Psychological support strategy generation engine for multi-modal data fusion

Abstract

The invention discloses a psychological support strategy generation engine based on multi-modal data fusion, and relates to the technical field of psychological health, the engine comprises a multi-modal data acquisition module, a multi-modal data fusion processing module, a dynamic recovery trajectory modeling module, a recovery process prediction and strategy generation module and a strategy evaluation and feedback adjustment module; multi-modal data such as physiology, behaviors and voice are deeply combined with core theories such as cognitive behavior therapy and positive psychology through an innovative mechanism, data processing is performed according to theoretical constraint cleaning, and data compliance is ensured; fusing the matching degree of the mid-quantized data and the theoretical dimension, and associating objective data with theoretical indexes; dynamic modeling strengthens sensitive features such as a critical period of wound recovery, so that evaluation reflects physiological and behavior changes and embodies intrinsic features such as psychological toughness, a real psychological state can be accurately captured, a recovery risk can be recognized in advance, a scientific basis is provided for accurate intervention, and pertinence and foresight of psychological support are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of mental health technology, specifically to a psychological support strategy generation engine based on multimodal data fusion. Background Technology

[0002] In today's society, psychological trauma has become one of the most important issues affecting individual mental health. After experiencing psychological trauma, the individual's recovery process is complex and varied, involving multiple reactions at the physiological, psychological, and behavioral levels. Traditional mental health services often rely on single-modal data collection methods such as regular interviews and questionnaires. With the rapid development of IoT, big data, and artificial intelligence technologies, multimodal data fusion technology has brought new opportunities to the field of mental health. By integrating multi-dimensional data such as physiological indicators, behavioral performance, voice expression, and psychological assessment results, it is possible to more accurately assess an individual's psychological state and provide more scientific and personalized support for the recovery from psychological trauma.

[0003] Traditional methods for modeling psychological trauma recovery trajectories rely primarily on single-modality data collection and analysis, which has certain limitations: First, single-modality data cannot fully reflect the complexity of a user's psychological state. For example, self-report questionnaires alone may not accurately capture changes in a user's behavior or physiological responses in real life. Second, traditional methods are inadequate in dealing with dynamically changing psychological states, especially in predicting relapses or plateaus in the recovery process, lacking sufficient sensitivity and accuracy. Furthermore, adjustments to traditional psychological support strategies are often based on empirical judgment, failing to translate core theories such as cognitive behavioral therapy and positive psychology into calculable parameters, resulting in a disconnect between theoretical application and real-time data, making precise intervention difficult. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies by providing a multimodal data fusion-based psychological support strategy generation engine. This engine comprehensively collects multimodal data from users recovering from psychological trauma, including physiological indicators, behavioral performance, verbal expression, and psychological assessment results based on quantitative psychological theory dimensions. This enables all-round monitoring of the user's psychological state. Through a multimodal data fusion processing module, the raw data is cleaned and combined with dynamic modal fusion under theoretical constraints to obtain a fused feature vector. Then, an improved LSTM-GRU hybrid network is used to establish a dynamic recovery trajectory model. This model embeds key dimensions of psychological theory to accurately track the user's recovery process from psychological trauma. The model can capture the temporal dependence of the fused feature vector, strengthen the features of key time nodes, and generate a recovery status index based on the psychological resilience coefficient, trajectory sensitivity coefficient, and recovery threshold of fused psychological theory. This predicts possible relapses or stagnation points. Based on the prediction results, the system can adjust psychological support strategies in a timely manner, such as adjusting treatment plans and recommending psychological counseling resources suitable for different recovery stages, thereby providing users with more scientific and personalized psychological support services.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a psychological support strategy generation engine based on multimodal data fusion, the engine comprising: (1) Multimodal data acquisition module: acquires users' physiological index data, behavioral performance data, voice data and psychological assessment data as multimodal raw input data; (2) Multimodal data fusion processing module: Cleans the original input data, performs dynamic modal fusion, and obtains the fused feature vector; (3) Dynamic recovery trajectory modeling module: An improved LSTM-GRU hybrid network is used to capture the temporal dependency of the fused feature vector and strengthen the features of key time nodes. The psychological elasticity coefficient is calculated based on the fused feature vector, and the recovery state index is generated by combining the trajectory sensitivity coefficient and the recovery threshold. (4) Recovery process prediction and strategy generation module: calculate the first and second derivatives of the recovery state index to obtain the recovery acceleration, determine modal consistency, predict the risk probability based on the recovery acceleration and modal consistency, calculate the strategy fit and screen the strategy; (5) Strategy evaluation and feedback adjustment module: used to calculate feedback gain and fitness gradient, and update new strategy parameters based on feedback gain, fitness gradient and initial strategy parameters.

