Psychological disease pre-diagnosis information processing method and system

Abstract

The invention discloses a psychological disease pre-diagnosis information processing method and system, and the method comprises the steps: collecting a multi-modal behavior signal and a physiological parameter signal of a user, and carrying out the differential privacy protection processing based on edge calculation and the noise separation processing based on wavelet transform, obtaining a standardized behavior feature sequence and a time sequence physiological feature vector; extracting behavior node features and psychological state markers from the standardized behavior feature sequence, and constructing a three-dimensional incidence matrix in combination with the time sequence physiological feature vector; and calculating a potential risk probability based on a graph neural network model, generating a pre-diagnosis grading result, fusing objective behavior environment data through a progressive protocol, outputting a personalized intervention strategy, and finally generating a comprehensive pre-diagnosis report. According to the method, accurate modeling of psychological-physiological-behavior dynamic association is realized through dynamic knowledge graph construction and cross-modal fusion analysis, meanwhile, data privacy protection is considered, and the accuracy and practicability of psychological health monitoring are effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of mental health monitoring, and in particular to a method and system for processing mental illness pre-diagnosis information. Background Art

[0002] In recent years, mental health monitoring technology has entered practical application. By integrating data from multiple sources, such as wearable devices and mobile terminals, it has provided a new technical path for early risk identification. Existing systems typically use time series analysis methods to process physiological signals and combine them with machine learning models to analyze user behavioral patterns. This has, to a certain extent, enabled the auxiliary diagnosis of common psychological issues such as depression and anxiety. However, in actual deployment, due to significant differences in the time scales and sampling frequencies of data from different modalities, it is difficult to accurately capture the dynamic correlation between behavioral events and physiological responses. Furthermore, privacy protection requirements require strict desensitization of raw data, and traditional encryption methods often destroy the continuity of behavioral characteristics, making it impossible to effectively extract some diagnostically valuable time series patterns. These limitations make the system prone to misjudgment in practical applications, either overinterpreting brief anomalies caused by environmental factors as psychological issues or ignoring potential risks suggested by weak multimodal signals. Summary of the Invention

[0003] In view of the above problems, the present invention provides a method and system for processing information on pre-diagnosis of mental illness. By establishing an analysis framework that can both protect user privacy and accurately reflect the dynamic correlation between psychology, physiology and behavior, it solves the problem of the difficulty in effectively fusing and analyzing multi-source heterogeneous data.

[0004] To achieve the above objectives, in a first aspect, the present application provides a method for processing mental illness prognostic information, comprising: Collect the user's multimodal behavioral signals and physiological parameter signals. The multimodal behavioral signals include keyboard interaction dynamics, screen touch trajectory characteristics, and environmental acoustic marker characteristics. The physiological parameter signals are indirectly obtained through the device's built-in sensors. Performing a first preprocessing on the multimodal behavioral signal to obtain a standardized behavioral feature sequence, wherein the first preprocessing is configured as differential privacy protection processing based on edge computing; and performing a second preprocessing on the physiological parameter signal to obtain a time-series physiological feature vector, wherein the second preprocessing is configured as noise separation processing based on wavelet transform; Extracting behavioral node features and psychological state markers from standardized behavioral feature sequences, and constructing a three-dimensional association matrix using a dynamic knowledge graph to combine behavioral node features, psychological state markers, and temporal physiological feature vectors. The dimensions of the three-dimensional association matrix include behavioral pattern dimension, time dimension, and physiological response dimension. Based on a graph neural network model, the potential risk probability of the three-dimensional association matrix is calculated to obtain a pre-diagnosis grading result. A progressive protocol is used to cross-modally integrate the pre-diagnosis grading result with the user's objective behavioral environment to generate a personalized intervention strategy. The pre-diagnosis grading result is configured as a multi-dimensional output including risk level, development trajectory, and vulnerability factor. A comprehensive pre-diagnosis report is generated based on behavioral node characteristics, pre-diagnosis grading results and personalized intervention strategies, including individual psychological portraits, environmental adaptability assessment and intervention effect prediction.

[0005] In some embodiments, a first preprocessing is performed on the multimodal behavior signal to obtain a standardized behavior feature sequence. The first preprocessing is configured as differential privacy protection processing based on edge computing, including: Performing keystroke interval analysis and pressure intensity analysis on the keyboard interaction dynamics characteristics to generate a first behavioral feature subvector. The keystroke interval analysis includes calculating the standard deviation of the time difference between adjacent keystrokes, and the pressure intensity analysis includes calculating the kurtosis characteristics of the pressure value. Perform sliding trajectory curvature analysis and contact point distribution analysis on the screen touch trajectory features to generate the second-behavior feature sub-vector. The sliding trajectory curvature analysis uses cubic spline interpolation to calculate the average curvature. The contact point distribution analysis includes calculating the spatial autocorrelation index of the contact point density. Perform voiceprint entropy analysis and silence duration analysis on the environmental acoustic marker features to generate a third behavioral feature subvector. Voiceprint entropy analysis uses Mel-frequency cepstral coefficients to calculate spectral entropy. Silence duration analysis includes statistically analyzing the duration distribution of silence segments within a unit time. Perform feature fusion processing based on attention weight on the first behavior feature sub-vector, the second behavior feature sub-vector, and the third behavior feature sub-vector to obtain a behavior feature intermediate vector; Perform noise addition processing based on the Laplace mechanism on the intermediate vector of the behavioral features. The noise scale parameter of the noise addition processing is dynamically adjusted according to the feature sensitivity to generate a noisy feature vector that meets the preset differential privacy requirements. The noisy feature vector is subjected to Z-score normalization transformation to obtain a standardized behavioral feature sequence.

[0006] In some embodiments, a second preprocessing is performed on the physiological parameter signal to obtain a time-series physiological feature vector. The second preprocessing is configured as a noise separation process based on wavelet transform, including: Performing multi-scale wavelet decomposition processing on the physiological parameter signal to generate a wavelet coefficient set including approximate coefficients and detail coefficients, wherein the multi-scale wavelet decomposition processing is configured to perform decomposition processing using an orthogonal wavelet basis function having a compact support characteristic; Performing adaptive threshold denoising on the wavelet coefficient set to obtain denoised wavelet coefficients, wherein the adaptive threshold denoising includes various scale threshold functions, and each scale threshold function is configured to be determined based on a statistical distribution characteristic of the wavelet coefficients; Performing signal reconstruction processing on the denoised wavelet coefficients to obtain a denoised parameter signal, wherein the signal reconstruction processing is configured to screen and retain scale coefficients reflecting physiological rhythm characteristics; The denoised parameter signal is subjected to time domain feature extraction processing to generate a time series physiological feature vector. The time domain feature extraction processing includes calculating time-varying feature parameters reflecting the activity characteristics of the autonomic nervous system.

[0007] In some embodiments, extracting behavior node features and mental state markers from a standardized behavior feature sequence includes: Performing high-information node detection processing on the standardized behavioral feature sequence to identify behavioral node features, wherein the high-information node detection processing is configured to perform detection based on a significance level of a change in the behavioral feature; Performing psychological state correlation analysis on the identified behavior node features to extract psychological state markers, wherein the psychological state correlation analysis is configured to perform feature matching according to a preset psychological-behavioral mapping relationship model; Perform temporal continuity verification on behavioral node features and psychological state markers to ensure temporal consistency of feature extraction. The temporal continuity verification process is configured to perform state transition probability analysis based on the Markov chain model. The verified behavioral node features and psychological state markers are feature encoded to generate structured features.

[0008] In some embodiments, a three-dimensional association matrix is constructed by combining behavior node features, psychological state markers, and temporal physiological feature vectors through a dynamic knowledge graph, including: Perform behavioral pattern dimension encoding on the structured features to generate behavioral pattern feature vectors; Performing physiological response dimension encoding processing on the time series physiological feature vector to generate a physiological response feature vector; Establishing dynamic associations among behavioral pattern dimensions, time dimensions, and physiological response dimensions. The dynamic associations are configured to perform multi-dimensional feature interactive learning based on a graph attention network. The behavioral pattern feature vectors, physiological response feature vectors and their dynamic correlation relationships are represented in a matrix to generate a three-dimensional correlation matrix.

