Psychological risk assessment method and system based on artificial intelligence

Abstract

The invention discloses a psychological risk assessment method and system based on artificial intelligence. The psychological risk assessment method comprises the steps of data acquisition, representation learning, risk assessment and assessment result output. The invention relates to the technical field of psychological risk intelligent assessment, in particular to a psychological risk assessment method and system based on artificial intelligence. Dynamic modal gating is introduced to weight modals, unreliable signals are eliminated, and prototype memory is matched to quickly adapt to typical high and low risk modes; a parallel self-attention and microstable state space mixed time sequence model is adopted, local and global time sequence dependence is captured, and the robustness, the few-sample adaptability and the interpretability of feature representation are improved; meanwhile, classification scores and instantaneous danger rates are output, confidence intervals are constructed through uncertainty decomposition, criteria are fused, parameters are fitted, posterior risk probabilities are provided in psychological risk scenes, early warning accuracy is improved, false and missing reports are reduced, and clinical practicability is enhanced.

Description

Technical Field

[0001] This invention relates to the field of intelligent psychological risk assessment technology, specifically to a psychological risk assessment method and system based on artificial intelligence. Background Technology

[0002] With the widespread adoption of wearable devices and smartphones, and the rapid advancements in artificial intelligence technologies such as natural language processing, deep temporal modeling and self-supervised learning, mental health monitoring is shifting from isolated questionnaire assessments to continuous, passive, and multimodal real-time perception. Research and industry are gradually fusing heterogeneous data such as conversational texts, diary emotions, heart rate, skin conductance, and behavioral usage trajectories to capture short-term emotional fluctuations and long-term risk trends. Meanwhile, hybrid methods based on attention mechanisms, state-space models, and prototype memory provide new pathways for few-sample adaptation and interpretability.

[0003] However, existing technologies still face several key limitations: multimodal data suffers from acquisition noise, missing data, and variations in modal reliability, leading to unstable performance of simple splicing or single pre-training methods under noisy conditions; scarce annotations and high label noise limit the generalization ability of supervised models; insufficient model interpretability and calibration, and incomplete uncertainty quantification make alarms prone to false alarms or difficult to gain clinical trust; in addition, domain migration and time drift, privacy and compliance requirements, and robustness and latency constraints in engineering deployment also hinder the widespread and reliable application of the system in real production environments. Summary of the Invention

[0004] To address the above issues and overcome the shortcomings of existing technologies, this invention provides an artificial intelligence-based psychological risk assessment method and system. Addressing the problems of existing multimodal representation learning methods, which typically employ simple concatenation or single-contrast pre-training, struggle to handle modal noise and missing data, suffer from poor adaptability, and lack interpretability, this solution combines interventional positive samples... Compared to pre-trained robust modal encoders, this approach introduces dynamic modal gating to weight modalities by confidence level and eliminates or weakens unreliable signals during fusion. Combined with differentiable prototype memory, it achieves rapid adaptation to typical risk patterns. Simultaneously, a hybrid temporal model of parallel self-attention and differentiable state space captures local and global temporal dependencies and maps structured interpretable features, significantly improving the robustness, temporal modeling ability, few-sample adaptability, and interpretability of the representation, thereby enhancing the accuracy and stability of downstream risk prediction. Addressing the issues that existing psychological risk assessment outputs often only provide single probability or threshold alarms, lack unified quantification of time-dynamic risks and multi-source uncertainties, and struggle to reliably fuse neural network results with rule-based or sample prototype-based criteria, this solution simultaneously outputs window-level instantaneous classification scores and time-varying instantaneous hazard rates, and employs... The study uses observation variance prediction and posterior variance tridecomposition to distinguish between model uncertainty, data uncertainty, and sample uncertainty, constructs confidence intervals to ensure frequency coverage, and introduces support vector machines and prototype-based fast prediction as complementary criteria in parallel. Then, it uses a small Bayesian network to structurally fuse this information and uses supervised learning to fit the fusion parameters. Finally, it provides calibrated posterior risk probabilities, confidence intervals, and time-sensitive hazard rates in psychological risk scenarios, thereby improving the accuracy and interpretability of early warnings, reducing false alarms and false negatives, and enhancing the trust and operability of clinical and intervention systems in alarms.

