in some implementations, after determining the target recommendation task corresponding to the target object based on the content recommendation request, and when the target recommendation task is a recommendation task in a proactive recommendation scenario, the historical physiological data and medical needs information of the target object in the current cognitive state within a preset time range can be obtained. Then, based on the target object's current cognitive state, historical physiological data, and medical needs information, a content recommendation model can be used to generate at least one modality of medical education content corresponding to the medical needs information. After that, the medical education content can be used as the target recommendation content and recommended to the target object.

[0150] The preset time range can be understood as a pre-defined historical data review interval, such as the last 7 days, the last 30 days, or since the last assessment. Historical physiological data can be understood as data collected and stored within this time range that reflects the physiological state of the target individual, such as heart rate, blood pressure, blood oxygen saturation, sleep data, and vascular risk factors. Medical needs information can be understood as descriptions of intervention needs or medical topics related to cognitive health or chronic disease management, derived from the analysis of the current cognitive state, such as guidance on improving sleep quality, support for mood regulation, reminders and popular science information on abnormal blood pressure, medication management, home safety reminders, and cognitive health education. Medical education content can be understood as multimodal information generated based on medical needs, aimed at improving the target individual's health awareness and self-management abilities, such as text health knowledge cards, audio explanations, and popular science animations.

[0151] There are several ways to obtain the target object's historical physiological data and medical needs information in the current cognitive state within a preset time range. For example, it can connect to wearable devices or health monitoring platforms to synchronize physiological data according to a preset time range; or it can analyze the physiological indicator dimensions in the current cognitive state and the historical baseline to identify abnormal trends and generate demand tags; or it can combine the object knowledge base and the medical knowledge base to match typical health education topics corresponding to the current cognitive state; or it can actively ask the target object about their current physical feelings through conversational interaction, and so on.

[0152] After obtaining the target's current cognitive state, historical physiological data, and medical needs information, the target's cognitive abilities and interests can be determined based on the target's current cognitive state. Then, a content recommendation model can be used to analyze the target's health trends based on historical physiological data. After that, a content recommendation model can be used to retrieve relevant knowledge based on medical needs information and generate health education content, thereby obtaining at least one modality of medical education content corresponding to the medical needs information.

[0153] After obtaining medical education content in at least one modality, this content can be filtered, and the filtered content can be used as target recommended content to the target audience. The filtering methods can be varied. For example, medical accuracy verification can be performed to ensure the content conforms to the latest medical guidelines; language can be age-friendly, converting technical terms into expressions easily understood by the elderly; emotional positivity testing can be conducted to ensure the content is positive and uplifting; or personalized suitability assessment can be performed to select suggestions that best match the target audience's lifestyle, and so on.

[0154] It should be noted that in this application, the proactive recommendation scenario is based on continuous monitoring and analysis of the target object's cognitive state, physiological indicators, and behavioral patterns. It automatically triggers health management-related recommendation tasks, directly generating and pushing appropriate target recommendation content to achieve timely early warning and proactive intervention for the target object's health risks. The interactive recommendation scenario, on the other hand, focuses on the target object's active participation and cognitive ability training. Through multiple rounds of guided interaction, it dynamically generates training tasks closely related to the target object's personal experiences and interests to enhance their cognitive function and emotional experience. These two scenarios complement each other, jointly constructing a dual-path training architecture that balances health management and cognitive training to adapt to personalized service needs at different stages and in different contexts.

[0155] Optionally, in some implementations, the system can also respond to inquiries initiated by the target object and perform thought chain reasoning based on the inquiries, the target object's current cognitive state, historical object information, current cognitive state, and cognitive intervention content set to obtain the current task interaction content, so as to establish a conversation with the target object and answer the inquiries initiated by it.

[0156] As an example, such as Figure 4 As shown, after obtaining the current task's interaction content, a conversation can be initiated with the target object. For example... Figure 4 As shown in the interactive interface on the left, after the target audience (such as an elderly person with cognitive impairment) initiates a conversation asking "What should I do if I'm feeling very anxious lately?", the system can identify the target audience's intention to seek help and their emotional state (anxiety). It can then connect to a medical knowledge base to obtain cognitive intervention strategies for anxiety (such as "diaphragmatic breathing training"). Next, considering the target audience's current cognitive state (such as short attention span), the standard breathing training steps can be broken down into simple instructions, with clear voice guidance and visual cues. This can then generate a visually appealing interactive message for the current task: "I understand that anxiety is really unpleasant, but please believe that it can be improved. Let's start with a few simple methods that can be taken immediately: Press the pause button: Now take a deep breath—inhale for 4 seconds, hold your breath for 4 seconds, and slowly exhale for 6 seconds (at the speed of your breath fogging up in front of a mirror). Repeat 3 times, and your heart rate will noticeably slow down. (Instructional video attached)." This guides the target audience to follow the instructional video and complete the breathing exercise.

