Suicide risk prediction method and system based on facial micro-expression recognition

Abstract

The invention relates to the technical field of crisis prediction, and discloses a suicide risk prediction method and system based on facial micro-expression recognition, and the method comprises the steps: collecting micro-expression sequences of MDD and BD patients and suicide risk crowds, carrying out the normalization and cutting of the micro-expression sequences, and carrying out the data dimension reduction processing through employing KPCA, and obtaining a processed micro-expression sequence; constructing a multi-label graph model based on a fuzzy measurement function, and performing fuzzy processing on the processed micro-expression sequence to obtain a facial feature set; a sparrow search algorithm is adopted to screen out features related to suicide risk prediction from the facial feature set, and a target feature set is obtained; constructing a suicide risk identification model based on fusion of a graph convolutional neural network and a long short-term memory network, inputting the target feature set into the suicide risk identification model, and outputting an identification result by the suicide risk identification model through space-time sequence feature fusion; according to the invention, effective prediction of the suicide risk is realized, early diagnosis of mental diseases is assisted, and the recognition rate and prediction precision are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of crisis prediction, and in particular to a suicide risk prediction method and system based on facial micro-expression recognition. BACKGROUND

[0002] Major depressive disorder (MDD) and bipolar disorder (BD) are common mental illnesses with high morbidity, high disability rate, high suicide risk, and frequent recurrence. The incomplete recovery during the inter-episode period causes continuous psychological and social function damage. MDD is considered one of the top ten major causes of disability worldwide and is associated with an increased risk of suicidal behavior. The suicide rate of MDD patients is 9%, and up to 15% of BD patients have suicidal behavior. The suicide risk of BD patients is 15 times that of the normal population, which brings a serious burden to families and society. Currently, the diagnosis of MDD and BD is mainly through inquiry. Observing the facial expressions, postures, voices, and tones of patients during communication is an important part of the inquiry process. People communicate a lot of psychological information through postures, facial expressions, and non-verbal sounds (such as sighs, groans, and cries) during communication. The inquiry diagnosis method has the problems of strong subjectivity, insufficient feature extraction, and weak model generalization ability, resulting in a low suicide risk recognition rate. SUMMARY

[0003] The present application relates to the technical field of crisis prediction, and in particular to a suicide risk prediction method and system based on facial micro-expression recognition.

[0004] The present application relates to the technical field of crisis prediction, and in particular to a suicide risk prediction method and system based on facial micro-expression recognition. Collecting micro-expression sequences of MDD, BD patients and suicide risk population, and performing data dimension reduction processing on the normalized and cropped micro-expression sequences using KPCA to obtain processed micro-expression sequences; Based on the fuzzy measurement function, a multi-label graph model is constructed to fuzz the processed micro-expression sequences, the correlation between features is quantified by the fuzzy measurement function, and multi-dimensional and multi-scale micro-expression features are extracted to obtain a facial feature set; The sparrow search algorithm is used to select features related to suicide risk prediction from the facial feature set to obtain a target feature set; Based on the fusion of graph convolutional neural network and long short-term memory network, a suicide risk identification model is constructed, the target feature set is input into the suicide risk identification model, the suicide risk identification model performs spatial-temporal feature fusion, and outputs an identification result.

[0005] Optionally, in the first implementation manner of the first aspect, the micro-expression sequence of the MDD, BD patient and suicide risk population is collected, normalized and cropped, and then KPCA is used for data dimension reduction processing to obtain the processed micro-expression sequence, including: The micro-expression video of the MDD, BD patient and suicide risk population is collected, and the micro-expression video is converted into a first image data set by using an OpenCV library; Based on the feature point coordinates, a face ROI region is extracted from the first image data set to obtain an ROI image; The histogram equalization is used to enhance the contrast of the ROI image, and the Z-score normalization is used to process the pixel value to obtain a second image data set; For the second image data set in each window, the Farneback algorithm is used to calculate the optical flow field between adjacent frames to extract motion information as a micro-expression dynamic feature; The second image data set of each frame is converted into a one-dimensional vector, and the micro-expression dynamic feature is combined to form a first feature vector; The KPCA is used for dimension reduction processing on the first feature vector to obtain the processed micro-expression sequence.

[0006] Optionally, in the second implementation manner of the first aspect, the micro-expression sequence of the MDD, BD patient and suicide risk population is collected, normalized and cropped, and then KPCA is used for data dimension reduction processing to obtain the processed micro-expression sequence, including: The micro-expression video of the MDD, BD patient and suicide risk population is collected, and the micro-expression video is converted into a first image data set by using an OpenCV library; Based on the feature point coordinates, a face ROI region is extracted from the first image data set to obtain an ROI image; The histogram equalization is used to enhance the contrast of the ROI image, and the Z-score normalization is used to process the pixel value to obtain a second image data set; For the second image data set in each window, the Farneback algorithm is used to calculate the optical flow field between adjacent frames to extract motion information as a micro-expression dynamic feature; The second image data set of each frame is converted into a one-dimensional vector, and the micro-expression dynamic feature is combined to form a first feature vector; The KPCA is used for dimension reduction processing on the first feature vector to obtain the processed micro-expression sequence.