[0005] The technical solution adopted in this invention is as follows: A psychological risk assessment method based on artificial intelligence, the method comprising the following steps: Step S1: Data collection, collecting multi-source, multimodal time-series data of assessors according to a unified timestamp; simultaneously extracting structured psychological and behavioral indicators from the raw signals at the window level; Step S2: Representation learning, first using interventional positive samples and Comparative pre-training is performed, and modal codes are fused by confidence-weighted modality gating and prototype memory. Then, parallel processing is used. A joint embedding is generated with a hybrid temporal model of differentiable state space, and finally mapped to interpretable structured features; Step S3: Risk assessment, obtaining both classification score and time-varying hazard rate from the temporal hidden state, and using... The model uncertainty is decomposed and the observation variance is output by the network to represent the data uncertainty, and a nonparametric confidence interval is constructed. The observation probability is generated by support vector machine decision and prototype prediction based on structured vectors. The three are input into a small Bayesian network and fused to obtain a window-level risk posterior and learn the fusion parameters. Step S4: Output the evaluation results. Combine the final posterior risk probability, normalized instantaneous hazard rate and uncertainty penalty into a composite risk score according to the weights, map it to the risk level and output it.

[0006] Further, in step S1, the data collection involves collecting multi-source data from historical assessors at a unified timestamp, including text data, physiological signal data, behavioral data, and whether psychological problems have occurred. The text data includes dialogue text and diary text. The physiological signal data includes heart rate, heart rate variability, and skin conductance. The behavioral data includes mobile phone screen time and social interaction frequency. Whether psychological problems have occurred is set as a label. Finally, each sample forms a unified multimodal sample and label, and a structured psychological and behavioral indicator vector is extracted from the original signal.

[0007] Further, in step S2, the representation learning specifically includes the following steps: Step S21: Compare with pre-training, using interventional positive samples and... Contrast training; specifically, for each original window, a modality is randomly selected for replacement intervention, and the original sample and the intervention sample pair are regarded as positive pairs, and other samples are regarded as negative pairs, thereby training the encoder family and the joint projection head; Step S22: Dynamic modal gating. A modal reliability gating network is introduced to output a confidence threshold for each modality. A differentiable prototype memory is introduced for rapid adaptation to obtain the fused joint embedding. ; Step S23: Construct a hybrid time series model and embed the fusion. Input to a hybrid temporal model: parallel self-attention and differentiable state-space modules, through learnable coefficients. Adaptive fusion output final temporal hidden state ; Step S24: Structured mapping, embedding multimodal modes jointly. Mapping to explanatory features yields a comprehensive structured vector. .

[0008] Furthermore, in step S3, the risk assessment specifically includes the following steps: Step S31: Joint output, the model simultaneously outputs instantaneous scores. and time-varying risk rate , means as follows: ; in, This represents the time-series latent variable in step S23. Represents the instantaneous raw score. and These represent the weight vector and bias of the classification head, respectively. Represents the actual label at the window level. Indicates the time-varying instantaneous hazard rate in the window The estimate; and These represent the survival head weight vector and the bias, respectively. For binary cross-entropy, This represents the discrete-time survival likelihood function; and This represents the task weight parameter. Represents an individual The interval indicator is 1 if an event occurs, and 0 otherwise. Step S32: Decompose uncertainty into three parts for prediction. The uncertainty is decomposed into three parts, and the model uncertainty is used. It is estimated that, given the uncertainty of the data, the network outputs an observation variance prediction. Sample uncertainty is supplemented by Bayesian posterior variance, which is then combined into the total variance. This combined estimate of model and data uncertainty is based on... Next forward, for : and Output from the model; , , Total variance: ;in, express The estimated number of forward passes is an integer. Indicates the first The original scores obtained by forward propagation of the sub-model. Indicates the first The variance of observation noise predicted by the sub-model. This represents the average of the model's output scores. Indicates the variance between models. express The mean variance of the observation noise in the forward pass of the sub-model; then, the fractional space variance is approximated by mapping it to the probability space, expressed as: ; This represents the point estimate probability. express The derivative of ; The approximate variance of the probability space is represented by... The method uses a first-order approximation to construct confidence intervals; Step S33: Generate confidence intervals, using a stratified order to ensure approximate frequency coverage of the confidence intervals in the production environment, and in the validation set. Calculate the residuals: ;make If the residuals are of the order of magnitude, then the confidence interval for the new sample is: ;in, Represents the set of validation set indices. For the size of the set, Indicates the true label, This indicates that the model performs well on the validation set. The predicted probability, Represents the absolute value of the residual, used in nonparametric classical quantiles. , Indicates the probability of not being covered. This indicates taking the first element after sorting by size. The value of the item; Step S34: Support Vector Machine (SVM) determination, using structured vectors. The training uses a support vector machine with radial basis function kernel to obtain the original decision function. Then use The method converts function values ​​into probabilities. ;in, This represents the decision probability of the support vector machine. This represents the original decision function of the support vector machine. and express Scaling factor, used by support vector machines Loss training, Cross-entropy fitting and ; Step S35: Construct a Bayesian network, establish a small Bayesian network, nodes: ;Target: ;Structural diagram: All are observed child nodes. These are potential risk nodes; among them, Estimated probability of corresponding points , Corresponding support vector machine decision probability , Corresponding to prototype-based fast prediction probability ; Calculate the posterior during reasoning The Bayesian form is: The conditional probability of each observation is given by a parameterized logistic conditional probability network, and is ultimately expressed as: ; in, Represents a probability operator. Indicates the name of the observation node in the Bayesian network. , and This represents the actual value of the corresponding observation node. Represents the posterior probability, in the case of observation. hour, The probability, Indicates proportionality. Indicates training set The prior probability, , , and Represents the conditional probability density; Indicates the fusion parameters; Step S36: Learning fusion parameters, fusing parameters Supervised learning is used to directly fit the model onto the training set, with the target being the window-level true label. The loss is cross-entropy.