[0157] like Figure 4As shown in the interactive interface on the right, after the target user initiates the question "I was recently diagnosed with mild cognitive impairment, how can I practice memory at home?" and selects today's recommended task, the system can identify the target user's cognitive training goal, link the user's knowledge base and medical knowledge base, obtain cognitive intervention strategies for anxiety and historical user information (such as "likes playing cards"), and then combine the target user's current cognitive state with clear voice guidance and visual cues to generate a visually appealing interactive content for the current task: "After being diagnosed with mild cognitive impairment, it is very important to practice memory at home. Here are some effective memory training methods: Playing cards: Take out a deck of cards, memorize the suit and number, starting with 2 cards and gradually increasing the number of cards. For example, memorize 2 cards first, then 4 cards, 5 cards, and gradually increase the difficulty. (See attached illustration)." Further interaction can then be conducted based on the target user's feedback, and the system can generate recommended content for the recommended task based on the interactive content and the cognitive intervention content set.

[0158] In this application, as Figure 5 As shown, in the case where the execution entity of this application embodiment is a server (where a content recommendation model is deployed), it can interact with the terminal. Furthermore, the terminal is equipped with an interactive interface to interact with the target object. The following is an example illustrating this application embodiment: it is applied to generative artificial intelligence, the target object is an elderly user with cognitive impairment, the target recommendation task in the interactive scenario is a "Today's Challenge" task, and the target recommendation task in the proactive scenario is a health education class. exist Figure 5 In this process, the terminal is equipped with an interactive interface to interact with the target object (hereinafter referred to as an elderly user with cognitive impairment). Based on step S201, the terminal can display the corresponding start control for cognitive intervention to the elderly user with cognitive impairment through the interactive interface, and receive the start operation of the start control by the elderly user with cognitive impairment on the interactive interface, so as to send the start signal corresponding to the elderly user with cognitive impairment to the server. After receiving the start signal, the server can send the corresponding cognitive assessment task of the elderly user to the terminal through a preset interface based on step S202. Afterwards, based on step S203, the terminal can display a cognitive assessment task to the elderly user with cognitive impairment through an interactive interface. The cognitive assessment task can be a 1-2 minute number recitation game (forward / backward recitation). The terminal acquires multimodal data (including any one of interactive behavior data, voice data, visual data, and physiological data) of the elderly user while performing the cognitive assessment task and sends the acquired multimodal data to the server. After receiving the multimodal data, the server can use a content recommendation model to extract features from the multimodal data based on step S204 to obtain the modal features corresponding to each modality. At least one modal feature is then fused to obtain a fused modal feature. Based on the fused modal feature, the server can perform a cognitive assessment on the elderly user with cognitive impairment to determine their initial cognitive state. The initial cognitive state is then fused with the historical static cognitive state to obtain the current cognitive state of the elderly user with cognitive impairment. Based on the current cognitive state of the elderly user with cognitive impairment, the server can generate a corresponding daily health briefing and personalized greeting and send it back to the terminal. Afterwards, based on step S205, the terminal can display a recommended task selection interface (such as...) to the elderly user with cognitive impairment through an interactive interface. Figure 3 The task selection interface can include health briefings, personalized greetings, interactive recommendation scenarios, and proactive recommendation scenarios. It can also receive task selection interaction content from elderly users with cognitive impairment based on step S206 and send the task selection interaction content to the server. Furthermore, after the server receives the task selection interaction content from elderly users with cognitive impairment, it can determine the target recommended task selected by the elderly user based on the task selection interaction content. When the target recommended task is a recommended task in an interactive recommendation scenario, it can obtain the content of the elderly user's interaction under the target interactive item based on the task attribute information of the target recommended task, and obtain the current interaction content. For example, if the target recommended task is the "Old Friends Card Game" in today's challenge task, then it can guide the elderly user with cognitive impairment to select the content scenario of this card game: Option A: Select a photo related to getting together with old friends (such as a group photo of a senior university activity or a group photo of a community competition), Option B: Directly enter the name or nickname of an old friend (such as Lao Li), Option C: Recall a specific place where you used to play cards with friends (such as the activity room of your old workplace or the senior center in the community). Next, based on step S207, the server can determine the cognitive training objective for the current target recommendation task for the elderly user with cognitive impairment, according to the current interaction content and current cognitive state. For example, if the target recommendation task is the "Friends Card Game" challenge in today's challenge, the cognitive training objective for the elderly user with cognitive impairment could be to train their cognitive control and conflict resolution abilities. Then, based on step S208, the server can use a content recommendation model to perform thought chain reasoning on the current cognitive state, current interaction content, and cognitive training objective to obtain the initial task framework information for the target recommendation task. Specifically, the content recommendation model can identify the content scenario of the current interaction content and, based on the content scenario, determine at least one knowledge base query condition (object information query condition for the object knowledge base and medical content query condition for the medical knowledge base). Then, based on the current cognitive state and cognitive training objective, the server can infer the interaction configuration information of the target recommendation task (interaction rounds and question types, etc.). Finally, based on the query conditions and interaction configuration information, the server can construct the task framework for the target recommendation task to obtain the initial task framework information.

