A method, apparatus, device, and medium for depression relapse risk intervention

Abstract

The application discloses a depression relapse risk intervention method, device, equipment and medium, and belongs to the medical technical field. The method comprises the following steps: integrating a gene risk score, a neural imaging brain region feature and clinical history data through a multi-modal fusion neural network to generate an individualized baseline risk score; dynamically correcting the baseline score based on self-evaluation data to enhance the real-time performance of risk assessment; analyzing the time sequence characteristics of physiological and behavioral data by using an LSTM time sequence model to output a short-term relapse early warning label; and combining real-time voice / text data emotion knowledge graph analysis and dynamic risk grading to trigger a hierarchical intervention strategy. The application solves the problems of single evaluation dimension and response lag of traditional methods through a multi-dimensional data fusion and dynamic calibration mechanism, realizes closed-loop management from risk warning to precise intervention on the premise of avoiding direct clinical diagnosis, and improves the comprehensiveness and timeliness of depression relapse prevention and control.

Description

Technical Field

[0001] This invention belongs to the field of medical technology, specifically relating to a method, device, equipment, and medium for intervening in the risk of relapse of depression. Background Technology

[0002] Depression, as a mental illness with a high relapse rate, presents a significant challenge in clinical practice in predicting and intervening in relapse risk. Current research indicates that approximately 50% of patients with depression will experience at least one relapse after their first episode, and the risk of subsequent relapses increases significantly with the number of relapses. Traditional relapse risk assessments primarily rely on the experience and judgment of clinicians and standardized scales. However, these methods suffer from high subjectivity and limited assessment dimensions, failing to comprehensively reflect the patient's biological characteristics and dynamic behavioral changes, thus hindering timely and accurate intervention in relapse situations. Summary of the Invention

[0003] The purpose of this invention is to provide a method, device, equipment, and medium for intervening in the risk of relapse of depression, so as to solve the problem that the existing technology is unable to fully reflect the biological characteristics and dynamic behavioral changes of patients.

[0004] To achieve the above objectives, the present invention adopts the following technical solution:

[0005] In a first aspect, the present invention provides a method for intervening in the risk of relapse of depression, comprising:

[0006] The system acquires the user's genetic risk score data, clinical history data, and neuroimaging data, and inputs them into a pre-trained personalized baseline risk model, which outputs the patient's baseline risk score.

[0007] Obtain the user's self-assessment data, determine the correction coefficient based on the self-assessment data, and correct the baseline risk score based on the correction coefficient to obtain the corrected risk score;

[0008] Acquire users' physiological and behavioral data, input the corrected risk score, physiological and behavioral data into a pre-trained short-term early warning model, and the short-term early warning model outputs a short-term early warning label;

[0009] The system acquires users' voice data and chat text data in real time, and inputs the short-term warning tags, voice data, and chat text data into a pre-trained real-time intervention model, which then generates a real-time intervention strategy.

[0010] Furthermore, the individualized baseline risk model is trained in the following manner:

[0011] Acquire basic data of patients with depression; wherein, the basic data includes genetic data, neuroimaging data, and clinical history data;

[0012] Features are extracted from the basic data, and a training set is constructed based on the extracted features; wherein, gene risk scores are calculated for gene data; brain region structural features are extracted from neuroimaging data, including the gray matter volume of the nucleus accumbens and the cortical thickness of the ventral tegmental area; and medical history features are extracted from clinical medical history data.

[0013] The architecture of the individualized baseline risk model is determined to be a multimodal fusion neural network. This network comprises an input layer, a fusion layer, and an output layer. The input layer receives multimodal features, including genetic risk scores, brain region structural features, and medical history features. The fusion layer combines these features using tensor concatenation, dynamically calculates the weights of each modality using a self-attention module, and generates hidden layer vectors through dimensionality reduction using a fully connected layer. The output layer uses a single-node fully connected layer with a sigmoid activation function to output the baseline risk score based on the hidden layer vectors.

[0014] The model is trained based on the training set and the multimodal fusion neural network, and an individualized baseline risk model is obtained after training.

[0015] Furthermore, the self-assessment data includes symptom indicators, immediate status, behavioral data, and derived characteristics;

[0016] Symptom indicators include the severity of depressive symptoms, sleep quality index, and suicidal ideation scale;

[0017] Real-time status includes real-time emotion score and attention score;

[0018] Behavioral data includes social time and moderate-to-vigorous intensity exercise time.

