A mental health assessment system based on big data

Abstract

The present invention provides a big data-based mental health assessment system, comprising 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 the user's heterogeneous data sources in real time through a distributed interface and generates structured data; the data processing module receives the structured data and generates a multimodal feature vector; the dynamic modeling module receives the multimodal feature vector, constructs a user psychological state map through a graph neural network, and generates a psychological state evolution signal by updating the topological structure of the map based on a time sliding window; the risk assessment module receives the psychological state evolution signal and forms an early warning signal; the intervention generation module receives the early warning signal, generates a personalized intervention instruction, and transmits it to the user terminal. The big data-based mental health assessment system of the present invention can solve the problems of computational redundancy, poor geometric adaptability, and low knowledge distillation efficiency of the traditional attention mechanism when processing circular areas.

Description

Technical Field

[0001] The present invention relates to the intersection of mental health monitoring and big data analysis technology, and specifically to a mental health assessment system based on big data. Background Art

[0002] Mental health assessments have long relied on psychological scales (such as the SCL-90 and PHQ-9) or clinical interviews. These are highly subjective, have long update cycles, and are unable to promptly reflect the dynamic fluctuations of a user's mental state. In recent years, some research has attempted to improve assessment efficiency by introducing technological means, such as analyzing emotional tendencies through social media text or using wearable devices to monitor physiological indicators such as heart rate variability. However, existing technologies have significant limitations: First, the data dimension is single, and it is difficult to fully characterize the psychological state by relying solely on text or physiological signals. For example, anxiety may be manifested simultaneously as negative semantics in social media, decreased sleep quality, and withdrawal from daily behaviors, 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 psychological states. For example, the symptoms of patients with post-traumatic stress disorder may show periodic exacerbations, 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, which faces the risk of re-identification attacks and lacks fine-grained control over data use by users; fourth, intervention lags. Traditional assessment results are usually presented in the form of reports, which 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 attempt to integrate multimodal data to improve assessment accuracy, such as combining speech emotion recognition with gait analysis. However, such solutions are mostly limited to laboratory environments and do not address heterogeneous data alignment, real-time computing load, and privacy compliance issues in real-world scenarios. At the algorithmic level, although some studies have used graph neural networks to model psychological feature associations, their node definitions are too general (for example, only distinguishing between "positive" and "negative" emotions) and lack a dynamic update mechanism, resulting in a disconnect between the graph and the actual evolution of psychological states. In terms of privacy protection, while traditional encryption technology can ensure the security of data transmission, it cannot support model calculations in an encrypted 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), which makes it difficult to adapt to the differentiated psychological needs and behavioral habits of individuals. Summary of the Invention

[0004] In light of the shortcomings of the aforementioned prior art, the present invention aims to provide a big data-based mental health assessment system to address the shortcomings of traditional mental health assessments, which rely on subjective scales and single-dimensional data, resulting in poor real-time performance, weak privacy protection, and disconnected interventions. This system fuses multi-source heterogeneous data to construct a dynamic mental state map to quantify the correlations between emotions, cognition, and behavior. It also combines time series prediction with reinforcement learning to generate graded warnings and personalized intervention plans. Furthermore, it employs differential privacy and blockchain technology to achieve data encryption and authorized traceability, balancing data utilization with privacy and security.

[0005] The present invention provides a big data-based mental health assessment system, comprising:

[0006] Data acquisition module, the data acquisition module obtains the user's heterogeneous data sources in real time through a distributed interface and generates structured data; the distributed interface of the data acquisition module includes a text acquisition interface for connecting to a third-party social media platform, obtaining a real-time updated text data stream through an open authentication protocol authorized by the user, and performing standardized conversion of data structures of different platforms; a signal synchronization interface for accessing wearable devices, receiving physiological signal data pushed by device manufacturers, including heart rate fluctuations, skin conductance changes and sleep quality indicators, and the interface has a built-in protocol adapter to be compatible with communication specifications of different manufacturers; an archive retrieval interface for connecting to medical institutions, periodically querying diagnostic records and medication information in electronic health records through a database connection protocol; the data output by the text acquisition interface, signal synchronization interface and archive retrieval interface are uniformly packaged into a structured data format and transmitted to the data processing module.

