Model for evaluating and predicting mild cognitive impairment risk of old people in nursing institution

Abstract

The invention relates to a model for evaluating and predicting mild cognitive impairment risk of old people in a pension institution. The model sequentially comprises a behavior analysis module, a language recognition module, a social modeling module, a toughness calculation module, a feature fusion module, a risk reasoning module and the like. Behavior deviation characteristics and abnormal time periods are extracted by collecting behavior data of daily life, diet, social contact and the like of old people and comparing the behavior data with an institution work and rest template; in combination with nursing records, extracting language anomaly features; analyzing social frequency and structure changes in the abnormal time period, and extracting social variation features; a cognitive toughness index is calculated by integrating the health archive and the recovery ability to the health event; and performing toughness weighting on the multi-dimensional features to construct a time sequence tensor, and inputting the time sequence tensor into a recursive model to predict a cognitive impairment risk value. And if the risk value suddenly changes, the system automatically backtracks the feature trajectory of nearly 7 days, constructs and screens a prediction path with the strongest interpretation force, outputs a dominant prediction result and a key factor sequence, and realizes high-interpretability and high-reliability early recognition and intervention reference.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence, and particularly to a risk assessment and prediction model for mild cognitive impairment of the elderly in pension institutions. Background Art

[0002] In recent years, with the continuous deepening of the aging degree of the population in China, the prevalence rate of cognitive impairment in the middle-aged and elderly population has been continuously rising. China has become the country with the largest number of Alzheimer's disease patients in the world, reaching approximately 15.07 million. Mild Cognitive Impairment (MCI), as a reversible stage between normal cognitive aging and Alzheimer's Disease (AD), is a key window period for implementing early prevention and control. Under the background that the Outline of the "Healthy China 2030" proposes to strengthen the early screening and intervention of mental disorders of the elderly and promote the application of technologies such as big data in the health field, the research on MCI risk prediction methods has grown rapidly.

[0003] In the prior art, for the early identification and risk prediction methods of mild cognitive impairment of the elderly, static prediction models are mainly constructed relying on data sources such as structured health records, questionnaire scales, medical images, and a small amount of physiological signals. There are still many limitations in the prior art in the specific scenario of pension institutions, mainly reflected in aspects such as limited data dimensions, insufficient systematic modeling ability for multi-modal information (such as behavior, language, social interaction), lack of interpretability of prediction results, and lack of a dynamic adjustment mechanism for individual differences. Most methods model behavioral deviations at the frequency statistics level, lacking interactive analysis between behavioral changes and language descriptions; unstructured information such as care records has not been effectively utilized; differences in the response capabilities of different individuals to the same risk factors have not been incorporated into the modeling framework; and the prediction results are difficult to trace, lacking factor-level explanations for medical staff to judge and intervene.

[0004] In view of this, there is an urgent need to propose a mild cognitive impairment risk assessment model that is oriented to the pension institution environment, can integrate multi-source heterogeneous data, dynamically evaluate individual cognitive risks, and has an interpretability ability. Summary of the Invention

[0005] The present application provides a risk assessment and prediction model for mild cognitive impairment of the elderly in pension institutions to achieve highly interpretable and personalized dynamic prediction of the risk of mild cognitive impairment of the elderly in the pension institution environment.

[0006] The present application provides a risk assessment and prediction model for mild cognitive impairment of the elderly in pension institutions, including:

[0007] A behavior analysis module, which is used to collect behavior data including the daily living, diet, walking and social interaction of the elderly, compare it with the institutional work and rest template, generate behavior deviation characteristics and mark abnormal behavior periods;

[0008] A language recognition module, which is used to analyze the descriptive statements in the nursing records, and combine the statistical keyword frequency changes during the abnormal behavior periods to obtain language abnormality characteristics and an enhanced abnormal period set;

[0009] A social modeling module, which is used to calculate the individual social frequency and structural changes according to the spatial positioning and voice interaction records in the enhanced abnormal period set, and generate social variation characteristics;

[0010] A resilience calculation module, which is used to calculate the cognitive resilience index based on the health records, behavior deviation characteristics, language abnormality characteristics and social variation characteristics, and combine the reaction speed of the elderly to health events;

[0011] A feature fusion module, which is used to perform weighted processing on the behavior deviation characteristics, language abnormality characteristics and social variation characteristics according to the cognitive resilience index to generate a time series tensor;

[0012] A risk inference module, which is used to output the cognitive impairment risk value through a time-recursive model according to the time series tensor. If the cognitive impairment risk value mutates, it automatically traces back the feature trajectory of the past 7 days, constructs multiple prediction paths and evaluates the interpretability, and selects the optimal path to generate the dominant prediction result and the key factor ranking.

[0013] The present application has the following beneficial technical effects: (1) Through the joint modeling of the behavior analysis, language recognition and social modeling modules, making full use of the daily activities, care records and social data of the elderly in the nursing home, breaking through the limitation of traditional models that only rely on static scales or structured health records, and improving the real-time performance of risk prediction and the breadth of data coverage. (2) By integrating the health background and recovery ability factors through the resilience calculation module, quantifying the individual's resistance ability to the causes of cognitive degradation, effectively reflecting the dynamic adaptability of the elderly individuals, and significantly enhancing the model's hierarchical judgment ability for the risk sensitivity of different populations. (3) When the risk value fluctuates abnormally, the risk inference module automatically triggers the feature trajectory backtracking and prediction path reconstruction mechanism, selects the dominant path output result and the key factor ranking through the interpretability comparison, and provides clear influencing factors and intervention reference basis for medical staff. (4) This model is specially designed for the nursing home environment, can run stably in situations such as complex population interaction, work and rest template standardization, data loss and behavior repetition, has good adaptability and deployment feasibility, and improves the practical application value of early identification of cognitive impairment in institutions. Brief Description of the Drawings

[0014] Figure 1It is a schematic diagram of a risk assessment and prediction model for mild cognitive impairment in the elderly in a pension institution provided by the first embodiment of the present application. Detailed implementation manners

[0015] Many specific details are set forth in the following description to facilitate a full understanding of the present application. However, the present application can be implemented in many other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the connotation of the present application. Therefore, the present application is not limited by the specific implementations disclosed below.

[0016] The first embodiment of the present application provides a risk assessment and prediction model for mild cognitive impairment in the elderly in a pension institution. Please refer to Figure 1 , which is a schematic diagram of the first embodiment of the present application. The following combines Figure 1 to detail the risk assessment and prediction model for mild cognitive impairment in the elderly in a pension institution provided by the first embodiment of the present application.

[0017] The risk assessment and prediction model for mild cognitive impairment in the elderly in the pension institution includes a behavior analysis module 101, a language recognition module 102, a social modeling module 103, a resilience calculation module 104, a feature fusion module 105, and a risk inference module 106.

[0018] The behavior analysis module 101 is used to collect behavior data including the daily living, diet, walking, and social activities of the elderly, and compare it with the institutional schedule template to generate behavior deviation features and mark abnormal behavior periods.

[0019] The behavior analysis module 101 aims to model the daily behavior patterns of the elderly in the pension institution through multi-source data collection and rule modeling, extract behavior deviation features related to mild cognitive impairment therefrom, and further mark abnormal behavior periods to provide input features with high time resolution for subsequent modules.

[0020] In the implementation process, first, the behavior analysis module 101 is configured with collection terminals integrated with the existing infrastructure of the pension institution, mainly including: wearable devices (such as smart bracelets, Bluetooth positioning tags) for collecting data such as walking paths, activity durations, and step frequencies; mattress pressure sensors and infrared night vision devices for recording wake-up and sleep times; cameras and area sensors for recording the times and frequencies of entering and leaving public areas such as rooms and restaurants; RFID tags for tracking diet punching times; if the pension institution has an electronic health record system (EHR), the daily diet intake records can be integrated, including meal times, types, and calorie estimation data.

[0021] After the collected data enters this module through a standardized interface, it is first organized into a daily time series format, divided into 1440 time segments in minutes, and within each segment, a mark indicating the occurrence of a certain behavioral event (such as eating, getting up, walking, social interaction, etc.) is recorded, and a behavioral timestamp sequence is established.

[0022] Subsequently, this module calls the preset daily routine template in the elderly care institution. This template can be provided by the institution management system or automatically generated through historical average behavioral data, covering standard behavioral time periods in daily life (such as getting up at 7:00, having lunch at 12:00, having activities at 15:30, going to bed at 21:30, etc.). The system compares the daily actual behavioral time series with the template one by one to judge the following three types of deviation characteristics:

[0023] (1) Temporal deviation: The occurrence time point of a certain behavior deviates from the template by more than a preset threshold (for example, the deviation is greater than 30 minutes);

[0024] (2) Behavioral absence: The behavior expected by the template does not occur on that day;

[0025] (3) Behavioral repetition or compression: The same behavior occurs repeatedly in a short period of time, or multiple behaviors occur in adjacent time slices, breaking the normal behavioral rhythm.

[0026] The system encodes the above comparison results into a behavioral deviation vector, and each vector element corresponds to a specific behavioral event, including the offset, absence mark, or repetition frequency. For time periods with two or more consecutive anomalies (such as the deviation of eating time and multiple walks in the evening at the same time), the behavior analysis module 101 will automatically mark this time period as a behaviorally abnormal time period.

[0027] In addition, to improve the long-term performance of the model, the behavior analysis module 101 supports recording the statistical indicators of behavioral deviations in the past week through a sliding window mechanism, including the average daily deviation times, the coverage rate of deviation behavior types, the total duration of abnormal time periods, etc., for subsequent cognitive resilience calculation and time series modeling.

[0028] Finally, the behavior analysis module 101 outputs two types of data structures: one is the behavioral deviation feature vector, as a structured input; the other is the set of behaviorally abnormal time periods, marked in the form of time periods (such as 09:15–10:00), providing a constraint range for the language recognition module and the social modeling module. The entire module has the characteristics of low latency, easy deployment, high compatibility, etc., and can adapt to the data collection environments of various elderly care institutions.