[0019] The derived features include dynamic change features and risk-related features; among them, dynamic change features are used to calculate short-term volatility indicators, and risk-related features are used to measure the increase in suicidal ideation.

[0020] Furthermore, the correction coefficient is generated through a gradient boosting tree regression model. The input features of the gradient boosting tree regression model include symptom indicators, immediate status, behavioral data, and derived features. The output is a correction coefficient λ∈[0.9,1.1]. The corrected risk score is calculated as the product of the baseline risk score and the correction coefficient λ.

[0021] Furthermore, the training method for the short-term early warning model is as follows:

[0022] Acquire historical physiological and behavioral data for model training; historical physiological data includes user heart rate variability, skin conductance, and diurnal temperature variation; historical behavioral data includes duration of social activities and exercise intensity.

[0023] The architecture of the short-term early warning model was determined to be an LSTM time series prediction model;

[0024] The model is trained using a modified risk score, historical physiological data, historical behavioral data, and an LSTM time series prediction model. After training, a short-term early warning model is obtained.

[0025] Furthermore, the method for generating immediate intervention strategies using the immediate intervention model is as follows:

[0026] The system acquires real-time voice data and real-time text data from users, extracts voice keywords from the real-time voice data, extracts text keywords from the real-time text data, and generates an original keyword list based on the voice keywords and text keywords. The original keyword list is then input into a depression emotion knowledge graph for implicit emotion inference to obtain implicit emotion labels.

[0027] Acoustic features are extracted from real-time speech data, and the acoustic features and implicit sentiment labels are input into a sentiment classification model for processing to obtain the user's real-time sentiment category; time-series features are extracted based on the user's historical sentiment categories; real-time sentiment categories and time-series features are input into a dynamic risk monitoring model for processing to obtain the user's risk level label.

[0028] Short-term warning labels, risk level labels, and implicit sentiment labels are input into the sentiment intensity quantification model for processing to obtain an immediate intervention intensity rating.

[0029] Based on risk level labels, implicit sentiment labels, and immediate intervention intensity classifications, immediate intervention strategies are generated.

[0030] Furthermore, the intensity grading of immediate intervention is obtained as follows:

[0031] Based on the risk level labels output by the dynamic risk monitoring model, risks are divided into three levels: low risk, medium risk, and high risk; among which, the weight of low risk is 0.2, the weight of medium risk is 0.5, and the weight of high risk is 0.8.

[0032] The emotional intensity values ​​are divided into three levels based on the implicit emotional tags of the depression emotional knowledge graph: mild, moderate, and severe. The intensity coefficient for mild is 1, for moderate it is 1.5, and for severe it is 2.

[0033] If the short-term warning label is high risk, the warning correction factor is 20%.

[0034] Intervention intensity value = (weight × 0.6 + intensity coefficient × 0.4) × (1 + early warning correction factor)

[0035] The calculated intervention intensity values ​​are then mapped to different levels to obtain an instantaneous intervention intensity level.

[0036] In a second aspect, the present invention provides a device for intervening in the risk of relapse of depression, comprising:

[0037] The first assessment module is used to acquire the user's genetic risk score data, clinical history data, and neuroimaging data, and input them into a pre-trained individualized baseline risk model, which outputs the patient's baseline risk score.

[0038] The second assessment module is used to acquire the user's self-assessment data, determine the correction coefficient based on the self-assessment data, and correct the baseline risk score based on the correction coefficient to obtain the corrected risk score.

[0039] The third assessment module is used to acquire the user's physiological and behavioral data, and input the corrected risk score, physiological data and behavioral data into a pre-trained short-term early warning model. The short-term early warning model outputs a short-term early warning label.

[0040] The fourth evaluation module is used to acquire users' voice data and chat text data in real time, and input the short-term warning tags, voice data and chat text data into a pre-trained real-time intervention model, which generates a real-time intervention strategy.

[0041] In a third aspect, the present invention provides an electronic device including a processor and a memory, the processor being configured to execute a computer program stored in the memory to implement the method described above.

[0042] In a fourth aspect, the present invention provides a computer-readable storage medium storing at least one instruction which, when executed by a processor, implements the method described above.

[0043] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0044] This approach integrates genetic risk scores, neuroimaging brain region structural features, and clinical history data, and uses a multimodal fusion neural network to dynamically allocate the weights of each modality. This overcomes the limitations of traditional methods that rely on single-dimensional data, enabling synergistic analysis of biological characteristics and clinical information, and providing a more comprehensive baseline for individualized risk assessment.