[0007] Data processing module: The data processing module receives structured data, desensitizes sensitive information in the text data through a noise filtering algorithm, and generates a multimodal feature vector;

[0008] Dynamic modeling module, the dynamic modeling module receives multimodal feature vectors and constructs a user psychological state map through a graph neural network, in which the nodes represent three psychological dimensions of emotion, cognition, and behavior, and the edge weights represent the dynamic correlation strength between the dimensions. The topological structure of the map is updated based on the time sliding window to generate a psychological state evolution signal; the graph neural network construction method 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, exercise frequency and work and rest 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 of historical correlation strength; the span of the time sliding window is dynamically adjusted according to the user data update frequency, and when the window slides, the old node connection relationship that exceeds the time range is removed, and an updated psychological state evolution signal is generated based on the latest multimodal feature vector.

[0009] The risk assessment module receives the psychological state evolution signal, calculates the probability value of psychological crisis within the future preset period through the time series prediction model, and generates an early warning signal;

[0010] The intervention generation module receives the early warning signal, 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.

[0011] In one embodiment of the present invention, the noise filtering algorithm of the data processing module includes: desensitizing personal identity information in text data using regular expression matching and mask replacement technology, and identifying negative emotional keywords in the text through sentiment dictionary and semantic analysis model and generating emotion intensity labels; using a sliding 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 user screen usage time, application switching frequency and geographic location change sequence from behavioral log data, aligning them with emotion labels and filtered physiological signals according to the time axis, and fusing them into a multimodal feature vector, which contains timestamp, data source type and normalized numerical features.

[0012] In one embodiment of the present invention, the processing process of the time series prediction model includes: decomposing the psychological state evolution signal into trend terms, periodic terms and residual terms, respectively using long-short-term memory networks to capture long-term dependencies, autoregressive models to fit periodic fluctuations and Gaussian process regression to estimate uncertainties; weighted fusion of the three prediction results to obtain the probability value of psychological crisis within a preset period in the future, where the weight coefficient is dynamically adjusted according to the prediction error of the user's historical data; when the probability value exceeds the threshold for a preset number of times continuously, the warning signal upgrade mechanism is triggered, and the original medium-risk warning signal is automatically upgraded to a high-risk warning signal.

[0013] In one 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 intensity in the intervention resource library, and the reward function based on the degree of improvement in the user's psychological state and compliance score after the intervention; iteratively optimizing the action selection strategy through the policy gradient algorithm so that high-risk warning signals give priority to triggering the matching of manual psychological counseling services, and medium-risk prompt signals give priority to triggering the push of self-service cognitive training courses; after each intervention instruction is issued, the click rate, completion rate and subsequent psychological state change data fed back by the user terminal are collected to update the reinforcement learning model parameters.

[0014] In one embodiment of the present invention, it also includes a privacy protection module. Before the data acquisition module transmits structured data to the data processing module, it adds Laplace noise that meets the requirements of differential privacy to the geographic location information and device identifier in the text data, so that the real user identity cannot be reversely inferred from a single piece of data; after the data processing module generates a multimodal 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 decryption; 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 erase instruction to all modules storing the user data.

[0015] In one embodiment of the present invention, the interaction method of the user terminal includes displaying an interactive virtual assistant interface after receiving a personalized intervention instruction, which provides graphic-guided cognitive behavioral training tasks, a psychological counselor video connection entrance, and a medication reminder setting panel; after the user completes the task, the terminal collects the real-time change data of the user's physiological signals and subjective emotion scores, and transmits them back to the data processing module through the data acquisition module; when the user refuses to execute the intervention instruction for a preset number of consecutive times, the terminal automatically triggers the emergency contact notification function and sends a behavioral abnormality mark to the risk assessment module to recalculate the psychological crisis probability value.

[0016] In one embodiment of the present invention, the operating architecture of the big data-based mental health assessment system includes a data acquisition module and a data processing module deployed on edge computing nodes for real-time processing of frequently updated text and physiological signal data; a dynamic modeling module and a risk assessment module deployed in a cloud computing center for executing large-scale graph construction and predictive model operations; an intervention generation module adopts a distributed microservice architecture to dynamically select the nearest service node to issue intervention instructions based on the user's geographic location; asynchronous communication is achieved between modules through 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.

[0017] In one embodiment of the present invention, the dynamic modeling module also includes a real-time feedback calibration mechanism. When the user terminal returns negative feedback for 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 behavioral dimension nodes in the psychological state map according to the feedback calibration signal, and recalculates the correlation strength between the emotion and cognitive dimensions based on the attenuated edge weights; at the same time, the recalculated correlation strength is compared and analyzed with the historical map data. If the correlation strength deviation exceeds the preset floating range, the map reconstruction instruction is triggered to initialize the new graph neural network model; during the execution of the map reconstruction instruction, the user's historical multimodal feature vectors within three months are retained as training data, and the transfer learning technology is used to inherit the topological features of the original map to accelerate model convergence.

