Psychological state dynamic evaluation and early warning system based on multi-modal behavior data

Abstract

The invention discloses a psychological state dynamic evaluation and early warning system based on multi-modal behavior data, which belongs to the field of medical care informatics and comprises a multi-modal behavior data hierarchical coding module, a time sequence causal atlas construction and reasoning module, a double-stage self-adaptive early warning decision module and a context awareness intervention strategy generation module. A cross-modal association mode is extracted through a double-layer coding mechanism, a time sequence graph containing a causal relationship is constructed, causal reasoning is performed, a double-stage mechanism of short-term mutation detection and long-term trend prediction is adopted to generate graded early warning, and an optimal intervention strategy is selected based on a deep Q network according to a user situation. According to the method, the accuracy, timeliness and intervention effectiveness of psychological health assessment are improved, and dynamic monitoring and early warning of the psychological state are realized.

Description

Technical Field

[0001] This invention relates to the fields of mental health monitoring and artificial intelligence technology, specifically to a dynamic assessment and early warning system for mental state based on multimodal behavioral data, which falls under the category of healthcare informatics technology. Background Technology

[0002] Mental health assessment is a crucial step in the prevention and intervention of mental illnesses. Traditional assessment methods mainly rely on psychological scales such as the SCL-90 and PHQ-9, or clinical interviews, which have limitations such as strong subjectivity, long update cycles, and difficulty in reflecting the dynamic fluctuations of users' mental states in real time. With the widespread use of smart devices, mental health monitoring based on multimodal behavioral data has gradually become a research hotspot. By collecting multi-dimensional data such as users' voice, facial expressions, social behavior, physical activity, and sleep quality, it is possible to objectively quantify the changing trends of mental states.

[0003] The prior art CN120260938A discloses a big data-based mental health assessment system. This system acquires heterogeneous data sources through a data acquisition module, generates multimodal feature vectors through a data processing module, constructs a mental state map through a graph neural network through a dynamic modeling module, calculates the probability of mental crisis through a time series prediction model through a risk assessment module, and matches intervention plans through a reinforcement learning strategy through an intervention generation module. However, this existing technology has the following shortcomings: First, the data processing module only performs simple feature vector fusion, failing to capture deep-seated correlation patterns and hierarchical semantic information between different modalities, resulting in insufficient feature representation; second, the graph constructed by the dynamic modeling module only contains the correlation between psychological dimensions, failing to clearly distinguish between causal and correlational relationships, and thus failing to identify the root causes of changes in psychological state; third, the risk assessment module uses a single time series prediction model, failing to consider the multi-timescale characteristics of changes in psychological state, resulting in insufficient ability to capture sudden psychological crisis events; fourth, the intervention generation module matches intervention plans from a predefined resource library, failing to fully consider the user's current specific situation, thus limiting the adaptability and effectiveness of the intervention strategy; fifth, the lack of an effective closed-loop feedback mechanism between the various modules of the system makes it impossible to dynamically optimize the feature extraction and model inference process based on the warning results and intervention effects, and the system's adaptability and robustness need to be improved.

[0004] In recent years, several new research findings have emerged in the field of multimodal mental health assessment. Digital phenotype studies objectively quantify and monitor mental health status using real-time behavioral indicators collected from personal digital devices; however, most studies focus on single-modal data, and research on multimodal fusion methods is relatively insufficient. Temporal graph neural networks have demonstrated the ability to model spatiotemporal dynamic changes in brain connectivity research and disease prediction, but their application to causal inference of mental states remains a gap. Furthermore, existing research mostly employs static early warning models, which are difficult to adapt to the dynamic evolution of individual mental states and lack comprehensive consideration of short-term abrupt events and long-term evolutionary trends. Summary of the Invention

[0005] The purpose of this invention is to provide a dynamic assessment and early warning system for psychological states based on multimodal behavioral data, which solves the technical problems in the prior art such as insufficient feature representation, inaccurate identification of causal relationships, insufficient timeliness of early warning, poor adaptability of intervention strategies, and lack of closed-loop optimization.

[0006] This invention achieves deep semantic extraction of behavioral features by setting up a multimodal behavioral data hierarchical encoding module. It employs a dual-layer encoding mechanism to capture single-modal atomic features at the bottom layer and fuse cross-modal correlation patterns at the top layer, generating a more comprehensive feature representation. A temporal causal graph construction and inference module accurately models the mechanisms influencing psychological states. Based on Granger causality tests, a causal graph is established, and temporal causal convolution is used for inference to identify key behavioral patterns leading to changes in psychological states. A two-stage adaptive early warning decision-making module enables risk assessment across multiple time scales. In the short-term early warning stage, mutation detection captures abnormal events, while in the long-term early warning stage, a Bayesian dynamic linear model predicts evolutionary trends, generating tiered early warnings by integrating short-term and long-term risks. A context-aware intervention strategy generation module enables personalized and precise intervention, dynamically selecting the optimal intervention strategy based on the user's current context using a deep Q-network. A closed-loop collaborative mechanism ensures continuous system optimization. The feature importance score in the early warning decision-making process is fed back to the encoding module to adjust attention weights, and the intervention effect evaluation results are fed back to the graph module to update the causal relationship strength, forming a closed loop of feature encoding optimization and causal relationship reinforcement, improving the system's adaptability and early warning accuracy.

[0007] The beneficial effects of this invention include: First, the dual-layer encoding mechanism can simultaneously extract fine-grained features of a single modality and high-order association patterns across modalities. Compared with simple feature concatenation methods, the completeness and discriminativeness of feature representation are significantly improved, providing a more reliable data foundation for subsequent causal inference and early warning decision-making. Second, the temporal causal graph clearly distinguishes between causal relationships and correlation relationships. By using Granger causality tests and temporal causal convolution to identify the root causes of changes in psychological state, it avoids the false positive problem of traditional association analysis methods and improves the accuracy of psychological state assessment. Third, the dual-stage early warning decision-making module comprehensively considers sudden events and long-term evolution. The system's short-term early warning system can promptly capture abnormal fluctuations, while long-term early warning can predict deterioration risks in advance. The fusion assessment across multiple time scales significantly improves the comprehensiveness and timeliness of early warnings. Fourth, the context-aware intervention strategy generation fully considers the user's current time, geographical location, and social status, making intervention measures more relevant to actual scenarios, improving user acceptance and compliance, and enhancing the effectiveness of interventions. Fifth, the closed-loop collaborative mechanism optimizes feature encoding through early warning feedback and strengthens causal relationships through intervention feedback, enabling continuous learning and dynamic adaptation of the system. As data accumulates, the system performance continuously improves, exhibiting good scalability and robustness. Attached Figure Description

[0008] Figure 1 This is a schematic diagram of the overall architecture of the system of the present invention;

[0009] Figure 2 This is a schematic diagram of the structure of the hierarchical coding module for multimodal behavioral data;

[0010] Figure 3 This is a schematic diagram of the structure of a time-series causal graph;

[0011] Figure 4 This is a schematic diagram of a two-stage adaptive early warning decision-making process;

[0012] Figure 5 A schematic diagram of the process for generating context-aware intervention strategies. Detailed Implementation

[0013] Please refer to the attached document. Figures 1-5 To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only for explaining the present invention and are not intended to limit the present invention.