[0045] This solution is based on real-time feedback from self-assessment data and dynamically calibrates the baseline risk score through a gradient boosting tree model. It can quickly respond to short-term fluctuations in patient behavior or symptoms and improve the timeliness of risk prediction.

[0046] This approach uses an LSTM time series model to analyze the temporal changes in continuous physiological and behavioral data, capture potential patterns before depression relapses, and identify early warning signals.

[0047] This solution uses a depression affective knowledge graph to analyze the implicit emotions in real-time voice / text, and combines dynamic risk levels with quantitative intervention intensity grading to automatically trigger matching intervention measures, ensuring rapid response to high-risk events without direct clinical diagnosis.

[0048] This solution forms a fully automated processing flow from multi-source data collection and dynamic risk assessment to tiered intervention implementation, reducing reliance on manual assessment and improving management efficiency in primary healthcare settings. At the same time, it protects user privacy through anonymized data processing and encrypted transmission. Attached Figure Description

[0049] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0050] Figure 1 This is a flowchart of a method for intervening in the risk of relapse of depression according to an embodiment of the present invention;

[0051] Figure 2 This is a structural block diagram of a depression relapse risk intervention device according to an embodiment of the present invention;

[0052] Figure 3 This is a structural block diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0053] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0054] The following detailed description is exemplary and intended to provide further detailed explanation of the invention. Unless otherwise specified, all technical terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. The terminology used in this invention is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention.

[0055] Example 1

[0056] like Figure 1 As shown, a method for intervening in the risk of relapse of depression includes:

[0057] S10. Obtain the user's genetic risk score data, clinical history data, and neuroimaging data, and input them into a pre-trained individualized baseline risk model, which outputs the patient's baseline risk score.

[0058] In one embodiment, the personalized baseline risk model integrates the Polygenic Risks Score (PRS), neuroimaging features, and clinical history data to output a quantitative score (range [0,1]) that reflects an individual's biological baseline risk, providing a basis for subsequent risk correction, short-term early warning, and other functions.

[0059] In one embodiment, the individualized baseline risk model is trained as follows:

[0060] S101. Obtain three types of basic data from a large number of patients with depression through clinical databases and construct a training set; the three types of basic data include genetic data, neuroimaging data, and clinical history data.

[0061] Specifically, the gene risk score (PRS) is calculated based on gene data, integrating multi-gene locus information related to depression to quantify the risk tendency at the genetic level; brain region structural features such as gray matter volume of the nucleus accumbens and cortical thickness of the ventral tegmental area are extracted based on neuroimaging data to reflect the biological basis of the brain's reward circuit and emotion regulation center; and clinical history data is processed into numerical (such as the number of episodes and medication dosage) and categorical (such as the age of onset and comorbidities) features, covering information on disease progression and treatment history.

[0062] As an example, a sample of 8000 patients was used, including 3000 in the relapse group and 5000 in the non-relapse group. Relapse was defined as the occurrence of a depressive episode meeting the DSM-5 criteria within 6 months of the last episode. Stratified sampling was used in a 7:2:1 ratio to divide the sample into a training set of 5600 cases, a validation set of 1600 cases, and a test set of 800 cases. Stratified sampling ensured a balanced relapse rate across the groups and avoided sample bias affecting the model's generalization ability.

[0063] It should be noted that the patient sample data can be set according to actual needs, such as 10,000 cases, 15,000 cases, etc.

[0064] S102. The architecture of the individualized baseline risk model is determined to be a multimodal fusion neural network.

[0065] Specifically, the multimodal fusion neural network uses a three-layer architecture to integrate genetic, imaging, and clinical data, including:

[0066] The input layer receives PRS (genetic modality), brain region structural features (imaging modality), and medical history features (clinical modality) respectively, and retains the original information of each modality.

[0067] The fusion layer is used to: directly merge three-modal data along the feature axis to form a unified feature vector; introduce a self-attention module to dynamically calculate the weights of each modality through a trainable weight matrix, such as the correlation importance of gene, image, and clinical features, and capture cross-modal interaction relationships; and compress the fusion feature dimension through a fully connected layer to output a hidden layer vector.

[0068] The output layer uses a single-node fully connected layer with a Sigmoid activation function to output a baseline risk score in the range [0,1], which directly maps to the individual's baseline relapse probability.

[0069] It should be noted that the self-attention module involved in this solution is an existing technology in the field of artificial intelligence.