[0018] The big data-based mental health assessment system provided by the present invention constructs a dynamic mental state map by fusing multi-source heterogeneous data to quantify the correlation between emotions, cognition, and behavior; combines time series prediction and reinforcement learning to generate graded 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

[0019] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0020] Figure 1 This is the system architecture diagram of the mental health assessment system based on big data. DETAILED DESCRIPTION

[0021] The following describes the embodiments of the present invention through specific examples. 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. The 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 the following embodiments and features in the embodiments can be combined with each other unless they conflict.

[0022] It should be noted that the illustrations provided in the following embodiments are merely schematic illustrations of the basic concept of the present invention. Therefore, the illustrations only show components related to the present invention and are not drawn according to the number, shape, and size of components in actual implementation. In actual implementation, the type, quantity, and proportion of each component may be changed arbitrarily, and the component layout may also be more complex.

[0023] In the following description, numerous details are discussed to provide a more thorough explanation of the embodiments of the present invention. However, it will be apparent to those skilled in the art that the embodiments of the present invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring the embodiments of the present invention.

[0024] See Figure 1 , shown is the big data-based mental health assessment system of the present invention. The big data-based mental health assessment system 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 the user's heterogeneous data source in real time through a distributed interface and generates structured data; the data processing module receives the structured data, desensitizes the sensitive information in the text data through a noise filtering algorithm, and generates a multimodal feature vector; the dynamic modeling module receives the multimodal feature vector, constructs the user's psychological state map through a graph neural network, in which the nodes represent the three psychological dimensions of emotion, cognition, and behavior, and the edge weights represent the dynamic correlation strength between the dimensions, and generates a psychological state evolution signal by updating the topological structure of the map based on the time sliding window; the risk assessment module receives the psychological state evolution signal, calculates the probability value of the psychological crisis in the future preset period through a time series prediction model, and forms an early warning signal; the intervention generation module receives the early warning signal, 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.

[0025] like Figure 1As shown in the figure, dynamic monitoring and personalized intervention of mental health status are achieved through the collaborative work of multiple modules. The system first uses the data acquisition module to acquire users' heterogeneous data sources in real time, including social media text, physiological signals from wearable devices, mobile terminal behavioral logs, and electronic health record data. The data acquisition module utilizes a distributed interface design, enabling it to connect to data sources across different platforms and devices. For example, it connects to social media platforms through open interface protocols to extract user-posted text content in real time. It communicates with wearable devices such as smart bracelets and heart rate monitors via Bluetooth or wireless network protocols to obtain physiological indicators such as heart rate, step count, and sleep quality. It uses mobile terminal sensors or application programming interfaces to collect behavioral data such as screen time, app switching frequency, and geographic movement trajectory. Furthermore, it connects to medical institutions' data systems to regularly access users' electronic health records, including structured data such as historical diagnosis records, medication regimens, and physical examination reports. This heterogeneous data undergoes preliminary cleaning and format standardization in the acquisition module. For example, data from different time zones is uniformly converted to Coordinated Universal Time (UTC), and unstructured text data is converted into structured data packets containing timestamps, user identifiers, and content bodies. All data is then encapsulated into a unified transmission format and sent to the data processing module. After receiving structured data from the acquisition module, the data processing module performs multi-level noise filtering and feature extraction. For text data, regular expression matching techniques are first used to identify personal identifying information (PII) such as mobile phone numbers, ID numbers, and addresses. Masking or hashing are then used to desensitize the data. Simultaneously, a pre-trained sentiment analysis model is used to classify the text content as positive, negative, or neutral, and the intensity of the sentiment is quantified using a sentiment dictionary. For physiological signal data, a sliding average filter algorithm is used to eliminate noise caused by device measurement errors or environmental interference. For example, a five-second moving average is applied to heart rate signals to smooth out abnormal fluctuations. Missing data segments caused by device offline or transmission interruptions are filled using linear interpolation or methods based on trend prediction of nearby time points. For behavioral log data, feature engineering is used to extract key metrics, such as total daily screen time, frequently used application categories, and time spent at a specific location. These metrics are then sliced ​​and aligned at a minute-level granularity. Finally, the processing module aligns the text emotion labels, filtered physiological signal slices and behavioral feature indicators along the time axis, and fuses them to generate multimodal feature vectors containing timestamps, data source types and multidimensional feature values. These vectors are organized through a specific data structure and transmitted to the dynamic modeling module.