[0014] Reference Figure 1The present invention provides a dynamic psychological state assessment and early warning system based on multimodal behavioral data, comprising a multimodal behavioral data hierarchical encoding module 1, a temporal causal graph construction and reasoning module 2, a two-stage adaptive early warning decision-making module 3, and a context-aware intervention strategy generation module 4. The system continuously collects multimodal behavioral data from users via smart terminals, extracts deep-level features through hierarchical encoding, constructs a temporal causal graph for causal reasoning, employs a two-stage mechanism for risk early warning, generates personalized intervention strategies based on user context, and achieves continuous system optimization through a closed-loop feedback mechanism, thereby realizing dynamic assessment and early warning of users' psychological states.

[0015] Reference Figure 2 The multimodal behavioral data hierarchical coding module 1 is responsible for collecting and encoding users' multimodal behavioral data. This module collects users' voice signals through the microphone of the smart terminal, extracting prosodic and spectral features including fundamental frequency, formants, and Mel-frequency cepstral coefficients. These features reflect the user's emotional state and psychological stress level. It also collects facial image sequences from the smart terminal's front-facing camera and extracts micro-expression action unit combinations based on the facial motion coding system. For example, AU1 indicates raised inner eyebrows, commonly associated with sadness; AU6 indicates raised cheek muscles, commonly associated with happiness; and AU4 indicates furrowed brows, commonly associated with anxiety. These action unit combinations can identify subtle emotional changes in the user. By analyzing users' call logs, text message records, and social application usage, it statistically analyzes call frequency, message interaction frequency, and social circle activity. These social characteristics can reflect the user's social withdrawal or social overload. Finally, it collects users' physical activity data through the accelerometer sensor of a smartwatch or wristband, calculating daily steps, exercise duration, and exercise intensity. A sustained decline in physical activity levels is often an early sign of depressive tendencies. By monitoring users' sleep quality through heart rate and body motion sensors in wearable devices, the distribution of sleep stages, including the proportions of light sleep, deep sleep, and REM sleep, was analyzed. Sleep structure disorders are closely related to psychological problems such as anxiety and depression.

[0016] The multimodal behavior data hierarchical encoding module 1 employs a two-layer encoding mechanism to process the aforementioned multimodal data. The bottom-layer encoder includes a speech feature encoding unit, a facial feature encoding unit, a social feature encoding unit, an activity feature encoding unit, and a sleep feature encoding unit. Each unit is responsible for processing data from one modality and extracting atomic features. The speech feature encoding unit uses a one-dimensional convolutional neural network to extract features from the speech spectrogram. Through multi-layer convolution and pooling operations, it captures the local patterns and global statistical characteristics of the speech signal, outputting a fixed-dimensional speech atomic feature vector. The facial feature encoding unit uses a pre-trained convolutional neural network to extract deep features from facial images. It focuses on key facial regions such as eyebrows, eyes, and mouth through an attention mechanism, outputting a facial atomic feature vector containing micro-expression information. The social feature encoding unit normalizes call frequency and message interaction frequency, calculates the deviation from historical averages, and outputs a social atomic feature vector representing changes in social activity. The activity feature encoding unit standardizes step count and exercise intensity, introduces time window statistical features such as the average step count and variance over the past three days, and outputs an activity atomic feature vector reflecting body activity patterns. The sleep feature encoding unit analyzes the duration and transition frequency of sleep stages, calculates sleep efficiency and sleep quality scores, and outputs a sleep atom feature vector containing sleep structure information.

[0017] The top-level encoder receives all five atomic feature vectors and constructs a feature association matrix to represent the interdependencies between features of different modalities. Each element of this matrix represents the similarity between two atomic feature vectors, obtained by calculating the cosine similarity or Euclidean distance of the vectors. The top-level encoder uses a multi-head self-attention mechanism to process the feature association matrix. Multi-head attention can learn different types of associations between features from multiple representation subspaces. In a preferred embodiment, eight attention heads are used, each independently calculating attention weights and extracting cross-modal association patterns. Specifically, for the h-th attention head, an attention score matrix is ​​calculated through linear transformations of the query matrix, key matrix, and value matrix. This matrix represents the contribution of each atomic feature to the fusion result. Multiplying the attention score matrix by the value matrix yields a weighted feature representation. The outputs of the eight attention heads are concatenated and nonlinearly transformed through a fully connected layer to generate the final fused feature vector. This fused feature vector contains both atomic feature information from each modality and high-order cross-modal association patterns, comprehensively reflecting the user's behavioral state and psychological tendencies.

[0018] Importantly, the attention weights in the top-level encoder are dynamically adjusted by the feature importance scores from the two-stage adaptive early warning decision module 3. During the early warning decision-making process, the two-stage adaptive early warning decision module 3 calculates the contribution of each dimension of the fused feature vector to the early warning result, obtaining the feature importance score through gradient backpropagation. This score is fed back to the multimodal behavioral data hierarchical encoding module 1 to adjust the weight parameters in the multi-head self-attention mechanism. Specifically, for feature dimensions with high importance scores, their corresponding attention weights are increased, giving the feature a greater influence during the fusion process; for feature dimensions with low importance scores, their attention weights are decreased to reduce their interference with the fusion result. Through this feedback adjustment mechanism, the multimodal behavioral data hierarchical encoding module 1 can gradually learn the most valuable cross-modal association patterns for psychological state early warning, making the feature representation more focused on key information and improving the early warning accuracy of subsequent modules.