[0070] S103. Based on the training set and the multimodal fusion neural network, the model is trained to obtain an individualized baseline risk model.

[0071] Specifically, before model training, Xavier uniform distribution is used to initialize weights, so that the activation values ​​of each layer are stable and gradient vanishing or exploding is avoided.

[0072] During training, the AUC-ROC metric of the validation set is continuously monitored. If there is no significant improvement for 10 consecutive rounds (the standard for no significant improvement is an improvement of <0.01), training is automatically terminated to prevent overfitting.

[0073] After the model is trained, its performance is tested based on the test set.

[0074] S20. Obtain the user's self-assessment data, determine the correction coefficient based on the self-assessment data, and correct the baseline risk score based on the correction coefficient to obtain the corrected risk score.

[0075] Specifically, self-assessment indicators are determined, and specific data of these indicators are obtained through daily self-assessment by patients or automatic data collection by devices. These self-assessment indicators include symptom indicators, current status, and behavioral data.

[0076] As an example, symptom indicators include the PHQ-9 total score (severity of depressive symptoms), PSQI (sleep quality index), and BSI (suicidal ideation scale).

[0077] As an example, immediate status includes: immediate emotion score (1-10 points) and attention score (reflecting cognitive function).

[0078] As an example, behavioral data includes: social time (including face-to-face social interactions) and moderate-to-high intensity exercise time (recorded via mobile phone accelerometer).

[0079] Specifically, derived characteristics are determined based on self-assessment data, including: dynamic change characteristics and risk-related characteristics; among which,

[0080] Dynamic change characteristics are used to calculate short-term fluctuation indicators, such as the rate of change of social activity ΔT (the percentage difference between today's duration and the average of the past 7 days) and the fluctuation of sleep quality ΔPSQI (the difference between today's PSQI and the historical average).

[0081] The risk-associated feature is the increase in suicidal ideation (the difference between the current BSI and the average of the past 7 days), which is used to identify a sudden increase in suicide risk.

[0082] Specifically, a regression model is constructed using gradient boosting trees (XGBoost), with the goal of generating correction coefficients λ (range [0.9, 1.1]) from self-assessment data to achieve dynamic calibration of the baseline risk score.

[0083] The input features of the regression model are 10-dimensional features, including 7 original indicators (PHQ-9, real-time sentiment, etc.) and 3 derived features (ΔT, ΔPSQI, BSI increase).

[0084] The output variable of the regression model is the correction coefficient λ, and the formula for calculating the corrected risk score is:

[0085] Adjusted risk score = Baseline risk score × λ

[0086] Where λ=1 indicates no correction.

[0087] S30. Obtain the user's physiological and behavioral data, and input the corrected risk score, physiological and behavioral data into a pre-trained short-term early warning model. The short-term early warning model outputs a short-term early warning label.

[0088] It should be noted that, since physiological data and behavioral data may have different time steps, in order to adapt the data to LSTM, high-frequency data can be aggregated by time to unify the time steps of the data input to the model.

[0089] In one embodiment, the short-term warning model receives a risk score corrected by an individualized baseline risk model, combines it with the user's physiological and behavioral data, and outputs a short-term warning label to predict whether the patient has a high risk of relapse of depression in the short term (e.g., in the next 1-2 weeks).

[0090] In one embodiment, physiological data is acquired, including user heart rate variability, skin conductance, and diurnal temperature variation.

[0091] As examples, heart rate variability (HRV) can include SDNN (standard deviation of normal RR intervals), RMSSD (root mean square of the difference between adjacent RR intervals), pNN50 (the proportion of differences >50ms), etc., to reflect autonomic nervous system regulation (HRV often decreases during depressive episodes); skin conductance (sampled every 5 minutes) is used to reflect the level of emotional arousal, and the higher the value, the greater the anxiety or stress; body temperature diurnal variation includes body temperature fluctuations over 24 hours.

[0092] Specifically, this involves acquiring behavioral data, including: obtaining information such as the duration of social activities and the intensity of physical activity through smartphone apps.

[0093] As an example, social activity duration is the cumulative time spent using social media apps (such as WeChat and Weibo) each day, and social withdrawal is an important signal of depression relapse; exercise intensity is the daily duration of moderate to high intensity exercise (>3 METs) calculated by accelerometer, and reduced exercise is associated with worsening depression.

[0094] Specifically, the short-term early warning model is trained based on the LSTM time series prediction model.

