Mental health assessment system based on big data

Abstract

The invention discloses a psychological health assessment system based on big data. The psychological health assessment system comprises a data acquisition module, a data processing module, a dynamic modeling module, a risk assessment module and an intervention generation module. The data acquisition module acquires a heterogeneous data source of a user in real time through a distributed interface and generates structured data; the data processing module receives the structured data to generate a multi-modal feature vector; the dynamic modeling module receives the multi-modal feature vector, constructs a user psychological state atlas through a graph neural network, and updates a topological structure of the atlas based on a time sliding window to generate a psychological state evolution signal; the risk assessment module receives the psychological state evolution signal and forms an early warning signal; and the intervention generation module receives the early warning signal, generates a personalized intervention instruction and transmits the personalized intervention instruction to the user terminal. The psychological health assessment system based on the big data can solve the problems of calculation redundancy, poor geometric adaptability and low knowledge distillation efficiency when a traditional attention mechanism processes a circular region.

Description

Technical Field

[0001] The present invention relates to the cross - field of mental health monitoring and big data analysis, and particularly to a mental health assessment system based on big data. Background Art

[0002] Mental health assessment has long relied on psychological scales (such as SCL - 90, PHQ - 9) or clinical interviews, which are highly subjective, have a long update cycle, and are difficult to reflect the dynamic fluctuations of users' mental states in a timely manner. In recent years, some studies have attempted to introduce technical means to improve the assessment efficiency, such as analyzing emotional tendencies through social media text or monitoring physiological indicators such as heart rate variability using wearable devices. However, the existing technologies have significant limitations: First, the data dimension is single. It is difficult to comprehensively depict mental states relying only on text or physiological signals. For example, anxiety may be manifested as negative semantics in social media, decreased sleep quality, and withdrawal from daily behaviors simultaneously, but a single data source cannot capture such complex associations. Second, the model is static. Most systems train fixed models based on historical data and cannot adapt to the temporal evolution of individual mental states. For example, the symptoms of patients with post - traumatic stress disorder may show periodic aggravation, and static models are prone to misjudging their risk levels. Third, privacy protection is insufficient. Mental health data is highly sensitive, but existing systems mostly use basic anonymization processing, facing the risk of re - identification attacks and lacking fine - grained control by users over data usage. Fourth, the intervention is lagging. Traditional assessment results are usually presented in the form of reports and fail to be linked in real - time with resources such as psychological counseling and cognitive training, making it difficult for users to obtain immediate support.

[0003] Furthermore, existing technologies have attempted to integrate multi - modal data to improve assessment accuracy, such as combining speech emotion recognition and gait analysis. However, such solutions are mostly limited to laboratory environments and do not solve the problems of heterogeneous data alignment, real - time computing load, and privacy compliance in actual scenarios. At the algorithm level, although some studies have used graph neural networks to model mental feature associations, their node definitions are too general (such as only distinguishing between "positive" and "negative" emotions) and lack a dynamic update mechanism, resulting in the disconnection between the graph and the actual evolution of mental states. In terms of privacy protection, traditional encryption technologies can ensure data transmission security but cannot support model calculations in the ciphertext state, forcing the system to choose between data utility and privacy security. In addition, existing intervention strategies are mostly driven by fixed rules (such as setting thresholds to trigger notifications) and are difficult to adapt to individual - specific mental needs and behavioral habits. Summary of the Invention

[0004] In view of the above disadvantages of the prior art, the object of the present invention is to provide a mental health assessment system based on big data, which is used to solve the problems that traditional mental health assessments rely on subjective scales and single-dimensional data, resulting in poor real-time performance, weak privacy protection, and disjointed interventions. The present invention constructs a dynamic mental state map through multi-source heterogeneous data fusion to quantify the associations among emotions, cognitions, and behaviors; combines time series prediction and reinforcement learning to generate hierarchical early warnings and personalized intervention plans; and uses differential privacy and blockchain technologies to achieve data encryption and authorized traceability, balancing data utilization and privacy security.

[0005] The present invention provides a mental health assessment system based on big data, comprising: A data acquisition module, which acquires heterogeneous data sources of users in real time through a distributed interface and generates structured data; A data processing module, which receives the structured data, desensitizes sensitive information in the text data through a noise filtering algorithm, and generates multi-modal feature vectors; A dynamic modeling module, which receives the multi-modal feature vectors, constructs a user mental state map through a graph neural network, where nodes represent three types of mental dimensions: emotions, cognitions, and behaviors, and edge weights represent the dynamic association strengths among the dimensions, and updates the topological structure of the map based on a time sliding window to generate mental state evolution signals; A risk assessment module, which receives the mental state evolution signals, calculates the probability value of a mental crisis within a preset future period through a time series prediction model, and forms an early warning signal; An intervention generation module, which receives the early warning signals, matches at least one intervention plan from a predefined intervention resource library through a reinforcement learning strategy, generates personalized intervention instructions, and transmits them to the user terminal.