[0007] Optionally, in a third implementation form of the first aspect of the present application, the multi-label graph model is constructed based on the fuzzy measurement function, the processed micro-expression sequence is fuzzed, the correlation between features is quantified by the fuzzy measurement function, multi-dimensional and multi-scale micro-expression features are extracted, and a facial feature set is obtained, including: The micro-expression feature sequence after KPCA dimension reduction is divided into a plurality of third feature vectors according to time steps, a node set is constructed, and an initial label set is assigned to each node in the node set; For the third feature vector of each node, the fuzzy membership degree of each label in the initial label set is calculated, and a fuzzy membership degree matrix is formed; A fuzzy similarity matrix between nodes is constructed, a weighted undirected graph is constructed based on the fuzzy similarity matrix, a fuzzy relationship type is assigned to each edge of the weighted undirected graph, and a fuzzy relationship graph is obtained; The constructed fuzzy relationship graph is processed by multi-scale graph convolution, and the feature tensors obtained by different scale graph convolutions are spliced in the channel dimension to form a multi-scale feature representation; An attention mechanism is applied, weights are assigned to each scale feature according to the output of the fuzzy measurement function, and a fused spatial feature sequence is obtained, wherein the fuzzy measurement function at least includes fuzzy entropy, fuzzy similarity and fuzzy correlation; The time convolution operation is applied to the fused spatial feature sequence, the dilated convolution is used to capture dynamic information of different time scales, and the feature tensor obtained by the time convolution is obtained through the gating mechanism; The feature tensor output by the time convolution and the fuzzy membership degree matrix are combined to construct a final facial feature set.

[0008] Optionally, in a fourth implementation form of the first aspect of the present application, the sparrow search algorithm is used to filter features related to suicide risk prediction from the facial feature set to obtain a target feature set, including: The facial feature set is mapped to a binary feature selection space, each feature in the facial feature set corresponds to a binary variable, and an initial population is generated by using chaotic mapping; The adaptive weight mechanism is used to dynamically assign sparrow roles, update the positions of producers, followers and sentinels, calculate the fitness values of all updated sparrow individuals, and reorder the population according to the fitness values; The elite pool is used to save elite solutions of all generations, the elite solutions are operated by chaotic mapping to generate new solutions, and the new solutions are crossed with the current population; Quantum behavior characteristics are introduced in the position update, each sparrow position is represented as a probability distribution of quantum state, the probability amplitude is updated by quantum rotation gate operation, when the iteration error is lower than the threshold 1x10 -4 and the difference between the probability amplitudes of quantum states is less than 5x10 -6 , the quantum state collapses into a certain solution; When the maximum number of iterations is reached, the algorithm terminates, and outputs the feature subset corresponding to the global optimal solution as the target feature set.

[0009] Optionally, in a fifth implementation form of the first aspect of the present application, the adaptive weight mechanism is used to dynamically assign roles to sparrows, update the positions of producers, followers and sentinels, calculate the fitness values of all updated sparrow individuals, and reorder the population according to the fitness values, including: The producer probability of each sparrow is calculated according to the individual fitness value, the probability is normalized by a Softmax function, a sentinel threshold is set, and individuals below the sentinel threshold are automatically converted to sentinels, and the rest are followers; An adaptive step length control is introduced to update the producer position using a double guide mechanism; The followers are divided into exploratory followers and developmental followers, the exploratory followers update the position by random disturbance using a differential evolution strategy, and the developmental followers update the position in the direction of the fastest fitness increase using a gradient guide strategy; The average mutual information of each feature is calculated as an importance index, and the sentinel updates the position using an adaptive disturbance strategy based on feature importance.

[0010] Optionally, in a sixth implementation form of the first aspect of the present application, the adaptive weight mechanism is used to dynamically assign roles to sparrows, update the positions of producers, followers and sentinels, calculate the fitness values of all updated sparrow individuals, and reorder the population according to the fitness values, including: The producer probability of each sparrow is calculated according to the individual fitness value, the probability is normalized by a Softmax function, a sentinel threshold is set, and individuals below the sentinel threshold are automatically converted to sentinels, and the rest are followers; An adaptive step length control is introduced to update the producer position using a double guide mechanism; The followers are divided into exploratory followers and developmental followers, the exploratory followers update the position by random disturbance using a differential evolution strategy, and the developmental followers update the position in the direction of the fastest fitness increase using a gradient guide strategy; The average mutual information of each feature is calculated as an importance index, and the sentinel updates the position using an adaptive disturbance strategy based on feature importance.

[0011] Optionally, in a seventh implementation form of the first aspect of the present application, the first GCN layer in the suicide risk identification model captures local feature interaction, the remaining GCN layers aggregate neighborhood information through a message passing mechanism, a nonlinear activation function is applied to each GCN layer to introduce a nonlinear transformation, and the feature map output by the GCN layer is reorganized in the time dimension to form the input of the bidirectional LSTM layer.

[0012] Optionally, in an eighth implementation form of the first aspect of the present application, the forward LSTM in the suicide risk identification model captures the time sequence dependence from the past to the future, and the backward LSTM captures the dependence from the future to the past, and a residual connection is added after each LSTM layer.

[0013] The second aspect of the present application provides a suicide risk prediction system based on facial micro-expression recognition, which comprises: A data preprocessing module is configured to collect micro-expression sequences of MDD, BD patients and suicide risk groups, perform normalization and cropping on the micro-expression sequences, and perform data dimension reduction processing on the micro-expression sequences by using KPCA to obtain processed micro-expression sequences. A feature extraction module is configured to construct a multi-label graph model based on a fuzzy measurement function, perform fuzzy processing on the processed micro-expression sequences, quantify the correlation between features by using the fuzzy measurement function, extract multi-dimensional and multi-scale micro-expression features, and obtain a facial feature set. A feature selection module is configured to filter out features related to suicide risk prediction from the facial feature set by using a sparrow search algorithm to obtain a target feature set. A risk identification module is configured to construct a suicide risk identification model based on the fusion of a graph convolutional neural network and a long short-term memory network, input the target feature set into the suicide risk identification model, and output an identification result by the suicide risk identification model through spatial-temporal feature fusion.