[0009] In some embodiments, a potential risk probability calculation is performed on a three-dimensional association matrix based on a graph neural network model to obtain a pre-diagnosis grading result, including: Performing graph structure modeling processing on the three-dimensional correlation matrix to construct a dynamic heterogeneous graph network, wherein the graph structure modeling processing is configured to model the behavioral pattern dimension, the time dimension, and the physiological response dimension as different types of nodes respectively; Multi-hop neighbor feature aggregation is performed on dynamic heterogeneous graph networks to generate node enhanced feature representations. The multi-hop neighbor feature aggregation process is configured to use a gated attention mechanism to achieve cross-dimensional feature propagation. Perform risk probability prediction on the node enhanced feature representation to calculate the potential risk probability of each mental illness category. The risk probability prediction process is configured as a multi-label classifier based on the sigmoid activation function. The potential risk probability is subjected to a grading threshold judgment process to generate a pre-diagnosis grading result. The grading threshold judgment process is configured to set a multi-level risk threshold according to clinical diagnostic standards.

[0010] In some embodiments, a progressive protocol is used to cross-modally integrate the pre-diagnosis grading results with the user's objective behavioral environment to generate a personalized intervention strategy, including: Performing risk level analysis on the pre-diagnosis grading results to generate pre-diagnosis risk level intervals, wherein the risk level analysis is configured to perform interval matching based on preset multi-level risk thresholds; Performing environmental feature extraction processing on the user's objective behavioral environment to generate an environmental feature vector, wherein the environmental feature extraction processing is configured to extract context features from the user's device usage records and application logs; Performing cross-modal alignment processing on the pre-diagnostic risk level interval and the environmental feature vector to establish a risk-environment association map. The cross-modal alignment processing is configured to use a cross-modal attention mechanism to achieve feature space alignment; Generate intervention strategies based on risk-environment correlation mapping and output personalized intervention strategies; The personalized intervention strategy is subjected to strategy encoding processing to generate an executable intervention instruction sequence. The strategy encoding processing is configured to encapsulate strategy parameters and execution conditions in JSON format.

[0011] In some embodiments, the method further comprises: The personalized intervention strategy is dynamically adjusted according to the anti-fragility feedback mechanism. When the user's psychological resilience is detected to be improved, negative weight decay is automatically triggered, and an adaptive execution plan is generated that includes digital companionship intensity, environmental adjustment gradient, and artificial intervention camouflage strategy.

[0012] In some embodiments, dynamically adjusting the parameters of the personalized intervention strategy based on the anti-fragility feedback mechanism includes: Collect the user's response behavior characteristics, and conduct psychological resilience assessment processing on the response behavior characteristics to generate resilience improvement indicators. The psychological resilience assessment processing includes multi-dimensional analysis based on the characteristics of behavioral pattern changes and physiological parameter stability characteristics; Obtaining the current personalized intervention strategy and performing utility evaluation processing to generate a strategy utility score, wherein the utility evaluation processing is configured to perform comprehensive calculations combining short-term effect indicators and long-term trend indicators; Optimize strategy parameters based on resilience improvement indicators and strategy utility scores, and generate parameter adjustment instructions. The strategy parameter optimization process is configured to use a dynamic weight adjustment algorithm. The parameter adjustment instructions are used to update the personalized intervention strategy and generate an adaptive execution plan.

[0013] In the second aspect, the present invention also provides a mental illness pre-diagnosis information processing system, which is applicable to the mental illness pre-diagnosis information processing method described in the first aspect. The system includes: a signal acquisition unit, a signal processing unit and a pre-diagnosis analysis unit. The signal acquisition unit is used to collect the user's multimodal behavior signals and physiological parameter signals, and the multimodal behavior signals include keyboard interaction dynamics characteristics, screen touch trajectory characteristics and environmental acoustic marker characteristics; the signal processing unit is used to perform a first pre-processing on the multimodal behavior signal to obtain a standardized behavior feature sequence, and the first pre-processing is configured as differential privacy protection processing based on edge computing, and the second pre-processing is configured to obtain a time series physiological feature vector on the physiological parameter signal, and the second pre-processing is configured as noise separation processing based on wavelet transform; the pre-diagnosis analysis unit is used to obtain a time series physiological feature vector from the physiological parameter signal. Behavioral node features and psychological state markers are extracted from the standardized behavioral feature sequence, and a three-dimensional association matrix is constructed by combining the behavioral node features, psychological state markers and time-series physiological feature vectors through a dynamic knowledge graph. The dimensions of the three-dimensional association matrix include behavioral pattern dimension, time dimension and physiological response dimension. The potential risk probability of the three-dimensional association matrix is calculated based on the graph neural network model to obtain the pre-diagnosis grading result. The pre-diagnosis grading result is cross-modally fused with the user's objective behavioral environment using a progressive protocol to generate a personalized intervention strategy. The pre-diagnosis grading result is configured as a multi-dimensional output including risk level, development trajectory and vulnerability factor. A comprehensive pre-diagnosis report is generated based on the behavioral node features, pre-diagnosis grading results and personalized intervention strategy, which includes individual psychological portrait, environmental adaptability assessment and intervention effect prediction.

[0014] Different from the existing technology, the above technical solution of the present invention provides a method and system for processing information on pre-diagnosis of mental illness, the method includes: collecting the user's multimodal behavioral signals and physiological parameter signals, and performing differential privacy protection processing based on edge computing and noise separation processing based on wavelet transform respectively, to obtain standardized behavioral feature sequences and time-series physiological feature vectors; extracting behavioral node features and psychological state markers from the standardized behavioral feature sequence, and constructing a three-dimensional correlation matrix of behavioral patterns, time dimensions and physiological response dimensions in combination with time-series physiological feature vectors; calculating the potential risk probability based on the graph neural network model, generating pre-diagnosis grading results, and fusing objective behavioral environment data through a progressive protocol to output personalized intervention strategies; and finally generating a comprehensive pre-diagnosis report including individual psychological portraits, environmental adaptability assessments and intervention effect predictions. The present invention realizes accurate modeling of the dynamic correlation between psychology, physiology and behavior through dynamic knowledge graph construction and cross-modal fusion analysis, while taking into account data privacy protection, effectively improving the accuracy and practicality of mental health monitoring.

[0015] The above-mentioned records related to the content of the invention are only an overview of the technical solution of this application. In order to enable ordinary technicians in this field to understand the technical solution of this application more clearly, and then implement it according to the text of the specification and the contents recorded in the drawings, and to make the above-mentioned purposes and other purposes, features and advantages of this application easier to understand, the following is an explanation in combination with the specific implementation methods and drawings of this application. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The drawings are only used to illustrate the principles, implementation methods, applications, features and effects of the specific embodiments of the present invention and other related contents, and are not to be considered as limiting the present application.

[0017] In the drawings of the specification: Figure 1 A method step diagram of steps S101 to S105 of the processing method described in the specific embodiment; Figure 2 A method step diagram of steps S201 to S206 of the processing method described in the specific embodiment; Figure 3 It is a structural diagram of the processing system described in the specific implementation method.

[0018] The reference numerals of the drawings in the above description are as follows: 1. Processing system; 11. Signal acquisition unit; 12. Signal processing unit; 13. Pre-diagnosis analysis unit. DETAILED DESCRIPTION

[0019] In order to explain in detail the possible application scenarios, technical principles, specific solutions that can be implemented, and the purpose and effects of this application, the following is a detailed description of the specific embodiments listed in conjunction with the accompanying drawings. The embodiments described herein are only used to more clearly illustrate the technical solutions of this application and are therefore only examples and are not intended to limit the scope of protection of this application.

[0020] References to "embodiments" herein mean that the specific features, structures, or characteristics described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the word "embodiment" in various places in the specification does not necessarily refer to the same embodiment, nor does it particularly limit its independence or relevance to other embodiments. In principle, in this application, as long as there are no technical contradictions or conflicts, the various technical features mentioned in the embodiments can be combined in any manner to form a corresponding implementable technical solution.

[0021] Unless otherwise defined, the technical terms used herein have the same meanings as those generally understood by those skilled in the art to which this application belongs; the use of relevant terms herein is only for describing specific embodiments and is not intended to limit this application.

[0022] In the description of this application, the term "and / or" is used to describe a logical relationship between objects, indicating that three relationships can exist. For example, A and / or B means: A exists, B exists, and both A and B exist. In addition, the character " / " in this document generally indicates that the objects before and after are in a logical "or" relationship.

[0023] In this application, terms such as "first" and "second" are merely used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual quantity, priority or sequence relationship between these entities or operations.

[0024] Without further limitations, in this application, the words "include", "comprise", "have" or other similar open-ended expressions used in sentences are intended to cover non-exclusive inclusion. These expressions do not exclude the presence of additional elements in the process, method or product that includes the elements, so that the process, method or product that includes a series of elements may include not only those defined elements, but also other elements that are not explicitly listed, or also include elements inherent to such process, method or product.