[0006] In an embodiment of the present invention, the distributed interface of the data acquisition module includes a text acquisition interface for connecting to a third-party social media platform, which obtains a real-time updated text data stream through an open authentication protocol authorized by the user and standardizes the data structures of different platforms; a signal synchronization interface for accessing wearable devices, which receives physiological signal data pushed by device manufacturers, and the signals include heart rate fluctuations, skin conductance changes, and sleep quality indicators, and a protocol adapter is built into the interface to be compatible with the communication specifications of different manufacturers; a file retrieval interface for docking with medical institutions, which periodically queries diagnostic records and medication information in the electronic health record through a database connection protocol; the data output by the text acquisition interface, the signal synchronization interface, and the file retrieval interface are uniformly encapsulated into a structured data format and then transmitted to the data processing module.

[0007] In an embodiment of the present invention, the noise filtering algorithm of the data processing module includes: desensitizing personal identity information in text data by using regular expression matching and mask replacement techniques, and at the same time identifying negative emotion keywords in the text through an emotion dictionary and a semantic analysis model and generating emotion intensity labels; using a moving average filtering algorithm to eliminate device measurement noise from physiological signal data, and filling in data missing segments caused by device offline through interpolation; extracting the user's screen usage duration, application switching frequency, and geographical location change sequence from behavior log data, and aligning them with emotion labels and filtered physiological signals along the time axis and fusing them into a multi-modal feature vector, where the feature vector includes a timestamp, a data source type, and normalized numerical features.

[0008] In an embodiment of the present invention, the method for constructing a graph neural network of the dynamic modeling module includes: dividing emotion dimension nodes into positive emotion, negative emotion, and neutral emotion sub-nodes, cognitive dimension nodes into attention level, memory ability, and decision-making tendency sub-nodes, and behavior dimension nodes into social activity, exercise frequency, and work and rest regularity sub-nodes; calculating the edge weights between different sub-nodes through an attention mechanism, and introducing a time decay factor to adjust the contribution degree of historical association strength; the span of the time sliding window is dynamically adjusted according to the user data update frequency. When the window slides, the connection relationships of old nodes beyond the time range are removed, and an updated psychological state evolution signal is generated based on the latest multi-modal feature vector.

[0009] In an embodiment of the present invention, the processing process of the time series prediction model includes: decomposing the psychological state evolution signal into a trend term, a periodic term, and a residual term, and respectively using a long short-term memory network to capture long-term dependencies, an autoregressive model to fit periodic fluctuations, and a Gaussian process regression to estimate uncertainties; weighting and fusing the three prediction results to obtain the probability value of a psychological crisis within a preset future period, where the weight coefficients are dynamically adjusted according to the prediction errors of the user's historical data; when the probability value continuously exceeds the threshold for a preset number of times, a warning signal upgrade mechanism is triggered, and the original medium-risk prompt signal is automatically upgraded to a high-risk warning signal.

[0010] In an embodiment of the present invention, the reinforcement learning strategy of the intervention generation module includes: defining the state space as a combination of the current user's psychological crisis probability value and historical intervention records, the action space as all available solution types and implementation intensities in the intervention resource library, and the reward function based on the improvement degree of the user's psychological state and compliance score after the intervention; iteratively optimizing the action selection strategy through a policy gradient algorithm, so that high-risk warning signals preferentially trigger the matching of artificial psychological counseling services, and medium-risk prompt signals preferentially trigger the push of self-help cognitive training courses; after each intervention instruction is issued, the click-through rate, completion rate, and subsequent psychological state change data feedback by the user terminal are collected to update the parameters of the reinforcement learning model.

[0011] In one embodiment of the present invention, it further includes a privacy protection module. Before the data acquisition module transmits structured data to the data processing module, Laplace noise that meets the requirements of differential privacy is added to the geographical location information and device identifier in the text data, so that the true user identity cannot be inferred reversely from a single piece of data. After the data processing module generates the multi-modal feature vector, the privacy protection module uses homomorphic encryption technology to encrypt the feature vector to ensure that the dynamic modeling module directly executes the graph construction operation without decrypting. When the user initiates a data deletion request, the privacy protection module verifies the user identity and operation authority through the hash value recorded in the blockchain, and then sends an irreversible erasure instruction to all modules storing the user's data.

[0012] In one embodiment of the present invention, the interaction method of the user terminal includes, after receiving a personalized intervention instruction, displaying an interactive virtual assistant interface, which provides a graphic and text-guided cognitive behavior training task, a video connection entry for a psychological counselor, and a panel for setting medication reminders. After the user completes the task, the terminal collects the real-time change data of their physiological signals and subjective emotion scores, and transmits them back to the data processing module through the data acquisition module. When the user continuously refuses to execute the intervention instruction up to a preset number of times, the terminal automatically triggers the emergency contact notification function and sends a behavior abnormality mark to the risk assessment module to recalculate the psychological crisis probability value.

[0013] In one embodiment of the present invention, the operation architecture of the mental health assessment system based on big data includes that the data acquisition module and the data processing module are deployed on edge computing nodes for real-time processing of high-frequency updated text and physiological signal data; the dynamic modeling module and the risk assessment module are deployed in the cloud computing center for executing large-scale graph construction and prediction model operations; the intervention generation module adopts a distributed microservice architecture and dynamically selects the nearest service node according to the user's geographical location to send intervention instructions; asynchronous communication is realized between each module through a message middleware. When any module fails, the middleware automatically caches the unprocessed data and redelivers it after the module recovers, ensuring the continuity of the assessment process.

