A dynamic mental health risk early warning assessment and intervention system for depression in the elderly

Abstract

This invention discloses a dynamic mental health risk early warning assessment and intervention system for geriatric depression, belonging to the field of medical information technology. Specifically, it includes: a data acquisition module for acquiring natural language dialogue content between the user and a smart home assistant, and acquiring standardized psychological scale response data filled in by the user; a dialogue recognition module for separating instruction and response statements from the dialogue content and extracting interaction feature vectors; a scale analysis module for extracting scale feature values ​​from the psychological scale response data; and a risk verification module for establishing a dynamic mapping relationship between the interaction feature vectors and scale feature values, calculating the degree of contradiction, initiating a risk verification process when the degree of contradiction exceeds a preset threshold, integrating the interaction feature vectors and the degree of contradiction index to output a depression risk value, and triggering an intervention command when the risk value continuously exceeds a set threshold. This invention constructs a multi-dimensional assessment system with dual-source mutual verification of objective dialogue interaction data and subjective scale data.

Description

Technical Field

[0001] This invention relates to the field of medical information technology, specifically to a dynamic mental health risk early warning assessment and intervention system for elderly people with depression. Background Technology

[0002] In current clinical practice, early identification of depression in the elderly still mainly relies on regular screening using psychological scales, such as the Geriatric Depression Scale (GDS) and the Patient Health Questionnaire-9 (PHQ-9). These tools depend on patients' subjective descriptions of their emotional state and assess the frequency and intensity of depressive symptoms through standardized questions.

[0003] In current clinical practice, regular psychological scale screening remains the primary identification method. These tools rely on individuals' subjective descriptions of their emotional state and assess the severity of depressive symptoms through standardized questions. However, for the elderly, this subjective response-dependent model presents several challenges. Influenced by traditional beliefs, some elderly individuals associate psychological distress with negative labels such as weakness of will, forming a strong sense of stigma. They may deliberately downplay or even conceal their true emotional experiences when answering scales, leading to biases in key information. This self-protective response tendency makes it difficult for screening results to accurately reflect their inner psychological state. Therefore, it is necessary to use non-invasive objective indicators to validate scale response data and construct a multidimensional assessment system. Summary of the Invention

[0004] The purpose of this invention is to provide a dynamic mental health risk early warning assessment and intervention system for elderly people with depression, and to solve the following technical problems:

[0005] The scale assessment relies on the patient's subjective response. Some elderly people deliberately conceal their illness due to the stigma associated with it, so the data from the scale response cannot reflect their true psychological state.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] A dynamic mental health risk early warning assessment and intervention system for depression in the elderly includes:

[0008] The data acquisition module is used to acquire the natural language dialogue content between the user and the smart home assistant, and to acquire the standardized psychological scale response data filled in by the user.

[0009] The dialogue recognition module is used to separate user-initiated command statements and response statements from natural language dialogue content and extract interaction feature vectors. The interaction feature vectors include generating interaction behavior feature vectors by identifying the behavior request type in the command statements and generating interaction sentiment feature vectors by analyzing sentiment words in the response statements.

[0010] The scale analysis module is used to extract scale feature values ​​from standardized psychological scale response data. The scale feature values ​​include scale behavioral feature values ​​and scale emotional feature values.

[0011] The risk verification module is used to establish a dynamic mapping relationship between the interaction feature vector and the scale feature value, calculate the contradiction degree of the dual-source features, start the risk verification process when the contradiction degree exceeds a preset threshold, integrate the interaction feature vector and the contradiction degree index to output the depression risk value, and trigger the intervention instruction when the risk value continues to exceed the set threshold.

[0012] As a further aspect of the present invention: the specific process of generating interactive behavior feature vectors by recognizing the behavior request type in the instruction statement in the dialogue recognition module is as follows:

[0013] Establish a three-tiered classification system for self-care instructions, social seeking instructions, and medical consultation instructions. Self-care instructions cover basic daily living operations requests, social seeking instructions include proactively initiating interpersonal interactions, and medical consultation instructions specifically refer to inquiries about non-emergency health issues.

[0014] Implement instruction validity filtering to exclude invalid instructions that the device cannot execute and repeated failed instructions; finally, calculate the distribution density of each type of instruction within a fixed period, and combine it with historical baseline data to calculate the deviation and generate a three-dimensional interactive behavior feature vector.

[0015] As a further aspect of the present invention: the specific process of generating interactive sentiment feature vectors by analyzing sentiment words in response statements in the dialogue recognition module is as follows:

[0016] Explicit sentiment words matched by the basic sentiment lexicon are extracted from the response statements, while implicit sentiment reversal words with the effect of negative modifiers are identified through the semantic association rule base.

[0017] The intensity weight calculation of emotional words implements a subject association strategy. When the subject of the emotional description is the user himself, a standard weight coefficient is used, and when the subject of the description is a specific person, an intimacy correction coefficient is loaded.

[0018] All emotional words are processed by a time-effect decay function to generate a time-weighted comprehensive emotional intensity value. The final output interactive emotional feature vector contains the weighted intensity difference between positive and negative emotions.