[0159] Subsequently, based on the initial task framework information in step S209, the server can filter out the historical object information and cognitive intervention content set of the target object under the target recommendation task from at least one knowledge base. For example, it can construct object information query conditions (requiring information related to Lao Li from the object knowledge base, such as former colleagues, bridge partners, etc.) and medical content query conditions (requiring task constraints corresponding to the current cognitive state of the elderly user from the medical knowledge base, such as prioritizing the use of error-independent learning methods and providing clear positive feedback after each round of interaction, etc.) based on the elderly user's choice of the card game. This allows the server to extract historical object information from the object knowledge base and cognitive intervention content set from the medical knowledge base.

[0160] Afterwards, the server can obtain interactive prompts based on historical object information, current cognitive state, and cognitive intervention content set from step S210. It then uses a content recommendation model to generate the current task interaction content for the target recommendation task based on the interactive prompts and sends this content to the terminal. This allows the terminal to display the current task interaction content to elderly users with cognitive impairment through an interactive interface in step S211, and receive feedback from the elderly users to send back to the server. For example, the current task interaction content could be: "A virtual card game opening interface, voice output: Okay, let's imagine we're playing cards with Lao Li in the activity room. I'll deal first. First question: I dealt the 5 of hearts, the King of diamonds, and the 8 of spades. What sport did Lao Li like best before? Also, please quickly calculate the total points of these three cards (K is worth 10 points)." The feedback from the elderly user could be: "He likes playing badminton, um, the total points are 5 + 10 + 8 = 23 points." When the server receives feedback from the elderly user with cognitive impairment regarding the current task interaction content based on step S212, it can update the interaction prompts based on the current task interaction content and the feedback content to obtain the current task interaction content for the next interaction round, until the interaction round ends. Then, the feedback content corresponding to multiple interaction rounds can be merged to obtain the target feedback content. Based on the target feedback content and the cognitive intervention content set, recommended prompts can be obtained. Then, a content recommendation model can be used to generate content based on preset content generation parameters and recommended prompts to obtain at least one candidate recommended content. Then, the candidate recommended content is filtered to obtain filtered content, and modal transformation is performed on the filtered content to obtain at least one transformed content. Then, the filtered content and the transformed content are merged to obtain the target recommended content, and the target recommended content is sent to the terminal so that the terminal can display the target recommended content to the elderly user with cognitive impairment through the interactive interface based on step S213.

[0161] Furthermore, on the server side, when the target recommendation task is a recommendation task in a proactive recommendation scenario, the historical physiological data of the elderly user with cognitive impairment within a preset time range, their current cognitive state and medical needs information, and their current cognitive state can be obtained based on step S214. Then, a content recommendation model is used to generate at least one modality of medical education content corresponding to the medical needs information. This medical education content can then be used as the target recommendation content and sent to the terminal, allowing the terminal to display the target recommendation content to the elderly user with cognitive impairment through an interactive interface based on step S215. For example, taking a health education class as the target recommendation task, the historical physiological data (such as vascular risk factors and historical sleep data) of the elderly user with cognitive impairment within a preset time range is obtained. Combined with the elderly user's current cognitive state, and through matching with a medical knowledge base, their current medical needs information can be identified as "autonomic nervous system dysfunction caused by anxiety and decreased sleep quality, requiring non-pharmacological intervention for emotional problems and sleep disorders." Based on this medical needs information, the content recommendation model can retrieve a set of relevant cognitive intervention content (i.e., task constraints and materials) from the medical knowledge base, including anxiety management, sleep hygiene, and relaxation training. Subsequently, the content recommendation model can also combine historical object information from the cognitively impaired elderly user's object knowledge base (showing the user's love for gardening and preference for content combining audio and text) to perform thought chain reasoning and content generation, ultimately generating a personalized multimodal medical education content as the target recommendation content. This target recommendation content may include: image and text cards: displaying several beautiful pictures of tranquil gardening scenes (such as potted plants under moonlight), accompanied by concise text explaining how gardening activities can help relax the nerves and improve sleep through sensory stimulation; and guided audio: a 3-minute guided meditation audio, using a gentle tone to guide the cognitively impaired elderly user to imagine themselves breathing in a quiet garden, feeling the vitality of plants, and incorporating their preferred classical background music.