[0014] In an embodiment of the present invention, the dynamic modeling module further includes a real-time feedback calibration mechanism. When the user terminal returns negative feedback on the intervention instruction, a feedback calibration signal is generated and transmitted to the dynamic modeling module. The dynamic modeling module dynamically attenuates the edge weights of the corresponding behavior dimension nodes in the mental state map according to the feedback calibration signal, and recalculates the association strength between the emotion and cognitive dimensions based on the attenuated edge weights. At the same time, the recalculated association strength is compared and analyzed with the historical map data. If the deviation of the association strength exceeds the preset floating range, a map reconstruction instruction is triggered to initialize a new graph neural network model. During the execution of the map reconstruction instruction, the historical multi-modal feature vectors of the user within three months are retained as training data, and transfer learning technology is used to inherit the topological features of the original map to accelerate model convergence.

[0015] The mental health assessment system based on big data provided by the present invention constructs a dynamic mental state map through multi-source heterogeneous data fusion to quantify the associations among emotions, cognitions, and behaviors; combines time series prediction and reinforcement learning to generate hierarchical early warnings and personalized intervention plans; and uses differential privacy and blockchain technology to achieve data encryption and authorized traceability, balancing data utilization and privacy security. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0017] Figure 1 It is a system architecture diagram of a mental health assessment system based on big data. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] The following specific examples illustrate the embodiments of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, without conflict, the following embodiments and the features in the embodiments can be combined with each other.

[0019] It should be noted that the diagrams provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner. Therefore, only the components related to the present invention are shown in the diagrams, rather than being drawn according to the number, shape, and size of the components in actual implementation. The type, quantity, and proportion of each component in actual implementation can be arbitrarily changed, and the component layout type may also be more complex.

[0020] In the following description, numerous specific details are set forth to provide a more thorough explanation of embodiments of the present invention. However, it will be apparent to those skilled in the art that embodiments of the present invention may be practiced without these specific details. In other instances, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the present invention.

[0021] Please refer to Figure 1 , which shows the mental health assessment system based on big data of the present invention. The mental health assessment system based on big data of the present invention includes a data acquisition module, a data processing module, a dynamic modeling module, a risk assessment module, and an intervention generation module. The data acquisition module obtains heterogeneous data sources of users in real time through a distributed interface and generates structured data; the data processing module receives the structured data, desensitizes sensitive information in the text data through a noise filtering algorithm, and generates a multi-modal feature vector; the dynamic modeling module receives the multi-modal feature vector, constructs a user mental state map through a graph neural network, where nodes represent three types of psychological dimensions: emotion, cognition, and behavior, and edge weights represent the dynamic association strength between each dimension, and updates the topological structure of the map based on a time sliding window to generate a mental state evolution signal; the risk assessment module receives the mental state evolution signal, calculates the probability value of a mental crisis within a preset future period through a time series prediction model, and forms a warning signal; the intervention generation module receives the warning signal, matches at least one intervention plan from a predefined intervention resource library through a reinforcement learning strategy, generates a personalized intervention instruction, and transmits it to the user terminal.

[0022] As Figure 1As shown in the figure, the dynamic monitoring and personalized intervention of mental health status are realized through multi-module collaboration. First, the system obtains the heterogeneous data sources of users in real time through the data acquisition module, including social media texts, physiological signals of wearable devices, behavior logs of mobile terminals, and electronic health record data. The data acquisition module adopts a distributed interface design and can connect to the data sources of different platforms and devices. For example, it can establish a connection with the social media platform through the open interface protocol to extract the text content published by users in real time; communicate with wearable devices such as smart bracelets and heart rate monitors through Bluetooth or wireless network protocols to obtain physiological indicators such as heart rate, steps, and sleep quality; collect behavior data such as the total screen usage time, application switching frequency, and geographical location movement trajectory of users through the sensors or application programming interfaces built into the mobile terminal; at the same time, by docking with the data system of medical institutions, regularly retrieve the electronic health records of users, including structured data such as historical diagnosis records, medication plans, and physical examination reports. These heterogeneous data are initially cleaned and format-standardized in the acquisition module. For example, data in different time zones are uniformly converted to Coordinated Universal Time, unstructured text data is converted into structured data packets containing timestamps, user identifiers, and content bodies, and all data is encapsulated into a unified transmission format and then sent to the data processing module. After receiving the structured data from the acquisition module, the data processing module performs multi-level noise filtering and feature extraction. For text data, first, personal identity information such as mobile phone numbers, ID numbers, and addresses is identified through regular expression matching technology, and desensitization processing is performed using mask replacement or hash encryption; at the same time, a pre-trained sentiment analysis model is used to classify the emotions of the text content, identify positive, negative, or neutral emotions, and quantify the emotion intensity in combination with the sentiment dictionary. For physiological signal data, a moving average filtering algorithm is used to eliminate noise caused by device measurement errors or environmental interference. For example, a moving average process with a window length of five seconds is performed on the heart rate signal to smooth abnormal fluctuations; for data missing segments caused by device offline or transmission interruption, linear interpolation or a method based on the data trend prediction of adjacent time points is used to fill in the gaps. Behavior log data extracts key indicators through feature engineering, such as the total daily screen usage time, high-frequency usage application categories, and geographical location stay time, and slices and aligns these indicators at a minute-level time granularity. Finally, the processing module aligns the text emotion labels, filtered physiological signal slices, and behavior feature indicators along the time axis, and fuses them to generate multi-modal feature vectors containing timestamps, data source types, and multi-dimensional feature values. These vectors are organized through a specific data structure and then transmitted to the dynamic modeling module.