[0029] Furthermore, the behavior analysis module is also used for:

[0030] Perform rhythm analysis on the timestamp sequence of the behavioral events of the elderly over consecutive days, calculate the individual dominant rhythm frequency and its variation index through time-frequency transformation, and construct the individual rhythm spectrum; the behavioral events include getting up, eating, using the toilet, walking, and socializing.

[0031] Screen the elderly with normal cognitive status from the same nursing home, construct a collective rhythm reference spectrum, and calculate the rhythm deviation amount of the individual rhythm spectrum relative to the collective reference spectrum.

[0032] Jointly model the rhythm deviation amount with the density and cross-day distribution law of the continuous abnormal behavior period, and dynamically divide the individual's current rhythm state into stable type, slow drift type, mutation type, or structural disconnection type through a set classification function or clustering algorithm.

[0033] Based on the classification result and the rhythm deviation amount, calculate the abnormal behavior rhythm disorder index, and use the abnormal behavior rhythm disorder index as a component of the enhanced behavior deviation feature, and output it uniformly with other behavior deviation features for subsequent module calls.

[0034] In the implementation process, the system first records the key behavioral events in the daily life of the elderly through the behavior analysis module, including but not limited to getting up, eating, using the toilet, walking, and socializing, etc. These events are all attached with accurate timestamps to form behavioral data in a time series format. To ensure the continuity and rhythmic expression of the data, this module will collect and organize the occurrence times of the same behavioral event on different days within a set observation period (such as the recent 7 days or 14 days) on a daily basis, and generate a behavioral timestamp sequence matrix. Each column represents the time records of the same type of behavioral event on multiple dates for subsequent rhythm modeling.

[0035] Subsequently, the system performs rhythm analysis on this timestamp sequence, preferably using time-frequency joint analysis methods such as fast Fourier transform (FFT), continuous wavelet transform (CWT), or empirical mode decomposition (EMD), etc., to extract the main frequency components and frequency domain energy distribution of each behavioral event. This process can extract the periodic indicators reflecting the repeatability and stability of the behavior from the original time records and construct the individual behavior rhythm spectrum. In the spectrum, the dominant rhythm frequency represents the daily repetition period of an individual in a certain behavior (such as whether getting up at approximately the same time every day), and the frequency bandwidth or coefficient of variation represents whether the rhythm is stable.

[0036] To determine whether the rhythm has potential abnormalities, the system further selects the elderly with normal cognitive status and no recent behavioral abnormalities in the same elderly care institution as the control group, and constructs a collective rhythm reference spectrum based on their behavioral event datasets. The reference spectrum is generated in the form of the mean, median, or co-spectrum average of multiple individual dominant frequencies and serves as the standard rhythm template. Then, the rhythm spectrum of the target individual is measured for the frequency domain distance from the collective reference spectrum, and its rhythm deviation is calculated. The deviation can be measured by metrics such as Euclidean distance, KL divergence, or spectral cross-entropy, and the result represents the degree of consistency between the individual and the group rhythm.

[0037] In addition to the frequency domain deviation features, the system also combines the generated information on abnormal behavior periods and extracts features such as the density, span, and coherence of abnormal behaviors on a cross-day scale. For example, it determines whether the abnormal behavior recurs at fixed times or is continuously concentrated starting from a certain day. After jointly modeling the rhythm deviation with these time domain behavioral abnormality indicators, the system classifies the current rhythm state of the individual through a preset classification function or semi-supervised clustering algorithms (such as K-Means, DBSCAN, Gaussian mixture model, or a custom rule set weighted by time windows). The classification results are divided into stable type, slow drift type, mutation type, and structural disconnection type. Among them, the stable type indicates that the rhythm is close to the group standard and the daily routine is regular; the drift type indicates that the rhythm frequency slowly shifts; the mutation type indicates that there are drastic rhythm changes in the short term; and the structural disconnection type represents that the rhythm structure is chaotic or the periodicity basically disappears.

[0038] After obtaining the rhythm classification label and rhythm deviation of the individual, the system further fuses these two indicators into an abnormal behavior rhythm disorder index with the ability of quantitative expression. The design goal of this index is to provide a continuous value ranging from 0 to 1, which is used to represent the comprehensive deviation degree between the individual's behavioral rhythm and the cognitive health baseline. The higher the value, the more serious the disorder, which is convenient for fusing with other behavioral deviation features.

[0039] In the specific implementation, the system first normalizes the rhythm deviation. Let the original rhythm deviation be Δf, which represents the main frequency difference between the individual spectrum and the reference group spectrum, with the unit of hertz. The system statistically calculates the maximum value Δf max and the minimum value Δf min of Δf in the reference group during the training phase. The normalization process uses a linear normalization function:

[0040]

[0041] This normalized Δf norm falls into the interval [0, 1], representing the relative position of the individual rhythm deviation degree.

[0042] Subsequently, the system introduces classification coefficients according to the rhythm classification tags to which the individual belongs, for adjusting the weight of the rhythm deviation value. This coefficient represents the contribution degree of different types of rhythm states to the cognitive risk, and the specific coefficient design is as follows:

[0043] Stable type: weight coefficient α = 0.2

[0044] Slow drift type: weight coefficient α = 0.5

[0045] Mutation type: weight coefficient α = 0.8

[0046] Structural disconnection type: weight coefficient α = 1.0

[0047] These coefficients can be calibrated according to a large amount of clinical follow-up data, and can also be automatically optimized by fitting during the model training stage.

[0048] Finally, the system multiplies the normalized offset value Δf norm by the rhythm classification weight α to obtain the abnormal behavior rhythm disorder index R:

[0049] R = α × Δf norm

[0050] If it is necessary to enhance the non-linear separability of discrete values, a sigmoid activation function or an exponential function can be introduced as a mapping during post-processing. For example:

[0051] R’ = 1 – e –λR

[0052] where λ is a regulation coefficient, and generally takes values of 2 - 5.

[0053] This disorder index R’ is the final output value, which will be concatenated with features such as the behavior offset and the abnormal duration to form a part of the behavior deviation feature vector, and is uniformly used by the feature fusion module as the input of the prediction model.

[0054] Through this design, the present invention can not only objectively model the individual rhythm changes based on timestamp data, but also perform inter-module data transmission in the form of a unified and standardized normalized index, realizing the promotion of the behavior deviation modeling from the point anomaly to the high-order feature of rhythm disorder, and significantly enhancing the early sensitivity and feature expressiveness of the cognitive impairment risk modeling.

[0055] The following is a specific example:

[0056] Taking an elderly Mr. Zhang living in a nursing home as an example, the wearable device he wears records the time of his behavioral events for 14 consecutive days, including the timestamps of getting up, eating, using the toilet, walking, and social behaviors every day. For example, the getting-up time gradually delays from 6:50 on the first day to 7:35 on the 14th day. The lunch time is relatively fixed, but the frequency of using the toilet significantly increases in the early morning, and social activities significantly decrease or are absent in the past three days.

[0057] The system first extracts the timestamp sequence of each type of behavioral event and performs a fast Fourier transform (FFT) on each type of event separately to obtain its main rhythm frequency. For example, the main frequency of the elderly's "getting up" behavior is 0.042 Hz (corresponding to once a day), while the main frequency of the "using the toilet" behavior drifts from the normal 0.125 Hz (about once every 8 hours) to 0.167 Hz (once every 6 hours), indicating that the cycle shortens and becomes frequent. The frequency energy of the "social" behavior decreases concentratedly, reflecting that its activity rhythm has significantly weakened.

[0058] Next, the system selects 30 elderly people with normal cognitive status and regular lifestyles in the nursing home as the reference group and calculates the average main rhythm frequency of the same type of behavioral events. For example, for the "using the toilet" behavior, the collective average frequency is 0.125 Hz, and the standard deviation is 0.015 Hz. The system calculates the rhythm deviation amount Δf between the individual and the group using the absolute value of the frequency difference:

[0059] Δf = |f individual – f reference | = |0.167 Hz – 0.125 Hz| = 0.042 Hz

[0060] The system presets the upper and lower bounds of the rhythm deviation amount of the reference population statistically during the training stage as Δf min =0 Hz, Δf max =0.05 Hz. Then the normalized rhythm deviation value of the elderly Mr. Zhang is:

[0061]

[0062] Meanwhile, the system combines the detection results of abnormal behavior time periods in the behavior analysis module and finds that the elderly Mr. Zhang gets up at night early many times in the recent three days, the time span of using the toilet extends, and social behaviors are significantly reduced. The system automatically classifies his rhythm state as "mutant" through a method based on density clustering (such as DBSCAN), and the classification weight coefficient corresponding to this type is:

[0063] α = 0.8

[0064] Finally, the system generates an abnormal behavior rhythm disorder index R according to the following calculation formula:

[0065] R = α × Δfnorm = 0.8 × 0.84 = 0.672

[0066] If it is necessary to enhance the discrimination of the index, the system can choose to perform a non - linear mapping on this value. For example, the following exponential function form can be used for compression and normalization:

[0067] R’ = 1 – e –λR = 1 – e –3×0.672 ≈ 1 – e –2.016 ≈ 1 – 0.133 = 0.867

[0068] where λ = 3 (empirical value)

[0069] Finally, the index R’ = 0.867 indicates that the recent behavior rhythm of the elderly with the surname Zhang has a relatively strong degree of disorder. This value will be used as one of the enhanced behavior deviation features, and together with other features such as behavior omission rate and abnormal behavior span, they will be jointly spliced to form a time - series feature vector and transmitted to the feature fusion module and the risk inference module.

[0070] Through the comprehensive implementation of the above steps, the behavior analysis module not only has the ability to capture short - term behavior anomalies, but also introduces a rhythm anomaly expression mechanism combining time and frequency, thereby constructing a more continuous, interpretable and predictive behavior deviation feature dimension, which helps to identify dynamic risks of mild cognitive impairment earlier and more accurately.