[0019] As a further aspect of the present invention: the specific process of extracting scale feature values ​​from standardized psychological scale response data in the scale analysis module is as follows:

[0020] First, a consistency check of the response logic is performed to analyze the conflicting options for questions of the same dimension within the scale; then, a reliability flag is added to the filling behavior, either for filling operations that are shorter or longer than the normal time.

[0021] The behavioral characteristics of the scale are derived from the aggregated scores of physiological activity questions, while the emotional characteristics of the scale are derived from the distribution analysis of emotional state questions. The extracted results include three data attributes: the original written values, the reliability indicator, and the list of logical conflict points.

[0022] As a further aspect of the present invention: the specific process for calculating the degree of contradiction of the dual-source features in the risk verification module is as follows:

[0023] The establishment of the dynamic mapping relationship includes a first verification channel and a second verification channel. The first verification channel aligns the self-care dimension of the interactive behavior feature vector with the scale behavior feature value by trend and calculates the directional deviation between the two within a continuous time window. The second verification channel matches the comprehensive intensity value of the interactive emotion feature vector with the scale emotion feature value by numerically detecting whether the absolute difference exceeds a reasonable fluctuation range.

[0024] The directional deviation and the absolute difference are normalized and fused to generate a primary contradiction index. When the scale response data carries a low reliability indicator, contradiction compensation calculation is added.

[0025] As a further aspect of the present invention: the risk verification module specifically includes the following risk verification process:

[0026] When the degree of contradiction exceeds the standard continuously, the interaction data is subjected to instruction feedback verification to analyze whether the user's instructions are correctly responded to by the device and the subsequent changes in the frequency of interaction; the scale data is subjected to logical chain verification to detect whether conflicting options form a contradictory logical chain.

[0027] The instruction feedback verification results and logic chain verification results are converted into data credibility level. The data credibility level is used to correct the primary contradiction index. The corrected contradiction index, along with the interaction behavior feature vector and the interaction sentiment feature vector, are synchronously input into the risk fusion engine.

[0028] As a further aspect of the present invention: the risk fusion engine first synthesizes a basic risk value by combining the interactive behavior feature vector and the interactive emotion feature vector according to a preset ratio;

[0029] Then, the contradiction degree conversion function is used to map the corrected contradiction degree index to a risk amplification coefficient, and the basic risk value is multiplied by the risk amplification coefficient to obtain the enhanced risk value;

[0030] Finally, a dynamic balancing factor for the scale's characteristic values ​​is introduced. When the scale data shows a clear tendency to depression, a positive calibration is applied to the reinforcement risk value; otherwise, a negative inhibition is applied. The final output is the reinforcement risk value processed by the dynamic balancing factor, which is then labeled as the depression risk value.

[0031] As a further aspect of the present invention: the contradiction degree conversion function specifically includes:

[0032] The contradiction index is divided into low contradiction zone, medium contradiction zone and high contradiction zone according to the value. When the contradiction index is in the low contradiction zone, the linear transformation mode is adopted. When the contradiction index is in the medium contradiction zone, the logarithmic growth mode is activated. When the contradiction index is in the high contradiction zone, the exponential amplification mode is activated.

[0033] The conversion process is dynamically adjusted by the data credibility level. If the data credibility level is greater than or equal to a preset threshold, the slope of the conversion curve is increased. If the data credibility level is less than the preset threshold, the range of output values ​​is compressed. The risk amplification coefficient is set with a safe range value.

[0034] As a further aspect of the present invention: in the risk verification module, the intervention instructions specifically include:

[0035] When the depression risk value is abnormal but the scale characteristic value is normal, a hidden depression intervention instruction is generated. The hidden depression intervention instruction includes an interactive characteristic analysis report and a set of suggested verification questions.

[0036] When both sources of data are abnormal, an emergency medical intervention instruction is generated. The execution process of the emergency medical intervention instruction is embedded with a feedback learning loop. After collecting the intervention results, the rules for generating interactive feature vectors, the logic for extracting scale feature values, and the parameters for calculating the degree of contradiction are automatically optimized.

[0037] The beneficial effects of this invention are:

[0038] This invention effectively addresses the limitations of traditional scales that rely on subjective responses by constructing a multi-dimensional assessment system that integrates dual-source data. Specifically targeting the information bias caused by stigma associated with illness among the elderly, the invention utilizes a data acquisition module to simultaneously collect natural language dialogue and scale response data. A dialogue recognition module extracts interaction feature vectors containing the distribution density of behavioral request types and the intensity of emotional tendencies from non-intrusive interactions. Simultaneously, a scale analysis module performs logical consistency checks and assigns credibility markers to the scale data, extracting scale feature values ​​containing reliability indicators. A dynamic mapping relationship is established between the two through a risk verification module, calculating the degree of contradiction between the dual-source features. When the degree of contradiction exceeds a certain threshold, instruction feedback verification and logical chain verification are initiated to generate a data credibility level. Finally, a risk fusion engine combines the interaction features with the modified degree of contradiction index to form a depression risk value that includes a basic risk value, a risk amplification coefficient, and a dynamic balance factor, thus achieving objective verification of subjective response bias. This system compensates for the deficiencies of subjective responses on the scale by objectively analyzing non-invasive interactive features. It constructs a multi-dimensional assessment system that verifies both objective dialogue interaction data and subjective scale data from two sources. This effectively solves the problem of distorted screening information caused by the stigma and cognitive bias of the elderly, improves the accuracy of depression risk assessment and early identification efficiency, and provides a reliable basis for dynamic early warning and precise intervention. Attached Figure Description

[0039] The invention will now be further described with reference to the accompanying drawings.