[0025] Consistent with the understanding in the Examination Guidelines, in this application, expressions such as "greater than," "less than," and "exceed" are understood to exclude the number itself; expressions such as "above," "below," and "within" are understood to include the number itself. Furthermore, in the description of the embodiments of this application, "multiple" means two or more (including two), and similar expressions related to "multiple," such as "multiple groups" and "multiple times," are also understood in this manner, unless otherwise specifically defined.

[0026] In the description of the embodiments of the present application, the space-related expressions used, such as "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "vertical", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise", "axial", "radial", "circumferential", etc., indicate the orientation or position relationship based on the orientation or position relationship shown in the specific embodiments or drawings, and are only for the convenience of describing the specific embodiments of the present application or facilitating the reader's understanding, and do not indicate or imply that the device or component referred to must have a specific position, a specific orientation, or be constructed or operated in a specific orientation. Therefore, it should not be understood as a limitation on the embodiments of the present application.

[0027] The processor described in the embodiments of the present application can be implemented by hardware, firmware, software or a combination thereof, and can use a circuit, a single or multiple application-specific integrated circuits (ASIC), a digital signal processor (DSP), a digital signal processing device (DSPD), a programmable logic device (PLD), a field programmable gate array (FPGA), a central processing unit (CPU), a controller, a microcontroller, at least one of a microprocessor, and also includes other physical, biological or chemical structures that can achieve similar or equivalent functions to the processors listed above, such as biological neurons, quantum computing units, DNA computing units, etc., so that the processor can execute some or all of the steps in the computer program or method involved in the various embodiments of the present application, or any combination of the steps mentioned therein.

[0028] The computer program involved in the embodiment can be stored in a computer device readable storage medium, which includes but is not limited to a disk, a tape, a magnetic card, a floppy disk, a flash memory, an optical disc, an optical card, a read-only memory (ROM), a random access memory (RAM), an erasable programmable ROM (EPROM) and an electrically erasable programmable ROM (EEPROM), etc., and also includes other biological, physical or chemical structures that can achieve similar or equivalent functions to the storage media listed above, such as DNA, RNA, protein and other units with information storage capabilities. In a specific embodiment, the storage medium involved can be one of the above-mentioned media types or a combination of the above-mentioned media types. In different embodiments, the computer program involved in the embodiment can be stored in a single medium in a centralized manner or in a distributed manner on multiple media. The memory containing the computer device readable storage medium can be a non-volatile memory or a random access memory. These computer device readable storage media can be built into the device or connected to the device involved in the embodiment as an external device or part of an external device. In some embodiments, a memory having a computer-readable storage medium is deployed locally; in other embodiments, a solution of deploying the memory away from the processor may also be adopted, such as a network-attached memory accessed via an RF circuit or an external port and a communication network, wherein the communication network may be the Internet, one or more intranets, a local area network (LAN), a wide area network (WLAN), a storage area network (SAN), etc., or a suitable combination thereof, as long as the computer device can access the memory. In addition, the computer program involved in the embodiment can be stored in plaintext / ciphertext form, or can be designed as training data, which can be integrated and reorganized through model training and implicitly stored in the parameter state of a deep neural network or other machine learning model.

[0029] See also Figure 1 In a first aspect, this embodiment provides a method for processing mental illness prognostic information, comprising: S101. Collect multimodal behavioral signals and physiological parameter signals of the user. The multimodal behavioral signals include keyboard interaction dynamics, screen touch trajectory characteristics, and environmental acoustic marker characteristics. The physiological parameter signals are indirectly obtained through built-in sensors of the device. S102: performing a first preprocessing on the multimodal behavioral signal to obtain a standardized behavioral feature sequence, wherein the first preprocessing is configured as differential privacy protection processing based on edge computing; and performing a second preprocessing on the physiological parameter signal to obtain a time-series physiological feature vector, wherein the second preprocessing is configured as noise separation processing based on wavelet transform; S103, extracting behavior node features and psychological state markers from the standardized behavior feature sequence, and constructing a three-dimensional association matrix using the dynamic knowledge graph to combine the behavior node features, psychological state markers, and time-series physiological feature vectors. The dimensions of the three-dimensional association matrix include a behavior pattern dimension, a time dimension, and a physiological response dimension. S104. Calculate the potential risk probability of the three-dimensional association matrix based on a graph neural network model to obtain a pre-diagnosis grading result. Use a progressive protocol to cross-modally integrate the pre-diagnosis grading result with the user's objective behavioral environment to generate a personalized intervention strategy. The pre-diagnosis grading result is configured as a multi-dimensional output including risk level, development trajectory, and vulnerability factor. S105. Generate a comprehensive pre-diagnosis report based on behavioral node characteristics, pre-diagnosis grading results, and personalized intervention strategies, including individual psychological portraits, environmental adaptability assessments, and intervention effect predictions.

[0030] In step S101, multimodal behavioral signals refer to feature data acquired through the user's daily digital interaction behavior. Keyboard interaction dynamics characterize operational characteristics such as key pressure and interval time when typing; screen touch trajectory characteristics reflect user touchscreen behavior patterns such as sliding and clicking; and environmental acoustic marker characteristics capture ambient sound characteristics through the device's microphone. Preferably, physiological parameter signals are acquired indirectly through sensors such as the accelerometer and gyroscope built into the smart terminal, including physiological indicators such as heart rate variability and galvanic skin response. These signals together constitute the basic data source for psychological state assessment.

[0031] In step S102, preferably, the first preprocessing adopts an edge computing architecture to implement differential privacy protection, and balances privacy and data availability by performing a two-stage process of noise addition and feature extraction locally on the user terminal: first, Laplace noise that meets the requirements of differential privacy is added to the original behavior signal, and then a standardized behavior feature sequence is generated through a noise-resistant feature extraction algorithm (such as local feature aggregation based on a sliding window). Through staged processing, the noise addition only affects the original signal while retaining the continuity at the feature level. At the same time, the edge computing architecture ensures that the original data does not leave the terminal device.

[0032] Preferably, the noise parameters are dynamically adjusted according to the type of behavioral features: the keyboard interaction dynamics feature uses a lower privacy budget to protect sensitive keystroke patterns, the screen touch trajectory feature uses medium noise to preserve the overall shape of the gesture, and the environmental acoustic marker feature uses a segmented noise addition strategy to distinguish between speech content and environmental noise.

[0033] In step S103, behavior node features refer to key behavioral event markers extracted from standardized behavioral feature sequences, and psychological state markers are combinations of behavioral features that reflect specific psychological states. The dynamic knowledge graph establishes a three-dimensional association between discrete behavior node features, psychological state markers, and continuous temporal physiological feature vectors through graph structure modeling. The behavior pattern dimension encodes the behavior feature type, the time dimension aligns the timestamps of multi-source data, and the physiological response dimension records the corresponding physiological change amplitude. This association method accurately captures the dynamic coupling relationship between psychology, physiology, and behavior.

[0034] In step S104, preferably, the progressive protocol implements cross-modal analysis through a three-level fusion architecture: the first level uses a dynamic time warping algorithm to align the timestamps of behavioral events (i.e., standardized behavioral feature sequences) and physiological responses (i.e., time-series physiological feature vectors) in the time dimension; the second level calculates the spatial correlation weights of behavioral node features and time-series physiological feature vectors through a graph attention mechanism to generate an inter-modal association matrix; the third level introduces an environmental context gateway to perform decision-level fusion of the pre-diagnosis grading results with the objective behavioral environment (including environmental metadata such as device usage time and interactive application type).

[0035] Optimally, the environmental context gateway uses a gating mechanism to filter out irrelevant environmental factors, retaining only environmental features significantly associated with psychological state (e.g., the correlation between frequent social media use at night and anxiety scores). The resulting personalized intervention strategy includes recommendations for environmental adaptation. This layered fusion mechanism mitigates the risk of privacy leakage of raw environmental data while ensuring the environmental relevance of intervention strategies.

[0036] In step S105, the individual psychological profile is constructed by integrating behavioral node characteristics and pre-diagnosis grading results into a user psychological profile. The environmental adaptability assessment analyzes the stability of the user's psychological state in different environments. The intervention effect prediction uses historical data to model the effectiveness probability of different intervention measures. The comprehensive pre-diagnosis report presents the multi-dimensional analysis results in a visual format, providing an objective basis for professional diagnosis.

[0037] This embodiment achieves accurate assessment of mental health status through a complete technical chain involving multi-source data collection, privacy-preserving processing, dynamic association modeling, and intelligent analysis and decision-making. First, data quality and privacy are ensured. Then, a dynamic model is established that reflects the inherent connections between psychology and physiology, ultimately outputting assessment results with clinical reference value. This embodiment places particular emphasis on protecting data privacy while preserving feature information. It also achieves the organic integration of multi-dimensional data through knowledge graphs, addressing the limitations of fragmented data analysis.