[0071] The language recognition module 102 is used to analyze the descriptive statements in the nursing records and obtain the language anomaly features and the enhanced anomaly time - period set by combining the statistical frequency changes of keywords during the abnormal behavior periods.

[0072] The language recognition module 102 is used to mine the language signals of possible mild cognitive impairment in the elderly from the nursing records of the elderly care institution, and perform time - series enhancement analysis by combining the abnormal behavior information, so as to form language anomaly features and output the structured data and time - period set that can be used for subsequent risk modeling. This module can process the text descriptions, voice inputs or voice texts after transcription input by the nursing staff daily, and has natural language processing capabilities, abnormal keyword extraction capabilities and time - period matching functions.

[0073] In implementation, the nursing system of a pension institution usually stores daily nursing logs, including text materials such as regularly filled nursing records, handover instructions, and care observation reports. The language recognition module 102 first accesses these text sources, unifies the text encoding format, and performs cleaning and preprocessing to remove table tags, invalid symbols, and redundant template content. Then, it classifies and organizes them according to the elderly individuals, dates, and time sequences to establish a multi-day nursing text sequence. For the voice input content, it can be transcribed through an integrated speech recognition engine (such as an ASR model based on the CTC or Transformer structure) to generate standard text for subsequent analysis.

[0074] Subsequently, the module calls the preset descriptive semantic vocabulary of the present invention, which is jointly constructed by medical experts and language models and covers expressions related to cognitive impairments such as memory problems, language confusion, distractibility, repetitive behaviors, and abnormal emotions, such as "asking repeatedly", "forgetting to eat", "speaking intermittently", "not being able to find the room", "not remembering what happened yesterday", "topic jumping", "excited", etc. The language recognition module 102 uses a method combining rule matching and machine learning to extract the above keywords and their variant expressions from the text, and counts their occurrence frequencies and context correlations in a single nursing record. If a keyword appears multiple times in the text or is concentrated in the dates corresponding to abnormal behavior periods, the language signal can be determined as a potential abnormal language feature.

[0075] To improve the temporal consistency and prediction correlation, the module further uses the "abnormal behavior period" marked by the behavior analysis module 101 as a time reference to compare the occurrence positions and timestamps of the keywords. If the keywords are concentrated in a certain time window (such as ±1 hour) before or after the abnormal behavior time period, then this period is strengthened as a "language-behavior linkage abnormal period". The module records these strengthened period sets and outputs them as a list of time periods (such as 10:30–11:15 on May 10, 2024, 16:00–16:45 on May 12) for use by the subsequent social modeling module.

[0076] Finally, the language recognition module 102 outputs two types of data: one is the language abnormal feature vector at the individual dimension of the elderly, including keyword categories, frequencies, change trends, etc.; the other is the set of strengthened abnormal periods. The deployment of this module does not depend on a complex medical corpus system and can be integrated into the existing nursing record platform of a pension institution, with good adaptability and feasibility. Through the above design, the module realizes extracting cognitive abnormality clues from care texts and combining them with behavioral data, enhancing the contribution of language signals in the risk prediction process.

[0077] Furthermore, the language recognition module is also used for:

[0078] When analyzing the descriptive statements in the nursing records, a semantic emergence matrix is constructed with the abnormal behavior period of an individual as the time anchor. The semantic emergence matrix records the emergence frequency of a specified semantic category during the abnormal period and the cross-category co-occurrence density.

[0079] Through a preset dictionary of semantic domains for cognitive impairment, the emerging words are classified into four high-risk semantic domains: memory impairment, mood swings, language disconnection, and life disorder, and the semantic category weight vector during the corresponding abnormal period is calculated.

[0080] Combining the nursing descriptive statements of multiple adjacent elderly people within the same time period, a context cross-reference vector group is constructed, and the target semantics with independent emergence characteristics are screened through similarity analysis, so as to exclude the noise expressions caused by collective behavior or nursing record habits.

[0081] According to the semantic category weight vector and the cross-reference result, a normalized language risk intensity score is generated and used as the semantic risk expression part in the language abnormality characteristics.

[0082] This module is used to extract the abnormal language features reflecting the cognitive state of the elderly from the unstructured descriptive statements written by the nursing staff in the elderly care institution, and construct a quantitative index for semantic risk expression, so as to provide higher-level interpretive information for subsequent risk assessment.

[0083] During the operation of the system, first, the behavior analysis module marks the abnormal behavior periods of the elderly every day, such as chaotic daily life, frequent nocturnal activities, disordered meal times, etc. After receiving the information of the abnormal period, the language recognition module extracts the observational descriptive statements recorded by the nursing staff within the same time range from the nursing information system according to the corresponding time anchor. The system aggregates each statement by day and establishes a semantic analysis window with the time anchor as the core. For the text within each window, the system first performs basic natural language processing operations such as word segmentation, part-of-speech tagging, entity recognition, and subject-predicate relationship extraction.

[0084] Based on this, the system constructs a semantic emergence matrix. The rows of the matrix represent semantic categories (such as cognition-related behaviors, language performance, emotional states, life rhythms, etc.), the columns represent time units within the abnormal period (such as every 30 minutes), and each cell in the matrix records the number of keywords identified in this semantic category during this time period and its relative frequency change. At the same time, the system also counts whether the keywords belonging to different semantic categories co-occur in the same time unit, and the co-occurrence frequency will be used as the cross-category co-occurrence density and input into the emergence matrix to identify the time points when multiple abnormal signals break out simultaneously.

[0085] After constructing the emergence matrix, the system calls a preset dictionary of semantic domains for cognitive impairment, which is jointly constructed by medical linguistics experts and engineering algorithms. Based on the early language manifestations of cognitive impairment, common expressions are collected and divided into four high-risk semantic domains, namely memory impairment (such as "forgot", "asked repeatedly", "couldn't remember clearly"), mood swings (such as "excited", "suddenly cried", "depressed"), language disconnection (such as "couldn't explain clearly", "jumped to a topic", "couldn't continue"), and life disorder (such as "mistook someone for another", "went to the wrong room", "didn't eat"). The system compares the words identified in the emergence matrix with the semantic domain dictionary, classifies the words according to the matching results, and counts the proportion, emergence frequency, and intensity change of high-frequency words in each semantic domain to form a semantic category weight vector for each elderly person during this abnormal period. For example, during the period from 10:00 to 12:00 on May 10th, descriptions such as "She asked the time to eat three times", "He said it again right after he finished", and "She didn't eat today" were identified. Among them, three high-risk semantic words, namely "asked repeatedly", "repeated", and "didn't eat", are involved, which can be classified into the semantic domains of memory impairment, language disconnection, and life disorder respectively. The corresponding weights will be weighted according to the appearance frequency and density.

[0086] To further eliminate the expression deviation caused by collective nursing operations or nursing language habits, after constructing the semantic category weight vector, the system introduces a context cross-reference mechanism. This mechanism extracts sentences from the nursing records of other adjacent elderly people within the same time period to construct multiple cross-reference vector groups. The system uses semantic similarity analysis algorithms (such as BERT-based embedded vector cosine distance, word vector clustering, semantic matching network, etc.) to compare the similarity between the emergence semantic vector of the target elderly person and the semantic vectors of other elderly people. If some keywords co-occur frequently in the descriptions of multiple elderly people, it is initially judged that the word may originate from the nursing record template, collective activities, or care situation, and does not have individual specificity. The system suppresses the weights of these "group background semantics", retains the words that only emerge in the records of the target elderly person and have unique context expressions, and adjusts their weights to form a screened "independent emergence semantic set".

[0087] Finally, the system comprehensively calculates the semantic category weight vector and the independent emergence semantic set to generate a normalized language risk intensity score R. The calculation method of this score can adopt the following formula:

[0088]

[0089] Among them, α i is the semantic category weight (such as memory impairment = 1.0, language disconnection = 0.8, etc.), w i is the emergence frequency of keywords in this category, s iThe keyword context consistency score (e.g., only the high score appears among the target elderly), and Z is the normalization factor, so that the score value finally falls into the [0,1] interval. The higher the score R, the more significant the language risk of the elderly in the abnormal period. This score will be used as the "semantic risk expression" dimension in the language abnormality feature, integrated with the original keyword frequency, text length, emotional tendency and other features, and input into the subsequent cognitive resilience calculation and risk reasoning model.

[0090] Through the above implementation method, the language recognition module can not only identify potential cognitive risk expressions in specific sentences, but also construct highly reliable and highly interpretable quantitative language features through steps such as corresponding to behavioral time periods, semantic domain modeling, and contextual cross-noise elimination, thereby significantly improving the sensitivity and predictive interpretability of the present invention to changes in early cognitive impairment risks.

[0091] The social modeling module 103 is used to calculate the individual social frequency and structural changes and generate social variation characteristics based on the spatial positioning and voice interaction records in the abnormal time period set after reinforcement.

[0092] The social modeling module 103 is used to analyze the social behavior structure and change trend of the elderly in a specific period of time based on the spatial activity trajectory and voice interaction data of the elderly in the nursing home environment, and to extract social variation characteristics related to the risk of mild cognitive impairment. The main goal of this module is to fill the gap in the traditional health assessment of the lack of understanding of the "social interaction status" of the elderly, and to mine potential signals of early cognitive deterioration through group interaction information.

[0093] In implementation, the social modeling module 103 first receives the "abnormal time period set after reinforcement" jointly output by the behavior analysis module 101 and the language recognition module 102. This set identifies the time period when behavioral abnormalities and language abnormalities may co-occur, and has a high risk signal density. The social modeling module uses these time periods as the core observation window and calls the spatial positioning and voice interaction records for analysis. Spatial positioning data can be obtained through wearable Bluetooth tags, RFID devices, or Wi-Fi / BLE positioning systems, and is used to mark the time and space trajectories of the elderly entering and staying in various public areas (such as restaurants, activity rooms, corridors, nursing stations, etc.) in real time. Voice interaction data is obtained through an embedded microphone array or a voice recognition module to detect language communication events between the elderly and others, including the number of speeches, interaction duration, and the number of participants.