[0014] In the technical solution of the present application, the micro-expression sequences of MDD, BD patients and suicide risk groups are collected, normalized and cropped, and then KPCA is used for data dimension reduction processing to obtain processed micro-expression sequences. A multi-label graph model is constructed based on a fuzzy measurement function, the processed micro-expression sequences are fuzzy processed, the correlation between features is quantified by using the fuzzy measurement function, multi-dimensional and multi-scale micro-expression features are extracted, and a facial feature set is obtained. The sparrow search algorithm is used to filter out features related to suicide risk prediction from the facial feature set to obtain a target feature set. The suicide risk identification model is constructed based on the fusion of a graph convolutional neural network and a long short-term memory network, the target feature set is input into the suicide risk identification model, and the suicide risk identification model outputs an identification result through spatial-temporal feature fusion. The present application realizes effective prediction of suicide risk, assists early diagnosis of mental illness, improves the recognition rate and prediction accuracy, provides an objective and automated prediction tool for mental illness diagnosis, assists doctors in early identification of suicide risk, reduces the rate of missed diagnosis, and promotes the development of precision medicine. BRIEF DESCRIPTION OF DRAWINGS

[0015] Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments. The drawings are for purposes of illustration only and are not considered a limitation of the present application.

[0016] Figure 1 The first embodiment schematic diagram of the suicide risk prediction method based on facial micro-expression recognition provided by the embodiment of the present application is shown in the figure. Figure 2 The second embodiment schematic diagram of the suicide risk prediction method based on facial micro-expression recognition provided by the embodiment of the present application is shown in the figure. Figure 3 The third embodiment schematic diagram of the suicide risk prediction method based on facial micro-expression recognition provided by the embodiment of the present application is shown in the figure. Figure 4 The structure schematic diagram of the suicide risk prediction system based on facial micro-expression recognition provided by the embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0017] The terms "first", "second", "third", "fourth" and the like in the description, claims, and drawings of the present application, if any, are used for distinguishing between similar objects and not necessarily for describing a particular sequential or chronological order. It is to be understood that the use of such terms, if any, are not intended to limit potential applicability of embodiments to a potentially differently ordered or sequenced embodiment to the embodiment depicted or described herein. Further, the term "comprise" or "comprising" or their variations, where appropriate, are intended to cover a process, method, device, product, or apparatus that comprises a list of elements but not other elements not expressly listed. Thus, the term "comprising" or "comprising" does not exclude the presence of additional elements or steps.

[0018] For the sake of understanding, the specific flow of the embodiment of the present application is described below, please refer to Figure 1 The first embodiment schematic diagram of the suicide risk prediction method based on facial micro-expression recognition provided by the embodiment of the present application is shown in the figure, which specifically includes the following steps: Step 101, collect micro-expression sequences of MDD, BD patients and suicide risk population, after normalization and cropping, use KPCA for data dimension reduction processing to obtain processed micro-expression sequences; Step 102, construct a multi-label graph model based on a fuzzy measurement function, fuzzy process the processed micro-expression sequences, quantify the correlation between features through the fuzzy measurement function, extract multi-dimensional and multi-scale micro-expression features, and obtain a facial feature set; Step 103, use the sparrow search algorithm to filter out features related to suicide risk prediction from the facial feature set to obtain a target feature set; In this embodiment, the face feature set is mapped to a binary feature selection space, each feature in the face feature set corresponds to a binary variable, and an initial population is generated using chaotic mapping; based on an adaptive weight mechanism, sparrow roles are dynamically assigned, the positions of producers, followers and sentinels are updated, the fitness values of all updated sparrow individuals are calculated, and the population is reordered according to the fitness values; an elite pool is used to save elite solutions of all generations, and new solutions are generated by performing chaotic mapping on the elite solutions and crossing with the current population; in the position updating, quantum behavior characteristics are introduced, each sparrow position is represented as a probability distribution of a quantum state, the probability amplitude is updated by quantum rotation gate operation, when the iteration error is less than a threshold value 1x10 -4 and the difference between the probability amplitudes of the quantum states is less than 5x10 -6 , the quantum state collapses into a determined solution; when the maximum number of iterations is reached, the algorithm terminates, and the feature subset corresponding to the global optimal solution is output as the target feature set.

[0019] In this embodiment, the producer probability of each sparrow is calculated according to the individual fitness value, the probability is normalized by a Softmax function, a sentinel threshold is set, individuals below the sentinel threshold are automatically converted to sentinels, and the rest are followers; an adaptive step length control is introduced to update the producer position using a double guide mechanism; the followers are divided into exploratory followers and developmental followers, the exploratory followers update their positions by random disturbance using a differential evolution strategy, and the developmental followers update their positions in the direction of the fastest fitness increase using a gradient guide strategy; the average mutual information of each feature is calculated as an importance index, and the sentinel updates its position using an adaptive disturbance strategy based on feature importance.