[0006] Furthermore, in the multimodal data acquisition module, physiological indicator data includes hormone levels, heart rate variability, skin conductance response, and electroencephalogram (EEG); behavioral performance data includes limb movement parameters, social interaction distance, social interaction duration, social interaction frequency, and daily behavior patterns; and psychological assessment data specifically includes the frequency of automatic thoughts and the degree of cognitive distortion in cognitive behavioral therapy, and the PERMA model (positive emotions, engagement, interpersonal relationships, meaning, achievement) score in positive psychology.

[0007] Furthermore, in the multimodal data fusion processing module, the cleaning of the original multimodal input data includes processing outliers and missing data. When processing outliers, an improved Z-score algorithm is used. When processing missing data, linear interpolation and LSTM prediction methods are used respectively according to the duration of the missing data. The missing data filling results must meet the value constraints of the corresponding psychological theory. Data with different sampling frequencies are normalized to a unified time slice through timestamp synchronization technology.

[0008] Furthermore, the dynamic modal fusion calculation formula in the multimodal data fusion processing module is as follows: ,in: for The fused feature vector at each time step; For the first The basic weighting coefficients of the modalities are preset constants; For the first Modal in Data reliability at any given time; For the first Modal in The dynamic adjustment factor at any given time; For the first The fit coefficient between modal data and psychological theories; For the first Modal The original input data at each moment; n is the modality number; n is the total number of modes; It is a time variable.

[0009] Furthermore, the reliability of the data The calculation formula is: ,in, For the first Modal in The variance of the sliding window at each time step; For the first The maximum variance threshold for modal data; Modal number; It is a time variable.

[0010] Furthermore, the dynamic adjustment factor The update rule is: when hour, ;when hour, ;when hour, .

[0011] Furthermore, in the improved LSTM-GRU hybrid network, LSTM units and GRU units are alternately distributed in the hidden layer at a ratio of 3:1. The attention gating mechanism strengthens recent features by calculating weight coefficients on the fused feature vectors of the first three time steps, while giving extra weight to data within the "theoretically sensitive time window" (such as the critical recovery period after trauma). The input layer is the fused feature vector, and there are three hidden layers.

[0012] Furthermore, the formula for calculating the psychological resilience coefficient in the dynamic recovery trajectory modeling module is as follows: ,in, For positive emotions, For the quality of interpersonal relationships, Score for meaning Weighting of positive psychology theories.

[0013] Furthermore, the risk probability calculation formula in the recovery process prediction and strategy generation module is as follows: ,in, This refers to the "frequency of negative automatic thoughts" in cognitive behavioral therapy. These are the theoretical factor weights; for The probability of risk at any given moment; This is a preset time interval; These are dynamic weighting coefficients; To recover the second derivative of the state exponent; for Modal consistency at any given moment; It is a time variable.