[0009] Further, in step S4, the evaluation result output specifically includes the following steps: Step S41: Calculate the composite risk score and construct three normalized quantities: the final posterior risk probability, the normalized instantaneous hazard rate, and the uncertainty penalty; finally, calculate the composite risk score. Step S42: Grade mapping, mapping the composite risk score to 4 discrete grades; Step S43: Set conservative rules to adjust the mapping level under fixed conditions.

[0010] The present invention provides an artificial intelligence-based psychological risk assessment system, including a data acquisition module, a representation learning module, a risk assessment module, and an assessment result output module; The data acquisition module collects multi-source, multimodal time-series data of the evaluators according to a unified timestamp; at the same time, it extracts structured psychological and behavioral indicators from the raw signals at the window level and sends the data to the representation learning module. The representation learning module receives data sent by the data acquisition module, and first uses interventional positive samples and... Comparative pre-training is performed, and modal codes are fused by confidence-weighted modality gating and prototype memory. Then, parallel processing is used. A joint embedding is generated with a hybrid temporal model of differentiable state space, which is then mapped to interpretable structured features, and the data is sent to the risk assessment module. The risk assessment module receives data from the representation learning module and simultaneously obtains the classification score and time-varying hazard rate from the temporal hidden state. The model uncertainty is decomposed and the observation variance is output by the network to represent the data uncertainty, and a nonparametric confidence interval is constructed. The observation probability is generated by support vector machine decision and prototype prediction based on structured vectors. The three are input into a small Bayesian network and fused to obtain a window-level risk posterior and learn the fusion parameters. The data is then sent to the evaluation result output module. The assessment result output module receives data sent by the risk assessment module, and combines the final risk posterior probability, normalized instantaneous hazard rate and uncertainty penalty into a composite risk score according to weights, maps it to a risk level and outputs it.

[0011] The beneficial effects achieved by the present invention using the above solution are as follows: (1) To address the problems of existing multimodal representation learning methods, which typically employ simple concatenation or single-contrast pre-training, struggle to handle modal noise and missing data, have poor adaptability, and lack interpretability, this approach combines interventional positive samples. Compared to pre-trained learning robust modal encoders, this paper introduces dynamic modal gating to weight each modality according to its credibility and removes or weakens unreliable signals during fusion. Combined with differentiable prototype memory, it enables rapid adaptation to typical risk patterns. At the same time, it uses a hybrid temporal model of parallel self-attention and differentiable state space to capture local and global temporal dependencies and map out structured interpretable features. This significantly improves the robustness of the representation, temporal modeling ability, few-sample adaptability and interpretability, and improves the accuracy and stability of downstream risk prediction.