[0040] Figure 1 This is a schematic diagram of the modules of the present invention. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] Please see Figure 1 As shown, this invention is a dynamic mental health risk early warning assessment and intervention system for elderly people with depression, comprising:

[0043] The data acquisition module collects two types of core data in real time: first, natural language dialogues between users and the smart home assistant, covering daily interactions such as device control, life consultation, and emotional communication; second, standardized psychological questionnaire responses completed by users, supporting electronic input via voice, touchscreen, or paper forms. All data is timestamped and device-identified to ensure spatiotemporal alignment and traceability.

[0044] The dialogue recognition module analyzes interaction features by separating commands and responses. Commands are categorized into three levels: self-care, social seeking, and medical consultation. Self-care requests include routine tasks like adjusting the temperature or turning on lights; social seeking requests encompass interpersonal interactions such as video calls and checking friends' or family updates; and medical consultations specifically refer to inquiries about non-emergency health issues like medication advice or symptom self-checks. After filtering invalid and duplicate commands, the system calculates the distribution density of each command type over seven days, combines this with historical baselines to calculate deviation, and generates a three-dimensional interaction behavior feature vector. For responses, explicit sentiment words and implicit sentiment words transformed by negation modifiers are extracted. Different weights are applied based on whether the subject is themselves or a related person. After time-based attenuation processing, an interaction sentiment feature vector is generated, containing the difference in intensity between positive and negative sentiment.

[0045] The scale analysis module performs in-depth processing of response data: First, it detects logical conflicts in questions within the same dimension, such as whether answers about sleep quality and difficulty falling asleep contradict each other; then, it marks the credibility of the responses, adding reliability indicators to abnormal response durations, i.e., those that are too short or too long. It extracts aggregated scores from physiological activity questions as behavioral feature values ​​and distribution analysis results from emotional state questions as affective feature values, outputting triple-attribute data including raw values, reliability indicators, and points of logical conflict.

[0046] The risk verification module establishes a dual-source data verification mechanism: the first channel aligns the self-care dimension of interactive behavior with the daily activity trends of the scale's behavioral characteristics, calculating the directional deviation over a continuous period; the second channel matches the numerical differences between the intensity of interactive emotions and the emotional characteristics of the scale, detecting whether they exceed the individual's reasonable fluctuation range. The two types of differences are normalized and fused to generate a preliminary degree of contradiction; if the scale data contains low-reliability indicators, compensation calculations are triggered.

[0047] When the inconsistency level continuously exceeds the limit, the system initiates dual verification: verifying the device response and subsequent interaction changes in the interactive data, and detecting cross-dimensional inconsistency logic chains in the scale data. The results are converted into data credibility levels to correct the inconsistency level. The risk fusion engine synthesizes interactive features proportionally to form a basic risk value, maps the inconsistency level to a risk amplification coefficient through a non-linear function, and performs dynamic calibration in conjunction with the scale's tendency to generate a final depression risk value. If the risk value continues to exceed the limit, a tiered intervention instruction is automatically triggered, achieving closed-loop management from data collection to intervention.

[0048] In a preferred embodiment of the present invention, the specific process of generating an interactive behavior feature vector by recognizing the behavior request type in the instruction statement in the dialogue recognition module is as follows:

[0049] The dialogue recognition module's behavior request type identification process is based on a three-level classification system, which categorizes user commands into three dimensions: self-care, social seeking, and medical consultation. Self-care commands cover basic daily operations, including but not limited to controlling home appliances, adjusting the environment, and checking the time. For example, commands such as "turn on the living room lights," "raise the air conditioner temperature by two degrees," and "what time is it?" are all classified as self-care commands. Social seeking commands involve proactive expressions of interpersonal interaction, involving scenarios such as contacting relatives and friends, arranging social activities, and sharing information. Commands such as "call my son," "check next week's community activity schedule," and "share today's weather with my neighbor Aunt Wang" fall into this category. Medical consultation commands specifically refer to inquiries about non-emergency health issues, including symptom self-checks, medication guidance, and health knowledge consultations. Questions such as "what should I do about my insomnia lately," "can patients with high blood pressure take this medication," and "how many steps should I walk each day for the best health?" are all identified as medical consultation commands.

[0050] To ensure the accuracy of command analysis, the system implements a command validity filtering mechanism. First, it filters commands based on the device capability list, eliminating invalid commands that devices cannot execute. For example, if a user issues a command to "turn on the kitchen oven," but the system detects that no oven is currently configured in the smart home environment, the command will be deemed invalid and filtered out. Second, the system identifies and removes duplicate failed commands. Commands that fail to execute three times consecutively within 24 hours are marked and will not be included in subsequent analyses to avoid interference from abnormal data caused by device malfunctions or network issues.