[0038] See also Figure 2In some embodiments, a first preprocessing is performed on the multimodal behavior signal to obtain a standardized behavior feature sequence. The first preprocessing is configured as differential privacy protection processing based on edge computing, including: S201, performing a keystroke interval analysis and a pressure intensity analysis on the keyboard interaction dynamics characteristics to generate a first behavioral feature subvector, wherein the keystroke interval analysis includes calculating the standard deviation of the time difference between adjacent keystrokes, and the pressure intensity analysis includes calculating the kurtosis characteristics of the pressure value; S202, performing sliding trajectory curvature analysis and contact point distribution analysis on the screen touch trajectory features to generate a second behavioral feature sub-vector. The sliding trajectory curvature analysis uses a cubic spline interpolation method to calculate the average curvature. The contact point distribution analysis includes calculating the spatial autocorrelation index of the contact point density. S203. Perform voiceprint entropy analysis and silence duration analysis on the environmental acoustic marker features to generate a third behavioral feature subvector. The voiceprint entropy analysis uses Mel-frequency cepstral coefficients to calculate spectral entropy. The silence duration analysis includes statistically analyzing the duration distribution of silence segments within a unit time. S204, performing feature fusion processing based on attention weight on the first behavior feature sub-vector, the second behavior feature sub-vector, and the third behavior feature sub-vector to obtain a behavior feature intermediate vector; S205: Performing a Laplace-based noise addition process on the intermediate vector of the behavioral feature. The noise scale parameter of the noise addition process is dynamically adjusted according to the feature sensitivity to generate a noisy feature vector that meets the preset differential privacy requirements. S206 , performing Z-score normalization conversion processing on the noise-added feature vector to obtain a normalized behavior feature sequence.

[0039] It should be noted that these steps are all completed locally on the user terminal device, and the original multimodal behavioral signal data is immediately destroyed after processing is completed. The noise scale parameter is inversely proportional to the user's privacy authorization level.

[0040] In step S201, the keystroke interval analysis of the keyboard interaction dynamics features adopts a dynamic baseline adjustment strategy, establishes a personalized reference threshold based on the user's historical interaction pattern, and automatically triggers the sensitivity adjustment of the feature extraction algorithm when a sudden change in the keystroke rhythm is detected in real time; the pressure intensity analysis combines the device hardware characteristics to normalize the pressure value, and establishes a pressure-time differential curve to identify the force fluctuation pattern with psychological state indicative significance, thereby enhancing the behavior characterization capability by capturing the nonlinear change pattern of force during the keystroke process.

[0041] In step S202, preferably, a speed weighted correction factor is introduced into the sliding trajectory curvature analysis to realize adaptive sampling rate processing, the spline interpolation density is dynamically adjusted according to the contact movement speed, and low-speed contacts are given higher weights when calculating the cubic spline interpolation curvature to capture hesitant gesture features, ensuring that both fast sliding and fine operations can be accurately modeled; preferably, the contact distribution analysis adopts an improved Moran's I index to calculate spatial autocorrelation, and the dynamic neighborhood radius is set to adapt to the contact distribution evaluation under different screen sizes.

[0042] In step S203, the voiceprint entropy analysis uses frequency band weighted entropy calculation based on the psychoacoustic model, focusing on enhancing the entropy contribution of the human voice sensitive frequency band; the silence duration analysis uses the hidden Markov model to identify effective silence segments and filter out pseudo-silence intervals caused by equipment operation noise.

[0043] In step S204, during the feature fusion phase, attention weight calculation can incorporate cross-modal relevance constraints. Feature importance is first assessed within each modality, followed by cross-modal weight distribution, creating a dual attention focus. This ensures that the weight distribution for keyboard interaction dynamics, screen touch trajectory features, and environmental acoustic marker features meets clinical prior knowledge for psychological state assessment. For example, in depressive tendency analysis, the fusion weight of acoustic features can be appropriately increased. This mechanism can better capture the changes in the relative importance of various behavioral features under different psychological states.

[0044] In step S205, noise addition is processed using an adaptive Laplace mechanism, with feature sensitivity grading criteria as follows: the keystroke interval feature is the most sensitive, followed by the pressure intensity feature, the touch curvature feature is intermediate, and the acoustic entropy feature is the least sensitive. Preferably, the noise scale parameter for the keystroke interval feature is determined through simulated attack experiments, effectively resisting privacy inference based on timing analysis; noise injection for the touch curvature feature is performed in the frequency domain, preserving the overall pattern of spatial distribution; and noise addition for the acoustic entropy feature uses segmented processing guided by voice activity detection, implementing strong noise protection only for non-speech segments. Furthermore, the noise scale parameter is exponentially attenuated according to the privacy level, ensuring stronger protection for highly sensitive features.

[0045] In step S206, the Z-score standardization conversion process uses the user's historical behavior data as a benchmark, and calculates the mean and variance of each feature through a sliding window for dynamic normalization to eliminate the baseline offset caused by device differences, while retaining the user's unique behavior pattern characteristics and avoiding behavior pattern distortion caused by over-standardization.

[0046] This embodiment achieves the coordinated optimization of accurate analysis of behavioral signals and privacy protection through multi-level feature processing. The keyboard interaction dynamics feature uses dynamic baseline adjustment and pressure normalization processing to capture the psychological indicative significance of input behavior; the screen touch trajectory feature realizes multi-scale gesture modeling through velocity-weighted curvature analysis and improved spatial autocorrelation index; the environmental acoustic marker feature enhances the accuracy of human voice frequency band analysis based on the psychoacoustic model. The feature fusion stage adopts a dual attention mechanism to dynamically weight each modal feature according to clinical priors. Privacy protection adopts a hierarchical differential privacy strategy, the keystroke interval feature implements time series analysis resistance processing, the touch curvature feature is noised in the frequency domain, and the acoustic feature is protected by voice activity detection segmentation. Finally, through dynamic normalization driven by user historical data, individual behavioral pattern characteristics are retained while eliminating device differences.

[0047] This embodiment achieves high-fidelity conversion of multimodal behavioral signals in an edge computing environment, addressing the issues of individual differences and noise interference in mobile device signal acquisition. The hierarchical privacy protection mechanism not only meets strict privacy requirements but also preserves the psychological discriminant information of behavioral characteristics, providing a privacy-compliant and clinically interpretable data foundation for subsequent knowledge graph construction.

[0048] In some embodiments, a second preprocessing is performed on the physiological parameter signal to obtain a time-series physiological feature vector. The second preprocessing is configured as a noise separation process based on wavelet transform, including: Performing multi-scale wavelet decomposition processing on the physiological parameter signal to generate a wavelet coefficient set including approximate coefficients and detail coefficients, wherein the multi-scale wavelet decomposition processing is configured to perform decomposition processing using an orthogonal wavelet basis function having a compact support characteristic; Performing adaptive threshold denoising on the wavelet coefficient set to obtain denoised wavelet coefficients, wherein the adaptive threshold denoising includes various scale threshold functions, and each scale threshold function is configured to be determined based on a statistical distribution characteristic of the wavelet coefficients; Performing signal reconstruction processing on the denoised wavelet coefficients to obtain a denoised parameter signal, wherein the signal reconstruction processing is configured to screen and retain scale coefficients reflecting physiological rhythm characteristics; The denoised parameter signal is subjected to time domain feature extraction processing to generate a time series physiological feature vector. The time domain feature extraction processing includes calculating time-varying feature parameters reflecting the activity characteristics of the autonomic nervous system.

[0049] In this embodiment, physiological parameter signals can be collected via biosensors built into wearable devices, with the sampling frequency dynamically adjusted based on the target physiological characteristics. For example, a higher sampling rate is used for heart rate variability analysis, while the sampling requirement is appropriately lowered for body temperature monitoring. In multi-scale wavelet decomposition, the selection of wavelet basis functions takes into account the waveform characteristics of specific physiological signals. For electrocardiogram (ECG) signals with distinct oscillation characteristics, a wavelet basis with good symmetry is preferred, while for transient signals such as galvanic skin response, a wavelet basis with better localization characteristics is more suitable.

[0050] Adaptive threshold denoising utilizes a dynamic adjustment mechanism. The determination of each scale threshold is based not only on the statistical distribution of wavelet coefficients but also on real-time optimization based on signal quality assessment results. The threshold strength is automatically increased when significant motion artifacts are detected, while the threshold is appropriately relaxed during periods of stable signal quality to preserve more physiological details. Signal reconstruction incorporates prior knowledge of physiological characteristic frequency bands. By establishing a frequency band feature library of typical physiological rhythms, it intelligently selects scale coefficients to be retained, prioritizing, for example, characteristic frequency bands associated with respiratory rhythm and heart rate variability.