[0019] Reference Figure 3 The temporal causal graph construction and reasoning module 2 receives the fused feature vector and constructs a heterogeneous temporal graph containing psychological state nodes and behavioral feature nodes. In a preferred embodiment, psychological states are divided into three categories: emotional stability state, cognitive acuity state, and behavioral regularity state. The emotional stability state reflects the amplitude and frequency of the user's emotional fluctuations; poor emotional stability is manifested as frequent emotional fluctuations and strong negative emotions. The cognitive acuity state reflects the user's level of concentration and decision-making ability; decreased cognitive acuity is manifested as inattentiveness and slow decision-making. The behavioral regularity state reflects the consistency and predictability of the user's daily behavioral patterns; disruption of behavioral regularity is manifested as disordered sleep patterns and chaotic activity patterns.

[0020] The temporal causal graph also includes five categories of behavioral feature nodes, corresponding to voice features, facial features, social features, activity features, and sleep features in the fused feature vector. The value of each behavioral feature node is determined by the feature value of the corresponding dimension in the fused feature vector and is continuously updated as new data arrives. Edges in the graph represent causal relationships between nodes. Directed edges from behavioral feature nodes to mental state nodes represent the influence of behavior on psychology; for example, an edge from a sleep feature node to an emotional stability state node represents the causal influence of sleep quality on emotional stability. Directed edges between mental state nodes represent the interaction of different psychological dimensions; for example, an edge from an emotional stability state node to a cognitive acuity state node represents how emotional fluctuations affect cognitive function.

[0021] The temporal causal graph construction and inference module 2 employs the Granger causality test to determine whether a causal relationship exists between nodes. The basic idea of ​​the Granger causality test is that if historical information about variable X significantly improves the prediction of variable Y, then X is considered a Granger cause of Y. In this invention, for any two nodes, regression models are established with and without the historical information of the source node. The F-test is used to determine whether the addition of historical information from the source node significantly reduces the prediction error. If the F-statistic exceeds a critical value, a Granger causal relationship is considered to exist between the source node and the target node, and corresponding directed edges are established in the graph. The edge weights are initialized to causal strength values, which are obtained by normalizing the F-statistic and reflect the strength of the causal relationship.

[0022] To capture the dynamic evolution of causal relationships, the temporal causal graph construction and inference module 2 introduces a time decay function to attenuate the edge weights at historical moments. In a preferred embodiment, an exponential decay function is used, with a time decay coefficient set to 0.95, meaning that the edge weight decays to 0.95 times its original value at each time step. This design makes recent causal relationships have a greater impact on the current state, while the impact of long-term causal relationships gradually weakens, consistent with the time-dependent characteristics of psychological state changes. The temporal causal graph construction and inference module 2 uses a sliding time window mechanism to periodically update the graph's topology. The sliding window length is set to 7 days. Whenever new day's data arrives, the oldest day's data is removed and the Granger causality test is re-executed. Edges whose causal strength decays below a threshold are removed from the graph; for newly emerging causal relationships, corresponding edges are added to the graph. Through this dynamic update mechanism, the temporal causal graph can adapt to changes in user psychological states and behavioral patterns, maintaining the graph's timeliness and accuracy.

[0023] The temporal causal graph construction and inference module 2 uses temporal causal convolution to infer the graph and identify key behavioral patterns that lead to changes in psychological state. Temporal convolutional layers are used to extract local temporal patterns of nodes over historical time series. The design of the convolutional kernels ensures that the current node state depends only on information from historical moments. In a preferred embodiment, dilated causal convolution is used, with dilation rates set to 1, 2, 4, and 8. Through exponentially increasing dilation rates, the temporal convolutional layers can cover a longer historical time span with fewer layers. In this embodiment, four layers of dilated causal convolution can cover historical information for 15 time steps.

[0024] For each mental state node in the temporal causal graph, the node's historical time series values ​​are first input into a temporal convolutional layer to extract temporal evolution features. Then, for each directed edge pointing to the mental state node, the edge weight is weighted and aggregated with the temporal features of the source node. Specifically, assuming a mental state node has N incoming edges, and the source node is... The edge weight is The temporal characteristics of the source node are A graph attention mechanism is used to dynamically adjust the aggregation weights of different source nodes, and the importance score of the source nodes to the target node is calculated through an attention function.

[0025] ,

[0026] in, Let be the attention weight of the i-th source node. This represents the current features of the target mental state node. This indicates a feature concatenation operation. This is a learnable attention parameter vector. Combining the attention weights with the edge weights yields the final aggregation coefficients.

[0027] ,

[0028] in, Let be the aggregation coefficient of the i-th source node. Then, perform a weighted summation of the temporal features of all source nodes:

[0029] ,

[0030] in, This is the aggregated neighbor information. A gating mechanism is used to fuse the node's own temporal evolution information with causal propagation information from neighboring nodes.

[0031] ,

[0032] ,

[0033] in, The temporal features are obtained by performing temporal convolution on the node itself. The gating coefficient, It is the Sigmoid activation function. , This is the weight matrix. For bias vectors, This represents element-wise multiplication. This represents the updated mental state node. The gating mechanism allows the model to adaptively balance the evolutionary trend of the node itself and the causal influence of neighboring nodes. When the gating coefficient is close to 0, it relies more on the node's own information, and when the gating coefficient is close to 1, it relies more on the causal propagation of neighbors.

[0034] A round of causal reasoning is completed by traversing all mental state nodes. The representations of each node are then input into the next temporal convolutional layer and graph aggregation layer. This process is repeated multiple times until the node representations converge. In a preferred embodiment, four rounds of reasoning are set up. Through multiple rounds of information propagation, the representation of each mental state node not only includes its own historical evolution information but also integrates the causal influences from behavioral feature nodes and other mental state nodes, forming a globally consistent mental state assessment. The temporal causal graph construction and reasoning module 2 outputs causal reasoning results containing the current values ​​and predicted trends of each mental state. These results serve as the input to the two-stage adaptive early warning decision module 3.

[0035] Reference Figure 4 The dual-stage adaptive early warning decision module 3 receives causal inference results and performs risk assessments in both the short-term and long-term early warning stages. The short-term early warning stage focuses on abrupt changes in user behavior characteristics, enabling timely detection of sudden psychological crisis signals. The dual-stage adaptive early warning decision module 3 arranges the fused feature vectors in chronological order to form a time-series feature sequence and uses a statistically based mutation detection algorithm to identify abrupt changes in feature values. In a preferred embodiment, a cumulative sum control chart method is used, which calculates the cumulative deviation between the feature sequence and reference values, identifying abrupt changes by detecting significant deviations from the cumulative sum curve.