[0051] After classifying and filtering the instructions, the system calculates the distribution density of each type of instruction over a fixed period. This period is set to 7 days to balance data timeliness and fluctuation smoothness. During the statistical process, the system records the frequency of each type of instruction at different times of the day, forming a distribution map along the time dimension. For example, self-care instructions typically appear more frequently between 7-9 am and 6-9 pm, while social seeking instructions tend to appear more frequently on weekend afternoons. By analyzing these distribution characteristics, the system can capture users' daily behavioral patterns and lifestyle rhythms.

[0052] Calculating deviation using historical baseline data is a crucial step in generating interactive behavior feature vectors. The system establishes a personalized command behavior baseline for each user, generated from the user's command data over the past 30 days, including the average frequency of each command type, its temporal distribution, and the correlation patterns between commands. When new periodic data is generated, the system compares the current period's command distribution with the historical baseline, calculating the deviation in each dimension. The deviation calculation employs a dynamic threshold mechanism, automatically adjusting the judgment criteria based on the stability of user behavior. For example, for users with relatively fixed command patterns, small frequency fluctuations may be judged as significant deviations; while for users with variable command behaviors, the system appropriately relaxes the judgment threshold. Finally, the system normalizes the deviations in the three dimensions of self-care, social seeking, and medical consultation, generating an interactive behavior feature vector containing these three dimensions. This vector not only reflects the user's current behavioral state but also reveals the trend and magnitude of behavioral changes by comparing with historical patterns, providing objective and dynamic behavioral indicators for subsequent risk assessment.

[0053] In another preferred embodiment of the present invention, the specific process of generating an interactive sentiment feature vector by analyzing sentiment words in the response statement in the dialogue recognition module is as follows:

[0054] The sentiment analysis process of the dialogue recognition module is implemented through multi-dimensional semantic processing technology, aiming to extract genuine emotional state information from user responses. The system first extracts explicit sentiment words from the response, matching them against a basic sentiment lexicon. This lexicon covers common positive sentiment words such as "happy," "joyful," and "satisfied," as well as negative sentiment words such as "sad," "heartbroken," and "anxious." For example, when a user answers "I had a very happy day," the system can directly identify the explicit sentiment word "happy."

[0055] To comprehensively capture the complexity of emotional expression, the system also identifies implicit sentiment reversal words through a semantic association rule base. This rule base contains a series of grammatical and semantic rules used to identify the combination relationships between negative modifiers such as "not," "no," and "none" and sentiment words. For example, when a user expresses "I haven't been interested lately," the system identifies the combination of the negative modifier "no" and "interest" and transforms it into a negative sentiment expression. For more complex semantic structures, such as the double negative expression "It's not that I'm unhappy," the system recursively analyzes the syntax tree to ultimately parse it into a positive sentiment expression.

[0056] The intensity weighting of sentiment words employs a subject-association strategy, which adjusts the weighting coefficient based on the subject of the sentiment description. When the subject is the user themselves, the system uses a standard weighting coefficient, directly reflecting the intensity of the user's emotional experience. For example, the word "sad" in "I'm especially sad today" will be included in the sentiment analysis with a standard weight. When the subject is a specific person with whom the sentiment is described, the system applies a closeness correction coefficient. This closeness correction coefficient is dynamically adjusted based on the degree of connection between the user and the person. For instance, the closeness correction coefficient is higher for core family members such as spouses and children, while it is relatively lower for ordinary friends and neighbors. For example, when a user expresses "I heard my daughter is sick, and I'm very worried," the system will identify "daughter" as the subject and adjust the weight of the sentiment word "worried" according to the preset closeness correction coefficient, giving it a more prominent position in the sentiment analysis.

[0057] All emotional terms are processed by a time-weighted decay function to generate a comprehensive emotional intensity value. The time-weighted decay function employs an exponential decay model, giving more weight to recent emotional expressions in the calculation. Specifically, the system converts the difference between the current time and the time of the emotional expression into a decay factor; as the time interval increases, the weight of the emotional expression decreases exponentially. For example, the weight of an emotional expression on the current day is 1, the weight of an emotional expression from 3 days ago drops to 0.7, and the weight of an emotional expression from 7 days ago is only 0.3. In this way, the system can more accurately reflect the user's current emotional state and avoid excessive interference from historical emotional expressions in the current analysis.

[0058] The final output interactive sentiment feature vector includes the weighted intensity difference between positive and negative sentiment. The system calculates the sum of the weighted intensities of all positive and negative sentiment words separately, and then subtracts the two to obtain the comprehensive sentiment intensity value. A positive value indicates that the user's current sentiment state is biased towards positive; a negative value indicates that the user's current sentiment state is biased towards negative; and a value close to zero indicates that the user's sentiment state is relatively stable. This feature vector not only quantifies the user's sentiment tendency, but also reflects the intensity and stability of the sentiment through weighted processing, providing a key sentiment dimension indicator for subsequent mental health risk assessment.