[0026] Furthermore, after receiving the multimodal feature vectors, the dynamic modeling module uses a graph neural network to construct a user psychological state map. The nodes in the map are organized into three main dimensions: the emotional dimension is further subdivided into positive, negative, and neutral emotion sub-nodes; the cognitive dimension includes sub-nodes for attention level, memory ability, and decision-making tendency; and the behavioral dimension includes sub-nodes for social activity, exercise frequency, and daily routine. The weight of each edge is dynamically calculated using an attention mechanism. For example, when calculating the correlation strength between a negative emotion node and a social activity node, the frequency and temporal proximity of the two nodes in historical data are considered, and an exponential decay function is introduced to reduce the influence of older historical data. The graph's topology is dynamically updated using a sliding window mechanism. The window span is adaptively adjusted based on the frequency of data updates: when user data updates frequently, the window is shortened to capture short-term changes; conversely, the window is lengthened to capture long-term trends. Whenever new data arrives, old node connections that fall outside the current window time range are removed, and the edge weights between each sub-node are recalculated based on the latest feature vector. The updated graph generates a psychological state evolution signal, which contains the temporal trends of the correlation strengths of each dimension and key inflection points. This signal is then transmitted to the risk assessment module. Upon receiving the psychological state evolution signal, the risk assessment module uses a time series prediction model to analyze the probability of a future psychological crisis. The model first decomposes the input signal into a trend term, a cycle term, and a residual term. The trend term uses a long-short-term memory network to capture the long-term evolution of psychological states. For example, a continuously rising correlation weight for negative emotions may indicate worsening depressive tendencies. The cycle term uses an autoregressive model to identify circadian or weekly fluctuations, such as peaks in anxiety on Sunday nights. The residual term uses Gaussian process regression to quantify the uncertainty of the model's predictions. The three prediction results are dynamically weighted based on historical prediction accuracy. For example, a more stable cycle term in recent predictions is given a higher weight. Finally, the weighted fusion results are combined to determine the probability of a psychological crisis within the next three days to one week. When the probability exceeds a preset first threshold, a high-risk warning signal is generated, triggering immediate manual intervention. When it falls between the first and second thresholds, a medium-risk warning signal is generated, initiating a self-help intervention process. The early warning signal includes the risk level, the main triggering dimensions (such as abnormal emotion-behavior correlation), and the confidence assessment, which is transmitted to the intervention generation module through an encrypted channel. After receiving the early 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 a variety of intervention types such as access to psychological counseling services, cognitive behavioral training courses, mindfulness meditation guidance, and medication compliance reminders, each with different implementation intensity and time schedules. The state space of the reinforcement learning model is composed of the current risk level, the user's historical intervention response records, and the environmental context (such as the current time period and geographic location). The action space is defined as all optional intervention plans and their combinations. The reward function is dynamically calculated based on the improvement of the user's physiological indicators, task completion rate, and subjective feedback score after the intervention.For example, for high-risk warnings, the model prioritizes matching real-time video psychological counseling and simultaneously pushes relaxation training; for medium-risk prompts, it sends a combination of cognitive behavioral exercise tasks and medication reminders. The generated intervention instructions are pushed through the user terminal in the form of text, voice, or video, and the user's interactive behavior is monitored in real time: if the user does not respond within the set time, the intervention intensity is automatically upgraded or the type of plan is switched. At the same time, the user feedback data collected by the terminal (such as the progress of training task completion and the heart rate change curve) is transmitted back to the dynamic modeling module to update the edge weight calculation of the psychological state map, forming a closed-loop optimization mechanism of evaluation-intervention-feedback.

[0027] In one embodiment of the present invention, the specific interface structure and data processing logic of the data acquisition module are further defined. This module includes three core interface components: a text acquisition interface, a signal synchronization interface, and a file retrieval interface. The text acquisition interface connects to mainstream social media platforms through an open authentication protocol and uses an asynchronous request mechanism to batch retrieve user-authorized posts, comments, and private messages. The interface includes a built-in flow control module that dynamically adjusts the request frequency based on platform interface restrictions to prevent high-frequency access from triggering flow control mechanisms. A data structure conversion layer is also designed to uniformly convert JSON or XML formatted data returned by different platforms into a standard structure containing user identifiers, posting times, text content, and emotional tags. The signal synchronization interface is responsible for connecting to multiple brands of wearable devices and receiving physiological signal data through the cloud push service provided by the device manufacturers. The interface includes a built-in protocol adapter that can parse binary or custom encoding formats from different manufacturers. For example, it can map the sleep stage code of a certain brand of wristband to standardized labels for light sleep, deep sleep, and REM sleep, and convert accelerometer data into step count values. The archive retrieval interface connects to the medical institution database via an encrypted tunnel and employs an incremental query strategy to regularly retrieve the user's latest electronic health record: the last acquisition timestamp is recorded for each query, and subsequent requests are made only for new data after that point in time, reducing network transmission load. The data output by all interface components undergoes format verification and outlier filtering before being packaged by the data encapsulation engine into a structured data packet containing a metadata header and a payload. The metadata includes the data source identifier, collection time, and a data integrity check code, while the payload arranges the data values ​​of each dimension in a pre-set field order, ensuring efficient parsing by subsequent processing modules.