[0095] As an example, LSTM includes:

[0096] The input layer receives multimodal time series data from the past 7 days. Each time step contains three types of features: modified risk score, physiological features (HRV index, skin conductance, body temperature difference, etc.), and behavioral features (social duration, exercise intensity, etc.).

[0097] The LSTM layer consists of two LSTM units (128 neurons per layer). It uses gating mechanisms (input gate, forget gate, output gate) to capture long-term dependencies and learn temporal patterns related to depression relapse (such as a continuous decrease in HRV for 3 days + a sudden reduction in social time may indicate an increased risk).

[0098] A fully connected layer is added after the LSTM layer, consisting of a 64-neuron ReLU layer, to further extract nonlinear features and enhance the model's ability to represent complex patterns.

[0099] The output layer, a single-node Sigmoid layer, outputs the recurrence probability in the interval [0,1]. A threshold (e.g., 0.6) is set to convert it into a binary warning label (1 = high risk, 0 = no risk).

[0100] S40. Obtain the user's real-time voice data and real-time chat text data, and input the short-term warning label, real-time voice data and real-time chat text data into a pre-trained real-time intervention model for processing. The real-time intervention model generates a real-time intervention strategy.

[0101] In one embodiment, the method for generating an immediate intervention strategy using an immediate intervention model is as follows:

[0102] S401. Obtain the user's real-time voice data and real-time text data; extract voice keywords from the user's real-time voice data and text keywords from the user's real-time text data; generate an original keyword list based on the voice keywords and text keywords; input the original keyword list into the depression emotion knowledge graph for implicit emotion reasoning, and obtain implicit emotion triples as implicit emotion labels.

[0103] Specifically, real-time speech data is converted into text using automatic speech recognition technology. High-frequency speech keywords are extracted using a domain keyword dictionary (e.g., depression-related words: "despair," "insomnia," "helplessness") and the TF-IDF algorithm. The intonation (e.g., slower speech rate, lower volume) and sentiment (positive / negative) of these high-frequency keywords are also labeled. The real-time chat text is segmented and stop words are removed. A BERT pre-trained model is used to extract semantic keywords (e.g., "suicide," "collapse," "don't want to live"), and sentiment analysis tools are used to label the explicit sentiment intensity of the text (e.g., "very painful" is stronger than "a little sad"), resulting in text keywords. The high-frequency speech keywords and text keywords are merged, duplicates are removed, and an original keyword list containing word, sentiment intensity, and context weights is generated (e.g., "suicide" is labeled as a high-risk keyword with a weight of +0.8).

[0104] Specifically, construct a knowledge graph of depressive emotions: build a triplet with "emotional state-triggering factor-physiological response" as the core, such as (despair, stemming from unemployment; despair, accompanied by insomnia), to form a knowledge graph of depressive emotions.

[0105] The original keywords are input into the graph, and the graph traversal algorithm is used to find the associated sentiment nodes. For example, the keywords "unemployment" and "insomnia" trigger the "despair" node in the graph, which further infers the implicit sentiment triple (current state, implicit, despair), and marks the duration of the sentiment (e.g., "frequently occurring recently") and the degree of impact (e.g., "seriously interferes with daily life").

[0106] It should be noted that the knowledge graph involved in this solution can be expanded in dimensions according to the actual situation.

[0107] S402. Extract acoustic features from the user's real-time voice data, input the acoustic features and implicit emotion labels into the emotion classification model for processing, and obtain the user's real-time emotion category; extract time series features based on the user's historical emotion categories; input the real-time emotion category and time series features into the dynamic risk monitoring model for processing, and obtain the user's risk level label.

[0108] Specifically, acoustic features are extracted from real-time speech data, including the fundamental frequency mean, formant frequencies, and energy entropy. The fundamental frequency mean reflects the degree of emotional distress, the formant frequencies reflect speech intelligibility, and the energy entropy reflects emotional stability. A CNN-LSTM hybrid model is used as the emotion classification model. Inputting acoustic features and implicit emotion labels, real-time emotion categories are obtained (such as "depressive episode", "emotional fluctuation period", and "relatively stable period").

[0109] Specifically, based on the user's emotional categories over the past 7 days, trend features, fluctuation features, and periodic features are extracted to form a time feature sequence of emotional states containing timestamps.

[0110] Specifically, using the Prophet time series prediction model, the input time feature series and real-time sentiment category output a risk level label for the next 24 hours. For example, two consecutive days of high sentiment volatility and a baseline risk score >0.8 are labeled as "high risk".