[0059] In another preferred embodiment of the present invention, the specific process of extracting scale feature values ​​from standardized psychological scale response data in the scale analysis module is as follows:

[0060] The system first conducts a comprehensive analysis of option conflicts within the same dimension of the scale. For example, in the Geriatric Depression Scale, questions related to sleep quality might include "Do you have difficulty falling asleep?", "Is your sleep restful?", and "Do you wake up too early?". The system checks for logical conflicts in the answers to these questions. If a user selects "no" for both "Do you have difficulty falling asleep" and "Is your sleep restful?", a conflict is identified. This conflict analysis is not limited to directly related questions but also includes question groups that can be linked through semantic association and logical reasoning. The system constructs a multidimensional semantic association network, clustering questions in the scale according to theme, semantics, and logical relationships to form multiple interconnected question clusters, thereby enabling in-depth mining of potential logical conflicts.

[0061] Credibility labeling of questionnaire completion is a crucial step in feature extraction. The system establishes a standard completion time model for each scale, generated based on historical completion data from a large number of users. This model includes average completion time, standard deviation, and the average response time distribution for each item. When users complete the scale, the system monitors completion time in real time, adding reliability indicators to completions shorter or longer than the standard time. Completions shorter than the standard time may indicate that the user did not carefully read the questions or answered carelessly, while completions longer than the standard time may suggest cognitive difficulties or indecisiveness. The system employs a dynamic threshold mechanism to determine abnormal completion times, adjusting it individually based on scale complexity, number of items, and the user's historical completion habits. For example, for users with good cognitive function and fast completion speed, the short-time threshold is lowered; while for users completing the scale for the first time or whose cognitive function has declined, the long-time threshold is appropriately increased.

[0062] The behavioral characteristics of the scale are derived from aggregated scores of physiological activity questions. Physiological activity questions typically involve aspects such as daily living abilities and physical symptoms, such as "Are you able to perform daily living activities independently?", "Do you often feel tired?", and "How is your appetite?". The system quantifies the answers to these questions and calculates the total score using a weighted aggregation method. The weights are set based on the clinical importance and relevance of the questions; for example, basic self-care ability questions such as "Can you bathe and dress independently?" have relatively high weights, while mild cognitive questions such as "Do you occasionally forget where you put things?" have relatively low weights. Through this weighted aggregation, the scale's behavioral characteristics can more accurately reflect the user's physiological health status and level of functional life.

[0063] The scale's affective characteristics are derived from the distribution analysis of emotion state-related questions. These questions primarily focus on the user's emotional experience and psychological state, such as "Do you often feel depressed?", "Are you hopeful about the future?", and "Have you lost interest in things you used to enjoy?". The system performs multidimensional distribution analysis on the answers to these questions, calculating not only the total score but also focusing on the distribution patterns of the answers. For example, by analyzing the score distribution across different emotion dimensions, the system can identify typical depressive mood patterns, such as "dominantly depressed," "dominantly lack of interest," or "mixed symptoms." This distribution analysis provides richer information on affective states for subsequent risk assessment, helping to more accurately identify depressive tendencies.

[0064] The final extracted results contain three data attributes: original input values, reliability indicators, and a list of logical conflicts. The original input values ​​retain the users' true responses, providing foundational data for subsequent analysis. The reliability indicators provide a basis for data quality assessment, helping the system determine the credibility of the scale results. The list of logical conflicts records all detected inconsistencies, providing clues for further data validation and risk assessment. These three data attributes complement each other, forming a complete scale feature description system, providing comprehensive and reliable data support for subsequent risk verification.

[0065] In a preferred embodiment, the specific process for calculating the degree of contradiction of the dual-source features in the risk verification module is as follows:

[0066] The risk verification module's contradiction calculation is based on the establishment of a dynamic mapping relationship, implemented through a first verification channel and a second verification channel. The first verification channel aligns the self-care dimension of the interactive behavior feature vector with the scale's behavioral feature values ​​in terms of trends. The self-care dimension reflects the user's autonomous activity ability and behavioral patterns in daily life, while the scale's behavioral feature values ​​quantify the user's physiological health status through aggregated scores of physiological activity-related questions. The system compares the data from these two dimensions over a continuous time window of four weeks to capture dynamic trends. By calculating the directional deviation between the two over time, the system can identify the degree of consistency between behavioral patterns and scale assessment results. For example, if the interactive behavior feature vector shows a declining trend in the user's self-care ability over the past four weeks, while the scale's behavioral feature values ​​show stability or increase, this indicates a directional deviation between the two.

[0067] The second verification channel numerically matches the comprehensive intensity value of the interactive emotional feature vector with the scale emotional feature value. The comprehensive intensity value of the interactive emotional feature vector reflects the user's emotional tendency and intensity in daily conversations, while the scale emotional feature value quantifies the user's psychological state through the distribution analysis of emotion state-related questions. The system calculates the absolute difference between these two values ​​and compares it with a reasonable fluctuation range established based on the user's historical data. The reasonable fluctuation range is dynamically adjusted according to the user's emotional stability and historical fluctuation range; for users with large emotional fluctuations, the range is relatively wide; while for users with relatively stable emotions, the range is relatively narrow. When the absolute difference exceeds the reasonable fluctuation range, it indicates a significant difference between the interactive emotional data and the scale emotional assessment results.