[0028] like Figure 1As shown, the desensitization process for text data consists of two stages: first, regular expressions are used to match predefined sensitive information patterns. For example, the regular pattern for mobile phone numbers is "1[3-9]\d{9}," and the first six digits of the area code and the last digit of the checksum are matched for ID card numbers. Matched content is either replaced with asterisks for the middle four digits or irreversibly encrypted using the SHA-256 hash algorithm. Second, a sentiment analysis model is used to identify potentially sensitive content. A text classification model based on the BERT architecture assesses the emotional polarity of input sentences. When high-risk content such as suicidal tendencies or descriptions of self-harm are detected, an emergency flag is added to the sentiment label and a real-time warning bypass channel is triggered. Physiological signal processing utilizes a sliding average filter combined with an outlier rejection strategy. For example, using the heart rate signal, a window size of five sampling points is set. The median and standard deviation of the values ​​within the window are calculated. Samples that deviate from the median by more than three standard deviations are considered noise and removed. The gap is then filled with the mean of the remaining data within the window. For extended periods of signal loss (e.g., a device offline for more than ten minutes), an autoregressive model-based prediction method is used to generate replacement data, and the source of the supplementation is indicated in the data flags. Behavioral log processing focuses on spatiotemporal feature extraction: The number of daily stops, categories of frequently visited areas, and entropy values ​​of movement trajectories are calculated from geolocation data to quantify behavioral regularity; app switching frequency, single-session duration distribution, and the proportion of nighttime usage duration are extracted from screen usage data. All features are aligned according to minute-level time windows. For data sources with different sampling frequencies (e.g., heart rate sampled per second and location recorded per minute), the nearest neighbor interpolation method is used to elevate low-frequency data to the high-frequency time axis. The resulting multimodal feature vector adopts a hierarchical storage structure: the first layer encodes the timestamp and data source type; the second layer contains the text's sentiment intensity, physiological indicator statistics, and behavioral feature vectors; and the third layer adds data quality flags (e.g., desensitization tags and supplementation tags) to ensure that downstream modules can accurately understand the data semantics and reliability.

[0029] like Figure 1As shown in the figure, it emphasizes a fine-grained classification of psychological state dimensions and a mechanism for calculating dynamic association strength. This module decomposes the emotion dimension into three sub-nodes: positive emotion, negative emotion, and neutral emotion. The positive emotion sub-node is further divided into sub-categories such as joy, satisfaction, and excitement. The negative emotion sub-node includes sub-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 attention level, memory ability, and decision-making tendency sub-nodes. The attention level sub-node quantifies the duration of user focus and frequency of distraction when completing tasks. The memory ability sub-node is based on short-term memory test results and the frequency of daily forgetfulness. The decision-making tendency sub-node is determined by analyzing the duration of user hesitation when faced with choices and risk preferences. The behavioral dimension is further divided into social activity, exercise frequency, and daily routine sub-nodes. The social activity sub-node counts the number and duration of users' online and offline social interactions. The exercise frequency sub-node records daily steps, exercise type, and duration. The daily routine sub-node analyzes the stability and degree of deviation in the sleep-wake cycle. Edge weights between sub-nodes are calculated using a multi-head attention mechanism, with each attention head focusing on correlation patterns at different timescales. For example, a short-term attention head analyzes the impact of emotional fluctuations within the past hour on decision-making tendencies, while a long-term attention head examines the correlation between behavioral patterns and cognitive abilities over the past week. A time decay factor is applied to historical correlation data in the form of an exponential function, so that interactions further in time contribute less to edge weights, ensuring that the graph reflects the latest changes in psychological states. The span of the sliding time window dynamically adjusts based on the frequency of data input: when the user data update interval is less than a preset threshold, the window span is automatically shortened to half its original value to capture rapid changes; when the data update interval exceeds the threshold, the window span is gradually expanded to double to incorporate long-term trends. As the window slides, the system identifies historical node connections that fall outside the current timeframe, such as the association between a high-intensity exercise session three days ago and negative emotions on the same day, and removes these outdated connections to free up computing resources. After removal, the connections between the affected sub-nodes are reinitialized based on the latest received multimodal feature vectors. For example, based on the newly detected intensity of anxiety and decreased social activity, edge weights between the corresponding sub-nodes are created or adjusted. When the updated mental state map generates evolutionary signals, topological data analysis methods are used to extract the persistent homological features of the map, identify strong correlation patterns that persist within the time window (such as the stable association between negative emotions and irregular work and rest) and short-lived weak correlation patterns (such as a brief decrease in attention caused by sudden stressful events), and encode these patterns into time series signals for transmission to the risk assessment module.