[0162] As can be seen from the above, after obtaining multimodal data of the target object performing at least one cognitive assessment task, the embodiments of this application can determine the current cognitive state of the target object based on the multimodal data. When a content recommendation request for the target object is received, the target recommendation task corresponding to the target object is determined according to the content recommendation request. When the target recommendation task is a recommendation task in an interactive recommendation scenario, the historical object information and cognitive intervention content set of the target object under the target recommendation task are obtained. Then, based on the historical object information, the current cognitive state, and the cognitive intervention content set, the current task interaction content of the target recommendation task can be generated. When the target object receives a content recommendation request for the current cognitive state, the current task interaction content of the target recommendation task can be generated. When providing feedback on task interaction content, the system recommends target content to the target user based on the feedback content and the cognitive intervention content set. Since this approach is applied to generative artificial intelligence, the current cognitive state determined by multimodal data can indicate the target user's real-time cognitive ability level, emotional valence, and cognitive load. This allows the generation of current task interaction content based on the current cognitive state combined with the target user's historical information to accurately match the target user's individual characteristics and immediate needs, avoiding the problem of severe homogenization of recommended content. This ensures that the recommended content always adapts to the target user's current cognitive state, thereby improving the adaptability and accuracy of the recommended content.

[0163] To better implement the above methods, this application also provides a generative content recommendation device for cognitive intervention in elderly people with cognitive impairment. This device can be integrated into electronic devices, such as servers or terminals, which may include tablet computers, laptops, and / or personal computers.

[0164] For example, such as Figure 6 As shown, the generative content recommendation device for cognitive intervention in elderly people with cognitive impairment may include an acquisition unit 301, a receiving unit 302, a query unit 303, a generation unit 304, and a recommendation unit 305, as follows: (1) Obtain unit 301; The acquisition unit 301 is used to acquire multimodal data of the target object when performing at least one cognitive assessment task, and determine the current cognitive state of the target object based on the multimodal data.

[0165] For example, acquisition unit 301 can be used to extract multi-dimensional features from interaction behavior records when the modal data is interaction behavior data, and aggregate the extracted behavior features to obtain aggregated behavior features of the session; then, based on the temporal information of the session and the aggregated behavior features, construct an object behavior feature sequence of the target object; perform feature transformation on the behavior features in the object behavior feature sequence to obtain a transformed object behavior feature sequence; then, perform dimensionality reduction on the transformed object behavior feature sequence and select at least one object behavior feature from the dimensionality-reduced object behavior feature sequence; or, when the modal data is speech data, use a content recommendation model to repair the speech data and extract features from the repaired speech data to obtain the object speech features of the target object; or, when the modal data is visual data, use a content recommendation model to perform object recognition on the visual data and extract features from the recognized object information to obtain the target object. The modal data can be either visual features of the object or, when the modal data is physiological data, a content recommendation model can be used to extract features from the physiological data to obtain heart rate variability features. Temporal modeling of the heart rate variability features can then be performed to obtain the object's physiological features. Next, any one of the object's behavioral features, speech features, visual features, and physiological features can be used as the modal feature. Then, the modal features are temporally aligned, and attention weighting is applied to the aligned modal features to obtain the fused modal features of the multimodal data. Based on the fused modal features, at least one cognitive dimension of the target object is evaluated to obtain an initial evaluation result for that cognitive dimension. Then, the initial evaluation results are fused to obtain the initial cognitive state of the target object under the cognitive evaluation task. Finally, the historical static cognitive state of the target object is obtained, and the initial and historical cognitive states are fused to obtain the current cognitive state of the target object.

[0166] (2) Receiving unit 302; The receiving unit 302 is configured to, when receiving a content recommendation request for the target object, determine the target recommendation task corresponding to the target object based on the content recommendation request.

[0167] (3) Query unit 303; The query unit 303 is used to obtain the historical object information and cognitive intervention content set of the target object under the target recommendation task when the target recommendation task is a recommendation task in an interactive recommendation scenario.