[0014] Furthermore, the formula for calculating the policy adaptability in the recovery process prediction and policy generation module is as follows: ,in, The degree of matching between the strategy and the user's current theoretical state; For theoretical matching weights; For the first This strategy in Time-based adaptability; This is the balance coefficient; for The recovery state index at time and the first The similarity of the target state indices of the two strategies; for Recovery status index at any given moment; For the first The target state index of the strategy; For the first Cost adaptation factor for this strategy; For the first Cultural adaptation factors for this strategy; The strategy number; It is a time variable.

[0015] Compared with existing technologies, this multimodal data fusion-based psychological support strategy generation engine has the following beneficial effects: I. This invention deeply integrates multimodal data such as physiological, behavioral, and speech data with core theories such as cognitive behavioral therapy and positive psychology through an innovative mechanism. In the data processing stage, the data is cleaned under constraints according to psychological theories to ensure that the data conforms to the reasonable range within the theoretical framework. In the fusion process, the degree of matching between data and theoretical dimensions is quantified to achieve an organic correlation between objective data and theoretical indicators. During dynamic modeling, the characteristics of theoretically sensitive stages such as the critical period of trauma recovery are emphasized, so that the assessment results can reflect both the objective changes in the user's physiological and behavioral aspects and the theoretically driven intrinsic characteristics such as psychological resilience, thereby capturing the user's true psychological state and identifying potential risk points in the recovery process in advance.

[0016] Second, this invention ensures that the intervention strategy conforms to the theoretical logic of cognitive behavioral therapy, positive psychology, etc. by accurately matching the intervention strategy with the user's current theoretical state; it incorporates theoretically relevant factors into the risk assessment to improve the response speed to high-risk users; in the strategy optimization stage, it quantifies the intervention effect based on theoretical effect indicators and dynamically adjusts the strategy parameters within the reasonable range of theoretical suggestions, forming a complete closed loop from data collection to strategy optimization. This mechanism ensures that the intervention measures are always highly adapted to the user's recovery stage, significantly improving the intervention effect and shortening the rehabilitation cycle.

[0017] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from an examination of the following, or may be learned from the practice of the invention. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0019] Figure 1 Generate a master flowchart for psychological support strategies; Figure 2 A detailed flowchart for multimodal data fusion processing; Figure 3 A flowchart for the generation and dynamic optimization of psychological support strategies. Detailed Implementation

[0020] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0021] Example 1: Generation of Psychological Support Strategies for Adolescents with Anxiety Disorders After the multimodal data acquisition module is activated, a comprehensive data acquisition network is constructed. Cortisol levels, heart rate variability (HRV), and skin conductance response (GSR) are collected every 5 minutes via a wrist-worn wearable device. Simultaneously, the power ratio of alpha waves (relaxation state) to beta waves (tension state) in the prefrontal cortex is recorded using an electroencephalogram (EEG) device. Using computer vision algorithms from a smart camera, limb movement parameters (such as the frequency of posture changes and gesture amplitude) are extracted every 30 seconds. Social interaction distance (average distance with family / classmates), daily social interaction duration (cumulative communication time), and frequency (number of times a conversation is initiated) are recorded using an indoor positioning system. Daily behavioral patterns (such as regularity of sleep patterns and fluctuations in study time) are statistically analyzed using smart calendar data. Speech fragments when patients mention anxiety-related events such as academic pressure and interpersonal conflicts are recorded using a smartphone microphone, and pitch frequency (average fundamental frequency), speech rate (syllables per second), and pauses are extracted. Features such as the number of pauses were collected; a standardized anxiety scale (such as GAD-7) was pushed daily using an online questionnaire platform, and data such as "frequency of automatic thoughts" (requiring patients to record the number of times negative thoughts occur in real time, such as the daily frequency of "I am sure I will not do well on the test"), scores from the positive psychology PERMA model (where the "positive emotion" dimension was obtained through three daily self-emotional assessments, the "engagement" dimension was calculated through the percentage of time spent on learning focus, the "interpersonal relationships" dimension was based on the social interaction quality score, the "meaning" dimension was assessed through the sense of identification with learning goals, and the "achievement" dimension was based on the statistics of the completion of small goals, with each dimension using a 0-10 scale), micro-expression analysis results (such as the frequency of frowning and the duration of drooping corners of the mouth) were obtained through facial recognition equipment, and assessment data (such as the number of avoidance behaviors and the duration of continuous participation) were recorded in a VR simulated social situation once a week (such as classroom speaking and group discussion), ultimately forming a multimodal raw input data matrix.