[0111] S403. Input the user's short-term warning label, risk level label and implicit sentiment label into the sentiment intensity quantification model for processing to obtain the real-time intervention intensity level.

[0112] Specifically, the intensity grading of immediate intervention is obtained according to the following method:

[0113] Based on the risk level labels output by the dynamic risk monitoring model, the risks are divided into three levels:

[0114] Low risk (weight 0.2): Probability of recurrence in the next 24 hours <40%;

[0115] Medium risk (weight 0.5): recurrence probability 40%-70%;

[0116] High risk (weight 0.8): recurrence probability > 70%.

[0117] Emotional intensity values ​​(range 0-1) are determined using reasoning based on knowledge graphs, and are categorized by intensity as follows:

[0118] Mild (coefficient 1.0): Intensity value ≤ 0.5 (e.g., "brief low mood");

[0119] Moderate (coefficient 1.5): 0.5 < intensity value ≤ 0.8 (e.g., "persistent anxiety");

[0120] Severe (coefficient 2.0): Intensity value > 0.8 (e.g., "despair and suicidal tendencies").

[0121] If the short-term warning label is "high risk", then the warning correction factor is 20%, which means the total intervention intensity is increased by 20%.

[0122] Therefore, the intervention intensity value = (risk level weight × 0.6) + (implicit emotion intensity coefficient × 0.4) × (1 + early warning correction factor)

[0123] The calculated intervention intensity values ​​are then hierarchically mapped:

[0124] Level 1 (Low Intensity): 1.0 ≤ Intensity value < 1.3, corresponding to scattered negative emotions without a continuous worsening trend;

[0125] Level 2 (Medium Intensity): 1.3 ≤ Intensity value < 1.7, requiring attention to emotional fluctuations and behavioral abnormalities;

[0126] Level 3 (High Intensity): Intensity value ≥1.7, indicating suicidal tendencies or risk of acute attacks.

[0127] S404. Generate immediate intervention strategies based on risk level labels, implicit sentiment labels, and immediate intervention intensity classifications.

[0128] Specifically, a strategy library and matching rules are constructed; combining risk level labels, implicit sentiment labels, and real-time intervention intensity levels, matching is performed in the strategy library according to the matching rules to obtain real-time intervention strategies.

[0129] Specifically, the input parameters for the matching rules include risk level labels, implicit sentiment labels, and immediate intervention intensity. Risk level labels include low risk, medium risk, and high risk; implicit sentiment labels include mild, moderate, and severe; and intervention intensity levels include Level 1 (low), Level 2 (medium), and Level 3 (high).

[0130] It should be noted that if the implied emotional label contains words related to "suicide," such as "don't want to live" or "end life," a Level 3 intervention will be forcibly triggered, overriding other low-priority operations and ignoring other parameters; if the pre-intensity rating is ≥ Level 2 and the risk level is high, manual intervention will be added, such as a real-time online consultation with a mental health counselor; if the risk level is low but the implied emotional label is moderate, preventive intervention will be initiated, such as pushing out emotion regulation tools.

[0131] As an example, the hierarchical strategy library and specific measures are as follows:

[0132] Level 1 intervention

[0133] Triggering conditions: Risk level = low risk, implied sentiment label = mild, and intervention intensity level = level 1.

[0134] Specific measures include: sending self-help documents on cognitive behavioral therapy, such as a "Negative Thought Record Sheet" and an "Emotional Diary Template"; pushing guided mindfulness breathing audio, 5-10 minutes long, at a speed of 90 words per minute, with a base frequency of 180Hz; recommending light daily social tasks, such as "talking to family and friends for 5 minutes" or "liking 3 posts on WeChat Moments"; setting exercise goals, such as "walking 3000 steps today," and syncing exercise data to the app's reward system. It can also push information about nearby parks or cafes based on real-time location, and push sleep music if nighttime activity is detected.

[0135] Level 2 intervention

[0136] Triggering conditions: Risk level = medium risk, or implied sentiment label = moderate, or intervention intensity level = level 2, as long as any one of these conditions is met.

[0137] Specific measures include: matching users with online psychological counselors within 24 hours; counselors developing a daily mood management plan through voice or text communication; dynamically increasing exercise goals (e.g., from 5000 steps to 6000 steps), with real-time monitoring via wristband data; and providing sleep restriction therapy guidance if sleep quality declines. With user authorization, encrypted prompts are sent to pre-defined family members, such as "Zhang San has been experiencing significant mood swings recently; we recommend spending more time with him / her"; the family member app simultaneously provides coping guidelines, such as "How to listen to a person with depression."