[0068] The directional deviation and absolute difference were normalized and fused to generate a primary inconsistency index. Normalization ensured the comparability of differences across different dimensions, and the system used Z-score standardization to convert directional deviation and absolute difference into standardized scores. The fusion process employed a weighted average, with directional deviation weighted at 0.6 and absolute difference weighted at 0.4. This weighting reflects the relative importance of behavioral trend consistency in risk assessment. Through this fusion, the primary inconsistency index can comprehensively reflect the degree of inconsistency between the two source data across both behavioral and emotional dimensions.

[0069] When scale response data carries a low reliability indicator, the system adds a contradiction compensation calculation. The compensation calculation is adjusted based on the type and severity of the reliability indicator. For responses shorter than the standard timeframe, the compensation coefficient is set to 0.1-0.2; for responses longer than the standard timeframe, the compensation coefficient is set to 0.2-0.3. These coefficients are dynamically adjusted according to the specific reliability indicator; for example, if the response time is significantly shorter or longer than the standard timeframe, the compensation coefficient will be increased accordingly. The purpose of the contradiction compensation calculation is to appropriately increase the sensitivity of the contradiction indicator when the scale data has low reliability, thereby avoiding missed judgments due to scale data distortion. Through this compensation mechanism, the system can more robustly handle potentially biased scale data, improving the accuracy of risk assessment.

[0070] In another preferred embodiment of the present invention, the risk verification process in the risk verification module specifically includes:

[0071] When the contradiction index continuously exceeds the preset threshold, the system initiates a dual-track parallel deep verification process, aiming to accurately identify the real risks behind data anomalies through a cross-validation mechanism. This process includes two core components: instruction feedback verification for interactive data and logical chain verification for scale data.

[0072] Command feedback verification focuses on the interaction data between users and smart home assistants, analyzing it from two dimensions: command execution and subsequent interaction behavior. First, the system traces the execution records of user commands to verify whether the commands received correct responses from the devices. For example, when a user issues the command "turn on the living room lights," the system not only confirms whether the lights are actually turned on but also verifies whether the device response time is within the normal range. If there are instances of command execution failure, excessive response delay, or execution results inconsistent with the command, the system will mark the command as abnormal interaction data. Second, the system monitors changes in the frequency of subsequent interactions after command execution, analyzing whether the user reduces the number of interactions or changes the interaction mode after command execution anomalies. For example, if a user fails to adjust the air conditioner temperature three times consecutively and then does not initiate any environmental control commands for a week, this significant decrease in interaction frequency will serve as an important basis for assessing the credibility of the interaction data. By integrating command execution status and subsequent interaction behavior, the system can identify deviations in interaction data caused by device malfunctions, network anomalies, or changes in user behavior.

[0073] Logical chain verification targets standardized psychological scale response data. By constructing a logical relationship network between questions, it detects whether conflicting options form contradictory logical chains. The system first establishes a semantic and causal relationship graph between questions based on the scale's design logic and clinical diagnostic criteria. For example, in a scale assessing sleep disorders and depressive symptoms, questions such as "difficulty falling asleep," "frequent sleep interruptions," and "morning fatigue" have a logically progressive relationship. If a user answers "no" to "difficulty falling asleep" but "yes" to "frequent sleep interruptions" and "morning fatigue," the system identifies a potential logical contradiction in this set of answers. Furthermore, the system uses a path search algorithm to find the propagation path of contradictory options in the question relationship graph, determining whether a cross-dimensional contradictory logical chain has formed. For example, if a user demonstrates good activity levels in physiological state questions but frequently expresses negative emotions in emotional state questions, and these contradictory answers form a continuous logical conflict path in the scale structure, it is judged as having a serious logical contradiction. This logical chain verification method not only discovers isolated answer conflicts but also captures deeper logical contradictions, effectively improving the verification depth of the scale data.

[0074] After completing the dual-track verification, the system converts the instruction feedback verification results and the logic chain verification results into a data credibility level. This level is divided into five tiers, from Level 1 (completely credible) to Level 5 (completely unreliable). The specific conversion rules are formulated based on the severity and frequency of the verification results: if the instruction execution is completely normal with no significant changes in interactive behavior, and the scale responses are logically consistent and conflict-free, the data credibility level is Level 1; if there are multiple instruction execution failures and a sharp drop in interaction frequency, and the scale responses exhibit serious cross-dimensional logical contradictions, the data credibility level is Level 5. The data credibility level not only reflects the reliability of the data but also provides a quantitative basis for subsequent contradiction correction.

[0075] In a preferred embodiment, the risk fusion engine deeply integrates the dual-source data features and the contradiction index through three progressive calculation steps to generate the final depression risk value.

[0076] The first step involves the engine synthesizing a baseline risk value by combining the interactive behavior feature vector and the interactive emotion feature vector according to a preset ratio. This preset ratio is dynamically configured based on the clinical importance of the two types of features in depression risk assessment, typically with the interactive behavior feature vector accounting for 60% and the interactive emotion feature vector accounting for 40%. This weighting is based on the correlation analysis between behavioral patterns and emotional expression in clinical studies of geriatric depression. Behavioral patterns reflect changes in a user's daily life functions, while emotional expression directly reflects fluctuations in psychological state. The synthesis process uses a weighted summation algorithm to integrate the three-dimensional interactive behavior feature vector and the one-dimensional interactive emotion feature vector into a single numerical value. This value initially reflects the user's depression risk level based on daily interaction data.