[0022] like Figure 1As shown, the multimodal data fusion processing module first cleans the raw data: an improved Z-score algorithm is used to handle outliers, such as marking heart rate data exceeding three times the standard deviation of the mean as outliers (possibly due to device detachment) and replacing them with the sliding window mean for that time period; for missing data, if the missing duration is ≤1 hour (e.g., a brief network outage), linear interpolation is used to fill in the missing data (based on data fitting before and after the missing time); if the missing duration is >1 hour (e.g., the device battery is depleted), a pre-trained LSTM prediction model is used to generate imputation values, and the imputation of "positive sentiment" scores is strictly constrained to 0-10 points (in line with the theoretical range of the PERMA model), and the imputation of "automatic thought frequency" values ​​cannot be negative; through timestamp synchronization technology, all data are regularized to a unified time slice of 5 minutes to ensure that different modal data are aligned in the time dimension; then dynamic modality fusion is performed, and the dynamic modality fusion calculation formula is: ,in: for The fused feature vector at each time step; For the first The basic weighting coefficients of the modality, For the first Modal in Data reliability at any given time; For the first Modal in The dynamic adjustment factor at any given time; For the first The fit coefficient between modal data and psychological theories; For the first Modal The original input data at each moment; n is the modality number; n is the total number of modes; Using time as the variable, a fusion feature vector is calculated, comprehensively considering the fundamental importance, real-time reliability, dynamic adjustment needs, and compatibility with psychological theories of each modality. For example, data highly relevant to theory, such as "automatic thought frequency," is given higher weight, and the influence of highly reliable speech data is strengthened. The final fusion feature vector can comprehensively reflect the patient's physiological stress level, behavioral pattern characteristics, speech emotion tendency, and psychological theory state at time t. Among these factors, data reliability is considered. The calculation formula is: ,in, For the first Modal in The variance of the sliding window at each time step; For the first The maximum variance threshold for modal data; Modal number; For time variables; dynamic adjustment factors The update rule is: when hour, ;when hour, ;when hour, .

[0023] The dynamic trajectory recovery modeling module utilizes an improved LSTM-GRU hybrid network: It has three hidden layers, each containing 128 neurons, with LSTM and GRU units alternating in a 3:1 ratio (the first three LSTM units capture long-term dependencies, and the last GRU unit improves computational efficiency). The attention gating mechanism calculates weight coefficients from the fused feature vectors of the first three time steps (within 15 minutes), reinforcing the impact of recent state changes. Extra weight is given to data from the theoretically sensitive time window of "three days before the exam" (based on theoretical research showing that anxiety patients are prone to symptom relapses before exams), focusing on capturing characteristic fluctuations during this period. The network output layer combines the fused feature vectors to calculate the psychological resilience coefficient, calculated using the following formula: ,in, For positive emotions, For the quality of interpersonal relationships, Score for meaning The system incorporates the core dimensions of the PERMA model (positive emotions, interpersonal relationships, and sense of meaning) of positive psychology, assigning weights based on theoretical importance. It then combines trajectory sensitivity coefficients and recovery thresholds to generate a recovery state index using the corresponding formula in the claims (which includes a cognitive flexibility index, quantified by the average time it takes for patients to adjust negative thoughts; for example, the shorter the reaction time from "I'm sure I won't do well on the test" to "I can do my best," the closer the index is to 1). For instance, when patients' positive emotion scores improve, social interactions increase, and cognitive flexibility improves, the recovery state index rises significantly, visually demonstrating their recovery progress.