[0030] The model first performs empirical mode decomposition on the input psychological state evolution signal, separating it into a trend term representing long-term trends, a cycle term reflecting cyclical fluctuations, and a residual term containing random noise and unexpected events. Trend term analysis utilizes a long-short-term memory network architecture. The network input layer receives slices of the evolution signal from the past thirty days. A forget gate and an input gate control the retention and updating of historical information. The output layer predicts the trend direction for the next seven days, such as the correlation weight of rising anxiety or the gradual recovery of sleep regularity indicators. Cycle term processing utilizes an autoregressive model with external variables, incorporating known cyclical factors (such as weekday and weekend divisions and seasonal labels) as covariates. The model automatically identifies inherent cycles in the user's psychological state (e.g., biweekly mood swings) and tests correlations with external cyclical factors to eliminate accidental cyclical patterns. Residual term analysis utilizes Gaussian process regression, using a kernel function to define the covariance relationship between different time points. This model then models the probability distribution of random fluctuations unexplained by the model and outputs confidence intervals for the probability of future psychological crises. The three prediction results are fused using dynamic weighting coefficients. The weighting coefficients are updated online based on the forecast errors over the past fifteen days. Error calculations use a weighted metric based on root mean square error and directional consistency. For example, when the trend term consistently demonstrates high accuracy in recent forecasts, its weight is increased to 60% of the total weight. When the confidence interval width of the residual term exceeds a threshold, its weight is reduced to avoid over-reliance on predictions with high uncertainty. The fused psychological crisis probability value is updated every six hours. When it exceeds the high-risk threshold three times in a row, the warning signal upgrade mechanism is triggered, automatically upgrading the original medium-risk warning to a high-risk warning. The user is also marked as a priority for monitoring, and high-frequency data collection is initiated (for example, reducing the physiological signal sampling interval from five minutes to thirty seconds). After the warning signal is generated, the system also generates a causal analysis report, using counterfactual reasoning to identify the association patterns that most contribute to the probability value (for example, if the edge weight of negative emotions and social activity suddenly becomes the primary causal factor, the report will highlight the urgency of social intervention). The report is then encrypted and transmitted to the intervention generation module for decision-making.

[0031] like Figure 1Figure 2 illustrates the construction and optimization process of the reinforcement learning strategy in the intervention generation module. This strategy defines the user's current psychological state as the core dimension of the state space, including the real-time probability of psychological crisis, intervention response records from the past seven days (e.g., click-through rate, task completion, and improvement in physiological indicators), and environmental context (e.g., whether it is nighttime or whether the user is at home). The action space encompasses all feasible options from the predefined intervention resource library and their combinations. For example, single actions might include sending a cognitive training task, scheduling an appointment with a counselor, or pushing a mindfulness breathing instructional video. Combined actions might include simultaneously sending a training task and a medication reminder. The reward function is designed using a multi-objective optimization framework, including short-term rewards (e.g., improvement in the user's heart rate variability within two hours of the intervention instruction), medium-term rewards (e.g., reduction in screen time over three consecutive days), and long-term rewards (e.g., decrease in the predicted probability of psychological crisis over the next week). The weights of each objective are dynamically adjusted based on the user's historical intervention effect preferences. During the training process of the policy gradient algorithm, the model maintains an exploration-exploitation balance: for new users or in data-sparse scenarios, random exploration is increased to try diverse intervention combinations; for users with sufficient data, historically high-reward actions are prioritized for exploitation. For high-risk warning signals, the policy sets intervention priority rules. For example, when the probability of a psychological crisis exceeds a high-risk threshold and the user is alone, a human counselor is forced to intervene in real time, and a notification is simultaneously sent to a preset emergency contact. For medium-risk warning signals, a progressive intervention strategy is adopted, initially delivering mild intervention content (such as a five-minute breathing exercise). If the user does not respond, the intervention is gradually upgraded to moderate intervention (such as a 20-minute cognitive training course). After each intervention instruction is issued, the system collects multi-dimensional feedback data from the user terminal: explicit feedback includes the user's rating of the intervention content, task completion progress, and subjective emotional self-assessment; implicit feedback is analyzed by analyzing changes in physiological signals (such as a decrease in skin conductance), behavioral changes (such as increased time spent on social apps), and the degree of improvement in the emotional tendency of the text after the intervention. This feedback data is processed and converted into a reinforcement learning reward signal, which is used to update the policy network's parameters online. For example, if a cognitive training session results in a reduction in negative emotion labels and a task completion rate exceeding 90%, the corresponding action's preference weight is significantly increased. Furthermore, the policy model undergoes offline batch retraining monthly, leveraging historical interaction data from all users to globally optimize network parameters and prevent local optimality traps caused by online learning.