[0094] The system aligns through timestamps, integrates the activity trajectories and voice data of a certain elderly person during an abnormal period, and constructs a social behavior graph for that period. Each node in the graph represents an interaction object, the edge represents the interaction relationship, and the weight of the edge can be determined by the interaction duration or voice frequency. By comparing this graph with the historical normal social patterns of the elderly person, the following change indicators can be extracted: a decrease in the total social frequency, a reduction in the number of social objects, a change in the core interaction objects, a decrease in the interaction concentration, an increase in the voice response delay, etc. The above indicators constitute multiple dimensions of the social variation characteristics. Each dimension can be numerically represented and form a structured feature vector, reflecting the change trends of an individual's stability, initiative, and response ability at the social level.

[0095] To enhance the time sensitivity of social modeling, the module can also model the continuity of social behaviors over a period of time, such as analyzing a continuous reduction in social activities over multiple days within the same time period, or the occurrence of a social silence period before and after a high-incidence day of language abnormalities. The system can update the social variation trend daily through a sliding window mechanism to capture weak but continuous degradation signals.

[0096] Finally, the social modeling module 103 outputs the social variation feature vector of each elderly person within the target time window. This vector will be transmitted to the resilience calculation module and the feature fusion module, serving as an important input for evaluating cognitive risks. The structure of this module is flexible, adaptable to various indoor positioning and voice acquisition technologies, and has high implementability. It is especially suitable for deployment in the environment of elderly care institutions and has practical engineering implementation value. Through the systematic analysis of social activity patterns, this module significantly improves the behavioral dimension integrity of cognitive impairment risk modeling and the accuracy of prediction results.

[0097] Furthermore, the social modeling module is also used for:

[0098] Based on the spatial positioning and voice interaction data within the enhanced abnormal period set, construct a social trajectory graph of the target elderly person for multiple consecutive days, and map this graph to the structure vector space respectively in three dimensions: social intensity, contact persistence, and interaction object diversity;

[0099] Construct a social isomorphism feature baseline graph of the elderly population with normal cognitive status within the same elderly care institution, and project the current social trajectory vector of the target individual into the isomorphism space of this baseline graph through the minimum distance matching algorithm to obtain a social path offset vector;

[0100] According to the amplitude, direction, and stability of the social path offset vector in the structure space, calculate the abnormal social path offset rate, and combine the voice participation frequency and spatial stay distribution during the abnormal period to generate a comprehensive social disorder index, which is output as part of the social variation characteristics.

[0101] In the present invention, the social modeling module not only focuses on the social frequency of the elderly with others, but also constructs a complete social trajectory graph starting from the structure, object, and spatio-temporal characteristics of the interaction. This trajectory graph is generated based on the spatial positioning data and voice interaction records within the enhanced abnormal time period set. In practical applications, each elderly person living in a nursing home is equipped with an indoor positioning tag, and voice pickup microphones are set in specific areas to capture voice interaction events.

[0102] The system first binds each voice interaction to specific spatial coordinates according to the time stamp. For example, an elderly person surnamed Zhang communicated with three other elderly people in the second-floor public reading area from 9:15 to 9:45 on the morning of May 10. The conversation lasted for 18 minutes, and during this period, the proportion of his speech was 45%. The system records this interaction as an event node with a time period, location, participants, voice intensity, and response rate.

[0103] By connecting all the social events of the elderly person surnamed Zhang within three consecutive days, the system constructs his social trajectory graph. In this graph, each node represents a social object or scenario (such as group activities, individual chats, conversations with caregivers), and the weight of each edge represents the interaction intensity between the elderly person surnamed Zhang and this node within a specific time window. The definition of the interaction intensity can be the total daily conversation duration, or a weighted value of the conversation frequency and the voice participation ratio. For example, if the elderly person communicates with object A 3 times, for a total of 30 minutes, and the voice participation rate is 50%, the system will calculate the weight of this edge as 0.5×30 = 15 score units.

[0104] Next, the system extracts three dimensions of social characteristics: social intensity refers to the sum of the weights of all edges, reflecting the overall social activity level of an individual; contact persistence is the concentration of interactions between the same objects, such as the proportion of the same object in the total interaction time; the diversity of interaction objects is equal to the number of independent nodes in the social graph divided by the total possible number of nodes.

[0105] Assume that the current three index values of the elderly person surnamed Zhang are: social intensity 16 (unit: minute·participation rate), contact persistence is 0.73 (that is, more than 70% of the interactions are concentrated on 2 acquaintances), and the diversity of interaction objects is 0.33 (3 objects / 9 permanent residents in the institution). The system normalizes these three values to form his social structure vector, denoted as {0.58, 0.73, 0.33}.

[0106] The system has previously collected social behavior data of cognitively normal elderly people in this institution for a long time and constructed a set of social baseline vectors. For example, the common value of social intensity is 22 - 30, the normal range of diversity is 0.55 - 0.75, and the contact persistence is generally lower than 0.55, indicating that normal people's social interactions are more balanced and they have a more diverse range of interaction partners. The system uses the Euclidean distance to match the social vector of the elderly person with the surname Zhang with multiple vectors in the baseline, and takes the closest reference vector as {0.82, 0.54, 0.68}. The difference between the two constitutes the social path deviation vector, that is:

[0107] Difference = Current vector - Baseline reference vector

[0108] ={0.58 - 0.82, 0.73 - 0.54, 0.33 - 0.68}

[0109] ={ - 0.24, + 0.19, - 0.35}

[0110] This deviation vector reflects that the social intensity of the elderly person with the surname Zhang has decreased, the diversity has decreased significantly, while the concentration of interaction has increased abnormally. The system calculates the modulus length of its vector (i.e., the deviation amplitude) as:

[0111]

[0112] If it exceeds the deviation threshold of 0.4 for three consecutive days, the system determines that the social pattern of this elderly person is stably deviated from the group standard.

[0113] Furthermore, the system combines the voice interaction records to calculate the change in the voice participation frequency during the abnormal period. If the average daily active speaking time of the elderly person with the surname Zhang in the same space is 12 minutes, but it drops to 2 minutes during the abnormal period, and at the same time the spatial positioning data shows that most of the areas where he stays are concentrated inside the room rather than public areas, then the system judges that his social withdrawal has language inhibition and spatial enclosure.

[0114] These indicators will be uniformly normalized and combined to generate a comprehensive social disorder index. It can be in the following form:

[0115] Social disorder index = α × Deviation amplitude + β × Voice participation decline rate + γ × Spatial stay deviation rate

[0116] Where α, β, and γ are the weighted factors after system optimization. For example, the deviation amplitude is 0.464, the voice participation drops by 80%, and the spatial deviation is 60%. If the weighted factors are 0.4, 0.3, 0.3, then the calculation result is:

[0117] Social disorder index = 0.4 × 0.464 + 0.3 × 0.8 + 0.3 × 0.6 = 0.1856 + 0.24 + 0.18 = 0.6056

[0118] If the imbalance index exceeds 0.6, the system records it as a moderate to severe social abnormality state and outputs this value as part of the social variation characteristics for reference by the cognitive impairment risk model.

[0119] Through the above processing flow, the present invention can not only identify the "decrease in social volume", but also distinguish the "restructuring of social structure", the "abnormality of object concentration", and the "covariation of social space and speech characteristics". This social deviation determination mechanism that integrates spatial trajectory, interaction network, and behavioral language significantly improves the interpretability and prediction timeliness of cognitive impairment risk assessment.

[0120] The resilience calculation module 104 is used to calculate a cognitive resilience index based on the health record, behavioral deviation characteristics, language abnormality characteristics, and social variation characteristics, and in combination with the response speed of the elderly to health events.

[0121] The core purpose of the resilience calculation module 104 is to evaluate the adaptation and recovery ability of the elderly in the face of risk factors related to cognitive impairment, so as to construct a dynamic adjustment factor, that is, the cognitive resilience index, for adjusting the influence weights of various abnormal characteristics in the subsequent model. By integrating the individual's basic health condition and the changing trends of current behavioral, language, and social characteristics, this module calculates a quantitative index reflecting the level of cognitive resistance, enhancing the personalized adaptation ability of the model.

[0122] In the actual implementation process, the resilience calculation module first receives the structured health record information of the elderly, including but not limited to previous diagnosis results, cognitive ability scale scores (such as MMSE), years of education, self-care ability assessment scores, previous history of stroke or cardiovascular disease, long-term medication records, and their corresponding treatment responses. The system standardizes this information and extracts static health characteristics related to cognitive stability, such as cognitive function score levels, educational attainment stratification, self-care ability classification, etc.

[0123] At the same time, this module also receives the behavioral deviation characteristics, language abnormality characteristics, and social variation characteristics output by the behavior analysis module, language recognition module, and social modeling module respectively. To avoid relying only on instantaneous abnormal signals, the system uses a sliding window mechanism to statistically analyze the fluctuation amplitude, duration, and fluctuation frequency of the above characteristics within the past week, generating a comprehensive dynamic risk stress factor.

[0124] Next, the resilience calculation module calculates the response speed of the elderly to risk fluctuations. Specifically, it combines historical data on emergency events (such as hospitalization, falls, infections, etc.) in the past health records with the subsequent daily behavior recovery trajectories, and evaluates the time, frequency, and amplitude of returning to normal daily routines after stress events. For example, if an individual resumes regular eating and activity rhythms within 48 hours after a fall, their recovery speed score is relatively high. This score participates in the comprehensive calculation of the cognitive resilience index as a dynamic recovery ability characteristic.

[0125] The system standardizes the static health characteristics, behavior anomaly trends, language fluctuation trends, social activity changes, and recovery speed scores respectively and inputs them into a linear weighted function or a multivariable regression model. According to the pre-set weight coefficients or model training parameters, it outputs the final cognitive resilience index. The value range of this index can be set as [0, 1], where 0 indicates being extremely sensitive to risk factors, and 1 indicates having a high buffering ability.