[0020] In this embodiment, statistical tests are performed on the feature subsets obtained by feature extraction, linear and nonlinear feature selection, and feature selection to verify the significant differences of micro-expression features in different groups. The target feature subset obtained by feature selection is divided into four groups according to the subject group label: MDD patient group, BD patient group, suicide risk group and normal control group. Normality test and homogeneity of variance test are performed on the data of each group. For non-normal distributed data, Box-Cox transformation or rank transformation is performed to make it close to normal distribution. Abnormal values are processed by Winsorization, such as replacing extreme values with values at a specified percentile. One-way ANOVA is performed on each feature dimension to compare the mean differences between the four groups. If the ANOVA shows significant differences, such as p<0.05, further tests are performed to determine which groups have differences. For data that do not meet the prerequisites for ANOVA, Dunn's test is used for multiple comparisons. Record the F value and p value of each feature, and screen out significant features with p < 0.05; construct a multivariate analysis of variance model, considering the interaction effects between multiple features; test the joint effect of group type as the independent variable on multiple feature dependent variables, such as MDD / BD / suicide risk / normal; calculate the statistical quantity to evaluate the overall effect significance; if there is a significant interaction effect, perform subsequent profile analysis to decompose the difference in the change pattern of different features among groups; for nonlinear features, use permutation-based nonparametric tests; perform random permutation on each sample to construct a zero distribution, and calculate the p value of the actual observation value in the zero distribution; use the bootstrap method to estimate the effect size to quantify the practical significance of the difference between groups; for high-dimensional nonlinear features, use distance-based test methods to consider the overall distribution difference in the feature space; For the subset of features selected, use random forest importance scores or support vector machine weight coefficients to determine the most discriminative feature combination; construct a logistic regression model with group type as the dependent variable and the feature subset as the independent variable, and test the overall significance of the model; calculate the area under the AUC-ROC curve to evaluate the model classification performance, and ensure the stability of the results through cross-validation; perform Wald test on the model coefficients to determine the independent contribution of each feature to classification; Calculate the effect size of significant features to distinguish statistical significance from practical significance; classify the effect size, such as small: d = 0.2, medium: d = 0.5, large: d = 0.8, and select features with medium or large effect sizes; construct an ROC curve to calculate the Youden index to determine the optimal classification threshold and evaluate the diagnostic accuracy of the feature; combine clinical data to calculate the positive predictive value, negative predictive value, sensitivity, and specificity of the feature to quantify its clinical application value; To address the multiple comparison problem caused by testing a large number of features, use the Benjamini-Hochberg method to control the false discovery rate; sort the original p values of all features and adjust the critical value according to the BH algorithm; retain features with adjusted p values < 0.05 as the final significant features; compare the number of significant features and the change in effect size before and after correction to evaluate the impact of multiple comparison correction; Draw a box plot to show the distribution difference of significant features among the four groups, and label the median, quartiles, and outliers; construct a heat map with features as rows and groups as columns, with color intensity representing feature values, and clustering showing feature expression patterns; use a volcano plot to show the relationship between effect size and p value of features, highlighting features with high effect size and high significance; generate a structured report containing a list of significant features, effect size, corrected p value, and clinical interpretation, providing a basis for subsequent model interpretation and clinical application.

[0021] In step 104, a suicide risk identification model is constructed based on the fusion of a graph convolutional neural network and a long short-term memory network. The target feature set is input into the suicide risk identification model. The suicide risk identification model outputs an identification result through space-time feature fusion.

[0022] In this embodiment, the target feature set is converted into graph structure data. Each feature in the target feature set is taken as a node in the graph. The weight of the edge between the nodes is calculated based on the mutual information between the features to construct a feature similarity matrix. The adjacent nodes of each node in the feature similarity matrix are determined through a K-nearest neighbor algorithm. The node representation is updated by aggregating neighborhood information to convert the high-dimensional sparse feature vector into a low-dimensional dense embedding vector. The embedding vectors are arranged in time sequence and input into the suicide risk identification model. The GCN layer of the suicide risk identification model captures the spatial dependency between the features through a message passing mechanism to obtain spatial features. The bidirectional LSTM layer of the suicide risk identification model processes time sequence information. The forward and backward propagation capture the time sequence dependency in different directions to obtain time sequence features. A multi-head attention mechanism is applied between the GCN layer and the bidirectional LSTM layer to dynamically adjust the weights of the spatial features and the time sequence features and perform feature fusion to obtain fused space-time features. The fused space-time features are input into a fully connected classifier. The Softmax function is used to generate a probability distribution of different suicide risk levels to output the identification result.

[0023] In this embodiment, a multi-layer GCN architecture is designed to process the feature graph. The first layer of GCN captures local feature interaction. The subsequent layers aggregate neighborhood information through a message passing mechanism. A nonlinear activation function is applied to each layer of GCN to introduce nonlinear transformation. Dropout technology is used to prevent overfitting. During the training process, part of the node connections are randomly ignored. Graph pooling operation is used to reduce the complexity of the graph and retain the most discriminative nodes. The feature graph output by GCN is reorganized in the time dimension. For the feature graph at each time step, the node embedding vector is extracted and flattened into a fixed-length feature vector. The feature vectors of consecutive time steps are arranged in sequence to form a time sequence input suitable for LSTM processing. According to the characteristics of micro-expression sequences, a suitable time window length, such as 15-30 frames, is set to ensure the capture of complete expression change cycles.

[0024] In this embodiment, a bidirectional LSTM network is constructed to process time sequence features. The forward LSTM captures the time sequence dependency from the past to the future. The backward LSTM captures the dependency relationship from the future to the past. Multiple LSTM layers are set to form a deep time sequence model. Each layer contains multiple hidden units. Batch normalization technology is applied to accelerate training and improve stability. Residual connections are added after each LSTM layer to enhance gradient flow.