[0051] Preferably, time-domain feature extraction utilizes a multi-granularity analysis approach, capturing transient changes in physiological parameters through short-term window analysis while preserving long-term trend characteristics. Feature parameter selection considers the needs of different application scenarios. Stress assessment focuses on sympathetic nervous system activity indicators, while sleep quality analysis focuses on parasympathetic nervous system-dominated feature parameters.

[0052] This embodiment achieves real-time processing through an edge computing architecture, providing high-quality physiological feature representation for subsequent behavioral analysis while ensuring privacy and security. A dynamic optimization mechanism based on physiological feature priors adapts to the differences in physiological features of different users, improving the accuracy and reliability of feature extraction.

[0053] In some embodiments, extracting behavior node features and mental state markers from a standardized behavior feature sequence includes: Performing high-information node detection processing on the standardized behavioral feature sequence to identify behavioral node features, wherein the high-information node detection processing is configured to perform detection based on a significance level of a change in the behavioral feature; Performing psychological state correlation analysis on the identified behavior node features to extract psychological state markers, wherein the psychological state correlation analysis is configured to perform feature matching according to a preset psychological-behavioral mapping relationship model; Perform temporal continuity verification on behavioral node features and psychological state markers to ensure temporal consistency of feature extraction. The temporal continuity verification process is configured to perform state transition probability analysis based on the Markov chain model. The verified behavioral node features and psychological state markers are feature encoded to generate structured features.

[0054] In this embodiment, high-information node detection processing adopts a multimodal fusion significance evaluation strategy, which not only analyzes the statistical changes of a single behavioral feature, but also comprehensively considers the coordinated change pattern of multi-dimensional features. In the specific implementation, when asynchronous changes in motion intensity and interaction frequency are detected, the system automatically raises the significance threshold to avoid false detection, and appropriately lowers the threshold for periods of synchronous feature changes to capture subtle behavioral changes. This dynamic adjustment mechanism enables node detection to adapt to the differences in behavioral habits of different users while maintaining sensitivity to key behavioral changes.

[0055] Mental state association analysis utilizes a hierarchical matching mechanism, first determining the broad categories of mental states through coarse-grained screening, and then matching specific markers based on fine-grained features. This process incorporates an attention mechanism to dynamically weight the mapping relationship model, assigning higher weights to mental state dimensions with a high correlation to the current behavior node, thereby improving the accuracy of mental state marker extraction.

[0056] The temporal continuity verification process further optimizes the application of the Markov chain model. It not only analyzes the transition probability of adjacent nodes, but also establishes a long-term dependency model and captures the medium- and long-term evolution of behavioral patterns by introducing a sliding window mechanism.

[0057] Feature encoding uses an adaptive encoding strategy, automatically adjusting the encoding granularity based on the temporal density of behavioral node features and psychological state markers. Fine-grained encoding is used in feature-dense areas to preserve detail, while coarse-grained encoding improves processing efficiency in sparse areas. Furthermore, a semantic similarity metric can be introduced into the encoding process to map nodes with similar distances in the behavioral feature space to adjacent regions in the encoding space.

[0058] This embodiment accurately captures key changes in behavioral features through a dynamically adjusted saliency detection mechanism, and combines it with hierarchical attention matching to improve the accuracy of psychological state marker extraction. Temporal continuity verification uses a sliding window mechanism to analyze the evolution of behavioral patterns, ensuring temporal consistency in feature extraction. An adaptive encoding strategy automatically adjusts the encoding granularity based on feature distribution and maintains the feature space topology through semantic similarity. This solution significantly improves the representation capabilities of behavioral node features and psychological state markers, providing a structured feature representation with temporal relevance and semantic consistency for psychological state analysis, enabling subsequent analysis processes to more accurately reflect the correlation between user behavior patterns and psychological states.

[0059] In some embodiments, a three-dimensional association matrix is constructed by combining behavior node features, psychological state markers, and temporal physiological feature vectors through a dynamic knowledge graph, including: Perform behavioral pattern dimension encoding on the structured features to generate behavioral pattern feature vectors; Performing physiological response dimension encoding processing on the time series physiological feature vector to generate a physiological response feature vector; Establishing a dynamic association relationship between the behavioral pattern dimension, the time dimension, and the physiological response dimension, wherein the time dimension feature is directly obtained by extracting timestamp information from the standardized behavioral feature sequence and the time-series physiological feature vector, and the dynamic association relationship is configured to perform multi-dimensional feature interactive learning based on a graph attention network; The behavioral pattern feature vectors, physiological response feature vectors and their dynamic correlation relationships are represented in a matrix to generate a three-dimensional correlation matrix.

[0060] In this embodiment, the behavioral pattern dimension encoding processing of structured features can be understood as the process of feature space conversion of structured behavioral features. This processing is implemented through a deep feature extraction network to map the original behavioral node features to a low-dimensional vector space with semantic consistency. The physiological response dimension encoding processing is used to convert the temporal physiological feature vector into a representation that matches the behavioral pattern features. This processing uses a temporal convolutional network to capture the multi-scale features of physiological signals. The time dimension feature is obtained by directly extracting the timestamp information from the standardized behavioral feature sequence and the temporal physiological feature vector to ensure strict alignment of the features of each dimension on the time axis.

[0061] Dynamic associations are learned through multi-dimensional feature interactions using a graph attention network, which is configured to adaptively calculate association weights between behavioral pattern dimensions, temporal dimensions, and physiological response dimensions. This network uses a multi-head attention mechanism to concurrently learn feature interaction patterns across different semantic spaces, ultimately fusing the outputs of each attention head to form a stable representation of associations. The generation of a three-dimensional association matrix utilizes tensor concatenation and compression techniques to integrate behavioral pattern feature vectors, physiological response feature vectors, and their dynamic associations into a matrix structure with clear physical meaning.

[0062] This embodiment achieves a unified representation of behavioral characteristics, psychological states, and physiological responses by constructing a three-dimensional correlation matrix. The behavioral pattern dimensional encoding preserves the semantic information of the original behavioral sequence, while the physiological response dimensional encoding extracts key features of the physiological signal. The establishment of dynamic correlations reveals potential connections between cross-modal features. This embodiment uses deep learning technology to automatically learn the complex mapping relationships between dimensions, providing a structured multimodal feature representation foundation for subsequent analysis.

[0063] In some embodiments, a potential risk probability calculation is performed on a three-dimensional association matrix based on a graph neural network model to obtain a pre-diagnosis grading result, including: Performing graph structure modeling processing on the three-dimensional correlation matrix to construct a dynamic heterogeneous graph network, wherein the graph structure modeling processing is configured to model the behavioral pattern dimension, the time dimension, and the physiological response dimension as different types of nodes respectively; Multi-hop neighbor feature aggregation is performed on dynamic heterogeneous graph networks to generate node enhanced feature representations. The multi-hop neighbor feature aggregation process is configured to use a gated attention mechanism to achieve cross-dimensional feature propagation. Perform risk probability prediction on the node enhanced feature representation to calculate the potential risk probability of each mental illness category. The risk probability prediction process is configured as a multi-label classifier based on the sigmoid activation function. The potential risk probability is subjected to a graded threshold judgment process to generate a pre-diagnosis grading result. The graded threshold judgment process is configured to set a multi-level risk threshold according to the clinical diagnostic standard, and the pre-diagnosis grading result is interpretedly associated with the original three-dimensional association matrix to generate a risk factor contribution analysis report. The interpretable association process is configured to be implemented using a graph attention weight backtracking method.

[0064] In this embodiment, the construction of a dynamic heterogeneous graph network can be understood as modeling the behavioral pattern dimension, time dimension, and physiological response dimension in the three-dimensional association matrix as graph nodes with different types of features. This processing is achieved through heterogeneous network embedding technology, in which the connection relationship between nodes of different dimensions is determined by the feature correlation in the original three-dimensional association matrix. Multi-hop neighbor feature aggregation processing is used to achieve cross-dimensional feature propagation, and a gated attention mechanism is used to dynamically adjust the information transmission strength of neighboring nodes with different hop counts to ensure effective information exchange between distant nodes.

[0065] The risk probability prediction process is implemented using a multi-label classifier based on the sigmoid activation function, which simultaneously outputs the potential risk probabilities of multiple mental illness categories. Preferably, the multi-level risk thresholds used in the grading threshold judgment process are set according to clinical diagnostic criteria, and the specific threshold divisions take into account factors such as symptom severity and intervention urgency. Interpretable association processing is implemented using a graph attention weighted backtracking method. Specifically, the graph attention weighted backtracking method reversely traces the contribution of each dimensional node to the final pre-diagnosis grading result along the feature propagation path, generating a risk factor contribution analysis report with a causal relationship.