[0012] (2) To address the problems that existing psychological risk assessments often only provide a single probability or threshold alarm, lack unified quantification of time-dynamic risks and multi-source uncertainties, and are difficult to reliably integrate neural network results with rule-based or sample prototype-based criteria, this solution simultaneously outputs window-level instantaneous classification scores and time-varying instantaneous risk rates, and adopts... The study uses observation variance prediction and posterior variance tridecomposition to distinguish between model uncertainty, data uncertainty, and sample uncertainty, constructs confidence intervals to ensure frequency coverage, and introduces support vector machines and prototype-based fast prediction as complementary criteria in parallel. Then, it uses a small Bayesian network to structurally fuse this information and uses supervised learning to fit the fusion parameters. Finally, it provides calibrated posterior risk probabilities, confidence intervals, and time-sensitive hazard rates in psychological risk scenarios, thereby improving the accuracy and interpretability of early warnings, reducing false alarms and false negatives, and enhancing the trust and operability of clinical and intervention systems in alarms. Attached Figure Description

[0013] Figure 1 A schematic diagram illustrating an artificial intelligence-based psychological risk assessment method provided by the present invention; Figure 2 A schematic diagram of an artificial intelligence-based psychological risk assessment system provided by the present invention; Figure 3 This is a schematic diagram of step S2; Figure 4 This is a schematic diagram of step S3; Figure 5 This is a schematic diagram of step S4.

[0014] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0015] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0016] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0017] Example 1, see Figure 1 This invention provides an artificial intelligence-based psychological risk assessment method, which includes the following steps: Step S1: Data collection, collecting multi-source, multimodal time-series data of assessors according to a unified timestamp; simultaneously extracting structured psychological and behavioral indicators from the raw signals at the window level; Step S2: Representation learning, first using interventional positive samples and Comparative pre-training is performed, and modal codes are fused by confidence-weighted modality gating and prototype memory. Then, parallel processing is used. A joint embedding is generated with a hybrid temporal model of differentiable state space, and finally mapped to interpretable structured features; Step S3: Risk assessment, obtaining both classification score and time-varying hazard rate from the temporal hidden state, and using... The model uncertainty is decomposed and the observation variance is output by the network to represent the data uncertainty, and a nonparametric confidence interval is constructed. The observation probability is generated by support vector machine decision and prototype prediction based on structured vectors. The three are input into a small Bayesian network and fused to obtain a window-level risk posterior and learn the fusion parameters. Step S4: Output the evaluation results. Combine the final posterior risk probability, normalized instantaneous hazard rate and uncertainty penalty into a composite risk score according to the weights, map it to the risk level and output it.

[0018] Example 2, see Figure 1This embodiment is based on the above embodiment. In step S1, the data collection involves collecting multi-source data from historical assessors at a unified timestamp, including text data, physiological signal data, behavioral data, and whether psychological problems have occurred. The text data includes dialogue text and diary text. The physiological signal data includes heart rate, heart rate variability, and skin conductance. The behavioral data includes mobile phone screen time and social interaction frequency. Whether psychological problems have occurred is set as a label. Finally, each sample forms a unified multimodal sample and label, and a structured psychological and behavioral indicator vector is extracted from the original signal, as shown below: ; in, Let i represent the dataset containing N evaluator samples, where i represents the sample index. Indicates a label, Indicates the first A multimodal time-series observation set of samples, with a length of [number missing]. ; Indicates the first Each sample in time Multimodal observation, Represents text data, Represents physiological signal data, Represents behavioral data; This indicates the extraction of a structured psychological and behavioral indicator vector from the raw signal, including average heart rate, peak skin conductance, emotional vocabulary density, diary emotional score, screen usage change rate, and social interaction decline rate. Indicates that sample i is in the window The structured feature vectors, Indicates the first A set of moments within a time window.