[0032] Furthermore, during the data collection phase, the module adds differentially private Laplace noise to the geolocation information and unique device identifiers in the text data. The noise level is adaptively adjusted based on the sensitivity of the field. For example, given the high sensitivity of geolocation latitude and longitude, a larger noise level is used to blur the frequently visited locations of a single user into a hundred-meter region. After noise processing, the device identifier cannot be mapped one-to-one to the actual device, but information such as device type (such as smartphone model) is retained for aggregate analysis. During the data processing phase, multimodal feature vectors are encrypted using a fully homomorphic encryption algorithm immediately after generation. This allows the dynamic modeling module to perform graph construction operations directly in the ciphertext state, such as calculating correlation weights between encrypted emotion intensity values ​​and physiological indicators, without decrypting the original data. Encryption keys are managed independently by the user, and the system can only access plaintext data for a specific time period (such as when emergency access is needed for crisis intervention). Data storage utilizes a distributed fragmentation scheme, dividing a single user's complete data record into multiple fragments, stored in different geographic nodes. Each fragment is individually encrypted and cannot be independently restored. When a user initiates a data deletion request, the privacy protection module initiates a blockchain verification process: the user digitally signs the deletion instruction with their private key. After the smart contract verifies the signature and user permissions, it sends an erase instruction containing the data hash value to all storage nodes. Upon receiving the instruction, the storage node retrieves the fragmented data matching the hash value from its local database, performs multiple irreversible overwrite and erase operations, and writes a log of the operation to the blockchain for audit tracking. Furthermore, the module implements a privacy leakage risk assessment subsystem that regularly simulates scenarios such as re-identification attacks and associative inference attacks to calculate the residual risk under the current data protection mechanism. If the risk exceeds a security threshold, it automatically triggers data re-anonymization or restricts access to high-risk data. During data transmission, all inter-module communications utilize a quantum key distribution protocol to establish a secure channel, ensuring data confidentiality even in the face of future quantum computing attacks. The privacy protection module also provides a transparent data usage dashboard, allowing users to view in real time the module's access history, processing purposes, and remaining storage period. It also supports one-click authorization revocation and data export, ensuring compliance with major global data protection regulations such as the EU General Data Protection Regulation.

[0033] The present invention proposes a big data-based mental health assessment system that constructs a dynamic mental state map by fusing multi-source heterogeneous data to quantify the correlation between emotions, cognition, and behavior; combines time series prediction with reinforcement learning to generate graded 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.

[0034] Therefore, the big data-based mental health assessment system of the present invention can solve the problems of traditional mental health assessment relying on subjective scales and single-dimensional data, poor real-time performance, weak privacy protection, and disconnected intervention.

[0035] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the present invention. Anyone skilled in the art may modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by one of ordinary skill in the art without departing from the spirit and technical principles disclosed herein are intended to be covered by the claims of the present invention.