[0077] The second step involves the engine using a contradiction degree conversion function to map the corrected contradiction degree index to a risk amplification coefficient. This conversion function employs a piecewise non-linear conversion strategy based on the contradiction degree value range: when the contradiction degree index is in the low contradiction degree range, a linear conversion mode is used to ensure the stability of risk assessment under low contradiction conditions; when it is in the medium contradiction degree range, a logarithmic growth mode is activated to moderately amplify the impact of contradiction on risk; and when it enters the high contradiction degree range, an exponential amplification mode is activated to significantly improve risk sensitivity. For example, when the contradiction degree index moves from the low to the medium range, the output value of the conversion function will grow rapidly in the form of a logarithmic curve, significantly increasing the risk amplification coefficient. Simultaneously, the conversion process is dynamically adjusted by the data credibility level. When the data credibility level is greater than or equal to a preset threshold, the system increases the slope of the conversion curve, enhancing the amplification effect of contradiction on risk; when the data credibility level is less than the preset threshold, the output value range is compressed to avoid misjudgment of risk due to unreliable data. Furthermore, the system sets a safe range for the risk amplification coefficient to prevent excessive amplification of risk values ​​in extreme cases, ensuring the rationality of the assessment results.

[0078] The third step involves the engine introducing a dynamic balancing factor for the scale's feature values ​​to perform a final calibration of the enhanced risk value. This dynamic balancing factor adjusts dynamically based on the degree of depressive tendency shown by the scale data: when both the behavioral and affective feature values ​​of the scale indicate a clear depressive tendency, a positive calibration is applied to the enhanced risk value, typically increasing it by 10%-20% to highlight the importance of the scale data in risk assessment; when the scale data does not show a significant depressive tendency, a negative suppression is applied, appropriately lowering the risk value to avoid over-assessment due to contradictions between the two data sources. For example, if the scale assessment results indicate a user has a mild depressive tendency, but the two data sources are highly contradictory, the dynamic balancing factor will perform a positive calibration on the enhanced risk value, making the final depressive risk value closer to the true risk level.

[0079] After iterative calculations through the three steps described above, the risk fusion engine outputs a reinforced risk value processed by a dynamic balancing factor, which is then labeled as the final depression risk value. This value integrates multiple dimensions of information, including the user's daily interaction behavior, emotional expression, scale assessment results, and data reliability. Through non-linear transformation and dynamic calibration mechanisms, it effectively compensates for the limitations of a single data source.

[0080] In another preferred embodiment of the present invention, the intervention instruction in the risk verification module specifically includes:

[0081] When the depression risk value is abnormal while the scale characteristic values ​​are normal, the system determines that there is a hidden depression risk. At this time, a hidden depression intervention instruction is generated, which includes an interaction feature analysis report and a set of suggested verification questions. The analysis report delves into the dialogue between the user and the smart home assistant, uncovering potential abnormal changes in the frequency of behavioral requests and emotional tendencies; the set of suggested verification questions is designed accordingly to help confirm depressive tendencies.

[0082] If the depression risk value is abnormal and both the interaction feature vector and the scale feature value show abnormalities, the system determines it to be a high-risk state, immediately generates an emergency medical intervention instruction, triggers a collaborative response from community medical institutions, family doctors, and other parties, and pushes a risk assessment report to the user. During execution, the system collects post-intervention data through a feedback learning loop, automatically optimizes the interaction feature vector generation rules, the scale feature value extraction logic, and the contradiction calculation parameters, improving the accuracy of the assessment and the effectiveness of the intervention, and achieving dynamic and precise management of depression risk in the elderly.