[0066] This embodiment realizes the deep fusion analysis of multi-dimensional features by constructing a dynamic heterogeneous graph network. Among them, the gated attention mechanism effectively solves the information attenuation problem of traditional graph neural networks in long-distance dependency modeling, and the multi-label classifier adapts to the clinical characteristics of comorbidity of mental illness. The risk factor contribution analysis not only provides pre-diagnosis results, but also reveals the specific impact path of each behavioral pattern and physiological indicator on the diagnostic conclusion, providing an explainable decision-making basis for the formulation of subsequent intervention measures. While maintaining computational efficiency, this embodiment realizes end-to-end mapping from multi-dimensional features to clinical diagnosis.

[0067] In some embodiments, a progressive protocol is used to cross-modally integrate the pre-diagnosis grading results with the user's objective behavioral environment to generate a personalized intervention strategy, including: Performing risk level analysis on the pre-diagnosis grading results to generate pre-diagnosis risk level intervals, wherein the risk level analysis is configured to perform interval matching based on preset multi-level risk thresholds; Performing environmental feature extraction processing on the user's objective behavioral environment to generate an environmental feature vector, wherein the environmental feature extraction processing is configured to extract context features from the user's device usage records and application logs; Performing cross-modal alignment processing on the pre-diagnostic risk level interval and the environmental feature vector to establish a risk-environment association map. The cross-modal alignment processing is configured to use a cross-modal attention mechanism to achieve feature space alignment; An intervention strategy generation process is performed based on the risk-environment association mapping to output a personalized intervention strategy. The intervention strategy generation process is configured to adopt differentiated strategy generation rules according to the risk level: Adopting environmental fine-tuning strategies for low-risk levels; Adopting a digital companionship strategy for medium risk levels; Use manual intervention camouflage strategy for high-risk levels; Performing feasibility verification on the generated personalized intervention strategy to ensure the strategy is executable, wherein the feasibility verification is configured to perform strategy simulation testing based on user historical behavior data; The personalized intervention strategy is subjected to strategy encoding processing to generate an executable intervention instruction sequence. The strategy encoding processing is configured to encapsulate strategy parameters and execution conditions in JSON format.

[0068] In this embodiment, risk level parsing refers to the process of mapping pre-diagnosis grading results to pre-set risk level intervals. This process is implemented through an interval matching algorithm, in which multi-level risk thresholds are set according to clinical psychology standards, discretizing continuous risk probabilities into clinically meaningful grading. Preferably, environmental feature extraction uses temporal feature coding technology to capture the spatiotemporal patterns of the user's behavioral environment. The generated environmental feature vector includes dimensional features such as device usage frequency, application switching patterns, and interaction time distribution.

[0069] The cross-modal alignment process uses a cross-modal attention mechanism to align the feature space of the pre-diagnostic risk level interval with the environmental feature vector. By automatically learning the potential correlation patterns between risk levels and environmental features, a risk-environment correlation mapping is established. The intervention strategy generation process uses differentiated strategy generation rules based on risk levels. Preferably, the environment fine-tuning strategy achieves imperceptible intervention by adjusting interface elements and notification frequency; the digital companionship strategy introduces virtual companion characters to provide emotional support; and the manual intervention disguise strategy disguises professional intervention as system-recommended content. Furthermore, the feasibility verification process constructs a strategy simulation environment based on historical user behavior data and evaluates the success rate of strategy execution through a behavior prediction model.

[0070] Policy encoding processing converts personalized intervention strategies into executable intervention instruction sequences, and uses JSON format to encapsulate elements such as strategy types, trigger conditions, and execution parameters to ensure the portability of strategies across different terminal devices.

[0071] This embodiment achieves seamless connection between psychological risk pre-diagnosis and actual intervention through a progressive protocol. The cross-modal fusion technology solves the problem of traditional intervention strategies being out of touch with the environment, while the layered strategy ensures the precise matching of intervention intensity and risk level. While maintaining a natural user experience, it achieves a closed-loop process from risk identification to precise intervention.

[0072] In some embodiments, the method further comprises: The personalized intervention strategy is dynamically adjusted according to the anti-fragility feedback mechanism. When the user's psychological resilience is detected to be improved, negative weight decay is automatically triggered, and an adaptive execution plan is generated that includes digital companionship intensity, environmental adjustment gradient, and artificial intervention camouflage strategy.

[0073] In this embodiment, the antifragility feedback mechanism is an adaptive system that dynamically adjusts parameters by monitoring user responses to intervention strategies. This mechanism quantifies user adaptability in real time using a resilience assessment model. Preferably, the resilience assessment model achieves real-time quantification of user adaptability through multimodal data fusion. This involves collecting user interaction behavior data (such as application usage time and response speed) and physiological indicator data (such as heart rate variability). After extracting features using a time-series neural network, a dynamic assessment matrix is constructed based on pre-diagnosis and grading results. This model uses a sliding window mechanism to process the real-time data stream and outputs a resilience index ranging from 0 to 1. An increase in the index triggers a negative weight decay mechanism. The assessment process comprehensively considers dimensions such as behavioral stability and physiological recovery rate. Negative weight decay refers to the process of automatically reducing the intervention intensity parameter when an increase in user resilience is detected. This process is implemented using a gradient descent algorithm to ensure that the intervention intensity remains in dynamic balance with the user's status.

[0074] Digital companionship intensity represents the frequency of interaction and emotional support provided by the virtual companion character. The environmental adjustment gradient reflects the gradual degree of interface adjustment and environmental optimization. The artificial intervention disguise strategy refers to the degree of natural presentation of professional intervention content. Adaptive implementation plans are generated through multi-parameter collaborative optimization, taking into account the current intervention effectiveness and user acceptance. Improved psychological resilience is detected based on a comprehensive assessment of changes in user behavior patterns and improvements in physiological indicators, using a sliding window statistical method to identify sustained positive trends.

[0075] This embodiment introduces an anti-fragile feedback mechanism to enable the intervention strategy to self-evolve. The system not only responds to the user's immediate state but also identifies changes in long-term adaptability, achieving a paradigm shift from passive intervention to proactive growth. The dynamic parameter adjustment process maximizes user autonomy while ensuring intervention effectiveness, creating a virtuous cycle of psychological development.

[0076] In some embodiments, dynamically adjusting the parameters of the personalized intervention strategy based on the anti-fragility feedback mechanism includes: Collect the user's response behavior characteristics, and conduct psychological resilience assessment processing on the response behavior characteristics to generate resilience improvement indicators. The psychological resilience assessment processing includes multi-dimensional analysis based on the characteristics of behavioral pattern changes and physiological parameter stability characteristics; Obtaining the current personalized intervention strategy and performing utility evaluation processing to generate a strategy utility score, wherein the utility evaluation processing is configured to perform comprehensive calculations combining short-term effect indicators and long-term trend indicators; Optimize strategy parameters based on resilience improvement indicators and strategy utility scores, and generate parameter adjustment instructions. The strategy parameter optimization process is configured to use a dynamic weight adjustment algorithm. The parameter adjustment instructions are used to update the personalized intervention strategy and generate an adaptive execution plan.

[0077] In this embodiment, response behavior features are a collection of interactive behavior data generated by users in response to the current intervention strategy. These features include dimensions such as response time, frequency of function usage, and emotional feedback intensity. These features are collected through smart terminal sensors and interaction logs. Resilience assessment processing involves feature extraction and quantitative analysis of raw behavioral data. Behavioral pattern change features use a Markov chain model to identify stable transitions in behavioral sequences. Physiological parameter stability features calculate autonomic nervous system regulation capacity based on biosignals such as galvanic skin response and heart rate variability.

[0078] Effectiveness evaluation is an analytical module that quantifies the effectiveness of ongoing personalized intervention strategies. Short-term effectiveness indicators reflect the magnitude of immediate behavioral change and can be calculated using before-and-after comparisons. Long-term trend indicators assess the effectiveness of ongoing interventions, using time series prediction models to analyze the cumulative effects of behavioral improvements. The strategy effectiveness score integrates these indicators using a weighted fusion algorithm, with weights dynamically adjusted based on the intervention phase.

[0079] Strategy parameter optimization involves generating control instructions based on resilience improvement indicators and strategy utility scores. The dynamic weight adjustment algorithm employs a dual-objective optimization framework to minimize user psychological burden while ensuring intervention effectiveness. Parameter adjustment instructions include a frequency correction value for digital companionship intensity, a step size parameter for the environmental adjustment gradient, and a trigger threshold for the manual intervention camouflage strategy. These parameters are solved within a pre-set feasible domain using a constrained optimization algorithm.