[0036] For feature sequences The reference value is set to the historical average. Calculate the cumulative sum sequence:

[0037] ,

[0038] in, Let be the cumulative sum at time t. Normally, the cumulative sum sequence fluctuates around 0. When the eigenvalues ​​deviate persistently, the cumulative sum sequence will show a clear upward or downward trend. Set an upper control limit. and lower control limit In a preferred embodiment, the upper control limit is set to The lower control limit is set to ,in is the standard deviation of the characteristic sequence. When or When a mutation event occurs, time t is marked as the mutation point.

[0039] For each detected mutation point, the two-stage adaptive early warning decision module 3 extracts the feature difference vector before and after the mutation. Specifically, it calculates the difference between the feature mean of the three time steps after the mutation point and the feature mean of the three time steps before the mutation point to obtain the feature difference vector. The feature difference vector is matched with various risk patterns in a predefined risk pattern library, which includes common psychological crisis signal patterns, such as a sharp decline in sleep quality, a sudden decrease in social activity, and abnormal emotional expression. Each risk pattern is represented by a feature change template. By calculating the cosine similarity between the feature difference vector and the template, the risk pattern with the highest similarity is selected as the current anomaly type. Based on the severity of the anomaly type, corresponding short-term risk index values ​​are predefined; for example, the risk index value for a sharp decline in sleep quality is 0.8, for a sudden decrease in social activity is 0.7, and for abnormal emotional expression is 0.9.

[0040] When the short-term risk indicator value exceeds the first threshold for a preset number of consecutive times, the dual-stage adaptive early warning decision module 3 generates a short-term early warning indicator and triggers an immediate intervention process. In a preferred embodiment, the first threshold is set to 0.7, and the preset number of times is set to 2, meaning that a short-term early warning is triggered when the short-term risk indicator value exceeds 0.7 twice consecutively. The short-term early warning mechanism can respond quickly after abnormal changes in user behavior patterns, providing timely intervention support and preventing further deterioration of psychological problems.

[0041] The long-term early warning phase focuses on the evolutionary trend of users' psychological states, enabling early prediction of mental health risks within future time windows. The dual-stage adaptive early warning decision module 3 employs a Bayesian dynamic linear model to model the time-series data of psychological states. This model can handle non-stationary time series, capturing long-term trends and periodic fluctuations in psychological states. The Bayesian dynamic linear model decomposes the time-series data of psychological states into horizontal, trend, and seasonal components. The horizontal component describes the baseline level of the psychological state, the trend component describes the growth or decline trend of the psychological state, and the seasonal component describes the periodic fluctuations of the psychological state.

[0042] In a preferred embodiment, the Bayesian dynamic linear model employs the following state-space representation. The observation equation is:

[0043] ,

[0044] in, Let be the observed value of the psychological state at time t. For horizontal components, For seasonal portions, To observe the noise, it follows a mean of 0 and a variance of 0. The normal distribution follows. The state evolution equation for the horizontal component is:

[0045] ,

[0046] in, As a trend component, The evolution noise of the horizontal component follows a mean of 0 and a variance of . The normal distribution follows. The state evolution equation for the trend component is:

[0047] ,

[0048] in, The evolutionary noise of the trend component follows a pattern with a mean of 0 and a variance of . The normal distribution is followed by the seasonal component. The state evolution equation of the seasonal component adopts a seasonal random walk model:

[0049] ,

[0050] in, The length of the seasonal cycle is set to 7 in this invention to represent a 7-day cycle, such as the difference in psychological state between weekdays and weekends. The seasonal component is the evolutionary noise, which follows a mean of 0 and a variance of . It follows a normal distribution.

[0051] The dual-stage adaptive early warning decision module 3 recursively updates the posterior distribution of model parameters using the Kalman filter algorithm. The Kalman filter consists of a prediction step and an update step. In the prediction step, the prior distribution of the current state is predicted based on the state evolution equation; in the update step, the posterior distribution of the state is updated using Bayesian rules, combined with the observations at the current time. Specifically, a state vector is defined... State transition matrix Describes the state evolution relationship, observation matrix This indicates that the observed value is obtained by adding the horizontal and seasonal components. The prediction step calculates the prior mean. and prior covariance ,in and The posterior mean and covariance are the values ​​from the previous time step. Let be the state evolution noise covariance matrix. Calculate the Kalman gain in the update step. posterior mean and posterior covariance .

[0052] Through Kalman filtering recursion, the two-stage adaptive early warning decision module 3 obtains the posterior distribution of the state at each time step. For the prediction of future time windows, a multi-step forward prediction is performed using the state evolution equation to predict the 1st... state mean at time 1 Covariance In a preferred embodiment, the prediction time window is set to the next 7 days, and the horizontal component, trend component and seasonal component of each day are predicted. The prediction results are then recombined to obtain the predicted trajectory and confidence interval of the psychological state.

[0053] Based on the predicted trajectory, the two-stage adaptive early warning decision module 3 calculates the probability of psychological state deterioration. Psychological state deterioration is defined as a predicted value falling below a preset health threshold. In a preferred embodiment, the psychological state score range is set to 0 to 100, and the health threshold is set to 60. Based on the posterior distribution of the prediction, the cumulative probability of a predicted value below 60 is calculated; this cumulative probability is the probability of psychological state deterioration. When the deterioration probability exceeds a second threshold, a long-term early warning indicator is generated. In a preferred embodiment, the second threshold is set to 0.3, meaning that a long-term early warning is triggered when the probability of psychological state deterioration exceeds 30% within the next 7 days.

[0054] The dual-stage adaptive early warning decision module 3 generates a comprehensive risk score by weighted summation of short-term and long-term risk indicators. Short-term risk indicators reflect the degree of abnormality at the current moment, while long-term risk indicators reflect the risk of deterioration within a future time window. Both characterize the user's mental health status from different time scales. The weighting coefficients are adjusted according to the actual application scenario. In the preferred embodiment, the weight of the short-term risk indicator is set to 0.6, and the weight of the long-term risk indicator is set to 0.4, resulting in the comprehensive risk score. ,in As a short-term risk indicator, This serves as a long-term risk indicator. Based on the comprehensive risk score, it is divided into low-risk, medium-risk, and high-risk levels. Low-risk corresponds to a score below 0.3, medium-risk corresponds to a score between 0.3 and 0.7, and high-risk corresponds to a score above 0.7. The dual-stage adaptive early warning decision module 3 outputs a graded early warning signal containing the risk level, which serves as the input to the context-aware intervention strategy generation module 4.