[0168] For example, query unit 303 can be specifically used to obtain the current interaction content of the target object for the target recommendation task, and determine the cognitive training target of the target object in the target recommendation task based on the current interaction content and the current cognitive state; then, a content recommendation model is used to identify the content scene of the current interaction content, and based on the content scene, at least one query condition for the knowledge base is determined; based on the current cognitive state and the cognitive training target, the interaction configuration information of the target recommendation task is inferred; based on the query conditions and the interaction configuration information, the task framework of the target recommendation task is constructed to obtain the initial task framework information; then, the query conditions corresponding to the knowledge base are identified in the initial task framework information, and based on the object information query conditions, the object preference information of the target object and the associated content corresponding to the current interaction content are filtered in the object knowledge base to obtain historical object information; based on the medical content query conditions, at least one dimension of task constraint conditions are filtered in the medical knowledge base to obtain a cognitive intervention content set.

[0169] (4) Generation unit 304; The generation unit 304 is used to generate the current task interaction content of the target recommendation task based on the historical object information, the current cognitive state, and the cognitive intervention content set.

[0170] For example, generation unit 304 can be used to generate initial interactive prompts for the target recommendation task based on the initial task framework information and the current cognitive state; then, historical object information and cognitive intervention content set are added to the initial interactive prompts to obtain interactive prompts for the target recommendation task; then, based on the interactive prompts, a content recommendation model is used to generate the current task interactive content for the target recommendation task.

[0171] (5) Recommended Unit 305; The recommendation unit 305 is used to recommend target content to the target object based on the feedback content and the cognitive intervention content set when it receives feedback content from the target object regarding the current task interaction content.

[0172] For example, recommendation unit 305 can be used to identify the interaction rounds of the target recommendation task in the initial task framework information; when there are multiple interaction rounds, the interaction prompt information is updated based on the feedback content, and the updated interaction prompt information is used as the interaction prompt information; the process returns to the step of generating the current task interaction content of the target recommendation task using a content recommendation model based on the interaction prompt information, until the interaction round is reached, and the target feedback content is obtained; then, recommendation prompt information of the target recommendation task is generated based on the target feedback content and the current task interaction content; at least one candidate recommendation content is generated using a content recommendation model based on preset content generation parameters and recommendation prompt information; the candidate recommendation content is filtered to obtain filtered content, and modal transformation is performed on the filtered content to obtain at least one transformed content; the filtered content and the transformed content are fused to obtain the target recommendation content, and the target recommendation content is recommended to the target object; when there is a single interaction round, the target recommendation content is generated using a content recommendation model based on the feedback content and the current task interaction content, and the target recommendation content is recommended to the target object.

[0173] For example, recommendation unit 305 can also be used to extract object attribute information of the target object from the current task interaction content and feedback content; perform feature encoding on the object attribute information to obtain incremental object features of the target object; incrementally update the object knowledge base based on the object attribute information and incremental object features, and use the updated object knowledge base as the object knowledge base.

[0174] For example, recommendation unit 305 can also be used to obtain historical physiological data and medical needs information of the target object in the current cognitive state within a preset time range when the target recommendation task is a recommendation task in an active recommendation scenario; based on the target object's current cognitive state, the historical physiological data, and the medical needs information, generate at least one modality of medical education content corresponding to the medical needs information using the content recommendation model; use the medical education content as the target recommendation content, and recommend the target recommendation content to the target object.

[0175] In practice, each of the above units can be implemented as an independent entity or can be arbitrarily combined to be implemented as the same or several entities. For the specific implementation of each of the above units, please refer to the previous method embodiments, which will not be repeated here.

[0176] As can be seen from the above, the embodiments of this application can obtain multimodal data of the target object performing at least one cognitive assessment task through the acquisition unit 301, and determine the current cognitive state of the target object based on the multimodal data; when the receiving unit 302 receives a content recommendation request for the target object, it can determine the target recommendation task corresponding to the target object according to the content recommendation request; when the target recommendation task is a recommendation task in an interactive recommendation scenario, the query unit 303 can obtain the historical object information and cognitive intervention content set of the target object under the target recommendation task; then, the generation unit 304 can generate the current task interaction content of the target recommendation task based on the historical object information, the current cognitive state, and the cognitive intervention content set; then, when the recommendation unit 305 receives the feedback content of the target object on the current task interaction content, it recommends the target recommendation content to the target object according to the feedback content and the cognitive intervention content set. Therefore, the adaptability and accuracy of the recommended content can be improved.