[0019] Example 3, see Figure 1 and Figure 3 This embodiment is based on the above embodiment. In step S2, the representation learning specifically includes the following steps: Step S21: Compare with pre-training, using interventional positive samples and... Comparative training; specifically, for each original window Randomly select modalities for replacement intervention The original sample and the intervention sample are treated as positive pairs, and other samples are treated as negative pairs, thereby training the encoder family. With combined projector head , means as follows: ; in, This represents the overall comparison loss, which needs to be minimized; Indicates the time window index. Indicates the number of modes. This represents the modality index, indicating the two modalities being compared. Denotes the set of positive sample pairs; in this method This includes modal pairs between the original window and the window after intervention; Indicates the original window intermediate mode The original characteristics; This indicates an intervention procedure, generating an intervention sample: ;exist Includes The combination; Representing modes The encoder output dimensional vector; This indicates a common projection head, using two layers. Output a p-dimensional comparison space vector; The similarity function is represented by cosine similarity. Indicates temperature parameter; This represents the set of negative sample pair indices, containing combinations of non-positive modalities and windows. ; Represents the natural logarithm function. Represents the natural exponential function; intervention operation The specific implementation is as follows: ; This represents a modality randomly selected from the same community of neighbors whose historical behavior is closest to the target. Features, used as substitutes; Indicates the mixing coefficient; Indicates small Gaussian noise; Step S22: Dynamic modal gating, introducing a modal reliability gating network. A confidence threshold is output for each modality, and a differentiable prototype memory is introduced. This indicates typical high- and low-risk patterns and enables rapid adaptation; the resulting joint embedding... Defined as a weighted mosaic followed by projection, represented as follows: ; in, This indicates joint embedding after fusion. Representing modes In the window The reliability gating value is determined by the gating network. calculate; Indicates vector concatenation; For small Projection layer, output dimensional vector, Output dimension for each modal encoder. To fuse vector dimensions; the gated network is defined as: ; Representing modes Gated feature extractor for small Output dimensional vector; and These represent the gating networks for different modes. Learnable weight vectors and gated bias scalars; express Activation function; prototype memory mechanism for fast adaptation; for memory Calculate similarity and use it for fine-tuning with few samples: ; This represents the r-th prototype vector, including high-risk and low-risk prototypes; Represents the similarity score. Representing the prototype Associated risk levels; This represents a rapid prediction probability based on the prototype; Step S23: Construct a hybrid time series model and embed the fusion. Input to a hybrid temporal model: parallel self-attention and differentiable state-space modules, through learnable coefficients. The final temporal hidden state of the adaptive fusion output is represented as follows: ; in, express Layer output; Indicates residuals and layer normalization Layer functions take the previous state and the current embedding as input and output a vector of the same dimension. Represents a smooth dynamic state. This represents a differentiable state-space update function, specifically a linear state transition. Output dimensional vector; The dimension is The adaptive fusion gate vector; This represents the linear weight vector of the adaptive gate. This indicates adaptive gate bias. This indicates element-wise multiplication. Indicates the final time-series hidden state; Represents the state transition matrix. This represents the input-to-state mapping matrix; Step S24: Structured mapping, embedding multimodal modes jointly. Mapping to explanatory features yields a comprehensive structured vector. , means as follows: ; in, and This represents the parameters of a trainable linear layer. This represents an intermediate quantity in the depth mapping. A hand-structured vector of indicators representing the raw signal, including mean heart rate and peak skin conductance. This represents a comprehensive structured vector.

[0020] By performing the above operations, this scheme addresses the problems of existing multimodal representation learning methods, which typically employ simple concatenation or single-contrast pre-training, struggle to handle modal noise and missing data, have poor adaptability, and lack interpretability. It utilizes an interventional positive sample combination... Compared to pre-trained learning robust modal encoders, this paper introduces dynamic modal gating to weight each modality according to its credibility and removes or weakens unreliable signals during fusion. Combined with differentiable prototype memory, it enables rapid adaptation to typical risk patterns. At the same time, it uses a hybrid temporal model of parallel self-attention and differentiable state space to capture local and global temporal dependencies and map out structured interpretable features. This significantly improves the robustness of the representation, temporal modeling ability, few-sample adaptability and interpretability, and improves the accuracy and stability of downstream risk prediction.