[0025] In this embodiment, a multi-head attention mechanism is introduced between the GCN and the LSTM, the attention weights between the spatial features output by the GCN and the time sequence features output by the LSTM are calculated, the correlation information of different subspaces is captured through the multi-head attention, the attention weights are applied to feature fusion, the contribution degrees of the spatial features and the time sequence features are dynamically adjusted, a self-attention mechanism is designed to process the output of the LSTM, and attention to key time sequence patterns is enhanced.

[0026] In this embodiment, GCN-LSTM branches of different scales are constructed in parallel, multiple GCN layers are set, different convolution kernel sizes are used to capture local and global feature interactions, each GCN branch is connected to an independent LSTM network to process time sequence information, and feature pyramid network architecture is used to fuse feature representations of different scales, so that bidirectional transmission of high-level semantic features and low-level detail features is performed.

[0027] In this embodiment, a multi-classifier architecture is designed to predict the suicide risk level, a fully connected layer is used to map the fused features to a low-dimensional decision space, output nodes corresponding to the number of risk levels are set, a Softmax activation function is used to generate a probability distribution of each risk level, a focal loss function is introduced to handle the class imbalance problem, and higher weights are given to difficult-to-classify samples.

[0028] Referring to Figure 2 The second embodiment of the suicide risk prediction method based on facial micro-expression recognition provided by the embodiment of the present application is schematically shown, and the method specifically includes the following steps: Step 201, collect micro-expression videos of MDD, BD patients and suicide risk groups, and convert the micro-expression videos into a first image data set using an OpenCV library; Step 202, extract a facial ROI region from the first image data set based on feature point coordinates to obtain an ROI image; Step 203, enhance the contrast of the ROI image by using histogram equalization, and normalize the pixel value by using Z-score normalization to obtain a second image data set; Step 204, calculate the optical flow field between adjacent frames in each window by using a Farneback algorithm to extract motion information as micro-expression dynamic features; Step 205, convert each frame of the second image data set into a one-dimensional vector, and combine the micro-expression dynamic features to form a first feature vector; Step 206, perform dimensionality reduction processing on the first feature vector by using KPCA to obtain a processed micro-expression sequence.

[0029] In this embodiment, Gaussian radial basis function is used as the kernel function of KPCA, and a kernel matrix is constructed according to the first eigenvector; the kernel matrix is first centralized, then the centralized kernel matrix is subjected to eigenvalue decomposition, the eigenvalue and the second eigenvector are calculated, and the second eigenvector is arranged in descending order of the eigenvalue; the second eigenvector corresponding to the first k largest eigenvalues is selected to form a projection matrix, and the projection of the first eigenvector in the projection matrix is calculated to obtain the processed micro-expression sequence, wherein k = 1, 2,..., k.

[0030] Referring to Figure 3 The third embodiment of the suicide risk prediction method based on facial micro-expression recognition provided by the embodiment of the present application is schematically shown, and the method specifically includes the following steps: Step 301, the micro-expression feature sequence after KPCA dimension reduction is divided into a plurality of third eigenvectors according to time steps, a node set is constructed, and an initial label set is assigned to each node in the node set; In this embodiment, the micro-expression feature sequence is modeled by a graph theory method, a node set containing time step eigenvectors is constructed, and a weighted undirected graph structure is established by using a fuzzy similarity matrix. This graph theory analysis can effectively capture the complex correlation between features, thereby improving the accuracy of facial micro-expression recognition.

[0031] Step 302, for the third eigenvector of each node, the fuzzy membership degree of each label in the initial label set is calculated to form a fuzzy membership degree matrix; Step 303, a fuzzy similarity matrix between nodes is constructed, a weighted undirected graph is constructed based on the fuzzy similarity matrix, a fuzzy relationship type is assigned to each edge of the weighted undirected graph, and a fuzzy relationship graph is obtained; In this embodiment, this step belongs to the central execution network.

[0032] Step 304, the constructed fuzzy relationship graph is subjected to multi-scale graph convolution processing, and the feature tensors obtained by different scale graph convolutions are spliced in the channel dimension to form a multi-scale feature representation; In this embodiment, the scale division of the multi-scale feature is quantitatively described based on the interlayer scaling coefficient of the Gaussian pyramid, such as 0.707 times down-sampling per layer.

[0033] Step 305, an attention mechanism is applied, weights are assigned to each scale feature according to the output of the fuzzy measurement function, and a fused spatial feature sequence is obtained, wherein the fuzzy measurement function at least includes fuzzy entropy, fuzzy similarity and fuzzy correlation degree; In this embodiment, the fuzzy measurement function is a Gaussian fuzzy function.