[0126] As a high-level dynamic adjustment factor, the cognitive resilience index does not directly replace specific behavior or language characteristics, but will be used as a control parameter for adjusting weights in the feature fusion module to adaptively adjust the importance of different types of abnormal signals. Through the design of this module, the adaptability of the entire prediction model to individual differences can be significantly enhanced, thereby improving the accuracy and clinical interpretation value of risk assessment, and ensuring its good deployability and practicality in the context of elderly care institutions.

[0127] Furthermore, the resilience calculation module is also used for:

[0128] After each behavior anomaly period, language anomaly period, or social anomaly period detected in the behavior deviation characteristics, language anomaly characteristics, and social variation characteristics, identify the continuous time period during which the behavior deviation characteristics, language anomaly characteristics, and social variation characteristics fall back within their individual historical fluctuation ranges and remain stable, as the corresponding recovery window;

[0129] For each pair of anomaly-recovery pairs composed of an anomaly period and its subsequent recovery window, calculate the response recovery offset value, which is defined as the product of the sum of the standardized deviation values of the behavior deviation characteristics, language anomaly characteristics, and social variation characteristics relative to their respective historical means during the anomaly period and the time interval from the end of the anomaly period to the start of the recovery window;

[0130] Perform a weighted average on the set of response recovery offset values of all anomaly-recovery pairs to obtain the average response recovery offset value, and use its reciprocal as the individual's recovery efficiency index, which is used to measure the overall response ability of the individual's self-repair behavior after multiple anomalies;

[0131] The recovery efficiency index is jointly modeled with the number of chronic disease interventions, medication compliance score, and historical mood fluctuation cycle index recorded in the health record of the elderly person to generate a cognitive resilience dynamic adjustment factor. The cognitive resilience dynamic adjustment factor is used to participate in the subsequent calculation of the cognitive resilience index and serve as the adjustment basis for weighting the behavioral deviation characteristics, language abnormality characteristics, and social variation characteristics.

[0132] In the context of a senior care institution, the system has respectively generated behavioral deviation characteristics, language abnormality characteristics, and social variation characteristics through a behavior analysis module, a language recognition module, and a social modeling module. These characteristics are all recorded in the form of time series, and a series of abnormal time periods have been marked in the previous processing stage. For example, the elderly person surnamed Zhang showed abnormal behavioral rhythm fluctuations, slow language expression, and a sharp drop in social frequency during the period from 08:00 to 12:00 on May 10, 2025. Therefore, the system marked this time period as a comprehensive abnormal time period.

[0133] The resilience calculation module first retrospectively examines each abnormal time period one by one. In the subsequent time series, it detects whether the three characteristics simultaneously fall back within the range of their respective individual historical fluctuation intervals. Specifically, the system calls the time series data of each characteristic in the past 14 days, calculates the individual median value and standard deviation of the characteristic, and defines the "individual historical fluctuation interval" as the range of the median value plus or minus one standard deviation. If, within a certain time period, the values of these three characteristics all stably fall within their corresponding intervals and the duration reaches a preset threshold (such as 30 minutes or 3 time slices), the system marks this time period as a recovery window. Continuing with the above example, if the three characteristics of the elderly person surnamed Zhang return to the stable interval between 13:00 and 15:00 on May 10 and the fluctuation range is small, the system can determine that this time period is the recovery window after this anomaly.

[0134] Next, the system regards each pair composed of an "abnormal time period - recovery window" as an abnormal recovery pair and conducts a quantitative calculation on it. The system first calculates the maximum standardized deviation value of each of the three characteristics within the abnormal time period, that is, the maximum deviation of each characteristic from its historical mean within this time period. After processing using the z - score standardization method and summing them up, it serves as the abnormal intensity value of this abnormal time period. Then, it calculates the time interval between the end time of the abnormal time period and the start time of the recovery window, and the unit can be minutes or the number of time slices. The product of this abnormal intensity value and the time interval is defined as the response recovery offset value. Taking the elderly person surnamed Zhang as an example, if the abnormal intensity is 2.5 and the recovery time required is 90 minutes, then the response recovery offset value for this time is 225.

[0135] After completing the calculation of multiple groups of anomaly recovery pairs, the system forms a set of all response recovery offset values, assigns a weighting factor based on the severity of each anomaly, and finally calculates the weighted average offset value. The recovery efficiency index is the reciprocal of this weighted average offset value. The higher this value, the faster the elderly person recovers from multiple abnormal states and the stronger their cognitive-behavioral self-regulation ability. This index will be further used in the subsequent adjustment mechanism of the resilience index.

[0136] In addition, to achieve differential modeling of an individual's long-term state, the system also incorporates individual health record data into the modeling. This includes the number of chronic disease interventions (such as diabetes management records), medication adherence scores (which can be statistically obtained from an electronic medication record system, such as calculated by the daily check-in rate), and the emotional fluctuation cycle index extracted by the system based on the historical amplitude and cycle of emotional states. After normalizing and encoding the above static health factors, they are jointly input into the modeler with the above-mentioned recovery efficiency index, and a dynamic adjustment factor for cognitive resilience can be generated using a regression function or a weighted rule method.

[0137] This dynamic adjustment factor for cognitive resilience will serve as the weighting basis for behavioral deviation features, language anomaly features, and social variation features, and dynamically participate in the final calculation of the cognitive resilience index. The introduction of the adjustment factor enables the risk prediction results generated by the same behavioral anomaly among individuals with different cognitive tolerances to be organically distinguished, significantly enhancing the individual adaptability and clinical interpretability of the prediction model.

[0138] For example, in a nursing home, assume that an elderly person surnamed Zhang showed significant abnormal behaviors on the morning of May 10, 2025: he got up one hour later than usual, refused breakfast, and showed obvious silence and avoidance in the next two hours, hardly communicating with other residents. The nursing record described his state during this period as "sitting by the window many times, unwilling to eat, and not responding to greetings from other elderly people". The system's behavior analysis module recorded a deviation in the daily routine rhythm, interruption of diet, and reduction in activities. The language recognition module detected an increase in negative keywords in the descriptive statements, and the social modeling module also marked the absence of effective social interactions during this period. Based on this, the system marked the period from 8 am to 11 am on this day as a comprehensive abnormal period.

[0139] The resilience calculation module then begins to search for the recovery window after this abnormal period. The system statistically analyzes the behavior deviation characteristics, language anomaly characteristics, and social variation characteristics of the elderly Mr. Zhang in the past 14 days, and calculates the individual median and standard deviation of each characteristic respectively. After 1 pm on May 10th, the behavior deviation characteristics of the elderly Mr. Zhang gradually declined, and the language interaction gradually recovered. He began to have short conversations with his familiar roommates and participated in the daily calligraphy group at 2:30 pm. The system found that from 1 pm to 4 pm, all three characteristics returned to the fluctuation range of "median ± one standard deviation" and remained relatively stable, so this time period was identified as the recovery window after the anomaly in the morning of May 10th.

[0140] Next, the system calculates the response recovery offset value of this abnormal recovery pair. The maximum standardized deviation values of the three characteristics of the elderly Mr. Zhang during the abnormal period in the morning are 2.2 (behavior deviation), 1.8 (language anomaly), and 1.5 (social variation) respectively. After standardization and summation, the total deviation value is 5.5. There is a 2-hour delay, that is, 120 minutes, from the end of the abnormal period to the start of the recovery window. The system multiplies the deviation value by the time interval to obtain the response recovery offset value of 5.5 multiplied by 120, which is 660.

[0141] Suppose the system has detected 5 similar abnormal recovery pairs of the elderly Mr. Zhang in the past 30 days, and the response recovery offset values are 420, 510, 600, 580, and 660 this time respectively. The system calculates the weighted average after assigning weights to each calculation according to the anomaly intensity, and obtains an average response recovery offset value of about 550. The system uses its reciprocal to obtain the recovery efficiency index, that is, 1 divided by 550, which is approximately equal to 0.0018. This value is considered to represent the recovery ability of the elderly Mr. Zhang to the disturbance of the life rhythm. The smaller the value, the better the ability to recover from the anomaly.

[0142] The system also further retrieves the health record data of the elderly Mr. Zhang. He is recorded as a diabetic patient and has received 5 structured management interventions for chronic diseases in the past year. His electronic medication record shows that the medication compliance score is 85%, and the emotional fluctuation cycle is about a medium fluctuation state every 10 days. The system normalizes these variables into comparable values respectively. For example, the intervention frequency is 0.5, the compliance is 0.85, and the emotional fluctuation intensity is 0.4. The system combines these values with the recovery efficiency index and inputs them into a weighted modeling function (such as support vector regression or rule tree model), and finally generates a normalized cognitive resilience dynamic adjustment factor, and the result is about 0.67.

[0143] The cognitive resilience dynamic adjustment factor will be input into the subsequent cognitive resilience index calculation module together with the behavioral deviation characteristics, language anomaly characteristics, and social variation characteristics of the elderly Zhang. This module will use the adjustment factor as the basis for feature weighting to reflect the individual differential adaptation ability of this elderly person in the face of cognitive function fluctuations.

[0144] Take a specific example. On May 11th, the behavioral deviation score of the elderly Zhang was 0.7, the language anomaly score was 0.4, and the social variation score was 0.6. These three scores respectively represent the anomaly intensity in each aspect on that day (the numerical range is uniformly normalized to 0 to 1). If the cognitive resilience dynamic adjustment factor of this elderly person is 0.67, the system will introduce an adjustment function in the calculation of the cognitive resilience index. For example, using the following formula:

[0145] Weighted score of behavioral deviation = original behavioral deviation score × (1 + α × adjustment factor)

[0146] Weighted score of language anomaly = original language anomaly score × (1 + β × adjustment factor)

[0147] Weighted score of social variation = original social variation score × (1 + γ × adjustment factor)

[0148] Among them, α, β, and γ are respectively the feature sensitivity parameters set in the system training stage. For example, set as α = 0.5, β = 0.3, γ = 0.4. The calculation process is as follows:

[0149] Weighted score of behavioral deviation = 0.7 × (1 + 0.5 × 0.67) ≈ 0.7 × 1.335 ≈ 0.934

[0150] Weighted score of language anomaly = 0.4 × (1 + 0.3 × 0.67) ≈ 0.4 × 1.201 ≈ 0.480

[0151] Weighted score of social variation = 0.6 × (1 + 0.4 × 0.67) ≈ 0.6 × 1.268 ≈ 0.761

[0152] Through such adjustment calculations, the original behavioral deviation score is increased from 0.7 to 0.934, the social variation score is increased from 0.6 to 0.761, and the language anomaly score is slightly increased. This result reflects that the system actively increases the contribution weights of various anomaly characteristics according to the relatively weak recovery efficiency of the elderly Zhang (the adjustment factor is 0.67, not close to 1), believing that the elderly person has an average ability to recover from the abnormal state. Therefore, the same intensity of anomaly has a higher risk significance for him.