[0080] The process of generating adaptive execution plans embodies the core characteristics of antifragile systems. Specifically, when resilience indicators indicate increased user adaptability, the system automatically reduces intervention intensity and shifts decision-making weight toward user autonomy. This mechanism, through the principle of negative feedback regulation, maintains a dynamic balance between intervention pressure and psychological resilience, avoiding both over-reliance on manual intervention and maladaptation caused by premature withdrawal of support. It also ensures that intervention strategies are always optimally aligned with the user's psychological state.

[0081] See also Figure 3In the second aspect, this embodiment further provides a mental illness pre-diagnosis information processing system 1, which is applicable to the mental illness pre-diagnosis information processing method described in the first aspect. The system includes: a signal acquisition unit 11, a signal processing unit 12 and a pre-diagnosis analysis unit 13. The signal acquisition unit 11 is used to collect the user's multimodal behavior signals and physiological parameter signals. The multimodal behavior signals include keyboard interaction dynamics characteristics, screen touch trajectory characteristics and environmental acoustic marker characteristics; the signal processing unit 12 is used to perform a first pre-processing on the multimodal behavior signal to obtain a standardized behavior feature sequence, and the first pre-processing is configured as differential privacy protection processing based on edge computing, and the second pre-processing is configured to obtain a time series physiological feature vector on the physiological parameter signal, and the second pre-processing is configured as noise separation processing based on wavelet transform; pre-diagnosis analysis Unit 13 is used to extract behavioral node features and psychological state markers from standardized behavioral feature sequences, and construct a three-dimensional association matrix of behavioral node features, psychological state markers, and time-series physiological feature vectors through a dynamic knowledge graph. The dimensions of the three-dimensional association matrix include behavioral pattern dimension, time dimension, and physiological response dimension; based on the graph neural network model, the potential risk probability of the three-dimensional association matrix is calculated to obtain the pre-diagnosis grading results, and a progressive protocol is used to cross-modally fuse the pre-diagnosis grading results with the user's objective behavioral environment to generate personalized intervention strategies. The pre-diagnosis grading results are configured as multi-dimensional outputs including risk level, development trajectory, and vulnerability factors; based on the behavioral node features, pre-diagnosis grading results, and personalized intervention strategies, a comprehensive pre-diagnosis report is generated including individual psychological portraits, environmental adaptability assessments, and intervention effect predictions.

[0082] In this embodiment, the signal acquisition unit 11 synchronously collects the user's multimodal behavior signals (including keyboard interaction dynamics characteristics, screen touch trajectory characteristics and environmental acoustic marker characteristics) and physiological parameter signals (such as heart rate variability and skin electrical response) through a high-precision sensor module, and ensures the timing synchronization of data acquisition.

[0083] The signal processing unit 12 performs edge computing-based differential privacy protection processing on the multimodal behavior signal at the edge computing node to generate a standardized behavior feature sequence, and at the same time performs wavelet transform-based noise separation processing on the physiological parameter signal to output a time series physiological feature vector.

[0084] The pre-diagnosis analysis unit 13 first extracts behavioral node features and psychological state markers from the standardized behavioral feature sequence, and then constructs a three-dimensional association matrix (behavioral pattern dimension × time dimension × physiological response dimension) with the time-series physiological feature vector through a dynamic knowledge graph, and uses a graph neural network model to calculate the potential risk probability to generate a pre-diagnosis grading result containing risk level, development trajectory and vulnerability factor; the unit further adopts a progressive protocol to cross-modally fuse the pre-diagnosis grading results with the user's objective behavioral environment to generate personalized intervention strategies, and finally integrates the behavioral node features, pre-diagnosis grading results and intervention strategies to form a comprehensive pre-diagnosis report that includes individual psychological portraits, environmental adaptability assessments and intervention effect predictions.

[0085] The mental illness pre-diagnosis information processing system 1 described in this embodiment executes the mental illness pre-diagnosis information processing method described in the first aspect. The specific content is detailed in the above embodiment and will not be repeated here.

[0086] By employing the above technical solutions, the present invention, unlike existing technologies, offers the following beneficial effects: It addresses the conflict between privacy protection and data availability in the pre-diagnosis of mental illnesses through the collaborative processing of multimodal behavioral signals and physiological parameter signals. Edge computing-based differential privacy protection locally adds noise and extracts features from keyboard interaction dynamics, screen touch trajectory features, and environmental acoustic markers. Dynamically adjusted privacy budgets ensure that raw behavioral data remains within the device, while the resulting standardized behavioral feature sequence preserves the continuity of behavioral patterns. Wavelet transform noise separation extracts time-series physiological feature vectors from physiological parameter signals, reflecting the activity characteristics of the autonomic nervous system. A dynamic knowledge graph constructs a three-dimensional correlation matrix of behavioral node features, psychological state markers, and time-series physiological feature vectors, modeling cross-dimensional feature interactions using a graph attention network. A progressive protocol employs a three-level process: time alignment, spatial correlation, and environmental fusion. It cross-modally matches pre-diagnosis grading results with the objective behavioral environment, generating a personalized intervention plan that includes environmental fine-tuning strategies, digital companionship strategies, and human intervention camouflage strategies. The anti-fragility feedback mechanism dynamically adjusts policy parameters based on the user's resilience index, forming an adaptive intervention loop. While meeting strict privacy requirements, this technical solution enables dynamic correlation analysis between behavioral characteristics and physiological responses, providing a multimodal fusion solution for the early identification of mental illness.

[0087] Finally, it should be noted that although the above embodiments have been described in the specification and drawings of this application, this does not limit the scope of patent protection of this application. All technical solutions generated by replacing or modifying equivalent structures or equivalent processes based on the essential concepts of this application using the contents recorded in the specification and drawings of this application, as well as directly or indirectly implementing the technical solutions of the above embodiments in other related technical fields, are included in the scope of patent protection of this application.

Claims

1. A method for processing mental illness prognostic information, characterized in that: include: Collecting the user's multimodal behavior signals and physiological parameter signals, the multimodal behavior signals include keyboard interaction dynamics characteristics, screen touch trajectory characteristics and environmental acoustic marker characteristics, and the physiological parameter signals are indirectly obtained through the device's built-in sensors; performing a first preprocessing on the multimodal behavioral signal to obtain a standardized behavioral feature sequence, wherein the first preprocessing is configured as differential privacy protection processing based on edge computing; and performing a second preprocessing on the physiological parameter signal to obtain a time-series physiological feature vector, wherein the second preprocessing is configured as noise separation processing based on wavelet transform; Extracting behavior node features and psychological state markers from the standardized behavior feature sequence, and constructing a three-dimensional association matrix using the behavior node features, psychological state markers, and time-series physiological feature vectors through a dynamic knowledge graph, wherein the dimensions of the three-dimensional association matrix include a behavior pattern dimension, a time dimension, and a physiological response dimension; A pre-diagnosis grading result is obtained by calculating the potential risk probability of the three-dimensional association matrix based on a graph neural network model. The pre-diagnosis grading result is cross-modally integrated with the user's objective behavioral environment using a progressive protocol to generate a personalized intervention strategy. The pre-diagnosis grading result is configured as a multi-dimensional output including risk level, development trajectory, and vulnerability factor. A comprehensive pre-diagnosis report including individual psychological portrait, environmental adaptability assessment and intervention effect prediction is generated based on the behavioral node characteristics, pre-diagnosis grading results and personalized intervention strategies.

2. The method for processing psychological disease prognosis information according to claim 1, characterized in that: Performing a first preprocessing on the multimodal behavior signal to obtain a standardized behavior feature sequence, wherein the first preprocessing is configured as differential privacy protection processing based on edge computing, including: Performing a keystroke interval analysis and a pressure intensity analysis on the keyboard interaction dynamics feature to generate a first behavioral feature subvector, wherein the keystroke interval analysis includes calculating a standard deviation of time differences between adjacent keystrokes, and the pressure intensity analysis includes calculating a kurtosis feature of the pressure value; Performing a sliding trajectory curvature analysis and a contact point distribution analysis on the screen touch trajectory feature to generate a second behavioral feature subvector, wherein the sliding trajectory curvature analysis uses a cubic spline interpolation method to calculate the average curvature, and the contact point distribution analysis includes calculating a spatial autocorrelation index of the contact point density; Performing voiceprint entropy analysis and silence duration analysis on the environmental acoustic marker features to generate a third behavioral feature subvector, wherein the voiceprint entropy analysis uses Mel-frequency cepstral coefficients to calculate spectral entropy, and the silence duration analysis includes statistically analyzing the duration distribution of silence segments within a unit time; Performing feature fusion processing based on attention weight on the first behavior feature sub-vector, the second behavior feature sub-vector, and the third behavior feature sub-vector to obtain a behavior feature intermediate vector; Performing a Laplace-based noise addition process on the intermediate vector of the behavioral feature, where the noise scale parameter of the noise addition process is dynamically adjusted according to the feature sensitivity, to generate a noisy feature vector that meets the preset differential privacy requirements; The noisy feature vector is subjected to Z-score normalization conversion processing to obtain the normalized behavior feature sequence.