[0024] like Figure 3 As shown, the recovery process prediction and strategy generation module first calculates the first derivative (recovery speed) and second derivative (recovery acceleration) of the recovery state index. When the second derivative is negative, it indicates a slowdown in recovery speed. Simultaneously, it analyzes modal consistency. For example, when physiological indicators show a decrease in heart rate (relaxation), does the behavioral data simultaneously show a shortening of social distance (initiative closeness)? If they are consistent, the modal consistency approaches 1; otherwise, it decreases. Based on recovery acceleration and modal consistency, the risk probability is calculated using the following formula: ,in, This refers to the "frequency of negative automatic thoughts" in cognitive behavioral therapy. These are the theoretical factor weights; for The probability of risk at any given moment; This is a preset time interval; These are dynamic weighting coefficients; To recover the second derivative of the state exponent; for Modal consistency at any given moment; The time variable is used to predict the risk probability 24 hours later (incorporating the theoretical factor of the frequency of negative automatic thoughts; a higher frequency significantly increases the risk probability). When the predicted risk probability exceeds a set threshold, the system activates a strategy selection mechanism, using the strategy fit calculation formula: ,in, The degree of matching between the strategy and the user's current theoretical state; For theoretical matching weights; For the first This strategy in Time-based adaptability; This is the balance coefficient; for The recovery state index at time and the first The similarity of the target state indices of the two strategies; for Recovery status index at any given moment; For the first The target state index of the strategy; For the first Cost adaptation factor for this strategy; For the first Cultural adaptation factors for this strategy; The strategy number; The time variable is used to calculate the fit, which comprehensively considers the similarity between the current recovery state and the strategy's target state, the cost and cultural compatibility of the strategy, and the matching degree between the strategy and the user's current theoretical state (e.g., for users with high negative thinking, "cognitive restructuring training" has a higher theoretical matching degree). Finally, the strategy combination with the highest fit is selected, such as "15 minutes of gamified mindfulness breathing training every day (pushed at 8 am) + online peer support group twice a week (communicating with patients of the same age)".

[0025] After the strategy is implemented, the strategy evaluation and feedback adjustment module continuously collects new data: recording the increase in heart rate variability after mindfulness training, changes in the frequency of social interaction after participating in peer groups, etc.; calculating feedback gain and fit gradient (if the fit decreases significantly after 3 days of strategy implementation, the adjustment mechanism is triggered); combining the initial strategy parameters (such as 15 minutes of mindfulness training duration), the training duration is updated to 20 minutes within the duration range recommended by positive psychology theory (because feedback shows that the patient's positive emotion score increases more significantly after extending the training), and the group theme is changed to "coping with academic stress" based on the interaction quality data of the peer group (the original theme "hobbies" had lower participation), so that the strategy continues to adapt to the changes in the patient's recovery status.

[0026] In summary, the multimodal data fusion-based psychological support strategy generation engine demonstrates precise and dynamic intervention capabilities in the process of generating psychological support strategies for adolescents with anxiety disorders. By comprehensively collecting physiological indicators, behavioral performance, vocal features, and assessment data including psychological theoretical dimensions, a multidimensional data matrix is ​​constructed. After cleaning and dynamic modality fusion, a fusion feature vector that comprehensively reflects the patient's state is generated. The dynamic modality fusion formula plays a crucial role, achieving scientific data integration by combining data reliability and dynamic adjustment factors. The dynamic recovery trajectory modeling module, using an improved LSTM-GRU hybrid network, captures temporal dependencies and strengthens theoretically sensitive time windows such as those before exams. The correlation coefficient is derived using the psychological resilience coefficient calculation formula, thereby generating a recovery state index. The recovery process prediction and strategy generation module accurately predicts risks and selects highly suitable strategies, such as the combination of gamified mindfulness training and peer support groups, using risk probability and strategy fit calculation formulas. The strategy evaluation and feedback adjustment module dynamically optimizes parameters based on feedback, ensuring that the strategy continuously adapts to the patient's recovery state. This fully demonstrates the engine's precision and timeliness in combining theory and practice in the intervention of adolescent anxiety disorders.