[0023] Further, after receiving the multi-modal feature vectors, the dynamic modeling module constructs a user mental state map using a graph neural network. The nodes in the map are divided into three main dimensions: The emotion dimension is further subdivided into sub-nodes of positive emotion, negative emotion, and neutral emotion. The cognitive dimension includes sub-nodes of attention level, memory ability, and decision-making tendency. The behavior dimension covers sub-nodes of social activity level, exercise frequency, and daily routine regularity. The weight of each edge is dynamically calculated through an attention mechanism. For example, when calculating the association strength between the negative emotion node and the social activity level node, the frequency of their co-occurrence and temporal proximity in historical data will be combined, and an exponential decay function is introduced to reduce the influence weight of long-term historical data. The topological structure of the map is dynamically updated based on the time-sliding window mechanism, and the window span is adaptively adjusted according to the data update frequency: when the user data is updated frequently, the window is shortened to capture short-term changes; otherwise, the window is extended to extract long-term trends. Each time new data arrives, the connection relationships of the old nodes outside the current window time range are removed, and the edge weights between each sub-node are recalculated based on the latest feature vectors. The updated map generates a mental state evolution signal, which includes the temporal change trend of the association strength of each dimension and the key inflection point markers, and is transmitted to the risk assessment module. After receiving the mental state evolution signal, the risk assessment module uses a time series prediction model to analyze the probability of future mental crises. The model first decomposes the input signal into a trend term, a periodic term, and a residual term: The trend term captures the long-term evolution direction of the mental state through a long short-term memory network. For example, a continuously increasing association weight of negative emotion may indicate an exacerbation of depressive tendencies; The periodic term uses an autoregressive model to identify fluctuations such as circadian rhythms or weekly regularities. For example, it is found that there is a peak in anxiety emotions on Sunday evenings for the user; The residual term quantifies the uncertainty of the model prediction through Gaussian process regression. The three prediction results are dynamically weighted according to the historical prediction accuracy. For example, when the periodic term performs stably in recent predictions, a higher weight is assigned. Finally, the weighted fusion obtains the probability value of mental crises within the next three days to one week. When the probability value exceeds the preset first threshold, a high-risk warning signal is generated to trigger immediate manual intervention; when it is between the first and second thresholds, a medium-risk reminder signal is generated to initiate a self-help intervention process. The warning signal includes the risk level, the main incentive dimension (such as abnormal emotion-behavior association), and the confidence assessment, and is transmitted to the intervention generation module through an encrypted channel. After receiving the warning signal, the intervention generation module matches a personalized intervention plan from a predefined resource library based on a reinforcement learning strategy. The resource library includes various intervention types such as access to psychological counseling services, cognitive-behavioral training courses, mindfulness meditation guidance, and medication compliance reminders. Each type is set with different implementation intensities and time arrangements. The state space of the reinforcement learning model is jointly composed of the current risk level, the user's historical intervention response records, and the environmental context (such as the current time period, geographical location). The action space is defined as all optional intervention plans and their combination methods, and the reward function is dynamically calculated according to the improvement degree of the user's physiological indicators, the task completion rate, and the subjective feedback score after the intervention.For example, for high-risk warnings, the model preferentially matches real-time video psychological counseling and synchronously pushes relaxation training; for medium-risk prompts, it combines and sends cognitive-behavioral practice tasks and medication reminders. The generated intervention instructions are pushed to the user terminal in the form of graphics, voice, or video, and the user interaction behavior is monitored in real time: if the user does not respond within the set time, the intervention intensity is automatically increased or the solution type is switched. At the same time, the user feedback data collected by the terminal (such as the completion progress of training tasks, the heart rate change curve) is transmitted back to the dynamic modeling module for updating the edge weight calculation of the mental state map, forming a closed-loop optimization mechanism of evaluation-intervention-feedback.

[0024] In an embodiment of the present invention, the specific interface composition and data processing logic of the data collection module are further defined. This module includes three core interface components: a text collection interface, a signal synchronization interface, and an archive retrieval interface. The text collection interface is docked with mainstream social media platforms through an open authentication protocol, and uses an asynchronous request mechanism to batch obtain user-authorized posts, comments, and private message data. The interface is built with a traffic control module that dynamically adjusts the request frequency according to the platform interface restrictions to avoid triggering the flow-limiting mechanism due to high-frequency access. At the same time, a data structure conversion layer is designed to uniformly convert JSON or XML format data returned by different platforms into a standard structure containing user identifiers, posting times, text content, and sentiment tags. The signal synchronization interface is responsible for accessing multi-brand wearable devices and receiving physiological signal data through the cloud push service provided by the device manufacturer. The interface is built with a protocol adapter that can parse binary or custom encoding formats of different manufacturers. For example, it maps the sleep stage encoding of a certain brand of smart bracelet to standardized light sleep, deep sleep, and REM sleep tags, and converts the acceleration sensor data into a step count value. The archive retrieval interface is connected to the medical institution database through an encrypted tunnel and uses an incremental query strategy to regularly obtain the latest electronic health records of users: the last acquisition timestamp is recorded during each query, and only new data after this time point is requested subsequently to reduce the network transmission load. After the data output by all interface components passes format verification and outlier filtering, it is uniformly packaged by the data encapsulation engine into a structured data packet containing a metadata header and a payload body. The metadata includes data source identification, collection time, and data integrity verification code, and the payload body arranges the data values of each dimension in the preset field order to ensure that the subsequent processing module can efficiently parse it.