[0055] Furthermore, during the generation of tiered early warning signals, the dual-stage adaptive early warning decision module 3 calculates the contribution of each dimension of the fused feature vector to the early warning decision. Specifically, it calculates the gradient of the comprehensive risk score with respect to each dimension of the fused feature vector through gradient backpropagation. The absolute value of the gradient reflects the degree of influence of that dimension of feature on the early warning result; the larger the absolute value of the gradient, the more important the feature. The absolute value of the gradient is normalized to obtain a feature importance score, which is fed back to the multimodal behavioral data hierarchical encoding module 1 to adjust the attention weights in the top-level encoder, thereby achieving continuous optimization of the feature encoding process.

[0056] Reference Figure 5The context-aware intervention strategy generation module 4 receives tiered early warning signals and generates personalized intervention strategies based on the user's current context. The core idea of ​​context awareness is that the effectiveness of an intervention strategy depends not only on the user's mental health but also on the context in which they are situated. For example, sending a long relaxation training message during work hours may be inappropriate, while sending a short message reminder during rest periods may be ineffective. The context-aware intervention strategy generation module 4 obtains the user's time context, geographical location context, and social status context, and selects the intervention strategy with the highest match to the current context, improving the feasibility of the intervention and user compliance.

[0057] The context-aware intervention strategy generation module 4 obtains the current time through the user terminal's clock module and determines whether the current time belongs to a work period or a rest period. In a preferred embodiment, the work period is defined as 9:00 to 18:00 from Monday to Friday, and the rest period is defined as other times. The current geographic coordinates are obtained through the user terminal's positioning module, and the system determines whether the user is in a home environment or a work environment based on the geographic coordinates. In a preferred embodiment, the user's home address and work address are pre-set using geofencing technology. When the user's location is within 500m of their home address, it is determined to be in a home environment; when it is within 500m of their work address, it is determined to be in a work environment. By analyzing the user's real-time communication records and application usage, the system determines whether the user is currently in a social interaction state. In a preferred embodiment, if the user has call records or social application activity records within the last 10 minutes, it is determined to be in a social interaction state; otherwise, it is determined to be in a solitary state.

[0058] The context-aware intervention strategy generation module 4 constructs a context feature vector containing time context identifiers, geographic location context identifiers, and social status context identifiers, represented using one-hot encoding. For example, the time context is encoded as [1, 0] for work hours and [0, 1] for rest hours; the geographic location context is encoded as [1, 0] for home environment and [0, 1] for work environment; and the social status is encoded as [1, 0] for social interaction and [0, 1] for being alone. The context feature vector is concatenated with the risk level in the graded early warning signal. The risk level is also encoded using one-hot encoding: low risk is encoded as [1, 0, 0], medium risk as [0, 1, 0], and high risk as [0, 0, 1]. The concatenated vector forms a state representation containing both risk and context information, which serves as the input to the deep Q-network.

[0059] The context-aware intervention strategy generation module 4 employs a deep Q-network for intervention strategy selection. A deep Q-network is a reinforcement learning-based decision-making method that learns a state-action value function to select actions that maximize long-term cumulative rewards. In this invention, the state space is defined as a representation of states containing risk and contextual information, and the action space is defined as all available strategies in the candidate intervention strategy set. Candidate intervention strategies include psychological counseling appointment strategies, cognitive training recommendation strategies, mindfulness practice guidance strategies, exercise suggestion strategies, and social activity suggestion strategies. The psychological counseling appointment strategy provides users with online or offline appointment services for psychological counselors, suitable for high-risk users in situations where they are at rest or at home; the cognitive training push strategy pushes cognitive behavioral exercise tasks that include attention training, memory training, or emotion regulation training, suitable for medium-risk users in situations where they are alone; the mindfulness practice guidance strategy provides users with audio or video guidance on mindfulness meditation or breathing exercises, suitable for users of all risk levels in quiet environments; the exercise suggestion strategy pushes appropriate exercise plans and exercise check-in reminders, suitable for users whose activity characteristics remain low; and the social activity suggestion strategy recommends users to participate in offline social activities or contact friends, suitable for users whose social characteristics have suddenly decreased.

[0060] The Deep Q-Network structure includes an input layer, hidden layers, and an output layer. The input layer receives the state representation. The hidden layer contains three fully connected layers, each with 128 neurons, using the ReLU activation function. The output layer contains five neurons corresponding to five candidate intervention strategies, outputting the Q-value of each action. The Q-value represents the expected long-term cumulative reward of performing the action in the current state; a higher Q-value indicates a better expected effect. The context-aware intervention strategy generation module 4 selects the action with the highest Q-value as the optimal intervention strategy, generating personalized intervention instructions that include specific intervention content and suggested execution times. For example, for high-risk users in rest periods and at home, the Deep Q-Network might choose the psychological counseling appointment strategy, generating intervention instructions that include a link to an online psychological counseling platform and a suggested consultation time slot.

[0061] The context-aware intervention strategy generation module 4 sends personalized intervention instructions to the user terminal, presenting them to the user via push notifications or in-app messages. The user terminal provides a user-friendly interface, allowing users to view intervention content, confirm their willingness to implement it, and provide feedback on the implementation status. The context-aware intervention strategy generation module 4 collects user feedback on the implementation of intervention instructions, including whether the user clicked on the intervention content, whether the intervention task was completed, and the user's subjective feeling rating after the intervention. Simultaneously, the context-aware intervention strategy generation module 4 continuously monitors the user's psychological state changes after the intervention, obtaining the user's latest psychological state assessment results through the multimodal behavioral data hierarchical coding module 1 and the temporal causal graph construction and inference module 2.

[0062] The context-aware intervention strategy generation module 4 calculates the intervention effect based on intervention feedback and psychological state change data. The immediate reward is determined based on the user's execution feedback; if the user clicks and completes the intervention task, the immediate reward is positive; if the user ignores or refuses the intervention instruction, the immediate reward is negative. The delayed reward is determined based on the user's degree of psychological state improvement within a subsequent time window. In a preferred embodiment, the delayed reward calculation window is set to 3 days after the intervention, comparing the average psychological state score of the 3 days after the intervention with the average psychological state score of the 3 days before the intervention; the greater the score improvement, the higher the delayed reward. The total reward is obtained by weighted summing of the immediate and delayed rewards, with the weight of the immediate reward set to 0.3 and the weight of the delayed reward set to 0.7. ,in For instant rewards, This is a delayed reward.