[0021] Example 4, see Figure 1 and Figure 4 This embodiment is based on the above embodiment. In step S3, the risk assessment specifically includes the following steps: Step S31: Joint output, the model simultaneously outputs instantaneous scores. and time-varying risk rate , means as follows: ; in, This represents the time-series latent variable in step S23. Represents the instantaneous raw score. and These represent the weight vector and bias of the classification head, respectively. Represents the actual label at the window level. Indicates the time-varying instantaneous hazard rate in the window The estimate; and These represent the survival head weight vector and the bias, respectively. For binary cross-entropy, This represents the discrete-time survival likelihood function; and This represents the task weight parameter. Represents an individual The interval indicator is 1 if an event occurs, and 0 otherwise. Step S32: Decompose uncertainty into three parts for prediction. The uncertainty is decomposed into three parts, and the model uncertainty is used. It is estimated that, given the uncertainty of the data, the network outputs an observation variance prediction. Sample uncertainty is supplemented by Bayesian posterior variance, which is then combined into the total variance. This combined estimate of model and data uncertainty is based on... Next forward, for : and Output from the model; , , Total variance: ;in, express The estimated number of forward passes is an integer. Indicates the first The original scores obtained by forward propagation of the sub-model. Indicates the first The variance of observation noise predicted by the sub-model. This represents the average of the model's output scores. Indicates the variance between models. express The mean variance of the observation noise in the forward pass of the sub-model; then, the fractional space variance is approximated by mapping it to the probability space, expressed as: ; This represents the point estimate probability. express The derivative of ; The approximate variance of the probability space is represented by... The method uses a first-order approximation to construct confidence intervals; Step S33: Generate confidence intervals, using a stratified order to ensure approximate frequency coverage of the confidence intervals in the production environment, and in the validation set. Calculate the residuals: ;make If the residuals are of the order of magnitude, then the confidence interval for the new sample is: ;in, Represents the set of validation set indices. For the size of the set, Indicates the true label, This indicates that the model performs well on the validation set. The predicted probability, Represents the absolute value of the residual, used in nonparametric classical quantiles. , Indicates the probability of not being covered. This indicates taking the first element after sorting by size. The value of the item; Step S34: Support Vector Machine (SVM) determination, using structured vectors. The training uses a support vector machine with radial basis function kernel to obtain the original decision function. Then use The method converts function values ​​into probabilities. ;in, This represents the decision probability of the support vector machine. This represents the original decision function of the support vector machine. and express Scaling factor, used by support vector machines Loss training, Cross-entropy fitting and ; Step S35: Construct a Bayesian network, establish a small Bayesian network, nodes: ;Target: ;Structural diagram: All are observed child nodes. These are potential risk nodes; among them, Estimated probability of corresponding points , Corresponding support vector machine decision probability , Corresponding to prototype-based fast prediction probability ; Calculate the posterior during reasoning The Bayesian form is: The conditional probability of each observation is given by a parameterized logistic conditional probability network, and is ultimately expressed as: ; in, Represents a probability operator. Indicates the name of the observation node in the Bayesian network. , and This represents the actual value of the corresponding observation node. Represents the posterior probability, in the case of observation. hour, The probability, Indicates proportionality. Indicates training set The prior probability, , , and Represents the conditional probability density; ; Represents the learnable fusion parameters; Step S36: Learning fusion parameters, fusing parameters Supervised learning is used to directly fit the model onto the training set, with the target being the window-level true label. The loss is the cross-entropy, expressed as follows: ; in, This represents the loss from learning the fusion parameters, which needs to be minimized.

[0022] By performing the above operations, this solution addresses the problems of existing psychological risk assessments, which often only provide single probability or threshold alarms, lack unified quantification of time-dynamic risks and multi-source uncertainties, and struggle to reliably integrate neural network results with rule-based or sample prototype-based criteria. This solution simultaneously outputs window-level instantaneous classification scores and time-varying instantaneous hazard rates, and employs... The study uses observation variance prediction and posterior variance tridecomposition to distinguish between model uncertainty, data uncertainty, and sample uncertainty, constructs confidence intervals to ensure frequency coverage, and introduces support vector machines and prototype-based fast prediction as complementary criteria in parallel. Then, it uses a small Bayesian network to structurally fuse this information and uses supervised learning to fit the fusion parameters. Finally, it provides calibrated posterior risk probabilities, confidence intervals, and time-sensitive hazard rates in psychological risk scenarios, thereby improving the accuracy and interpretability of early warnings, reducing false alarms and false negatives, and enhancing the trust and operability of clinical and intervention systems in alarms.

[0023] Example 5, see Figure 1 and Figure 5 This embodiment is based on the above embodiment. In step S4, the evaluation result is output, which specifically includes the following steps: Step S41: Calculate the composite risk score and construct three normalized quantities: the final posterior risk probability. Normalized instantaneous risk rate: Uncertainty penalty: ;in, This is a constant for the window length, defaulting to 1; weights are introduced. , , Finally, the composite risk score is calculated. ; Step S42: Level mapping, will Mapped to 4 discrete levels Low risk: Medium risk: 0.2 High risk: Extremely high risk: ; Step S43: Set a conservative rule, add the conservative rule: if the upper bound of the confidence interval... satisfy: If the risk level is not high enough, the risk level will be directly increased by one level; the final output sample will be the risk level mapped by the composite risk score as the assessment result.