[0025] Such as Figure 1As shown in the figure, for text data, the desensitization process is divided into two stages: First, use regular expressions to match predefined sensitive information patterns. For example, the regular pattern for mobile phone numbers is "1[3-9]\d{9}", and the rules for matching the first six-digit area code and the last digit verification code of the ID card number are used; the content that is matched is replaced with asterisks for the middle four digits or encrypted irreversibly using the SHA-256 hashing algorithm. Second, identify potential sensitive content through a sentiment analysis model: A text classification model based on the BERT architecture judges the emotional polarity of the input statement. When detecting high-risk content such as suicidal tendencies and self-harm descriptions, an emergency mark is appended to the emotion label, and a real-time warning bypass channel is triggered. Physiological signal processing uses a moving average filter combined with an outlier rejection strategy: Taking the heart rate signal as an example, set the window size to 5 sampling points, calculate the median and standard deviation of the values within the window. When a sampling point deviates from the median by more than three standard deviations, it is regarded as noise and eliminated, and the mean value of the remaining data within the window is used to fill the gap. For long-term signal loss (such as the device being offline for more than ten minutes), a prediction method based on an autoregressive model is used to generate substitute data, and the source of the complement is noted in the data flag bit. Behavior log processing focuses on spatio-temporal feature extraction: Calculate the number of daily stay points, the category of frequently visited areas, and the entropy value of the movement trajectory from the geographical location data to quantify the behavioral regularity; extract the application switching frequency, the distribution of single-use duration, and the proportion of night-time usage duration from the screen usage data. All features are aligned according to a minute-level time window. For data sources with different sampling frequencies (such as the heart rate sampled per second and the location recorded per minute), the nearest neighbor interpolation method is used to up-sample the low-frequency data to the high-frequency time axis. The finally generated multi-modal feature vector adopts a hierarchical storage structure: The first layer is the timestamp and data source type encoding, the second layer contains the text sentiment intensity, physiological index statistical values, and behavioral feature vectors, and the third layer is appended with data quality flag bits (such as desensitization marks, complement marks) to ensure that downstream modules can accurately understand the data semantics and reliability.

[0026] As Figure 1As shown, it emphasizes the fine-grained division of the psychological state dimension and the calculation mechanism of dynamic association strength. This module decomposes the emotion dimension into three sub-nodes: positive emotion, negative emotion, and neutral emotion. Among them, the positive emotion sub-node is further divided into secondary categories such as pleasure, satisfaction, and excitement. The negative emotion sub-node includes secondary categories such as anxiety, depression, and anger. The neutral emotion sub-node covers states such as calmness and indifference. The cognitive dimension is divided into sub-nodes of attention level, memory ability, and decision-making tendency. The attention level sub-node is quantified according to the attention concentration duration and distraction frequency when the user completes tasks. The memory ability sub-node is evaluated based on the results of short-term memory tests and the frequency of daily forgetting events. The decision-making tendency sub-node is determined by analyzing the hesitation duration and risk preference of the user when facing choices. The behavior dimension is subdivided into sub-nodes of social activity level, exercise frequency, and work and rest regularity. The social activity level sub-node counts the number and duration of online and offline social interactions of the user. The exercise frequency sub-node records the daily step count, exercise type, and duration. The work and rest regularity sub-node analyzes the stability and deviation degree of the sleep-wake cycle. The edge weight calculation between each sub-node adopts the multi-head attention mechanism, and each attention head focuses on the association patterns at different time scales. For example, the short-term attention head analyzes the impact of emotional fluctuations in the last hour on the decision-making tendency, and the long-term attention head examines the correlation between the behavior patterns in the past week and the cognitive ability. The time decay factor acts on the historical association data in the form of an exponential function, making the contribution degree of interaction events farther from the current time to the edge weight lower, ensuring that the graph reflects the latest changes in the psychological state. The span of the time sliding window is dynamically adjusted according to the data input frequency: when the user data update interval is less than the preset threshold, the window span is automatically shortened to half of the original value to capture rapid changes; when the data update interval exceeds the threshold, the window span is gradually expanded to twice to integrate long-term trends. When the window slides, the system identifies the historical node connections that exceed the current time range, such as the association record between a high-intensity exercise three days ago and the negative emotion on the current day, and removes these expired connections to release computing resources. After the removal operation, the connection relationship of the affected sub-nodes is re-initialized based on the latest received multi-modal feature vectors. For example, according to the latest detected anxiety emotion intensity and the decrease in social activity level data, the edge weights between the corresponding sub-nodes are newly established or adjusted. When the updated psychological state graph generates an evolution signal, the persistent homology features of the graph are extracted using topological data analysis methods, identifying the strongly associated patterns that persist within the time window (such as the stable association between negative emotion and disrupted work and rest) and the weakly associated patterns that appear briefly (such as the short-term decrease in attention caused by a sudden stress event), and encoding these patterns as time series signals and transmitting them to the risk assessment module.