[0063] The context-aware intervention strategy generation module 4 updates the parameters of the deep Q-network using the total reward. An experience replay mechanism is used to store historical interaction data; the experience replay pool stores a quadruple of state, action, reward, and next state for each intervention. During parameter updates, a batch of samples is randomly sampled from the experience replay pool, and the temporal difference error (TD error) is calculated. The TD error represents the difference between the current Q-value estimate and the target Q-value.

[0064] ,

[0065] in, This is the current state. For the selected action, The next state after the action is performed. This is the discount factor, which is set to 0.95 in the preferred embodiment. For the current network parameters, For the target network parameters, For the current network, estimate the Q-value of the state-action pair. Let Q be the maximum Q-value of the target network for all actions in the next state. Update the network parameters using gradient descent by minimizing the squared TD error.

[0066] ,

[0067] in, The learning rate is set to 0.001 in the preferred embodiment. Indicates the parameter The gradient. To stabilize the training process, a target network mechanism is adopted, with target network parameters... The target network is copied from the current network at regular intervals, and in a preferred embodiment, it is updated every 100 steps. Through experience replay and the target network, the deep Q-network can gradually learn the optimal intervention strategy under different states and situations, and the intervention effect continuously improves as interactive data accumulates.

[0068] This invention establishes a closed-loop collaborative mechanism for feature encoding optimization and causal relationship reinforcement, enabling continuous learning and dynamic adaptation of the system. The closed-loop collaborative mechanism includes two feedback paths: the first path feeds the feature importance score of the early warning decision-making process to the multimodal behavioral data hierarchical encoding module 1; the second path feeds the intervention effect evaluation results to the time-series causal graph construction and inference module 2.

[0069] In the first feedback path, the dual-stage adaptive early warning decision module 3 calculates the contribution of each dimension of the fused feature vector to the comprehensive risk score through gradient backpropagation during the generation of hierarchical early warning signals, thus obtaining a feature importance score. This score reflects the degree of influence of different behavioral features on psychological state early warning; features with high importance scores play a key role in early warning decisions, while features with low importance scores have a smaller impact. The multimodal behavioral data hierarchical encoding module 1 receives the feature importance scores and adjusts the attention weight allocation of the multi-head self-attention mechanism in the top-level encoder. Specifically, for feature dimensions with high importance scores, their corresponding attention weights are increased, giving the feature a greater weight in the fusion process; for feature dimensions with low importance scores, their attention weights are decreased, reducing their impact on the fusion result. Through iterative optimization, the multimodal behavioral data hierarchical encoding module 1 gradually learns the most valuable cross-modal association patterns for psychological state early warning, making the feature representation more focused on key information and improving the accuracy and timeliness of early warning decisions.

[0070] In the second feedback path, the context-aware intervention strategy generation module 4, after collecting user feedback on the execution of intervention instructions, assesses the magnitude of changes in psychological state nodes before and after the intervention. Specifically, it compares psychological state scores for the three days before and after the intervention, calculating the improvement magnitude of each psychological state node. It identifies psychological state dimensions that show significant improvement during the intervention; for example, if the emotional stability node significantly improves after the intervention, the intervention is considered to have a significant effect on emotional stability. The temporal causal graph construction and reasoning module 2 adjusts the edge weights in the causal graph based on the intervention effect evaluation results. For causal edges pointing to psychological state nodes that show significant improvement, their edge weights are increased to reflect the strengthening effect of the causal relationship under intervention regulation; for causal edges pointing to psychological state nodes that do not show improvement, their edge weights are decreased to reflect the decline in the influence of the causal relationship. Furthermore, it updates the time evolution parameters of the edge weights and adjusts the decay rate of the time decay function to make the evolution of causal relationships more consistent with actual observation data. Through the accumulation of multiple intervention feedbacks, the temporal causal graph construction and reasoning module 2 continuously corrects the structure and parameters of the causal graph, making the graph more accurately reflect the actual psychological state influence mechanism, thereby supporting more accurate causal reasoning and early warning decisions.

[0071] The closed-loop collaborative mechanism design enables the system of this invention to continuously learn and self-optimize. As the system's runtime increases and user data accumulates, the feature representation capability of the multimodal behavioral data hierarchical encoding module 1 continuously enhances, the causal reasoning accuracy of the temporal causal graph construction and reasoning module 2 continuously improves, the warning sensitivity and specificity of the two-stage adaptive early warning decision module 3 are continuously optimized, and the intervention effect of the context-aware intervention strategy generation module 4 is continuously improved. The entire system forms a positive feedback loop, with system performance continuously improving over time, ultimately achieving the goal of personalized, precise, and intelligent dynamic assessment and early warning of mental health.

[0072] In a specific application example, the user is a 30-year-old professional. The system continuously collects their multimodal behavioral data via smartphone and smartwatch. During one week, the multimodal behavioral data hierarchical encoding module 1 detected that the user's voice tone became lower and slower, facial micro-expressions (AU4) showed frequent furrowed brows, social activity frequency decreased from an average of 5 calls per day to 1 call per day, physical activity (steps) decreased from an average of 8000 steps per day to 3000 steps per day, and sleep quality score decreased from 80 to 50. The bottom encoder extracts atomic features from each modality, while the top encoder, through a self-attention mechanism, identifies a high correlation between voice features, facial features, social features, and activity features. The generated fused feature vector shows a significant change in the user's overall behavioral pattern.

[0073] The temporal causal graph construction and inference module 2 receives the fused feature vector and updates the state of each node in the graph. Through Granger causality tests, it was found that sleep feature nodes have a strong causal relationship with emotional stability state nodes, activity feature nodes have a strong causal relationship with behavioral regularity state nodes, and emotional stability state nodes have a strong causal relationship with cognitive acuity state nodes. Temporal causal convolution is used to infer the graph, identifying that decreased sleep quality is a key factor leading to emotional instability, which further affects cognitive function and behavioral regularity. Causal inference results show that the user's emotional stability state score decreased to 45 points, cognitive acuity state score decreased to 50 points, and behavioral regularity state score decreased to 40 points.