[0034] Step 306, applying a time convolution operation to the fused spatial feature sequence, using dilated convolution to capture dynamic information of different time scales, and obtaining a feature tensor through a gating mechanism; Step 307, combining the feature tensor output by the time convolution and the fuzzy membership matrix to construct the final facial feature set; In this embodiment, each node is assigned an initial label set, including emotion category labels and micro-expression type labels. The emotion category labels are, for example, sadness, fear, disgust, and remorse, and the micro-expression type labels are, for example, furrowed eyebrows, drooping mouth corners, and contracted eyelids. A Gaussian membership function is used, with the feature center of each label category as the mean value and the intra-class variance as the standard deviation. The Mahalanobis distance between the feature vector and each category center is calculated to convert it to a membership value in the [0, 1] interval, forming a fuzzy membership matrix; A fuzzy similarity matrix between nodes is constructed. First, the cosine similarity between feature vectors is calculated, then the time proximity weight is combined, and finally the label membership similarity is integrated through fuzzy logic operations to obtain a comprehensive similarity value; A weighted undirected graph is constructed based on the similarity matrix. The edge set represents the association between nodes, and each edge is assigned a fuzzy relationship type, such as temporal precedence, causality, and co-occurrence. The relationship type is determined by the joint distribution of node label membership, for example, both nodes have high membership in the "sadness" and "furrowed eyebrows" labels, and are assigned a strong causal relationship weight; A multi-scale graph convolution operation is performed on the constructed fuzzy relationship graph. Different neighborhood radii are defined, and for each node, the subgraph structure within its neighborhood is extracted. Graph convolution operations are applied to each subgraph to aggregate the feature information of the neighborhood nodes. The convolution kernel parameters are dynamically adjusted based on the node label membership, for example, for nodes with high membership in the "sadness" label, the convolution weight for extracting eyebrow region features is increased; Fuzzy entropy is used to quantify the uncertainty of node features, and the calculation formula is based on the extension of Shannon entropy. Fuzzy similarity measures the similarity between features of different scales using weighted cosine similarity. Fuzzy correlation assesses the relevance of features to the suicide risk label by calculating conditional fuzzy entropy. Features with low fuzzy entropy and high fuzzy correlation receive higher weights. Finally, feature fusion is achieved through weighted summation; Dilated convolution is used to capture dynamic information of different time scales, such as the occurrence, development, and regression process of micro-expressions. Different dilation rates are set to capture both short-term rapid changes and long-term slow changes in micro-expression features. The gating mechanism selectively retains key patterns in the time series information; The feature tensor output by time convolution is flattened into a feature vector, and a final facial feature set is constructed by combining the label membership matrix of the original node. The feature set includes multi-scale spatial features, time series dynamic features, fuzzy label membership features, and statistical features calculated based on the fuzzy measurement function. The feature set is normalized to ensure dimensional consistency between feature dimensions. The correlation-based feature screening method is applied to calculate the mutual information of each feature with the suicide risk label. Features with mutual information values higher than a threshold are retained. The screened feature set is then subjected to PCA dimension reduction to compress the dimension to a size suitable for subsequent classification model processing while retaining variance information. Finally, the facial feature set for suicide risk prediction is output.

[0035] In this embodiment, the collection of facial micro-expressions is combined with the family history of the subject: offspring of parents who have committed suicide are more likely to die by suicide (RR = 2.97) and attempt suicide (RR = 1.76), a prospective study with a follow-up of 5.6 years showed that offspring of suicide attempters are 5 times more likely to attempt suicide than the general population; highlighting the importance of genetic factors in suicide risk, although there is no direct suicide gene, certain genetic traits such as impulse control disorders, emotional regulation problems or mental illnesses can be passed down through generations, thereby indirectly increasing the risk of suicide; for example, if one or both parents have a pattern of behavior that is difficult to control impulses or have diseases that affect emotional stability, their children may be more likely to inherit similar biochemical imbalances or brain structural differences, making them less able to cope with life stressors or more emotionally volatile, which increases their likelihood of experiencing psychological distress and, in extreme cases, suicidal ideation and behavior.

[0036] In this embodiment, 1. In the psychiatric outpatient screening scenario, after the patient enters the waiting area, the system automatically collects 1-2 minutes of micro-expression video through the camera, without the need for active cooperation to complete data collection, reducing patient resistance, and the collected micro-expression sequence is normalized and key emotional features are extracted using KPCA dimension reduction. The correlation between expressions is quantified using a multi-label graph model combined with a fuzzy measurement function to further explore potential emotional fluctuation patterns. Subsequently, the Sparrow Search Algorithm selects features highly correlated with suicide risk, inputs the suicide risk identification model to perform joint analysis of spatial-temporal features, and outputs the risk level (low / medium / high). This scheme can quickly generate a preliminary risk report before consultation, providing objective evidence for doctors and significantly improving screening efficiency.

[0037] 2. In psychiatric wards or psychological crisis intervention wards, patients' micro-expression videos are collected regularly (e.g., daily), such as in the morning or after lunch break. By tracking the dynamic changes in emotional state through long-term time-series data, and based on KPCA dimensionality reduction and multi-label graph models, subtle differences in patients' micro-expressions can be captured, such as the persistent enhancement of high-risk emotions like anger and disgust. The target feature set is dynamically updated through a sparrow search algorithm. The suicide risk identification model combines spatial expression features with temporal evolution patterns, which can not only identify the current risk level but also predict short-term emotional fluctuation trends, such as the critical point of transition from low risk to high risk. In this scenario, doctors can develop personalized intervention plans, such as adjusting medication dosages or increasing the frequency of psychological counseling. Real-time early warning functions can be used to focus on monitoring high-risk patients, reducing the probability of suicide attempts.