[0027] Example 2: Generation of Psychological Support Strategies for Survivors of Major Accidents After the multimodal data acquisition module is activated, a specialized data acquisition system is built for survivors of major accidents (such as traffic accidents and natural disasters). Hormone levels, such as adrenaline and cortisol, are collected every 10 minutes using medical wearable devices (detected using electrochemical sensors). Simultaneously, peak heart rate variability (HRV), skin conductance response (GSR), and the intensity of theta waves (trauma-related brain activity) are recorded. Indoor and outdoor monitoring systems and motion sensors are used to record bradykinesia (the ratio of walking speed to pre-accident baseline), social interaction distance with others (whether a safe distance of >1.5 meters is intentionally maintained), daily social interaction duration (cumulative time of proactive communication), and frequency (number of responses to others). Combined with smart home data, daily eating and sleeping patterns (such as fluctuations in food intake and the number of times one wakes up at night) are statistically analyzed. The study used specialized recording equipment to record audio clips of survivors mentioning the accident scene, extracting trauma-related features such as the frequency of trembling tone, the number of sudden drops in speech rate, and the duration of silence. A standardized post-traumatic stress disorder (PTSD) scale (PCL-5) was pushed out every two days through an encrypted questionnaire platform. Simultaneously, data from cognitive behavioral therapy were collected on "frequency of traumatic memory flashbacks" (requiring recording the daily number of intrusive memories) and "safety perception score" (a self-assessment of environmental safety from 0-10). Eye-tracking devices were used to record the duration of avoidance gaze (the percentage of time spent avoiding trauma-related images) when mentioning the accident. Assessment data (such as the threshold for sudden heart rate spikes and the duration of avoidance behavior) were recorded in VR simulations of accident-related scenarios (such as recreating the accident location or similar modes of transportation), forming multimodal raw input data.

[0028] like Figure 2 As shown, the multimodal data fusion processing module cleans the raw data: an improved Z-score algorithm is used to handle outliers, such as marking a sudden increase in skin conductance (exceeding 5 times the standard deviation of the mean) as an anomaly (possibly caused by accidental device touch) and replacing it with the median of the same time period; for missing data, linear interpolation is used for short-term missing data (≤2 hours), and LSTM prediction model is used to fill in long-term missing data (>2 hours), with the "safety perception score" filling value strictly ≥2 points (theoretical minimum threshold, to avoid illogical extreme low scores); the data is normalized to a unified time slice of 10 minutes using timestamp synchronization technology. When performing dynamic modal fusion, the dynamic modal fusion calculation formula is used: The process calculates a fusion feature vector, assigning higher weights to high-theoretical-fit data such as "frequency of traumatic memory flashbacks." Simultaneously, it comprehensively calculates the fusion feature vector based on the fundamental importance, real-time reliability, and dynamic adjustment needs of each modality, generating a fusion feature vector that comprehensively reflects physiological stress response, behavioral avoidance level, vocal emotion, and theoretical state of psychological trauma. Data reliability is also considered. The calculation formula is: Dynamic adjustment factor Adjustments are made according to its update rules, with dynamic adjustment factors. The update rule is: when hour, ;when hour, ;when hour, .

[0029] The dynamic recovery trajectory modeling module utilizes an improved LSTM-GRU hybrid network: each of the three hidden layers contains 160 neurons, with LSTM and GRU units alternating in a 3:1 ratio. An attention gating mechanism enhances the features of the first three time steps (within 30 minutes), and additional weights are assigned to the theoretically sensitive time window of "7 days before and after the anniversary of the incident" (based on the annual response theory of post-traumatic stress disorder). The psychological resilience coefficient is calculated based on the fused feature vectors, using the following formula: It incorporates the "interpersonal relationships" and "sense of meaning" dimensions of the PERMA model and assigns weights according to their theoretical importance; then, combined with the trajectory sensitivity coefficient and recovery threshold, it generates a recovery state index through the corresponding formula for the recovery state index in the claims, and tracks the trajectory from traumatic stress to gradual recovery in real time.