Claims

1. A mental health assessment system based on big data, characterized in that: include: A data acquisition module, which acquires the user's heterogeneous data sources in real time through a distributed interface and generates structured data; The distributed interface of the data acquisition module includes a text acquisition interface for connecting to a third-party social media platform, obtaining a real-time updated text data stream through an open authentication protocol authorized by the user, and performing standardized conversion of data structures across different platforms; A signal synchronization interface for accessing wearable devices to receive physiological signal data pushed by device manufacturers, including heart rate fluctuations, skin conductance changes, and sleep quality indicators. The interface also has a built-in protocol adapter to be compatible with communication specifications from different manufacturers. A file retrieval interface for connecting to medical institutions periodically queries diagnostic records and medication information in electronic health records through a database connection protocol. The data output by the text acquisition interface, signal synchronization interface, and file retrieval interface is uniformly packaged into a structured data format and transmitted to the data processing module. A data processing module receives the structured data, desensitizes sensitive information in the text data using a noise filtering algorithm, and generates a multimodal feature vector; A dynamic modeling module receives the multimodal feature vector and constructs a user psychological state map through a graph neural network, wherein nodes represent three psychological dimensions: emotion, cognition, and behavior, and edge weights represent the dynamic correlation strength between the dimensions. The module also generates a psychological state evolution signal by updating the topology of the map based on a time sliding window; The graph neural network construction method 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, exercise frequency and work and rest 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 of historical correlation strength; the span of the time sliding window is dynamically adjusted according to the user data update frequency, and when the window slides, old node connection relationships that are out of the time range are removed, and an updated psychological state evolution signal is generated based on the latest multimodal feature vector; a risk assessment module, which receives the psychological state evolution signal, calculates the probability value of psychological crisis within a future preset period through a time series prediction model, and generates an early warning signal; An intervention generation module receives the early warning signal, 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.

2. The big data-based mental health assessment system according to claim 1, characterized in that: The noise filtering algorithm of the data processing module includes: desensitizing personal identity information in text data using regular expression matching and mask replacement technology, and identifying negative emotional keywords in the text through a sentiment dictionary and semantic analysis model to generate emotional intensity labels; using a sliding 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 user screen usage time, application switching frequency and geographic location change sequences from behavioral log data, aligning them with emotional labels and filtered physiological signals along the time axis, and fusing them into a multimodal feature vector, wherein the feature vector includes a timestamp, data source type and normalized numerical features.

3. The big data-based mental health assessment system according to claim 1, characterized in that: The processing process of the time series prediction model includes: decomposing the psychological state evolution signal into trend term, cycle term and residual term, using long-short-term memory network to capture long-term dependency, autoregressive model to fit periodic fluctuations and Gaussian process regression to estimate uncertainty; weighted fusion of the three prediction results to obtain the probability value of psychological crisis within a preset period in the future, where the weight coefficient is dynamically adjusted according to the prediction error of the user's historical data; when the probability value exceeds the threshold for a preset number of times in a row, the warning signal upgrade mechanism is triggered, and the original medium-risk warning signal is automatically upgraded to a high-risk warning signal.

4. The big data-based mental health assessment system 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 solution types and implementation intensities in the intervention resource library, and the reward function based on the degree of improvement in the user's psychological state and compliance score after the intervention; iteratively optimizing the action selection strategy through the policy gradient algorithm so that high-risk warning signals prioritize triggering matching with manual psychological counseling services, and medium-risk warning signals prioritize triggering the push of self-service cognitive training courses; after each intervention instruction is issued, collecting the click rate, completion rate and subsequent psychological state change data fed back by the user terminal to update the reinforcement learning model parameters.

5. The big data-based mental health assessment system according to claim 4, characterized in that: It also includes a privacy protection module. Before the data acquisition module transmits structured data to the data processing module, it adds Laplace noise that meets the requirements of differential privacy to the geographic location information and device identifier in the text data, so that the real user identity cannot be reversely inferred from a single piece of data; after the data processing module generates a multimodal 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 decryption; 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 erase instruction to all modules storing the user data.

6. The big data-based mental health assessment system according to claim 1, characterized in that: The user terminal's interactive mode includes displaying an interactive virtual assistant interface after receiving personalized intervention instructions, the interface providing graphic-guided cognitive behavioral training tasks, a psychological counselor video connection portal, and a medication reminder setting panel; After the user completes the task, the terminal collects the real-time change data of the physiological signal and the subjective emotion score, and transmits it back to the data processing module through the data collection module; When the user refuses to execute the intervention instruction for a preset number of consecutive times, the terminal automatically triggers the emergency contact notification function and sends a behavioral abnormality mark to the risk assessment module to recalculate the psychological crisis probability value.

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

8. The big data-based mental health assessment system 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 for 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 behavioral dimension nodes in the psychological state map according to the feedback calibration signal, and recalculates the correlation strength between the emotion and cognitive dimensions based on the attenuated edge weights; at the same time, the recalculated correlation strength is compared and analyzed with the historical map data. If the correlation strength deviation exceeds the preset floating range, the map reconstruction instruction is triggered to initialize the new graph neural network model; during the execution of the map reconstruction instruction, the user's historical multimodal feature vectors within three months are retained as training data, and the transfer learning technology is used to inherit the topological features of the original map to accelerate model convergence.

Images