[0038] 3. In remote screening scenarios, patients upload micro-expression videos recorded at home through the hospital's internet medical platform, such as those taken with a mobile phone's front-facing camera. The system automatically performs data preprocessing and feature analysis. Based on KPCA dimensionality reduction and fuzzy measurement functions, it can effectively eliminate interference factors in the home environment, such as changes in lighting. The system also uses a sparrow search algorithm to focus on micro-expression features strongly correlated with suicide risk, such as brief expressions of disgust or persistent low-valence emotions. The suicide risk identification model analyzes the videos in real time in the cloud, outputs the risk level, and synchronizes it to the doctor's end. At the same time, it triggers a differentiated response mechanism based on the risk level. For example, low-risk patients only need regular follow-ups, while high-risk patients need to be given priority for online consultations or offline intervention. This solution breaks through the dependence of traditional screening on physical medical institutions and is especially suitable for patients in remote areas or with limited mobility. It enables early detection and early intervention of mental health problems and reduces the pressure of uneven distribution of social medical resources.

[0039] Please see Figure 4 A schematic diagram of the structure of a suicide risk prediction system based on facial micro-expression recognition provided in this embodiment of the invention. The system includes: The data preprocessing module is used to collect micro-expression sequences from MDD, BD patients and people at risk of suicide. After normalization and cropping, KPCA is used to perform data dimensionality reduction to obtain the processed micro-expression sequences. The feature extraction module is used to construct a multi-label graph model based on a fuzzy measurement function, fuzzify the processed micro-expression sequence, quantify the correlation between features through the fuzzy measurement function, extract multi-dimensional and multi-scale micro-expression features, and obtain a facial feature set. The feature selection module is used to filter features related to suicide risk prediction from the facial feature set using the sparrow search algorithm to obtain the target feature set. A risk identification module is configured to construct a suicide risk identification model based on a graph convolutional neural network and a long short-term memory network, input a target feature set into the suicide risk identification model, and output an identification result through space-time feature fusion.

[0040] The above shows and describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the present application is not limited to the above-mentioned embodiments, and the above-mentioned embodiments and descriptions in the specification are only preferred examples of the present application and are not intended to limit the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for predicting suicide risk based on facial micro-expression recognition, characterized in that, The method comprises the following steps: Collecting micro-expression sequences of MDD, BD patients and suicide risk groups, performing normalization and cropping, and then performing data dimension reduction processing on the micro-expression sequences by KPCA to obtain processed micro-expression sequences; Based on a fuzzy measurement function, a multi-label graph model is constructed to fuzz the processed micro-expression sequences, the correlation between features is quantified by the fuzzy measurement function, multi-dimensional and multi-scale micro-expression features are extracted, and a facial feature set is obtained; A sparrow search algorithm is used to select features related to suicide risk prediction from the facial feature set to obtain a target feature set; A suicide risk identification model is constructed based on the fusion of graph convolutional neural network and long short-term memory network, the target feature set is input into the suicide risk identification model, the suicide risk identification model performs spatial-temporal feature fusion, and an identification result is output. 2.The suicide risk prediction method based on facial micro-expression recognition of claim 1, wherein, The micro-expression sequences of MDD, BD patients and suicide risk groups are collected, normalized and cropped, and then KPCA is used for data dimension reduction processing to obtain processed micro-expression sequences, which comprises: Collecting micro-expression videos of MDD, BD patients and suicide risk groups, and converting the micro-expression videos into a first image data set using an OpenCV library; Based on feature point coordinates, a facial ROI region is extracted from the first image data set to obtain an ROI image; The contrast of the ROI image is enhanced by histogram equalization, and the pixel value is normalized by Z-score to obtain a second image data set; For each window of the second image data set, the Farneback algorithm is used to calculate the optical flow field between adjacent frames to extract motion information as micro-expression dynamic features; Each frame of the second image data set is converted into a one-dimensional vector, and the micro-expression dynamic features are combined to form a first feature vector; KPCA is used for dimension reduction processing on the first feature vector to obtain processed micro-expression sequences. 3.The suicide risk prediction method based on facial micro-expression recognition of claim 2, wherein, The KPCA is used for dimension reduction processing on the first feature vector to obtain processed micro-expression sequences, which comprises: A Gaussian radial basis function is used as the kernel function of KPCA, and a kernel matrix is constructed according to the first feature vector; The kernel matrix is first centered, and then the centered kernel matrix is feature-decomposed to calculate the eigenvalues and a second feature vector, and the second feature vector is arranged in descending order of eigenvalues; The second feature vectors corresponding to the first k largest eigenvalues are selected to form a projection matrix, the projection of the first feature vector in the projection matrix is calculated, and the processed micro-expression sequences are obtained, wherein k=1, 2,..., k. 4.The suicide risk prediction method based on facial micro-expression recognition of claim 1, wherein, The multi-label graph model is constructed based on the fuzzy measurement function, the processed micro-expression sequences are fuzzed, the correlation between features is quantified by the fuzzy measurement function, multi-dimensional and multi-scale micro-expression features are extracted, and a facial feature set is obtained, which comprises: The KPCA dimension-reduced micro-expression feature sequences are divided into multiple third feature vectors according to time steps, a node set is constructed, and each node in the node set is assigned an initial label set; For the third feature vector of each node, the fuzzy membership degree of each label in the initial label set is calculated to form a fuzzy membership degree matrix; constructing a fuzzy similarity matrix between nodes, constructing a weighted undirected graph based on the fuzzy similarity matrix, assigning a fuzzy relationship type to each edge of the weighted undirected graph, and obtaining a fuzzy relationship graph; performing multi-scale graph convolution processing on the constructed fuzzy relationship graph, concatenating feature tensors obtained by different scale graph convolutions in the channel dimension to form a multi-scale feature representation; applying an attention mechanism to assign weights to each scale feature according to the output of a fuzzy measurement function, wherein the fuzzy measurement function at least includes fuzzy entropy, fuzzy similarity, and fuzzy correlation degree, and obtaining a fused spatial feature sequence; applying a time convolution operation to the fused spatial feature sequence, capturing dynamic information of different time scales using dilated convolution, and obtaining a feature tensor obtained by time convolution through a gating mechanism; combining the feature tensor output by the time convolution and the fuzzy membership matrix to construct a final facial feature set. 5.The suicide risk prediction method based on facial micro-expression recognition of claim 1, wherein, The sparrow search algorithm is used to select features related to suicide risk prediction from the facial feature set to obtain a target feature set, including: mapping the facial feature set to a binary feature selection space, each feature in the facial feature set corresponding to a binary variable, and generating an initial population using chaotic mapping; dynamically assigning sparrow roles based on an adaptive weight mechanism, updating the positions of producers, followers, and sentinels, calculating the fitness values of all updated sparrow individuals, and reordering the population according to the fitness values; saving the elite solutions of the previous generations in an elite pool, performing chaotic mapping on the elite solutions to generate new solutions, and crossing with the current population; In the position update, the quantum behavior characteristics are introduced, each sparrow position is expressed as a probability distribution of quantum state, and the probability amplitude is updated through quantum rotation gate operation. When the iteration error is lower than the threshold 1×10 -4 and the difference of quantum state probability amplitude is less than 5×10 -6 , the quantum state collapses into a certain solution. when the maximum number of iterations is reached, the algorithm terminates, and the feature subset corresponding to the global optimal solution is output as the target feature set. 6.The suicide risk prediction method based on facial micro-expression recognition of claim 5, wherein, The adaptive weight mechanism is used to dynamically assign sparrow roles, update the positions of producers, followers, and sentinels, calculate the fitness values of all updated sparrow individuals, and reorder the population according to the fitness values, including: calculating the producer probability of each sparrow according to the individual fitness value, normalizing the probability through a Softmax function, setting a sentinel threshold, and automatically converting individuals below the sentinel threshold to sentinels, and the rest to followers; introducing an adaptive step length control to update the producer position using a double guidance mechanism; dividing the followers into exploratory followers and developmental followers, updating the position of the exploratory followers through random disturbance using a differential evolution strategy, and updating the position of the developmental followers along the direction of the fastest fitness increase using a gradient guidance strategy; calculating the average mutual information of each feature as an importance indicator, and updating the position of the sentinels using an adaptive disturbance strategy based on feature importance.