[0153] Conversely, if the adjustment factor of another elderly person is 0.2, that is, his recovery ability is strong, then the same behavioral deviation score of 0.7 will only be weighted to:

[0154] Behavior deviation weighted score = 0.7×(1 + 0.5×0.2) = 0.7×1.1 = 0.77

[0155] This means that the system automatically reduces the impact of its anomalies on the overall risk assessment. Finally, these weighted scores will be fed into a time series modeling structure (such as a time recurrent network) together to participate in the calculation of the cognitive impairment risk value.

[0156] Through such a mechanism, the system no longer treats the behavior anomaly scores of all the elderly equally with a fixed formula, but adjusts the weights individually according to their actual resilience ability, enabling the prediction model to not only quantify the anomaly intensity but also reasonably explain the clinical phenomenon that "the impacts of anomalies on different people are not equal", greatly enhancing the interpretability and practicality of the model.

[0157] The feature fusion module 105 is used to perform weighted processing on the behavior deviation features, language anomaly features, and social variation features according to the cognitive resilience index to generate a time series tensor.

[0158] The feature fusion module 105 aims to structurally integrate the multi-modal features output by the previous modules, perform weighted processing on various features in combination with the cognitive resilience index, and finally generate a unified input tensor that meets the requirements of the time series structure as the input data for the cognitive impairment risk prediction model. This module is not only an information convergence point for multiple heterogeneous feature paths but also a key processing unit for realizing personalized dynamic adjustment, and its processing result directly determines the effectiveness and interpretability of the subsequent prediction model.

[0159] In the implementation process, this module first receives the behavior deviation features from the behavior analysis module, the language anomaly features from the language recognition module, and the social variation features from the social modeling module. These features may have different time dimensions, value types, and scale ranges in terms of structure. To achieve unified integration, the feature fusion module first performs standardization and time alignment processing on the three types of features. Standardization can adopt the maximum-minimum normalization or the Z-score method to map the input features with different dimensions to the interval [0,1] or the interval with a mean of 0 and a variance of 1. Time alignment is based on a unified time step (such as daily or hourly), and various features are resampled into a time series on the same time axis through interpolation or a sliding window method.

[0160] Subsequently, the module calls the cognitive resilience index output by the resilience calculation module to dynamically weight the above three types of standardized features. Specifically, the module presets three groups of weighting functions or weight parameters, and these weight coefficients can be adjusted according to the value of the cognitive resilience index. For example, when the cognitive resilience index of an elderly person is low (indicating poor resistance to cognitive risks), the system can automatically increase the weights of behavioral deviation and language abnormality features; while when the resilience index is high, the system may moderately reduce the importance of abnormal features to avoid misjudgment caused by short-term fluctuations. The weighting method can be channel-level weighting (i.e., the whole category of features is adjusted according to a unified coefficient), or element-level weighting (i.e., each dimension is refined and weighted according to the feature importance map). In actual deployment, linear weights, attention mechanisms, or gating functions can be selected as implementation methods according to the training data.

[0161] For example, the elderly person surnamed Zhang is a case user in a nursing home. The system extracts three types of features from his behavior monitoring device, voice collection system, and indoor positioning system every day: behavioral deviation features, language abnormality features, and social variation features. All original data will be standardized before entering the fusion module, so that these three types of features can be compared and weighted under a unified dimension.

[0162] On May 12th, the system completed the data collection of the elderly person surnamed Zhang and called the resilience calculation module to update the cognitive resilience index. The module calculated the cognitive resilience index for that day as 0.33 by combining factors such as the recovery speed of the behavioral deviation of the elderly person surnamed Zhang in the past week, the change trend of the language state, and the intervention response recorded in the health record. This value indicates that the elderly person surnamed Zhang has poor self-regulation ability for cognitive risks recently.

[0163] After receiving this value, the fusion module will judge that the index is lower than the preset threshold (such as 0.5), and the system will then select the "high-sensitivity weighting template" and start dynamically weighting the three types of standardized features of this elderly person.

[0164] At this time, the three types of features of the elderly person surnamed Zhang are input as follows:

[0165] The standardized behavioral deviation feature is: 0.62

[0166] The standardized language abnormality feature is: 0.47

[0167] The standardized social variation feature is: 0.51

[0168] According to the "high-sensitivity" strategy, the system increases the importance of the behavioral deviation feature by 1.5 times, the language abnormality feature by 1.3 times, and keeps the social variation feature unchanged to avoid over-sensitivity.

[0169] The input features after dynamic weighting become:

[0170] Behavior deviation: 0.62 × 1.5 = 0.93 (system processes as key observation)

[0171] Language anomaly: 0.47 × 1.3 = 0.61 (system regards as medium-weight feature)

[0172] Social variation: 0.51 (system maintains default attention level)

[0173] The three types of weighted feature vectors are reorganized into a weighted time series tensor and input into the subsequent risk inference module. Since the behavior deviation is magnified significantly, the system will respond more sensitively to risk fluctuations related to it. For example, the behavior of the elderly Mr. Zhang getting up one hour later than usual in the morning on that day will have a stronger impact on the model, which may lead to a higher risk score. On the contrary, if the cognitive resilience index is 0.8, even if these behaviors are abnormal, they will not be magnified by the system, thus reducing the false alarm risk.

[0174] After weighting, the three types of features are concatenated into a unified time series tensor. The shape of the tensor can be T×D, where T represents the number of time steps and D represents the total feature dimension of each time step. As the direct input of the risk inference module, the structure of the tensor needs to meet the format requirements of recurrent neural networks, temporal convolutional networks, or other time-aware models.

[0175] This module supports an online update mechanism. That is, when the resilience index is updated or new feature inputs are added from the upstream module, it can automatically trigger tensor reconstruction and weight adjustment calculations to ensure that the model always uses the current information flow that most truly reflects the cognitive state of the elderly. Through the above mechanism, on the basis of maintaining information integrity, the feature fusion module 105 fully considers the differences in individual risk tolerance, making the final input have higher personalized expression ability and prediction effectiveness.

[0176] Furthermore, the feature fusion module is also used for:

[0177] Dynamically construct a multi-dimensional weight adjustment kernel function according to the cognitive resilience index, which acts on each dimensional sub-index of the behavior deviation feature, language anomaly feature, and social variation feature within the current time slice, and calculate the adjustment coefficient based on the ratio between the short-term change rate of each index and its historical average change rate, for non-linearly amplifying or suppressing instantaneous anomalies;

[0178] The weight adjustment kernel function selects the kernel function type and response curvature within a preset range according to the value of the individual's cognitive resilience index. When the cognitive resilience index is lower than the set threshold, an exponential enhancement type weight adjustment kernel is adopted to exponentially amplify minor anomalies; when the cognitive resilience index is higher than the threshold, a smoothing suppression type weight adjustment kernel is adopted to buffer and compress instantaneous anomalies;

[0179] The behavior deviation features, language anomaly features, and social variation features processed by the weight-adjusting kernel function are respectively transformed into weighted feature vectors, while maintaining their original time series structure, and stitched together to form a three-dimensional fusion time series tensor, which is used to provide an interpretable and personalized adjustable unified input to the subsequent risk inference module.

[0180] In the present invention, the key task of the feature fusion module is to integrate the behavior deviation features, language anomaly features, and social variation features into a unified structure in the time dimension for the dynamic assessment of cognitive impairment risk. To avoid generalization problems caused by static weighting or linear fusion, the system introduces the cognitive resilience index to participate in the fusion process. By dynamically constructing a multi-dimensional weight-adjusting kernel function, the weights of each-dimensional sub-index in the three feature categories are adjusted in a personalized and non-linear manner. This weight-adjusting kernel function will enhance or compress the original features before fusion according to the speed of feature change and the individual's cognitive tolerance, thereby constructing a three-dimensional fusion time series tensor that can reflect time dynamics and individual differences.

[0181] First, in daily data processing, the system extracts the behavior deviation features, language anomaly features, and social variation features of the current day respectively. Each type of feature consists of several sub-indicators. For example, the behavior deviation features may include the daily rhythm fluctuation amount, abnormal behavior density, cross-day irregular frequency, etc. For each sub-indicator, the system first calculates its short-term change rate, that is, the difference between the value of this indicator in the current time slice and the previous time slice, and then calculates the average change rate of this indicator in the past several days (such as 7 days) as the reference interval. The ratio of the two is called the change rate ratio, which reflects whether there is a drastic fluctuation in this feature on the current day. For example, if the daily rhythm fluctuation amount of a certain elderly person is 0.68 on the current day, 0.55 on the previous day, the short-term change is +0.13, and the average daily change in the past week is 0.04, then its change rate ratio is 0.13÷0.04 = 3.25, indicating that the fluctuation rate is more than 3 times that of usual.