[0177] This application also provides an electronic device, such as... Figure 7 As shown, it illustrates a structural schematic diagram of the electronic device involved in the embodiments of this application, specifically: The electronic device may include components such as a processor 401 with one or more processing cores, a memory 402 with one or more computer-readable storage media, a power supply 403, and an input unit 404. Those skilled in the art will understand that... Figure 7 The electronic device structure shown does not constitute a limitation on the electronic device and may include more or fewer components than shown, or combine certain components, or have different component arrangements. Wherein: The processor 401 is the control center of the electronic device, connecting various parts of the device via various interfaces and lines. It executes software programs and / or modules stored in the memory 402, and calls data stored in the memory 402, to perform various functions and process data. Optionally, the processor 401 may include one or more processing cores; preferably, the processor 401 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, user interface, and applications, and the modem processor mainly handles wireless communication. It is understood that the modem processor may not be integrated into the processor 401.

[0178] The memory 402 can be used to store software programs and modules. The processor 401 executes various functional applications and data processing by running the software programs and modules stored in the memory 402. The memory 402 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, application programs required for at least one function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the electronic device, etc. In addition, the memory 402 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device. Accordingly, the memory 402 may also include a memory controller to provide the processor 401 with access to the memory 402.

[0179] The electronic device also includes a power supply 403 that supplies power to the various components. Preferably, the power supply 403 can be logically connected to the processor 401 through a power management system, thereby enabling functions such as charging, discharging, and power consumption management through the power management system. The power supply 403 may also include one or more DC or AC power supplies, recharging systems, power fault detection circuits, power converters or inverters, power status indicators, and other arbitrary components.

[0180] The electronic device may also include an input unit 404, which can be used to receive input digital or character information, and generate keyboard, mouse, joystick, optical or trackball signal inputs related to user settings and function control.

[0181] Although not shown, the electronic device may also include a display unit, etc., which will not be described in detail here. Specifically, in this embodiment, the processor 401 in the electronic device loads the executable files corresponding to the processes of one or more applications into the memory 402 according to the following instructions, and the processor 401 runs the applications stored in the memory 402 to realize various functions, as follows: The system acquires multimodal data of the target object performing at least one cognitive assessment task, and determines the target object's current cognitive state based on the multimodal data. When a content recommendation request for the target object is received, the system determines the target recommendation task corresponding to the target object based on the content recommendation request. When the target recommendation task is a recommendation task in an interactive recommendation scenario, the system acquires the target object's historical object information and cognitive intervention content set under the target recommendation task. Based on the historical object information, current cognitive state, and cognitive intervention content set, the system generates the current task interaction content for the target recommendation task. When feedback content from the target object regarding the current task interaction content is received, the system recommends the target recommendation content to the target object based on the feedback content and the cognitive intervention content set.

[0182] For details on the implementation of each of the above operations, please refer to the previous examples, which will not be repeated here.

[0183] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be performed by instructions, or by instructions controlling related hardware. These instructions can be stored in a computer-readable storage medium and loaded and executed by a processor.

[0184] Therefore, embodiments of this application provide a computer-readable storage medium storing a plurality of instructions that can be loaded by a processor to execute steps in any of the generative content recommendation methods for cognitive intervention in elderly people with cognitive impairment provided in embodiments of this application. Specific implementations of the above operations can be found in the preceding embodiments and will not be repeated here.

[0185] The computer-readable storage medium may include: read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0186] Since the instructions stored in the computer-readable storage medium can execute the steps in any of the generative content recommendation methods for cognitive intervention of elderly people with cognitive impairment provided in the embodiments of this application, the beneficial effects that any of the generative content recommendation methods for cognitive intervention of elderly people with cognitive impairment provided in the embodiments of this application can achieve can be realized. For details, please refer to the previous embodiments, which will not be repeated here.

[0187] According to one aspect of this application, a computer program product or computer program is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the electronic device to perform the methods provided in various alternative implementations of the generative content recommendation aspect of cognitive intervention for older adults with cognitive impairment described above.

[0188] The foregoing describes a generative content recommendation method and related equipment for cognitive intervention in elderly people with cognitive impairment, as provided in the embodiments of this application. The related equipment may include a generative content recommendation method apparatus, electronic equipment, computer program products, and computer-readable storage media. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A generative content recommendation method for cognitive intervention in elderly people with cognitive impairment, characterized in that, include: Acquire multimodal data of the target object while performing at least one cognitive assessment task, and determine the current cognitive state of the target object based on the multimodal data; When a content recommendation request for the target object is received, the target recommendation task corresponding to the target object is determined based on the content recommendation request; When the target recommendation task is a recommendation task in an interactive recommendation scenario, obtain the historical object information and cognitive intervention content set of the target object under the target recommendation task; Based on the historical object information, the current cognitive state, and the set of cognitive intervention content, the current task interaction content of the target recommendation task is generated; When feedback is received from the target object regarding the current task interaction content, target recommended content is recommended to the target object based on the feedback content and the cognitive intervention content set.