[0074] The dual-stage adaptive early warning decision module 3 receives causal inference results. In the short-term early warning stage, it detects abrupt changes in sleep quality and social activity. The cumulative sum of the mutation detection algorithm exceeds the upper control limit, matching a pattern of sharp decline in sleep quality and a pattern of sudden reduction in social activity. The short-term risk index is calculated to be 0.85, exceeding the first threshold of 0.7 twice consecutively, generating a short-term early warning indicator. In the long-term early warning stage, the Bayesian dynamic linear model predicts that the user's emotional stability score will continue to decline to 35 points in the next 7 days, with a probability of psychological deterioration of 0.65, exceeding the second threshold of 0.3, generating a long-term early warning indicator. The comprehensive risk score is calculated as follows: It is determined to be at a high risk level, and a high risk warning signal is issued.

[0075] The context-aware intervention strategy generation module 4 receives a high-risk warning signal and identifies the user's current context as Saturday evening at 8:00 PM, in a home environment, and alone. It constructs a context feature vector and concatenates it with the risk level, then inputs it into a deep Q-network. The deep Q-network outputs the Q-values ​​of each strategy: the Q-value for the psychological counseling appointment strategy is 0.92, the Q-value for the cognitive training push strategy is 0.65, the Q-value for the mindfulness practice guidance strategy is 0.78, the Q-value for the exercise suggestion strategy is 0.55, and the Q-value for the social activity suggestion strategy is 0.60. The psychological counseling appointment strategy with the highest Q-value is selected, generating a personalized intervention instruction containing a link to an online psychological counseling platform and a suggestion to consult at 10:00 AM on Sunday. The intervention instruction is sent to the user's mobile phone via push notification; the user clicks the link to complete the appointment and receives psychological counseling on Sunday.

[0076] Within three days of the intervention, the user's sleep quality score gradually recovered to 65 points, social activity frequency returned to three calls per day, and emotional stability score improved to 60 points. The context-aware intervention strategy generation module 4 calculated an immediate reward of 0.8, a delayed reward of 0.9, and a total reward of [missing information]. The total reward was used to update the parameters of the deep Q-network, further improving the Q-value of the psychological counseling appointment strategy in similar situations. Simultaneously, the intervention effect evaluation results were fed back to the temporal causal graph construction and inference module 2, enhancing the causal edge weights pointing to nodes of emotional stability and optimizing the causal graph structure. Through this closed-loop feedback, the system can more accurately identify risk signals and select effective strategies in subsequent early warning and intervention.

[0077] The present invention provides a dynamic assessment and early warning system for mental state based on multimodal behavioral data. This system achieves full representation of behavioral characteristics through a two-layer coding mechanism, accurate modeling of the psychological state influence mechanism through a time-series causal graph, risk assessment at multiple time scales through a two-stage early warning system, personalized matching of intervention strategies through context awareness, and continuous optimization of the system through closed-loop collaboration. This significantly improves the accuracy, timeliness, and effectiveness of mental health assessment and early warning, providing strong technical support for the early detection and timely intervention of mental illnesses.

[0078] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can 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 those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. A dynamic assessment and early warning system for psychological states based on multimodal behavioral data, characterized in that, include: The multimodal behavioral data hierarchical coding module collects users' voice tone features, facial micro-expression features, social activity features, physical activity features, and sleep quality features. It generates hierarchical behavioral feature representations through a two-layer coding mechanism. The bottom encoder extracts atomic features of a single modality, while the top encoder fuses cross-modal association patterns based on a self-attention mechanism and outputs a fused feature vector. The temporal causal graph construction and reasoning module receives the fused feature vector, constructs a heterogeneous temporal graph containing psychological state nodes and behavioral feature nodes, wherein the edges between nodes represent causal relationships and the edge weights evolve dynamically over time, performs causal reasoning on the graph through temporal causal convolution, identifies key behavioral patterns that lead to changes in psychological state, and generates causal reasoning results. The dual-stage adaptive early warning decision module receives the causal reasoning results, identifies mutation events of user behavior characteristics based on the mutation detection algorithm and calculates short-term risk indicators in the short-term early warning stage, and uses a Bayesian dynamic linear model to predict the evolution trend of psychological state within the future time window and calculates long-term risk indicators in the long-term early warning stage. It generates a graded early warning signal based on the weighted fusion result of short-term risk indicators and long-term risk indicators. The context-aware intervention strategy generation module receives the graded early warning signal, obtains the user's current time context, geographical location context, and social status context, selects the intervention strategy with the highest matching degree with the current context from the candidate intervention strategy set based on a deep Q network, generates a personalized intervention instruction containing intervention content and execution timing and sends it to the user terminal, and collects user's execution feedback and psychological state change data at the same time. Specifically, the dual-stage adaptive early warning decision module feeds back the feature importance score in the early warning decision process to the multimodal behavior data hierarchical encoding module, adjusts the attention weight of cross-modal association patterns in the top-level encoder, and the context-aware intervention strategy generation module feeds back the intervention effect evaluation result to the temporal causal graph construction and reasoning module, updates the evolution parameters of the edge weights in the causal graph, and forms a closed-loop collaborative mechanism for feature encoding optimization and causal relationship reinforcement.

2. The system according to claim 1, characterized in that, The two-layer coding mechanism of the multimodal behavioral data hierarchical coding module includes: The underlying encoder includes a speech feature encoding unit, a facial feature encoding unit, a social feature encoding unit, an activity feature encoding unit, and a sleep feature encoding unit. The speech feature encoding unit extracts prosodic and spectral features of the user's speech to generate speech atomic feature vectors. The facial feature encoding unit extracts micro-expression action units through a facial motion encoding system to generate facial atomic feature vectors. The social feature encoding unit calculates call frequency and message interaction frequency to generate social atomic feature vectors. The activity feature encoding unit calculates step count and exercise intensity based on accelerometer data to generate activity atomic feature vectors. The sleep feature encoding unit analyzes the distribution of sleep stages to generate sleep atomic feature vectors. The top-level encoder receives all atomic feature vectors, constructs a feature correlation matrix to represent the interdependencies between different modal features, calculates the contribution weight of each atomic feature to the fusion result through a multi-head self-attention mechanism, performs a nonlinear transformation on the weighted atomic feature vectors and concatenates them to generate a fused feature vector, wherein the contribution weight is dynamically adjusted by the feature importance score from the two-stage adaptive early warning decision module.