[0030] like Figure 3 As shown, the recovery process prediction and strategy generation module calculates the recovery acceleration (second derivative) and modal consistency (such as whether avoidance occurs synchronously when physiological stress increases), using the risk probability calculation formula, which is: The system predicts risk 72 hours later (incorporating the theoretical factor of the frequency of traumatic memory flashbacks; a higher frequency increases the risk probability). When the risk probability exceeds a set threshold, the strategy fit is calculated using the following formula: The optimal strategy was selected, and "virtual reality exposure therapy" was given a high degree of theoretical fit (in line with the theory of exposure therapy). Finally, the strategy of "cognitive behavioral therapy twice a week + virtual reality exposure therapy assisted training (gradually increasing from 5 minutes to 30 minutes)" was determined.

[0031] The strategy assessment and feedback adjustment module calculates feedback gain (including "flashback frequency reduction value", such as reducing from 4 times / day to 2 times / day) and fit gradient. Within the frequency range recommended by exposure therapy theory, the virtual reality training frequency is adjusted to 3 times per week (because feedback shows that moderately increasing the exposure duration can accelerate the decline in flashback frequency). The communication methods of cognitive behavioral therapy are adjusted according to the patient's cultural background (such as using more concrete metaphors to explain the trauma response) to ensure that the strategy is always in sync with the recovery progress.

[0032] In summary, this engine also demonstrates remarkable adaptability and effectiveness in providing psychological support to survivors of major accidents. Through a specialized data collection system, it acquired data covering physiological stress, behavioral avoidance, vocal emotion, and trauma-related theoretical assessments. After cleaning and dynamic modality fusion, particularly the weighting of key data such as the frequency of traumatic memory flashbacks based on the dynamic modality fusion formula, a comprehensive fusion feature vector was generated. The dynamic recovery trajectory modeling module, while focusing on theoretically sensitive time windows such as accident anniversaries, uses the psychological resilience coefficient calculation formula to derive correlation coefficients, tracking the survivor's trajectory from traumatic stress to recovery in real time. The recovery process prediction and strategy generation module accurately predicts risks and selects optimal strategies, such as cognitive behavioral therapy combined with virtual reality exposure therapy, based on risk probability and strategy fit calculation formulas. The strategy evaluation and feedback adjustment module optimizes parameters such as training frequency within the theoretically recommended range based on feedback, ensuring that intervention measures are always synchronized with the survivor's recovery progress. This highlights the engine's professionalism and targeted approach in psychological trauma intervention after major accidents.

[0033] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A psychological support strategy generation engine based on multimodal data fusion, characterized in that, The engine includes: (1) Multimodal data acquisition module: acquires users' physiological index data, behavioral performance data, voice data and psychological assessment data as multimodal raw input data; (2) Multimodal data fusion processing module: Cleans the original input data, performs dynamic modal fusion, and obtains the fused feature vector; (3) Dynamic recovery trajectory modeling module: An improved LSTM-GRU hybrid network is used to capture the temporal dependency of the fused feature vector and strengthen the features of key time nodes. The psychological elasticity coefficient is calculated based on the fused feature vector, and the recovery state index is generated by combining the trajectory sensitivity coefficient and the recovery threshold. (4) Recovery process prediction and strategy generation module: calculate the first and second derivatives of the recovery state index to obtain the recovery acceleration, determine modal consistency, predict the risk probability based on the recovery acceleration and modal consistency, calculate the strategy fit and screen the strategy; (5) Strategy evaluation and feedback adjustment module: used to calculate feedback gain and fitness gradient, and update new strategy parameters based on feedback gain, fitness gradient and initial strategy parameters.