[0027] Furthermore, the model first performs empirical mode decomposition on the input psychological state evolution signal, separating the trend term representing the long-term development trend, the periodic term reflecting periodic fluctuations, and the residual term containing random noise and unexpected events. The trend term analysis uses a long short-term memory network architecture. The input layer of the network receives slices of the evolution signal for the past 30 days, controls the retention and update of historical information through the forget gate and the input gate, and the output layer predicts the trend in the next seven days, such as the correlation weight of continuously rising anxiety or the sleep regularity index that gradually recovers. The periodic term processing uses an autoregressive model with exogenous variables, taking known periodic factors (such as the division of weekdays and weekends, seasonal change labels) as covariates. The model automatically identifies the inherent cycle of the user's psychological state (such as the emotional trough that occurs every two weeks) and conducts a correlation test with external periodic factors to eliminate accidental periodic patterns. The residual term analysis uses Gaussian process regression, defines the covariance relationship between different time points through the kernel function, models the probability distribution of the random fluctuations unexplained by the model, and outputs the confidence interval of the future psychological crisis probability. The three prediction results are fused through dynamic weight coefficients: the weight coefficients are updated online according to the prediction errors in the past 15 days, and the error calculation uses the root mean square error and the direction consistency weighted index. For example, when the trend term has been continuously highly accurate in recent predictions, its weight is increased to 60% of the total weight; when the width of the confidence interval of the residual term exceeds the threshold, its weight is reduced to avoid over-reliance on predictions with high uncertainty. The fused psychological crisis probability value is refreshed every six hours. When this value exceeds the high-risk threshold three times in a row, the early warning signal upgrade mechanism is triggered, automatically upgrading the original medium-risk prompt to a high-risk early warning, marking this user as a priority monitoring object at the same time, and starting the high-frequency data collection mode (such as shortening the physiological signal sampling interval from five minutes to 30 seconds). After the early warning signal is generated, the system additionally generates an incentive analysis report, identifies the associated pattern that contributes the most to the probability value through counterfactual reasoning methods (for example, if the edge weight between negative emotions and social activity suddenly becomes the main incentive, the report highlights the urgency of social intervention), and encrypts and transmits the report to the intervention generation module for decision-making reference.

[0028] Such as Figure 1As shown, it is the construction and optimization process of the reinforcement learning strategy in the intervention generation module. This strategy defines the user's current mental state as the core dimension of the state space, including the real-time psychological crisis probability value, the intervention response records in the past seven days (such as click-through rate, task completion rate, improvement range of physiological indicators), and environmental context information (such as whether the current time period is at night, whether the user is in a home environment). The action space covers all feasible solutions and their combination methods in the predefined intervention resource library. For example, single actions include sending cognitive training tasks, booking a psychological counselor, and pushing a mindful breathing guidance video. The combined action may be to send a training task and a medication reminder simultaneously. The design of the reward function adopts a multi-objective optimization framework, including short-term rewards (such as the degree of improvement in the user's heart rate variability within two hours after the intervention instruction is issued), medium-term rewards (such as the decreasing trend of screen usage time within three consecutive days), and long-term rewards (such as the decreasing range of the predicted psychological crisis probability value in the next week). The weights of each objective are dynamically adjusted according to the user's intervention effect preferences in historical data. During the training process of the policy gradient algorithm, the model maintains an exploration-exploitation balance mechanism: in the case of new users or data-sparse scenarios, increase random exploration to try diverse intervention combinations; in users with sufficient data, preferentially select historical high-reward actions for exploitation. For high-risk warning signals, the policy sets intervention priority rules. For example, when the psychological crisis probability value exceeds the high-risk threshold and the user is in a solitary state, force the real-time intervention action of a human psychological counselor to be triggered and send a notification to the preset emergency contact simultaneously; for medium-risk warning signals, adopt a progressive intervention strategy. In the first round, push mild intervention content (such as a five-minute breathing exercise). If the user does not respond, gradually upgrade to medium intervention (such as a 20-minute cognitive training course). After each intervention instruction is issued, the system collects multi-dimensional feedback data through the user terminal: explicit feedback includes the user's rating of the intervention content, task completion progress, and subjective emotion self-evaluation; implicit feedback is obtained by analyzing the changes in physiological signals (such as the decreasing range of skin conductance), adjustment of behavior patterns (such as an increase in the usage time of social applications), and improvement degree of text sentiment tendency after the intervention. These feedback data are processed and converted into reward signals for reinforcement learning to update the parameters of the policy network online. For example, when a certain cognitive training course results in a reduction in the user's negative emotion label and the task completion rate reaches more than 90%, the priority selection weight of the corresponding action is significantly increased. In addition, the policy model performs an offline batch retraining once a month, using the historical interaction data of all users to globally optimize the network parameters and prevent the local optimum trap caused by online learning.