[0138] Level 3 intervention

[0139] Triggering conditions: Risk level = high risk, or implied emotional label = severe, or intervention intensity level = level 3, or detection of suicide-related keywords, as long as any one of these conditions is met.

[0140] Specific measures include: automatically triggering emergency contact notifications, including the user's real-time location and recent emotional fluctuation curve; initiating device location tracking, updating every 5 minutes until safety is confirmed manually; direct connection to a psychological crisis intervention hotline; hotline specialists can communicate directly via the device's voice channel; and it can also loop professionally recorded reassuring voice messages, such as: "Please follow my voice and take deep breaths, you are not alone," with a base frequency of 150Hz and a 1.5-second interval between each sentence; if rapid breathing is detected (breathing rate > 20 breaths / minute), the device will vibrate to guide a deep breathing rhythm.

[0141] It should be noted that the intervention level is mainly determined by the highest level indicator. For example, when the risk level is medium risk and the implied emotional label is severe, a level 3 intervention is triggered.

[0142] In an optional embodiment, user profile parameters can also be introduced to fine-tune the immediate intervention strategy. For example, gamified psychological intervention tools can be prioritized for adolescent users, while telephone follow-ups can be prioritized for elderly users; user acceptance of the intervention strategy can be obtained as input for subsequent strategy optimization, and the strategy matching weight can be updated every 24 hours.

[0143] It should be noted that the above scheme involves genetic data and real-time location data, which must comply with data compliance and privacy protection regulations.

[0144] Based on the same inventive concept as the above-described solution, the present invention also provides a wearable device, which includes:

[0145] The shell is designed to be worn on the user's wrist or chest;

[0146] The multimodal sensor array, integrated within the housing, includes: a physiological sensor module, a motion sensor module, and a voice acquisition module; the physiological sensor module is used to collect real-time data on heart rate variability (HRV), skin conductance, and body temperature diurnal rhythm; the motion sensor module incorporates a three-axis accelerometer and a gyroscope to monitor the duration and intensity of moderate to high-intensity exercise; the voice acquisition module includes a directional microphone and a noise reduction chip to acquire real-time voice data;

[0147] The data processing module includes a baseline risk assessment unit, a dynamic correction unit, and a short-term early warning unit. The baseline risk assessment unit pre-stores an individualized baseline risk model and acquires the user's genetic risk score (PRS), neuroimaging brain region features, and clinical history data through a communication interface, outputting the baseline risk score. The dynamic correction unit connects to the user's terminal App, receives self-assessment data (PHQ-9, PSQI, BSI), calculates the correction coefficient λ, and generates a corrected risk score. The short-term early warning unit has a built-in LSTM time series model and outputs a relapse warning label for the next 1-2 weeks based on the corrected risk score, physiological data, and exercise data from the past 7 days.

[0148] The intervention execution module includes a multi-level strategy library, a real-time analysis engine, and an output interface. The multi-level strategy library stores intervention strategies at levels 1-3 and matching rules. The real-time analysis engine connects to the voice acquisition module, analyzes the implicit sentiment tags of real-time voice / text using a sentiment knowledge graph, and triggers corresponding intervention strategies based on warning tags. The output interface is used to execute at least one of the following intervention actions:

[0149] a) Biofeedback regulation via a vibration motor and LED light strip;

[0150] b) Push cognitive behavioral therapy documents or mindfulness audio to the paired mobile phone via Bluetooth module;

[0151] c) When suicide-related keywords are detected, automatically dial the preset emergency contact and send the location.

[0152] This wearable device collects physiological, behavioral, and voice data in real time using multimodal sensors. Combined with a personalized baseline risk assessment model and dynamic correction algorithms, it achieves precise quantification of relapse risk. The device incorporates a short-term warning LSTM model that can identify high-risk states 1-2 weeks in advance. It also analyzes implicit suicidal tendencies in real-time voice through an emotion knowledge graph, triggering tiered intervention strategies such as biofeedback regulation and emergency contact. This device deeply integrates genetic, imaging, and clinical data with wearable sensing technology, providing a closed-loop management solution from risk warning to immediate intervention while avoiding direct risk diagnosis.