7. The method of claim 1, wherein the method is based on facial micro-expression recognition. The graph convolutional neural network and long short-term memory network are fused to construct a suicide risk identification model, the target feature set is input into the suicide risk identification model, the suicide risk identification model performs spatial-temporal feature fusion, and outputs an identification result, including: converting the target feature set to graph structure data, each feature in the target feature set serving as a node in the graph, calculating the weight of the edge between the nodes based on the mutual information between the features, and constructing a feature similarity matrix; The K-neighbor algorithm is used to determine the adjacent nodes of each node in the feature similarity matrix, and the node representation is updated by aggregating the neighborhood information, so as to convert the high-dimensional sparse feature vector into a low-dimensional dense embedding vector; The embedding vectors are arranged in time sequence and input into the suicide risk identification model, the GCN layer of the suicide risk identification model captures the spatial dependency between features through the message passing mechanism, obtains spatial features, the bidirectional LSTM layer of the suicide risk identification model processes the time sequence information, and captures the time sequence dependency in different directions through forward and backward propagation, obtains time sequence features, and applies a multi-head attention mechanism between the GCN layer and the bidirectional LSTM layer, dynamically adjusts the weights of the spatial features and the time sequence features, and performs feature fusion to obtain fused spatio-temporal features. The fused spatio-temporal features are input into the full connection classifier, the probability distribution of different suicide risk levels is generated through the Softmax function, and the identification result is output. 8.The suicide risk prediction method based on facial micro-expression recognition of claim 7, wherein, The first GCN layer in the suicide risk identification model captures local feature interaction, and the remaining GCN layers aggregate neighborhood information through the message passing mechanism, and a nonlinear activation function is applied to each GCN layer to introduce nonlinear transformation, the feature map output by the GCN layer is reorganized in the time dimension to form the input of the bidirectional LSTM layer. 9.The suicide risk prediction method based on facial micro-expression recognition of claim 7, wherein, The forward LSTM in the suicide risk identification model captures the time sequence dependency from the past to the future, and the backward LSTM captures the dependency from the future to the past, and a residual connection is added after each LSTM layer. 10.A suicide risk prediction system based on facial micro-expression recognition, characterized in that, The system comprises: A data preprocessing module is configured to collect micro-expression sequences of MDD, BD patients and suicide risk groups, normalize and crop the micro-expression sequences, and perform dimension reduction processing on the micro-expression sequences by using KPCA to obtain processed micro-expression sequences. A feature extraction module is configured to construct a multi-label graph model based on a fuzzy measurement function, perform fuzzy processing on the processed micro-expression sequences, quantify the correlation between features by using the fuzzy measurement function, extract multi-dimensional and multi-scale micro-expression features, and obtain a facial feature set. A feature selection module is configured to select features related to suicide risk prediction from the facial feature set by using a sparrow search algorithm to obtain a target feature set. A risk identification module is configured to construct a suicide risk identification model based on a graph convolutional neural network and a long short-term memory network, input the target feature set into the suicide risk identification model, and output an identification result by the suicide risk identification model through spatial-temporal feature fusion.

Images