[0182] The system then reads the cognitive resilience index of this elderly person. The range of this index is from 0 to 1, which is used to characterize the individual's ability to withstand and recover from cognitive function fluctuations. If the resilience index is lower than the set threshold (such as 0.5), it means that the individual may be difficult to recover even when there are minor anomalies. The system will select an exponential enhancement type weight-adjusting kernel function to amplify its short-term changes; if the resilience index is higher than this threshold, indicating that it has good adjustment ability, then a smoothing suppression type weight-adjusting kernel function is selected to compress high-frequency short-term anomalies. The form of the exponential enhancement type function is:

[0183] Weighted value = Original value × exp(Adjustment coefficient × Rate ratio)

[0184] Among them, the adjustment coefficient is usually set between 0.2 and 0.3, which is used to control the amplification degree. The smoothing suppression function form is as follows:

[0185] Weighted value = original value ÷ (1 + compression coefficient × rate ratio)

[0186] The compression coefficient is usually set between 0.2 and 0.4, which is used to limit the influence of high-frequency anomalies. The selection of this kernel function and its response curvature are both dynamically determined by the cognitive resilience index. Different elderly people, different characteristic dimensions, and different time slices can use different kernel functions, so as to construct a fully time-domain adjustable fusion mechanism.

[0187] Taking the elderly person surnamed Zhang as an example, assuming that his cognitive resilience index is 0.32 (lower than the threshold of 0.5), the daily rhythm fluctuation on May 14 is 0.68, the short-term rate is 0.13, and the historical average rate is 0.04, then the rate ratio is 3.25. The system adopts an exponential enhancement type weight-adjusting kernel function, and the adjustment coefficient is 0.2. The calculation result is:

[0188] 0.68 × exp(0.2 × 3.25) ≈ 0.68 × 1.91 ≈ 1.30

[0189] It shows that the system magnifies this feature by nearly twice as a compensation process for its low resilience.

[0190] On the contrary, if the cognitive resilience index of the elderly person surnamed Wang is 0.85, under the same feature input, the system will adopt a smoothing suppression function, and the compression coefficient is set to 0.3. The result is:

[0191] 0.68 ÷ (1 + 0.3 × 3.25) ≈ 0.68 ÷ 1.975 ≈ 0.344

[0192] The system compresses high-frequency anomalies to avoid misjudging their healthy short-term fluctuations as cognitive risks.

[0193] All sub-indicators in each category of features are calculated through the above method to form three weighted vectors, which respectively represent the daily behavior weighted vector, language weighted vector, and social weighted vector. These vectors maintain the original time series order and are spliced day by day in the time dimension to form a fused three-dimensional time series tensor. The structure of this tensor is (time step × feature dimension), where the feature dimension is the weighted splicing result of the three categories of features. Finally, this tensor is sent as a unified input into the time-recurrent neural network model to complete the dynamic inference of the cognitive impairment risk.

[0194] Through this multi-dimensional kernel function weighted mechanism based on cognitive resilience index regulation and combined with short-term fluctuation and long-term trend analysis, the present invention constructs a feature fusion path with both time sensitivity and individual differences, providing a more accurate and early recognition capable basic data structure for the cognitive state monitoring of the elderly in nursing homes.

[0195] A risk inference module 106 is configured to output a cognitive impairment risk value based on the time series tensor through a time recurrent model. If the cognitive impairment risk value mutates, it automatically retraces the feature trajectory of the past 7 days, constructs multiple prediction paths and evaluates the interpretability, and selects the optimal path to generate a dominant prediction result and a ranking of key factors.

[0196] The function of the risk inference module 106 is to receive the time series tensor generated by the feature fusion module, dynamically predict the occurrence risk of future cognitive impairment in the elderly through a time-aware model such as a recurrent neural network, and automatically perform retrospective analysis and path reconstruction when the risk value mutates, so as to enhance the interpretability of the results and the stability of decision-making. This module combines multiple mechanisms such as prediction, anomaly detection, path generation, and feature attribution, and is the core component for achieving high-confidence intelligent evaluation in the present invention.

[0197] This module first receives an input tensor of shape T×D, where T represents the number of time steps and D represents the fused feature dimension. The input data is generated by weighted fusion of behavioral deviation features, language anomaly features, social variation features, etc. with a cognitive resilience index and has complete time dependence. The risk inference module uses a recurrent neural network (such as LSTM or GRU) to model this tensor. The network structure usually includes an input layer, one or more recurrent hidden units, and an output layer, which are used to gradually learn the change trend of the multimodal features of the elderly in the time dimension, and then output a risk score sequence aligned with time. The module can be configured to output the current risk level at each time step, or output the probability or risk level of mild cognitive impairment occurring in a future period at the end.

[0198] To improve the system's response ability to abnormal signals, this module integrates a mutation detection mechanism. This mechanism analyzes the change amplitude of the consecutive two risk score outputs and sets a sensitivity threshold (for example, 0.3). When it is detected that the current prediction result deviates significantly from the previous one, or the score suddenly rises above this threshold, the system determines that the risk value has mutated. At this time, the module will call the internal retrospective mechanism to automatically retrieve the input feature trajectory within the past 7 days and start the prediction path reconstruction process.

[0199] The path reconstruction mechanism constructs multiple hypothetical input paths based on the principle of feature perturbation. Based on the original features of the past 7 days, the module constructs multiple slightly different feature sequences through operations such as random masking, channel scaling, or feature smoothing respectively, to simulate the risk change trends under different input hypotheses. For each reconstructed path, the module executes the prediction process again to obtain its corresponding risk value and intermediate state. To screen the most representative path, the module evaluates the interpretability of the outputs of each path, and uses methods such as feature contribution distribution, input perturbation sensitivity, and Shapley value approximation to quantify the factor attribution quality and stability of the prediction results under this path.

[0200] Finally, the module selects the path with the strongest interpretability and the highest rationality of the prediction result from all alternative paths as the dominant path, outputs the prediction result of this path, and generates a key factor ranking table. This ranking table lists the top several features that contribute the most to the risk score in this prediction, which can be identified as the concentration of abnormal behaviors, the burst frequency of specific language keywords, the sudden drop in social participation, etc., and can be accompanied by a time index for subsequent manual intervention or model annotation.

[0201] The risk inference module 106 is designed to run in real time or in periodic batch processing, and can be adapted according to the deployment environment and data flow rate. Its core advantage is that it not only provides score prediction, but also can self-correct and interpret the output results, making the entire cognitive risk assessment mechanism have clinical auxiliary significance and management intervention operability, meeting the intelligent health management needs of actual nursing homes.

[0202] Furthermore, the risk inference module is also used for:

[0203] During the process that the time recursive model infers the time series tensor and outputs the cognitive impairment risk value, continuously monitor the gradient contribution distribution of the three types of inputs, namely the behavior deviation feature, the language anomaly feature, and the social variation feature, to the change rate of the risk value under abnormal states. When it is found that the risk contribution change trends corresponding to different feature sources show directional divergence, automatically trigger the path divergence discrimination mechanism;

[0204] The path divergence discrimination mechanism is used to judge whether the inference paths between the behavior deviation feature and the language anomaly feature, or between the language anomaly feature and the social variation feature, produce opposite trend judgments on the risk value orientation within the same time window. If there is a divergence, automatically trace back the historical sub-index trajectories of each type of feature within this time window, and construct at least two alternative inference chains based on different feature dominant paths;

[0205] Each alternative path is locally re-predicted, and a ranking function based on explanatory power score is introduced. The performance of each path in terms of prediction coherence, indicator consistency and actual feedback consistency is comprehensively considered, the dominant path is dynamically adjusted, and the key factor ranking results are generated to make the final risk prediction results more individually targeted and the feature source transparent.

[0206] In the risk reasoning module of the present invention, the system not only outputs the risk value of cognitive impairment, but also continuously monitors the changing trends of various input features and their actual impact on the model output during the reasoning process. This module tracks the contribution gradients of three types of data sources: behavioral deviation features, language abnormality features, and social variation features, to determine whether the change in risk value is driven in a consistent manner, or whether there is a sign of cognitive abnormality, namely "input path divergence".

[0207] In each time slice, the system records the three types of feature vectors in the input tensor and calculates their risk-oriented gradients in the model. For example, the model can calculate the risk impact of each type of feature in the recursive unit under the time slice by chain backpropagation. If the gradient of the behavioral deviation feature is negative, it means that this type of data reduces the risk value; the gradient of the language abnormality feature is positive, indicating that it drives the risk value up; the social variation feature may be unchanged or slightly declining. At this point, the system determines that the contribution direction of the three types of features to the risk value is divergent.

[0208] Once a divergence is triggered, the system will locate the time window where the divergence occurs and determine whether the phenomenon is time-persistent. For example, in the three-day window from May 10 to May 12, the behavior indicators of Mr. Zhang improved day by day (such as the behavioral deviation score dropped from 0.8 to 0.6), but the language indicators showed an abnormal increase (such as the abnormal keyword density increased from 0.3 to 0.6), and the risk gradients corresponding to the behavior and language continued to be in opposite directions in the model, so the time window was determined to be a path divergence period.

[0209] The system then automatically backtracks the input data within the time window and constructs multiple alternative reasoning paths. Taking the behavior-dominated path as an example, the system only retains the original value of the behavior deviation feature, and replaces the language and social features with their 7-day sliding average, indicating that the language and social features are stable and unchanged, and only observes the risk evolution under the dominance of behavior; taking the language-dominated path as an example, the original fluctuation of the language feature is retained, and the other two categories are set to be stable.

[0210] Each alternative path will be separately input into the risk assessment sub-model with the same structure as the main model for local re-forecasting. The forecast output includes: the risk value trend corresponding to the path (for example, rising, flat or falling); key feature change nodes (for example, mutation time point, amplification factor); and the degree of consistency with the actual subsequent risk trend (for example, whether the risk value increase on May 13 was successfully predicted).

[0211] The system will calculate an explanatory power score for all alternative paths. The score combines three dimensions: first, whether the risk changes within the path are coherent and stable, for example, whether it shows a reasonable upward or downward trend, and whether there are violent shocks; second, whether the dominant characteristics on the path are consistent with historical risk events. For example, the five previous risk increases for Mr. Zhang were all dominated by language abnormalities, so the language path score increased; third, the degree of consistency between the path output and the actual observation data for the next 2 to 3 days, for example, whether the predicted value reveals the upcoming risk in advance.