[0029] Furthermore, in the data collection phase, the module adds Laplace noise that meets the requirements of differential privacy to the geographical location information and device unique identifier in the text data. The noise magnitude is adaptively adjusted according to the field sensitivity: for example, the sensitivity of geographical location longitude and latitude is relatively high, and a relatively large magnitude of noise perturbation is used to blur the frequently visited locations of individual users into areas at the 100-meter level; after noise processing, the device identifier cannot form a one-to-one mapping with the real device, but information required for aggregation analysis such as the device type (e.g., smartphone model) is retained. In the data processing phase, the multi-modal feature vectors are immediately encrypted using the fully homomorphic encryption algorithm after generation, enabling the dynamic modeling module to directly perform graph construction operations in the ciphertext state. For example, the correlation weight is calculated between the encrypted emotion intensity value and physiological indicators without decrypting the original data. The encryption key is independently managed by the user, and the system can only access the plaintext data for a specific period when obtaining temporary decryption authorization (e.g., emergency calls for recent records are required for crisis intervention). Data storage adopts a distributed fragmentation scheme, splitting the complete data record of a single user into multiple fragments and storing them separately in different geographical nodes. Each fragment is encrypted individually and cannot independently restore valid information. When the user initiates a data deletion request, the privacy protection module starts the blockchain verification process: the user digitally signs the deletion instruction with the private key, and after the smart contract verifies the signature validity and user permissions, it sends an erasure instruction containing the data hash value to all storage nodes. After receiving the instruction, the storage nodes retrieve the fragmented data with the matching hash value in the local database, perform irreversible multiple overwrite erasure operations, and write the operation logs to the blockchain for audit tracking. In addition, the module sets up a privacy leakage risk assessment subsystem, regularly simulating scenarios such as re-identification attacks and association inference attacks, calculating the residual risk value under the current data protection mechanism. If the risk value exceeds the security threshold, it automatically triggers the data re-anonymization process or restricts the query permissions for high-risk data. During data transmission, all communication between modules uses the quantum key distribution protocol to establish a secure channel to ensure data confidentiality even in the face of future quantum computing attacks. The privacy protection module also provides a user-transparent data usage dashboard, allowing users to view in real-time the call records, processing purposes, and remaining storage periods of each module for their data, and supporting one-click authorization revocation and data export functions to ensure compliance with the requirements of major global data protection regulations (such as the EU General Data Protection Regulation).

[0030] A mental health assessment system based on big data according to the present invention constructs a dynamic mental state graph through multi-source heterogeneous data fusion to quantify the associations among emotions, cognitions, and behaviors; generates hierarchical early warnings and personalized intervention plans by combining time series prediction and reinforcement learning; and uses differential privacy and blockchain technologies to achieve data encryption and authorization traceability, balancing data utilization and privacy security.

[0031] Therefore, through a big data-based mental health assessment system of the present invention, the problems of poor real-time performance, weak privacy protection, and disconnection of intervention existing in the traditional mental health assessment relying on subjective scales and one-dimensional data can be solved.

[0032] The above embodiments merely illustrate the principles and effects of the present invention, rather than limiting the present invention. Any person familiar with this technology can modify or change the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or changes completed by those with ordinary knowledge in the technical field without departing from the spirit and technical idea disclosed by the present invention should still be covered by the claims of the present invention.

Claims

1. A mental health assessment system based on big data, characterized in that, Including: A data acquisition module that obtains the heterogeneous data sources of users in real time through a distributed interface and generates structured data; A data processing module that receives the structured data, desensitizes sensitive information in the text data through a noise filtering algorithm, and generates a multi-modal feature vector; A dynamic modeling module that receives the multi-modal feature vector, constructs a user mental state map through a graph neural network, where the nodes represent three psychological dimensions of emotion, cognition, and behavior, the edge weights represent the dynamic association strength between each dimension, and updates the topological structure of the map based on a time sliding window to generate a mental state evolution signal; A risk assessment module that receives the mental state evolution signal, calculates the probability value of a mental crisis within a preset future period through a time series prediction model, and forms a warning signal; An intervention generation module that receives the warning signal, matches at least one intervention plan from a predefined intervention resource library through a reinforcement learning strategy, generates a personalized intervention instruction, and transmits it to the user terminal.

2. The mental health assessment system based on big data according to claim 1, wherein The distributed interface of the data acquisition module includes a text acquisition interface for connecting to a third-party social media platform, obtains a real-time updated text data stream through an open authentication protocol authorized by the user, and performs a standardized conversion on the data structures of different platforms; A signal synchronization interface for accessing wearable devices, receives physiological signal data pushed by device manufacturers, the signals include heart rate fluctuations, skin conductance changes, and sleep quality indicators, and the interface is built with a protocol adapter to be compatible with the communication specifications of different manufacturers; a file retrieval interface for docking medical institutions, periodically queries the diagnosis records and medication information in the electronic health record through a database connection protocol; the data output by the text acquisition interface, signal synchronization interface, and file retrieval interface are uniformly encapsulated into a structured data format and then transmitted to the data processing module.

3. The mental health assessment system based on big data according to claim 1, characterized in that, The noise filtering algorithm of the data processing module includes: desensitizing personal identity information in the text data using regular expression matching and mask replacement techniques, and at the same time identifying negative emotion keywords in the text through an emotion dictionary and a semantic analysis model to generate emotion intensity labels; using a moving average filtering algorithm for physiological signal data to eliminate device measurement noise, and filling in the data missing segments caused by device offline through interpolation; extracting the user's screen usage duration, application switching frequency, and geographical location change sequence from the behavior log data, and aligning and fusing them with the emotion labels and filtered physiological signals along the time axis into a multi-modal feature vector, the feature vector contains a timestamp, data source type, and normalized numerical features.