3. The method for processing psychological disease prognosis information according to claim 1, characterized in that: Performing a second preprocessing on the physiological parameter signal to obtain a time-series physiological feature vector, wherein the second preprocessing is configured as a noise separation process based on wavelet transform, including: Performing multi-scale wavelet decomposition processing on the physiological parameter signal to generate a wavelet coefficient set including approximate coefficients and detail coefficients, wherein the multi-scale wavelet decomposition processing is configured to perform decomposition processing using an orthogonal wavelet basis function having a compact support characteristic; performing adaptive threshold denoising on the wavelet coefficient set to obtain denoised wavelet coefficients, wherein the adaptive threshold denoising includes various scale threshold functions, and each scale threshold function is configured to be determined based on a statistical distribution characteristic of the wavelet coefficients; performing signal reconstruction processing on the denoised wavelet coefficients to obtain a denoised parameter signal, wherein the signal reconstruction processing is configured to screen and retain scale coefficients reflecting physiological rhythm characteristics; The denoised parameter signal is subjected to a time domain feature extraction process to generate the time series physiological feature vector, wherein the time domain feature extraction process includes calculating a time-varying feature parameter reflecting the activity characteristics of the autonomic nervous system.

4. The method for processing psychological disease prognosis information according to claim 1, characterized in that: Extracting behavior node features and psychological state markers from the standardized behavior feature sequence includes: Performing high-information node detection processing on the standardized behavior feature sequence to identify behavior node features, wherein the high-information node detection processing is configured to perform detection based on a significance level of a change in the behavior feature; Performing psychological state association analysis on the identified behavior node features to extract psychological state markers, wherein the psychological state association analysis is configured to perform feature matching according to a preset psychology-behavior mapping relationship model; Performing temporal continuity verification processing on the behavior node features and psychological state markers to ensure temporal consistency of feature extraction, wherein the temporal continuity verification processing is configured to perform state transition probability analysis based on a Markov chain model; The verified behavioral node features and psychological state markers are feature encoded to generate structured features.

5. The method for processing psychological disease prognosis information according to claim 4, characterized in that: The behavior node features, psychological state markers and temporal physiological feature vectors are constructed into a three-dimensional association matrix through a dynamic knowledge graph, including: Performing behavioral pattern dimension encoding processing on the structured features to generate a behavioral pattern feature vector; Performing physiological response dimension encoding processing on the time series physiological feature vector to generate a physiological response feature vector; Establishing a dynamic association relationship between a behavioral pattern dimension, a time dimension, and a physiological response dimension, wherein the dynamic association relationship is configured to perform multi-dimensional feature interactive learning based on a graph attention network; The behavioral pattern feature vector, the physiological response feature vector and their dynamic correlation are processed into a matrix representation to generate the three-dimensional correlation matrix.

6. The method for processing psychological disease prognosis information according to claim 1, characterized in that: The potential risk probability of the three-dimensional association matrix is calculated based on the graph neural network model to obtain a pre-diagnosis grading result, including: Performing graph structure modeling processing on the three-dimensional correlation matrix to construct a dynamic heterogeneous graph network, wherein the graph structure modeling processing is configured to model the behavior pattern dimension, the time dimension, and the physiological response dimension as different types of nodes respectively; performing multi-hop neighbor feature aggregation processing on the dynamic heterogeneous graph network to generate node enhanced feature representation, wherein the multi-hop neighbor feature aggregation processing is configured to use a gated attention mechanism to achieve cross-dimensional feature propagation; Performing risk probability prediction processing on the node enhanced feature representation to calculate the potential risk probability of each mental illness category, wherein the risk probability prediction processing is configured to be implemented as a multi-label classifier based on a sigmoid activation function; A grading threshold judgment process is performed on the potential risk probability to generate a pre-diagnosis grading result, wherein the grading threshold judgment process is configured to set a multi-level risk threshold according to clinical diagnosis standards.

7. The method for processing psychological disease prognosis information according to claim 6, characterized in that: A progressive protocol is used to cross-modally integrate the pre-diagnosis grading results with the user's objective behavioral environment to generate a personalized intervention strategy, including: Performing risk level analysis on the pre-diagnosis grading result to generate pre-diagnosis risk level intervals, wherein the risk level analysis is configured to perform interval matching based on preset multi-level risk thresholds; Performing an environmental feature extraction process on the objective behavioral environment of the user to generate an environmental feature vector, wherein the environmental feature extraction process is configured to extract context features from user device usage records and application logs; Performing cross-modal alignment processing on the pre-diagnostic risk level interval and the environmental feature vector to establish a risk-environment association map, wherein the cross-modal alignment processing is configured to achieve feature space alignment using a cross-modal attention mechanism; Performing intervention strategy generation processing based on the risk-environment association mapping and outputting a personalized intervention strategy; The personalized intervention strategy is subjected to strategy coding processing to generate an executable intervention instruction sequence, wherein the strategy coding processing is configured to encapsulate strategy parameters and execution conditions in JSON format.

8. The method for processing psychological disease prognosis information according to claim 1, characterized in that: The method further comprises: The personalized intervention strategy is dynamically adjusted according to the anti-fragility feedback mechanism. When it is detected that the user's psychological resilience has improved, negative weight decay is automatically triggered, and an adaptive execution plan is generated that includes digital companionship intensity, environmental adjustment gradient, and artificial intervention camouflage strategy.

9. The method for processing psychological disease prognosis information according to claim 8, characterized in that: The personalized intervention strategy is dynamically adjusted based on the anti-fragility feedback mechanism, including: Collecting the user's response behavior characteristics and performing a psychological resilience assessment on the response behavior characteristics to generate a resilience improvement index. The psychological resilience assessment includes a multi-dimensional analysis based on the characteristics of behavioral pattern changes and the stability characteristics of physiological parameters. Obtaining a current personalized intervention strategy and performing a utility evaluation process to generate a strategy utility score, wherein the utility evaluation process is configured to perform a comprehensive calculation combining short-term effect indicators and long-term trend indicators; performing a strategy parameter optimization process based on the resilience improvement index and the strategy utility score, and generating a parameter adjustment instruction, wherein the strategy parameter optimization process is configured to be implemented using a dynamic weight adjustment algorithm; The parameter adjustment instruction is used to update the personalized intervention strategy to generate the adaptive execution plan.

10. A mental illness pre-diagnosis information processing system, characterized in that: The method for processing psychological disease prognostic information according to any one of claims 1 to 9, wherein the system comprises: A signal acquisition unit, configured to acquire multimodal behavioral signals and physiological parameter signals of the user, wherein the multimodal behavioral signals include keyboard interaction dynamics characteristics, screen touch trajectory characteristics, and environmental acoustic marker characteristics; a signal processing unit, configured to perform a first preprocessing on the multimodal behavioral signal to obtain a standardized behavioral feature sequence, wherein the first preprocessing is configured as differential privacy protection processing based on edge computing; and to perform a second preprocessing on the physiological parameter signal to obtain a time-series physiological feature vector, wherein the second preprocessing is configured as noise separation processing based on wavelet transform; A pre-diagnosis analysis unit is used to extract behavioral node features and psychological state markers from the standardized behavioral feature sequence, and construct a three-dimensional association matrix of the behavioral node features, psychological state markers and time-series physiological feature vectors through a dynamic knowledge graph, wherein the dimensions of the three-dimensional association matrix include behavioral pattern dimension, time dimension and physiological response dimension; based on the graph neural network model, the potential risk probability of the three-dimensional association matrix is calculated to obtain a pre-diagnosis grading result, and a progressive protocol is used to cross-modally fuse the pre-diagnosis grading result with the user's objective behavioral environment to generate a personalized intervention strategy, wherein the pre-diagnosis grading result is configured as a multi-dimensional output including risk level, development trajectory and vulnerability factor; based on the behavioral node features, the pre-diagnosis grading result and the personalized intervention strategy, a comprehensive pre-diagnosis report including individual psychological portrait, environmental adaptability assessment and intervention effect prediction is generated.

Images