[0083] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A geriatric depressive dynamic mental health risk early warning assessment and intervention system, characterized in that, The method comprises the following steps: a data acquisition module is configured to acquire natural language dialogue content of a user with an intelligent home assistant and to acquire standardized psychological scale response data filled in by the user; a dialogue recognition module is configured to separate instruction sentences and response sentences initiated by the user from the natural language dialogue content, and to extract an interaction feature vector, wherein the interaction feature vector comprises an interaction behavior feature vector generated by recognizing a behavior request type in the instruction sentences and an interaction emotion feature vector generated by analyzing emotion tendency words in the response sentences; a scale analysis module is configured to extract scale feature values from the standardized psychological scale response data, wherein the scale feature values comprise scale behavior feature values and scale emotion feature values; a risk verification module is configured to establish a dynamic mapping relationship between the interaction feature vector and the scale feature values, to calculate a contradiction degree of the double-source features, to start a risk verification process when the contradiction degree exceeds a preset threshold, to output a depression risk value by fusing the interaction feature vector and the contradiction degree index, and to trigger an intervention instruction when the risk value continuously exceeds a set threshold; In the risk verification module, the specific process of calculating the contradiction degree of the double-source features is as follows: The establishment of the dynamic mapping relationship comprises a first verification channel and a second verification channel, the first verification channel aligns the life self-care dimension of the interaction behavior feature vector with the scale behavior feature values in terms of trends, and calculates the directional deviation amount of the two in a continuous time window; the second verification channel matches the comprehensive intensity value of the interaction emotion feature vector with the scale emotion feature values in terms of numerical values, and detects whether the absolute difference amount exceeds a reasonable floating interval; The directional deviation amount and the absolute difference amount are fused to generate a primary contradiction degree index after normalization, and contradiction degree compensation calculation is added when the scale response data carries a low reliability identifier; In the risk verification module, the risk verification process specifically comprises: When the contradiction degree continuously exceeds the threshold, instruction feedback verification is performed on the interaction data to analyze whether the user instructions are correctly responded by the device and the subsequent interaction frequency changes; logical chain verification is performed on the scale data to detect whether the conflict options form a contradictory logical chain; The instruction feedback verification result and the logical chain verification result are converted into a data credibility level, which is used to correct the primary contradiction degree index; the corrected contradiction degree index, the interaction behavior feature vector and the interaction emotion feature vector are synchronously input into a risk fusion engine; The risk fusion engine first synthesizes a basic risk value from the interaction behavior feature vector and the interaction emotion feature vector according to a preset proportion; Then, a contradiction degree conversion function is used to map the corrected contradiction degree index to a risk amplification coefficient, and the basic risk value is multiplied by the risk amplification coefficient to obtain an enhanced risk value; Finally, a dynamic balance factor of the scale feature values is introduced, the enhanced risk value is positively calibrated when the scale data shows an obvious depression tendency, otherwise, a negative inhibition is loaded; finally, the enhanced risk value processed by the dynamic balance factor is output as a depression risk value.

2. The dynamic mental health risk early warning assessment and intervention system for geriatric depression of claim 1, wherein, In the dialogue recognition module, the specific process of generating the interaction behavior feature vector by recognizing the behavior request type in the instruction sentences is as follows: A three-level classification system of life self-care instructions, social seeking instructions, and medical consultation instructions is established, the life self-care instructions cover daily living basic operation requests, the social seeking instructions include interpersonal interaction active initiation expressions, and the medical consultation instructions specifically refer to non-emergency health problem inquiries; An instruction effectiveness filtering is implemented to exclude invalid instructions and repeated failed instructions that cannot be executed by the device; and finally, the distribution density of each type of instruction in a fixed period is counted, the deviation degree is calculated by combining historical baseline data to generate a three-dimensional interactive behavior feature vector.

3. The geriatric depressive dynamic mental health risk early warning assessment and intervention system according to claim 1, characterized in that, In the conversation recognition module, the specific process of generating an interactive emotion feature vector by analyzing the sentiment tendency words in the response statement is: Extracting explicit sentiment words matched by the basic sentiment word library from the response statement, and simultaneously identifying implicit sentiment reversal words affected by negative modifiers through a semantic association rule library; The intensity weight of the sentiment words is calculated by implementing a subject association strategy, and when the sentiment description subject is the user himself / herself, a standard weight coefficient is adopted, and when the description subject is a specific relationship person, a closeness correction coefficient is loaded; All sentiment words are processed by a time decay function to generate a time-weighted comprehensive sentiment intensity value, and the final output interactive emotion feature vector includes the weighted intensity difference value of positive and negative emotions.

4. The geriatric depressive dynamic mental health risk warning evaluation and intervention system according to claim 1, characterized in that, In the scale analysis module, the specific process of extracting scale feature values from standardized psychological scale response data is: Firstly, the response logic consistency is detected to analyze the option conflict of the same dimension problems in the scale; and then the filling behavior credibility is marked, and the reliability identification is added to the filling operation shorter than the regular time length or longer than the regular time length; The scale behavior feature value is derived from the aggregated score of the physiological activity type problem, and the emotion feature value is derived from the distribution analysis of the emotion state type problem, and the extraction result includes three data attributes of original filling value, reliability identification and logic conflict point list.

5. The geriatric depressive dynamic mental health risk warning evaluation and intervention system according to claim 1, characterized in that, The contradiction degree conversion function specifically includes: The contradiction degree index is divided into low contradiction region, medium contradiction region and high contradiction region according to the numerical value, when the contradiction degree index is located in the low contradiction region, the linear conversion mode is adopted, when the contradiction degree index is located in the medium contradiction region, the logarithmic growth mode is enabled, and when the contradiction degree index is located in the high contradiction region, the exponential amplification mode is activated; The conversion process is dynamically regulated by the data credibility level, if the data credibility level is greater than or equal to the preset threshold, the slope of the conversion curve is increased, if the data credibility level is less than the preset threshold, the output value range is compressed, and the risk amplification coefficient is set with a safety interval value.

6. The geriatric depressive dynamic mental health risk warning evaluation and intervention system according to claim 1, characterized in that, In the risk verification module, the intervention instruction specifically includes: When the depression risk value is abnormal and the scale feature value is normal, a hidden depression intervention instruction is generated, which includes an interactive feature analysis report and a recommended verification problem set; When both the source data are abnormal, an emergency medical intervention instruction is generated, and a feedback learning loop is embedded in the execution process of the emergency medical intervention instruction, and after collecting the intervention results, the interactive feature vector generation rules, the scale feature value extraction logic and the contradiction degree calculation parameters are automatically optimized.

Images