[0024] Example 6, see Figure 2 Based on the above embodiments, this embodiment provides an artificial intelligence-based psychological risk assessment system, including a data acquisition module, a representation learning module, a risk assessment module, and an assessment result output module. The data acquisition module collects multi-source, multimodal time-series data of the evaluators according to a unified timestamp; at the same time, it extracts structured psychological and behavioral indicators from the raw signals at the window level and sends the data to the representation learning module. The representation learning module receives data sent by the data acquisition module, and first uses interventional positive samples and... Comparative pre-training is performed, and modal codes are fused by confidence-weighted modality gating and prototype memory. Then, parallel processing is used. A joint embedding is generated with a hybrid temporal model of differentiable state space, which is then mapped to interpretable structured features, and the data is sent to the risk assessment module. The risk assessment module receives data from the representation learning module and simultaneously obtains the classification score and time-varying hazard rate from the temporal hidden state. The model uncertainty is decomposed and the observation variance is output by the network to represent the data uncertainty, and a nonparametric confidence interval is constructed. The observation probability is generated by support vector machine decision and prototype prediction based on structured vectors. The three are input into a small Bayesian network and fused to obtain a window-level risk posterior and learn the fusion parameters. The data is then sent to the evaluation result output module. The assessment result output module receives data sent by the risk assessment module, and combines the final risk posterior probability, normalized instantaneous hazard rate and uncertainty penalty into a composite risk score according to weights, maps it to a risk level and outputs it.

[0025] It should be noted that, in this document, the terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0026] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0027] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A psychological risk assessment method based on artificial intelligence, characterized in that: The method includes the following steps: Step S1: Data collection, collecting multi-source, multimodal time-series data of assessors according to a unified timestamp; simultaneously extracting structured psychological and behavioral indicators from the raw signals at the window level; Step S2: Representation learning, first using interventional positive samples and Comparative pre-training is performed, and modal codes are fused by confidence-weighted modality gating and prototype memory. Then, parallel processing is used. A joint embedding is generated with a hybrid temporal model of differentiable state space, and finally mapped to interpretable structured features; Step S3: Risk assessment, obtaining both classification score and time-varying hazard rate from the temporal hidden state, and using... The model uncertainty is decomposed and the observation variance is output by the network to represent the data uncertainty, and a nonparametric confidence interval is constructed. The observation probability is generated by support vector machine decision and prototype prediction based on structured vectors. The three are input into a small Bayesian network and fused to obtain a window-level risk posterior and learn the fusion parameters. Step S4: Output the evaluation results. Combine the final posterior risk probability, normalized instantaneous hazard rate and uncertainty penalty into a composite risk score according to the weights, map it to the risk level and output it.

2. The psychological risk assessment method based on artificial intelligence according to claim 1, characterized in that: In step S2, the representation learning specifically includes the following steps: Step S21: Compare with pre-training, using interventional positive samples and... Contrast training; specifically, for each original window, a modality is randomly selected for replacement intervention, and the original sample and the intervention sample pair are regarded as positive pairs, and other samples are regarded as negative pairs, thereby training the encoder family and the joint projection head; Step S22: Dynamic modal gating. A modal reliability gating network is introduced to output a confidence threshold for each modality. A differentiable prototype memory is introduced for rapid adaptation to obtain the fused joint embedding. ; Step S23: Construct a hybrid time series model and embed the fusion. Input to a hybrid temporal model: parallel self-attention and differentiable state-space modules, through learnable coefficients. Adaptive fusion output final temporal hidden state ; Step S24: Structured mapping, embedding multimodal modes jointly. Mapping to explanatory features yields a comprehensive structured vector. .