2. The psychological support strategy generation engine for multimodal data fusion according to claim 1, characterized in that, The multimodal data acquisition module includes physiological indicators such as hormone levels, heart rate variability, skin conductance, and electroencephalogram (EEG); behavioral data includes limb movement parameters, social interaction distance, social interaction duration, social interaction frequency, and daily behavior patterns; and psychological assessment data specifically includes the frequency of automatic thoughts and the degree of cognitive distortion in cognitive behavioral therapy, and the PERMA model score in positive psychology.

3. The psychological support strategy generation engine for multimodal data fusion according to claim 1, characterized in that, The data cleaning in the multimodal data fusion processing module includes handling outliers and missing data. An improved Z-score algorithm is used to handle outliers. When handling missing data, linear interpolation and LSTM prediction methods are used according to the duration of the missing data. The missing data filling results must meet the value constraints of the corresponding psychological theory. Data with different sampling frequencies are normalized to a unified time slice through timestamp synchronization technology.

4. The psychological support strategy generation engine for multimodal data fusion according to claim 1, characterized in that, The dynamic modal fusion calculation formula in the multimodal data fusion processing module is as follows: ,in: for The fused feature vector at each time step; For the first The basic weighting coefficients of the modalities are preset constants; For the first Modal in Data reliability at any given time; For the first Modal in The dynamic adjustment factor at any given time; For the first The fit coefficient between modal data and psychological theories; For the first Modal The original input data at each moment; n is the modality number; n is the total number of modes; It is a time variable.

5. The psychological support strategy generation engine for multimodal data fusion according to claim 4, characterized in that, The data reliability The calculation formula is: ,in, For the first Modal in The variance of the sliding window at each time step; For the first The maximum variance threshold for modal data; Modal number; It is a time variable.

6. The psychological support strategy generation engine for multimodal data fusion according to claim 4, characterized in that, The dynamic adjustment factor The update rule is: when hour, ;when hour, ;when hour, .

7. The psychological support strategy generation engine for multimodal data fusion according to claim 1, characterized in that, In the improved LSTM-GRU hybrid network, LSTM units and GRU units are alternately distributed in the hidden layer in a 3:1 ratio. The attention gating mechanism strengthens recent features by calculating weight coefficients on the fused feature vectors of the first three time steps, while assigning extra weights to data within the "theoretical sensitive time window". The input layer is the fused feature vector, and there are three hidden layers.

8. The psychological support strategy generation engine for multimodal data fusion according to claim 1, characterized in that, The formula for calculating the psychological resilience coefficient in the dynamic recovery trajectory modeling module is as follows: ,in, For positive emotions, For the quality of interpersonal relationships, Score for meaning Weighting of positive psychology theories.

9. The psychological support strategy generation engine for multimodal data fusion according to claim 1, characterized in that, The risk probability calculation formula in the recovery process prediction and strategy generation module is as follows: ,in, This refers to the "frequency of negative automatic thoughts" in cognitive behavioral therapy. These are the theoretical factor weights; for The probability of risk at any given moment; This is a preset time interval; These are dynamic weighting coefficients; To recover the second derivative of the state exponent; for Modal consistency at any given moment; It is a time variable.

10. The psychological support strategy generation engine for multimodal data fusion according to claim 1, characterized in that, The formula for calculating the policy adaptability in the recovery process prediction and policy generation module is as follows: ,in, The degree of matching between the strategy and the user's current theoretical state; For theoretical matching weights; For the first This strategy in Time-based adaptability; This is the balance coefficient; for The recovery state index at time and the first The similarity of the target state indices of the two strategies; for Recovery status index at any given time; For the first The target state index of the strategy; For the first Cost adaptation factor for this strategy; For the first Cultural adaptation factors for this strategy; The strategy number; It is a time variable.

Images