4. The mental health assessment system based on big data according to claim 1, characterized in that, The method for constructing the graph neural network of the dynamic modeling module includes: dividing the emotion dimension nodes into positive emotion, negative emotion, and neutral emotion sub-nodes, dividing the cognitive dimension nodes into attention level, memory ability, and decision-making tendency sub-nodes, and dividing the behavior dimension nodes into social activity level, exercise frequency, and daily routine regularity sub-nodes; calculating the edge weights between different sub-nodes through the attention mechanism and introducing a time decay factor to adjust the contribution degree of the historical correlation strength; the span of the time sliding window is dynamically adjusted according to the user data update frequency. When the window slides, the old node connection relationships beyond the time range are removed, and updated psychological state evolution signals are generated based on the latest multi-modal feature vectors.

5. The mental health assessment system based on big data according to claim 1, characterized in that The processing process of the time series prediction model includes: decomposing the psychological state evolution signal into a trend term, a periodic term, and a residual term, respectively using a long short-term memory network to capture long-term dependencies, an autoregressive model to fit periodic fluctuations, and a Gaussian process regression to estimate uncertainties; weighting and fusing the three prediction results to obtain the probability value of a psychological crisis within a preset future period, where the weight coefficients are dynamically adjusted according to the prediction errors of the user's historical data; when the probability value continuously exceeds the threshold for a preset number of times, a warning signal upgrade mechanism is triggered to automatically upgrade the original medium-risk warning signal to a high-risk warning signal.

6. The mental health assessment system based on big data according to claim 1, characterized in that The reinforcement learning strategy of the intervention generation module includes: defining the state space as a combination of the current user's psychological crisis probability value and historical intervention records, the action space as all available program types and implementation intensities in the intervention resource library, and the reward function based on the improvement degree of the user's psychological state and compliance score after the intervention; iteratively optimizing the action selection strategy through the policy gradient algorithm, so that the high-risk warning signal preferentially triggers the matching of artificial psychological counseling services, and the medium-risk warning signal preferentially triggers the push of self-help cognitive training courses; after each intervention instruction is issued, the click-through rate, completion rate, and subsequent psychological state change data feedback by the user terminal are collected to update the parameters of the reinforcement learning model.

7. The mental health assessment system based on big data according to claim 6, characterized in that, It also includes a privacy protection module. Before the data collection module transmits structured data to the data processing module, Laplace noise that meets the requirements of differential privacy is added to the geographical location information and device identifiers in the text data, so that the true user identity cannot be inferred from a single piece of data; after the data processing module generates multi-modal feature vectors, the privacy protection module uses homomorphic encryption technology to encrypt the feature vectors to ensure that the dynamic modeling module directly executes the graph construction operation without decryption; when the user initiates a data deletion request, the privacy protection module verifies the user identity and operation permissions through the hash value recorded in the blockchain, and sends an irreversible erasure instruction to all data collection modules storing the user's data.

8. The mental health assessment system based on big data according to claim 1, characterized in that The interaction method of the user terminal includes, after receiving a personalized intervention instruction, displaying an interactive virtual assistant interface, which provides a graphic and text-guided cognitive behavior training task, a video connection entry for a psychological counselor, and a drug-taking reminder setting panel. After the user completes the task, the terminal collects the real-time change data of their physiological signals and subjective emotion scores, and transmits them back to the data processing module through the data collection module; When the user continuously refuses to execute the intervention instruction up to the preset number of times, the terminal automatically triggers the emergency contact notification function and sends a behavior anomaly mark to the risk assessment module to recalculate the psychological crisis probability value.

9. The mental health assessment system based on big data according to claim 1, characterized in that, The operation architecture of the big data-based mental health assessment system includes that the data collection module and the data processing module are deployed on the edge computing node for real-time processing of high-frequency updated text and physiological signal data; the dynamic modeling module and the risk assessment module are deployed in the cloud computing center for performing large-scale graph construction and prediction model operations; the intervention generation module adopts a distributed microservice architecture to dynamically select the nearest service node according to the user's geographical location to issue intervention instructions; asynchronous communication is realized between modules through the message middleware. When any module fails, the middleware automatically caches the unprocessed data and redelivers it after the module recovers to ensure the continuity of the assessment process.

10. The mental health assessment system based on big data according to claim 1, characterized in that, The dynamic modeling module also includes a real-time feedback calibration mechanism. When the user terminal returns negative feedback on the intervention instruction, a feedback calibration signal is generated and transmitted to the dynamic modeling module; the dynamic modeling module dynamically attenuates the edge weights of the corresponding behavior dimension nodes in the mental state graph according to the feedback calibration signal, and recalculates the association strength between the emotion and cognitive dimensions based on the attenuated edge weights; at the same time, the recalculated association strength is compared and analyzed with the historical graph data. If the deviation of the association strength exceeds the preset floating range, a graph reconstruction instruction is triggered to initialize a new graph neural network model; during the execution of the graph reconstruction instruction, the historical multi-modal feature vectors of the user within three months are retained as training data, and transfer learning technology is used to inherit the topological features of the original graph to accelerate model convergence.

Images