3. The system according to claim 1, characterized in that, The graph construction process of the time-series causal graph construction and inference module includes: The psychological state is divided into three categories of psychological state nodes: emotional stability, cognitive acuity, and behavioral regularity. The voice features, facial features, social features, activity features, and sleep features in the fused feature vector are corresponding to five categories of behavioral feature nodes. Directed edges from behavioral feature nodes to psychological state nodes represent the influence of behavior on psychology, and directed edges between psychological state nodes represent the interaction between different psychological dimensions. Granger causality test is used to determine whether there is a causal relationship between nodes. For node pairs with causal relationship, directed edges are established, and the edge weights are initialized to the causal strength value. A time decay function is introduced to decay the edge weights at historical moments. A sliding time window mechanism is used to periodically update the topology of the graph, remove edges whose causal relationship has weakened to below the threshold, and add newly emerging causal relationships.

4. The system according to claim 3, characterized in that, The temporal causal convolution includes: For each mental state node in the heterogeneous temporal graph, a temporal convolutional layer is used to extract the local temporal pattern of the node in the historical time series. A causal convolutional kernel is used to ensure that the node state at the current moment depends only on the information of the historical moment and not on the future moment, so as to avoid information leakage. For each directed edge pointing to the mental state node, the edge weight is weighted and aggregated with the temporal features of the source node. The aggregation weights of different source nodes are dynamically adjusted using a graph attention mechanism. The node's own temporal evolution information and the causal propagation information from neighboring nodes are fused through a gating mechanism to generate an updated representation of the mental state node. Iterate through all mental state nodes to complete one round of causal reasoning. Repeat multiple rounds of reasoning until the node representations converge. Output the causal reasoning results, which include the current value of each mental state and the predicted trend of change.

5. The system according to claim 1, characterized in that, The short-term early warning phase of the two-stage adaptive early warning decision module includes: The fused feature vectors are arranged in chronological order to form a time-series feature sequence. A mutation detection algorithm based on statistical tests is used to identify mutation points in the feature values. The mutation detection algorithm calculates the cumulative sum of the feature sequence and detects significant deviations from the cumulative sum curve. When a mutation point is detected, it is marked as an abnormal event. For each abnormal event, extract the feature difference vector before and after the abnormal event, calculate the similarity between the feature difference vector and each risk pattern in the predefined risk pattern library, select the risk pattern with the highest similarity as the current abnormal type, and assign the corresponding short-term risk index value according to the severity of the abnormal type. When the short-term risk indicator value exceeds the first threshold for a preset number of consecutive times, a short-term warning sign is generated and an immediate intervention process is triggered.

6. The system according to claim 5, characterized in that, The long-term early warning phase includes: The time-series data of psychological states in the causal inference results are decomposed into horizontal components, trend components, and seasonal components. A Bayesian dynamic linear model is used to model the three components respectively. The evolution equation of the horizontal component describes the baseline level of the psychological state, the evolution equation of the trend component describes the growth or decline trend of the psychological state, and the evolution equation of the seasonal component describes the periodic fluctuation of the psychological state. The posterior distribution of the model parameters is updated recursively by the Kalman filter algorithm. The expected value and variance of the three components within the future time window are predicted. The prediction results are then recombined to obtain the predicted trajectory and confidence interval of the psychological state. The probability of psychological state deterioration is calculated based on the predicted trajectory. When the probability of deterioration exceeds the second threshold, a long-term warning sign is generated. Short-term risk indicators and long-term risk indicators are weighted and summed to generate a comprehensive risk score. Based on the comprehensive risk score, the risk is divided into low-risk, medium-risk, and high-risk levels, and a graded warning signal containing the risk level is output.

7. The system according to claim 1, characterized in that, The context acquisition of the context-aware intervention strategy generation module includes: The current time is obtained through the clock module of the user terminal to determine whether the current time belongs to the working period or the rest period to determine the time context. The current geographical coordinates are obtained through the positioning module of the user terminal to determine whether the user is in the home environment or the work environment to determine the geographical context. By analyzing the user's real-time communication records and application usage, it is determined whether the user is currently in a social interaction state to determine the social status context. The constructed context feature vector includes time context identifiers, geographic location context identifiers, and social status context identifiers. The context feature vector is then concatenated with the risk level in the graded early warning signal to form a state representation that includes both risk information and context information.

8. The system according to claim 7, characterized in that, The strategy selection based on deep Q-networks includes: The state space is defined as a state representation containing risk information and situational information, and the action space is defined as all selectable strategies in the candidate intervention strategy set. The candidate intervention strategies include psychological counseling appointment strategy, cognitive training push strategy, mindfulness practice guidance strategy, exercise suggestion strategy, and social activity suggestion strategy. The deep Q-network is used to estimate the long-term cumulative reward of each action in the current state, and the action with the largest Q value is selected as the optimal intervention strategy to generate personalized intervention instructions that include specific intervention content and suggested execution time. Data on changes in the user's psychological state after executing intervention instructions are collected as immediate rewards, and the degree of improvement in the user's psychological state within a subsequent time window is collected as delayed rewards. The immediate and delayed rewards are weighted and summed to obtain the total reward. The parameters of the deep Q network are updated using the total reward. The parameter update adopts an experience replay mechanism to store historical interaction data and randomly sample it for batch training.

9. The system according to claim 1, characterized in that, The feature encoding optimization of the closed-loop collaborative mechanism includes: The dual-stage adaptive early warning decision module calculates the contribution of each dimension of the features in the fused feature vector to the early warning decision during the generation of hierarchical early warning signals, and calculates the feature importance score through gradient backpropagation. The multimodal behavioral data hierarchical encoding module receives feature importance scores, adjusts the attention weight allocation of the multi-head self-attention mechanism in the top-level encoder, increases the weight of important features and decreases the weight of secondary features, so that the subsequent encoding process pays more attention to behavioral features that have a significant impact on psychological state warnings. Iterative optimization gradually focuses the feature encoding process on the most predictive cross-modal association patterns, improving the accuracy and timeliness of early warning decisions.

10. The system according to claim 1, characterized in that, The reinforcement of causality includes: The context-aware intervention strategy generation module, after collecting user feedback on the execution of intervention instructions, assesses the magnitude of changes in psychological state nodes before and after intervention and identifies psychological state dimensions that show significant improvement during the intervention process. The temporal causal graph construction and reasoning module enhances the causal edge weights pointing to nodes of psychological state that have significantly improved, weakens the causal edge weights pointing to nodes of psychological state that have not improved, and updates the temporal evolution parameters of the edge weights to reflect the moderating effect of the intervention on the causal relationship based on the intervention effect evaluation results. By accumulating multiple interventions and feedback to continuously revise the structure of the causal graph, the graph can more accurately reflect the actual psychological state's influence mechanism, thereby supporting more precise causal reasoning and early warning decisions.

Images