3. The psychological risk assessment method based on artificial intelligence according to claim 1, characterized in that: In step S3, the risk assessment specifically includes the following steps: Step S31: Joint output, the model simultaneously outputs instantaneous scores. and time-varying risk rate , means as follows: ； in, This represents the time-series latent variable in step S23. Represents the instantaneous raw score. and These represent the weight vector and bias of the classification head, respectively. Represents the actual label at the window level. Indicates the time-varying instantaneous hazard rate in the window The estimate; and These represent the survival head weight vector and the bias, respectively. For binary cross-entropy, This represents the discrete-time survival likelihood function; and This represents the task weight parameter. Represents an individual The interval indicator is 1 if an event occurs, and 0 otherwise. Step S32: Decompose uncertainty into three parts for prediction. The uncertainty is decomposed into three parts, and the model uncertainty is used. It is estimated that, given the uncertainty of the data, the network outputs an observation variance prediction. Sample uncertainty is supplemented by Bayesian posterior variance, which is then combined into the total variance. This combined estimate of model and data uncertainty is based on... Next forward, for : and Output from the model; , , Total variance: ;in, express The estimated number of forward passes is an integer. Indicates the first The original scores obtained by forward propagation of the sub-model. Indicates the first The variance of observation noise predicted by the sub-model. This represents the average of the model's output scores. Indicates the variance between models. express The mean variance of the observation noise in the forward pass of the sub-model; then, the fractional space variance is approximated by mapping it to the probability space, expressed as: ; This represents the point estimate probability. express The derivative, ; The approximate variance of the probability space is represented by... The method uses a first-order approximation to construct confidence intervals; Step S33: Generate confidence intervals, using a stratified order to ensure approximate frequency coverage of the confidence intervals in the production environment, and in the validation set. Calculate the residuals: ;make If the residuals are of the order of magnitude, then the confidence interval for the new sample is: ;in, Represents the set of validation set indices. For the size of the set, Indicates the true label, This indicates that the model performs well on the validation set. The predicted probability, Represents the absolute value of the residual, used in nonparametric classical quantiles. , Indicates the probability of not being covered. This indicates taking the first element after sorting by size. The value of the item; Step S34: Support Vector Machine (SVM) determination, using structured vectors. The training uses a support vector machine with radial basis function kernel to obtain the original decision function. Then use The method converts function values ​​into probabilities. ;in, This represents the decision probability of the support vector machine. This represents the original decision function of the support vector machine. and express Scaling factor, used by support vector machines Training on losses, Cross-entropy fitting and ; Step S35: Construct a Bayesian network, establish a small Bayesian network, nodes: ;Target: ;Structural diagram: All are observed child nodes. These are potential risk nodes; among them, Estimated probability of corresponding points , Corresponding support vector machine decision probability , Corresponding to prototype-based fast prediction probability ; Calculate the posterior during reasoning The Bayesian form is: The conditional probability of each observation is given by a parameterized logistic conditional probability network, and is ultimately expressed as: ； in, Represents a probability operator. Indicates the name of the observation node in the Bayesian network. , and This represents the actual value of the corresponding observation node. Represents the posterior probability, in the case of observation. hour, The probability, Indicates proportionality. Indicates training set The prior probability, , , and Represents the conditional probability density; Indicates the fusion parameters; Step S36: Learning fused parameters, fusing parameters Supervised learning is used to directly fit the model onto the training set, with the target being the window-level true label. The loss is cross-entropy.

4. The psychological risk assessment method based on artificial intelligence according to claim 1, characterized in that: In step S4, the evaluation result is output, specifically including the following steps: Step S41: Calculate the composite risk score and construct three normalized quantities: the final posterior risk probability, the normalized instantaneous hazard rate, and the uncertainty penalty; finally, calculate the composite risk score. Step S42: Grade mapping, mapping the composite risk score to 4 discrete grades; Step S43: Set conservative rules to adjust the mapping level under fixed conditions.

5. The psychological risk assessment method based on artificial intelligence according to claim 1, characterized in that: In step S1, the data collection involves collecting multi-source data from historical assessors at a unified timestamp, including text data, physiological signal data, behavioral data, and whether psychological problems have occurred. The text data includes dialogue text and diary text. The physiological signal data includes heart rate, heart rate variability, and skin conductance. The behavioral data includes mobile phone screen time and social interaction frequency. Whether psychological problems have occurred is set as a label. Finally, each sample forms a unified multimodal sample and label, and a structured psychological and behavioral indicator vector is extracted from the original signal.

6. An artificial intelligence-based psychological risk assessment system, used to implement the artificial intelligence-based psychological risk assessment method as described in any one of claims 1-5, characterized in that: It includes a data acquisition module, a representation learning module, a risk assessment module, and an assessment result output module.

7. The artificial intelligence-based psychological risk assessment system according to claim 6, characterized in that: The data acquisition module collects multi-source, multimodal time-series data of the evaluators according to a unified timestamp; at the same time, it extracts structured psychological and behavioral indicators from the raw signals at the window level and sends the data to the representation learning module. The representation learning module receives data sent by the data acquisition module, and first uses interventional positive samples and... Comparative pre-training is performed, and modal codes are fused by confidence-weighted modality gating and prototype memory. Then, parallel processing is used. A joint embedding is generated with a hybrid temporal model of differentiable state space, which is then mapped to interpretable structured features, and the data is sent to the risk assessment module. The risk assessment module receives data from the representation learning module and simultaneously obtains the classification score and time-varying hazard rate from the temporal hidden state. Decompose the model uncertainty and use the network output observation variance to represent the data uncertainty, and construct nonparametric confidence intervals; The observation probabilities are generated by support vector machine decision-making and prototype prediction based on structured vectors, respectively. The three are then input into a small Bayesian network for fusion to obtain a window-level risk posterior and learn the fusion parameters. The data is then sent to the evaluation result output module. The assessment result output module receives data sent by the risk assessment module, and combines the final risk posterior probability, normalized instantaneous hazard rate and uncertainty penalty into a composite risk score according to weights, maps it to a risk level and outputs it.

Images