[0212] For example, if the behavior-dominant path shows that the risk value continues to decline but deviates seriously from the actual data, and the language-dominant path successfully predicts the slight risk surge on May 13, and can accurately point out the key words "can't find the key" and "repeat words" and other language features as key factors, then the language path will get the highest explanatory power score. The system uses this path as the dominant path, outputs the final predicted risk value, and uses "keyword emergence frequency" and "language disconnection degree" as the top features for key factor sorting.

[0213] The final output not only contains a risk value, but also comes with a feature explanation label with a clear source, so that medical staff can clearly understand the main reason for the current increase in risk. This result not only improves the system's accuracy in distinguishing the early evolution trend of cognitive impairment, but also significantly improves the clinical interpretability of risk results and supports individualized intervention decisions.

[0214] The system can be implemented based on any neural network platform that supports back-propagation and differentiable path tracing, such as PyTorch or TensorFlow. The embedded "middle layer contribution tracking mechanism" calls the derivative values of each feature channel, and cooperates with the sliding window strategy to construct multiple "path shielding input versions" to quickly complete path re-prediction and score ranking. The whole mechanism does not require the introduction of additional external models, and can be implemented only by relying on the existing data and computing power within the system structure.

[0215] Compared with the risk models in the prior art that only focus on single feature trends or simple mutation judgments, the present invention provides a technical foundation for early intervention of cognitive impairment with traceable structure, decomposable logic and calibratable response by introducing the recognition logic of path divergence, a constructed re-prediction strategy and an explanatory power-driven dominant path reconstruction mechanism.

[0216] Although the present application is disclosed as above in the form of a preferred embodiment, it is not intended to limit the present application. Any technical personnel in this field may make possible changes and modifications without departing from the spirit and scope of the present application. Therefore, the scope of protection of the present application shall be based on the scope defined by the claims of the present application.

Claims

1. A risk assessment and prediction model for mild cognitive impairment in the elderly in a pension institution, characterized in that, Including: A behavior analysis module, which is used to collect behavior data including the daily living, diet, walking and social activities of the elderly, compare it with the institutional work and rest template, generate behavior deviation characteristics and mark the abnormal behavior time periods; A language recognition module, which is used to analyze the descriptive statements in the nursing records, and combine the frequency change of keywords statistically during the abnormal behavior time periods to obtain language abnormality characteristics and an enhanced abnormal time period set; A social modeling module, which is used to calculate the individual social frequency and structural changes according to the spatial positioning and voice interaction records in the enhanced abnormal time period set, and generate social variation characteristics; A resilience calculation module, which is used to calculate the cognitive resilience index based on the health record, behavior deviation characteristics, language abnormality characteristics and social variation characteristics, and combine the response speed of the elderly to health events; A feature fusion module, which is used to perform weighted processing on the behavior deviation characteristics, language abnormality characteristics and social variation characteristics according to the cognitive resilience index to generate a time series tensor; A risk inference module, which is used to output the cognitive impairment risk value through a time recurrence model according to the time series tensor. If the cognitive impairment risk value mutates, it automatically traces back the feature trajectory of the recent 7 days, constructs multiple prediction paths and evaluates the explanatory power, and selects the optimal path to generate the dominant prediction result and the key factor ranking.

2. The risk assessment and prediction model for mild cognitive impairment of the elderly in a pension institution according to claim 1, wherein The behavior analysis module is also used for: Performing rhythm analysis on the time stamp sequence of the behavior events of the elderly for consecutive days, calculating the individual main rhythm frequency and its variation index through time-frequency transformation, and constructing an individual rhythm spectrum; the behavior events include getting up, dining, using the toilet, walking and social activities; Selecting the elderly with normal cognitive status from the same elderly care institution to construct a collective rhythm reference spectrum, and calculating the rhythm deviation amount of the individual rhythm spectrum relative to the collective reference spectrum; Jointly modeling the rhythm deviation amount with the density and cross-day distribution law of the continuous abnormal behavior time periods, and dynamically dividing the individual's current rhythm state into stable type, slow drift type, mutation type or structural disconnection type through a set classification function or clustering algorithm; Calculating the abnormal behavior rhythm disorder index based on the classification result and the rhythm deviation amount, and using the abnormal behavior rhythm disorder index as a component of the enhanced behavior deviation characteristics.

3. The risk assessment and prediction model for mild cognitive impairment of the elderly in a pension institution according to claim 1, characterized in that, The language recognition module is also used for: When analyzing the descriptive statements in the nursing records, constructing a semantic emergence matrix with the individual's abnormal behavior time periods as time anchors, and the semantic emergence matrix records the emergence frequency and cross-class co-occurrence density of the specified semantic categories during the abnormal time periods; Through a preset cognitive impairment semantic domain dictionary, classifying the emerging words into four high-risk semantic domains of memory impairment, mood fluctuation, language disconnection and life disorder, and calculating the semantic category weight vector of them during the corresponding abnormal time periods; Combining the nursing description statements of multiple adjacent elderly people in the same time period, constructing a context cross-reference vector group, and screening the target semantics with independent emergence characteristics through similarity analysis, so as to exclude the noise expressions caused by collective behaviors or nursing record habits; Generating a normalized language risk intensity score according to the semantic category weight vector and the cross-reference result, and using it as the semantic risk expression part in the language abnormality characteristics.

4. The risk assessment and prediction model for mild cognitive impairment of the elderly in a pension institution according to claim 1, characterized in that, The social modeling module is further configured to: Based on the spatial positioning and voice interaction data within the enhanced abnormal time period set, construct a social trajectory map of the target elderly person for multiple consecutive days, and map this map to the structural vector space in three dimensions: social intensity, contact persistence, and interaction object diversity; Construct a social isomorphism feature baseline map of the elderly population with normal cognitive status within the same elderly care institution, and project the current social trajectory vector of the target individual into the isomorphism space of this baseline map through the minimum distance matching algorithm to obtain a social path deviation vector; According to the amplitude, direction, and stability of the social path deviation vector in the structural space, calculate the abnormal social path deviation rate, and combine the voice participation frequency and spatial stay distribution during the abnormal time period to generate a comprehensive social disorder index, which is used as part of the social variation feature.

5. The risk assessment and prediction model for mild cognitive impairment of the elderly in a pension institution according to claim 1, wherein The resilience calculation module is further configured to: After each behavior abnormal time period, language abnormal time period, or social abnormal time period detected in the behavior deviation feature, language abnormal feature, and social variation feature, identify the continuous time period during which the behavior deviation feature, language abnormal feature, and social variation feature fall back within their individual historical fluctuation ranges and remain stable, as the corresponding recovery window; For each pair of abnormal recovery pairs consisting of an abnormal time period and its subsequent recovery window, calculate the response recovery offset value, which is defined as the product of the sum of the standardized deviation values of the behavior deviation feature, language abnormal feature, and social variation feature relative to their respective historical means during the abnormal time period and the time interval from the end of the abnormal time period to the start of the recovery window; Perform a weighted average on the set of response recovery offset values of all abnormal recovery pairs to obtain the average response recovery offset value, and use its reciprocal as the individual's recovery efficiency index, which is used to measure the overall response ability of the individual's self-repair behavior after multiple abnormalities; Jointly model the recovery efficiency index with the number of chronic disease interventions, medication compliance score, and historical mood fluctuation cycle index recorded in the health record of this elderly person to generate a cognitive resilience dynamic adjustment factor, which is used to participate in the subsequent calculation of the cognitive resilience index and serve as the adjustment basis for weighting the behavior deviation feature, language abnormal feature, and social variation feature.

6. The risk assessment and prediction model for mild cognitive impairment of the elderly in a pension institution according to claim 1, characterized in that, The feature fusion module is further configured to: Dynamically construct a multi-dimensional weighted kernel function according to the cognitive resilience index, which acts on each dimensional sub-index of the behavior deviation feature, language abnormal feature, and social variation feature within the current time slice, and calculate the adjustment coefficient based on the ratio between the short-term change rate of each index and its historical average change rate, for non-linearly amplifying or suppressing instantaneous abnormalities; The weighted kernel function selects the kernel function type and response curvature within a preset range according to the value of the individual's cognitive resilience index. When the cognitive resilience index is lower than the set threshold, an exponential enhancement type weighted kernel is used to perform exponential amplification on minor abnormalities; when the cognitive resilience index is higher than the threshold, a smoothing suppression type weighted kernel is used to buffer and compress instantaneous abnormalities; The behavior deviation features, language anomaly features, and social variation features processed by the weight-adjusted kernel function are respectively transformed into weighted feature vectors, while maintaining their original time series structure, and are concatenated to form a three-dimensional fusion time series tensor, which is used to provide an interpretable and personalized adjustable unified input to the subsequent risk inference module.

7. The risk assessment and prediction model for mild cognitive impairment of the elderly in a pension institution according to claim 1, characterized in that, The risk inference module is further configured to: During the process that the time recurrence model infers the time series tensor and outputs the cognitive impairment risk value, continuously monitor the gradient contribution distribution of the three types of inputs, namely the behavior deviation features, language anomaly features, and social variation features, to the change rate of the risk value in the abnormal state. When it is found that the risk contribution change trends corresponding to different feature sources show directional divergence, automatically trigger the path divergence discrimination mechanism; The path divergence discrimination mechanism is used to determine whether the inference paths between the behavior deviation features and the language anomaly features, or between the language anomaly features and the social variation features, produce opposite trend judgments on the risk value orientation within the same time window. If there is a divergence, automatically trace back the historical sub-index trajectories of each type of feature within this time window, and construct at least two alternative inference chains based on different feature-dominated paths; Perform local re-prediction on each alternative path respectively, and introduce a sorting function based on the interpretability score. Considering the performance of each path in terms of prediction coherence, index consistency, and actual feedback matching degree, dynamically adjust the dominant path, and generate the key factor sorting result, so that the final risk prediction result is more individualized and the feature source is more transparent.

Images