2. The generative content recommendation method for cognitive intervention in elderly people with cognitive impairment according to claim 1, characterized in that, The multimodal data includes modal data of at least two modalities of the conversation generated by the target object during the cognitive assessment task. Determining the current cognitive state of the target object based on the multimodal data includes: The modal data are subjected to feature extraction using a content recommendation model to obtain the modal features of the modality. The modal features are temporally aligned, and the aligned modal features are attention-weighted to obtain the fused modal features of the multimodal data. Based on the fused modal features, a cognitive assessment is performed on the target object to obtain the target object's current cognitive state.

3. The generative content recommendation method for cognitive intervention in elderly people with cognitive impairment according to claim 2, characterized in that, The modal data includes any one of the following: interactive behavior data, voice data, visual data, and physiological data of the target object performing at least one cognitive assessment task. The modal data is then used with a content recommendation model to extract features, resulting in modal features, including: When the modal data is interactive behavior data, the content recommendation model is used to extract features from the interactive behavior data to obtain the object behavior features of the target object. When the modal data is speech data, the content recommendation model is used to repair the speech data, and feature extraction is performed on the repaired speech data to obtain the object speech features of the target object; When the modal data is visual data, the content recommendation model is used to perform object recognition on the visual data, and the recognized object information is used to extract features to obtain the object visual features of the target object. When the modal data is physiological data, a content recommendation model is used to extract features from the physiological data to obtain heart rate variability features, and time-series modeling is performed on the heart rate variability features to obtain the object's physiological features; The modal feature is any one of the object's behavioral features, the object's voice features, the object's visual features, and the object's physiological features.

4. The generative content recommendation method for cognitive intervention in elderly people with cognitive impairment according to claim 3, characterized in that, The interaction behavior data includes at least one interaction behavior record of the session. The step of using the content recommendation model to extract features from the interaction behavior data to obtain the object behavior features of the target object includes: Multidimensional feature extraction is performed on the interaction behavior records, and the extracted behavioral features are aggregated to obtain the aggregated behavioral features of the session; Based on the temporal information of the session and the aggregated behavioral characteristics, construct the object behavior feature sequence of the target object; The behavioral features in the object behavior feature sequence are transformed to obtain the transformed object behavior feature sequence. The transformed object behavior feature sequence is dimensionality reduced, and at least one object behavior feature is selected from the dimensionality-reduced object behavior feature sequence.

5. The generative content recommendation method for cognitive intervention in elderly people with cognitive impairment according to claim 2, characterized in that, The step of performing a cognitive assessment of the target object based on the fused modal features to obtain the current cognitive state of the target object includes: Based on the fused modal features, at least one cognitive dimension of the target object is evaluated to obtain an initial evaluation result under the cognitive dimension. The initial assessment results are fused to obtain the initial cognitive state of the target object when performing the cognitive assessment task; The historical static cognitive state of the target object is obtained, and the initial cognitive state and the historical cognitive state are fused to obtain the current cognitive state of the target object. The historical static cognitive state is evaluated by aggregating historical multimodal data from multiple historical sessions.

6. The generative content recommendation method for cognitive intervention in elderly people with cognitive impairment according to claim 2, characterized in that, The acquisition of the target object's historical object information and cognitive intervention content set under the target recommendation task includes: Obtain the current interaction content of the target object in response to the target recommendation task, and determine the cognitive training target of the target object in the target recommendation task based on the current interaction content and the current cognitive state; The content recommendation model is used to perform thought chain reasoning on the current cognitive state, the current interactive content, and the cognitive training target to obtain the initial task framework information of the target recommendation task; Based on the initial task framework information, the historical object information and cognitive intervention content set of the target object under the target recommendation task are filtered out from at least one knowledge base.

7. The generative content recommendation method for cognitive intervention in elderly people with cognitive impairment according to claim 6, characterized in that, The step of using the content recommendation model to perform thought chain reasoning on the current cognitive state, the current interactive content, and the cognitive training objective to obtain the initial task framework information for the objective recommendation task includes: The content recommendation model is used to identify the content scenario of the current interactive content, and based on the content scenario, at least one query condition for the knowledge base is determined; Based on the current cognitive state and the cognitive training objective, the interaction configuration information of the objective recommendation task is inferred; Based on the query conditions and the interaction configuration information, a task framework for the target recommendation task is constructed, and the initial task framework information is obtained.

8. The generative content recommendation method for cognitive intervention in elderly people with cognitive impairment according to claim 6, characterized in that, The knowledge base includes an object knowledge base for the target object and a medical knowledge base for the current cognitive state. The step of filtering historical object information and a set of cognitive intervention content for the target object under the target recommendation task from at least one knowledge base, based on the initial task framework information, includes: The query conditions corresponding to the knowledge base are identified in the initial task framework information. The query conditions include object information query conditions and medical content query conditions. Based on the object information query conditions, the object preference information of the target object and the associated content corresponding to the current interaction content are filtered out in the object knowledge base to obtain the historical object information; Based on the medical content query conditions, at least one dimension of task constraints is selected from the medical knowledge base to obtain the cognitive intervention content set.

9. The generative content recommendation method for cognitive intervention in elderly people with cognitive impairment according to claim 6, characterized in that, The step of generating the current task interaction content for the target recommendation task based on the historical object information, the current cognitive state, and the cognitive intervention content set includes: Based on the initial task framework information and the current cognitive state, generate initial interactive prompt information corresponding to the target recommendation task; The historical object information and the cognitive intervention content set are added to the initial interactive prompt information to obtain the interactive prompt information for the target recommendation task; Based on the interactive prompts, the current task interaction content for the target recommendation task is generated using the content recommendation model.

10. The generative content recommendation method for cognitive intervention in elderly people with cognitive impairment according to claim 8, characterized in that, The step of recommending target content to the target object based on the feedback content and the cognitive intervention content set includes: The interaction rounds of the target recommendation task are identified in the initial task framework information; When the interaction rounds are multiple, the interaction prompt information is updated based on the feedback content in order to recommend target content to the target object; When the interaction round is a single round, the target recommended content is generated using the content recommendation model based on the feedback content and the current task interaction content, and the target recommended content is recommended to the target object.

11. The generative content recommendation method for cognitive intervention in elderly people with cognitive impairment according to claim 10, characterized in that, The step of updating the interactive prompt information based on the feedback content to recommend target content to the target object includes: Based on the feedback, the interactive prompt information is updated, and the updated interactive prompt information is used as the interactive prompt information. Return to the step of generating the current task interaction content of the target recommendation task based on the interaction prompt information and using the content recommendation model, until the interaction round is reached, and obtain the target feedback content; Based on the target feedback content and the current task interaction content, the target recommended content is generated using the content recommendation model, and then recommended to the target object.

12. The generative content recommendation method for cognitive intervention in elderly people with cognitive impairment according to claim 11, characterized in that, The step of generating the target recommended content using the content recommendation model based on the target feedback content and the current task interaction content includes: Based on the target feedback content and the current task interaction content, generate recommendation prompts for the target recommended task; Based on preset content generation parameters and the recommendation prompt information, at least one candidate recommendation content is generated using the content recommendation model; The candidate recommended content is filtered to obtain filtered content, and the filtered content is modally transformed to obtain at least one transformed content. The filtered content is merged with the transformed content to obtain the target recommended content, and the target recommended content is recommended to the target object.

13. The generative content recommendation method for cognitive intervention in elderly people with cognitive impairment according to claim 11, characterized in that, Also includes: Extract the object attribute information of the target object from the current task interaction content and the feedback content; The object attribute information is feature-encoded to obtain the incremental object features of the target object; Based on the object attribute information and the incremental object characteristics, the object knowledge base is incrementally updated, and the updated object knowledge base is used as the object knowledge base.

14. The generative content recommendation method for cognitive intervention in elderly people with cognitive impairment according to claim 2, characterized in that, After determining the target recommendation task corresponding to the target object based on the content recommendation request, the method further includes: When the target recommendation task is a recommendation task in an active recommendation scenario, the historical physiological data of the target object within a preset time range and the medical needs information of the current cognitive state are obtained; Based on the target object's current cognitive state, the historical physiological data, and the medical needs information, the content recommendation model is used to generate at least one modality of medical education content corresponding to the medical needs information. The medical education content is used as the target recommended content, and the target recommended content is recommended to the target audience.

15. A generative content recommendation device for cognitive intervention in elderly people with cognitive impairment, characterized in that, include: An acquisition unit is used to acquire multimodal data of a target object performing at least one cognitive assessment task, and to determine the current cognitive state of the target object based on the multimodal data. The receiving unit is configured to, upon receiving a content recommendation request for the target object, determine the target recommendation task corresponding to the target object based on the content recommendation request; The query unit is used to obtain the historical object information and cognitive intervention content set of the target object under the target recommendation task when the target recommendation task is a recommendation task in an interactive recommendation scenario. The generation unit is used to generate the current task interaction content of the target recommendation task based on the historical object information, the current cognitive state, and the cognitive intervention content set. The recommendation unit is used to recommend target content to the target object based on the feedback content and the cognitive intervention content set when it receives feedback content from the target object regarding the current task interaction content.

Images