[0153] Example 2

[0154] like Figure 2 As shown, based on the same inventive concept as the above embodiments, the present invention also provides a device for intervening in the risk of depression relapse, characterized in that it includes:

[0155] The first assessment module is used to acquire the user's genetic risk score data, clinical history data, and neuroimaging data, and input them into a pre-trained individualized baseline risk model, which outputs the patient's baseline risk score.

[0156] The second assessment module is used to acquire the user's self-assessment data, determine the correction coefficient based on the self-assessment data, and correct the baseline risk score based on the correction coefficient to obtain the corrected risk score.

[0157] The third assessment module is used to acquire the user's physiological and behavioral data, and input the corrected risk score, physiological data and behavioral data into a pre-trained short-term early warning model. The short-term early warning model outputs a short-term early warning label.

[0158] The fourth evaluation module is used to acquire users' voice data and chat text data in real time, and input the short-term warning tags, voice data and chat text data into a pre-trained real-time intervention model, which generates a real-time intervention strategy.

[0159] Example 3

[0160] like Figure 3 As shown, the present invention also provides an electronic device 100 for implementing a method for intervening in the risk of relapse of depression;

[0161] The electronic device 100 includes a memory 101, at least one processor 102, a computer program 103 stored in the memory 101 and executable on at least one processor 102, and at least one communication bus 104.

[0162] The memory 101 can be used to store computer program 103. The processor 102 implements the steps of the method for intervening in the risk of relapse of depression in Embodiment 1 by running or executing the computer program stored in the memory 101 and calling the data stored in the memory 101.

[0163] The memory 101 may primarily include a program storage area and a data storage area. The program storage area may store the operating system, application programs required for at least one function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created based on the use of the electronic device 100 (such as audio data), etc. In addition, the memory 101 may include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other non-volatile solid-state storage device.

[0164] At least one processor 102 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Processor 102 may be a microprocessor or any conventional processor. Processor 102 is the control center of electronic device 100, connecting various parts of electronic device 100 via various interfaces and lines.

[0165] The memory 101 in the electronic device 100 stores multiple instructions to implement a method for intervening in the risk of relapse of depression, and the processor 102 can execute multiple instructions to achieve the following:

[0166] The system acquires the user's genetic risk score data, clinical history data, and neuroimaging data, and inputs them into a pre-trained personalized baseline risk model, which outputs the patient's baseline risk score.

[0167] Obtain the user's self-assessment data, determine the correction coefficient based on the self-assessment data, and correct the baseline risk score based on the correction coefficient to obtain the corrected risk score;

[0168] Acquire users' physiological and behavioral data, input the corrected risk score, physiological and behavioral data into a pre-trained short-term early warning model, and the short-term early warning model outputs a short-term early warning label;

[0169] The system acquires users' voice data and chat text data in real time, and inputs the short-term warning tags, voice data, and chat text data into a pre-trained real-time intervention model, which then generates a real-time intervention strategy.

[0170] Example 4

[0171] If the modules / units integrated in the electronic device 100 are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, and read-only memory (ROM).

[0172] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0173] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0174] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0175] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0176] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0177] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A method for intervening in the risk of relapse of depression, characterized in that, include: The system acquires the user's genetic risk score data, clinical history data, and neuroimaging data, and inputs them into a pre-trained personalized baseline risk model, which outputs the patient's baseline risk score. Obtain the user's self-assessment data, determine the correction coefficient based on the self-assessment data, and correct the baseline risk score based on the correction coefficient to obtain the corrected risk score; Acquire users' physiological and behavioral data, input the corrected risk score, physiological and behavioral data into a pre-trained short-term early warning model, and the short-term early warning model outputs a short-term early warning label; The system acquires users' voice data and chat text data in real time, and inputs the short-term warning tags, voice data, and chat text data into a pre-trained real-time intervention model, which then generates a real-time intervention strategy. The individualized baseline risk model is trained as follows: Acquire basic data of patients with depression; wherein, the basic data includes genetic data, neuroimaging data, and clinical history data; Features are extracted from the basic data, and a training set is constructed based on the extracted features; wherein, gene risk scores are calculated for gene data; brain region structural features are extracted from neuroimaging data, including the gray matter volume of the nucleus accumbens and the cortical thickness of the ventral tegmental area; and medical history features are extracted from clinical medical history data. The architecture of the individualized baseline risk model is determined to be a multimodal fusion neural network. This network comprises an input layer, a fusion layer, and an output layer. The input layer receives multimodal features, including genetic risk scores, brain region structural features, and medical history features. The fusion layer combines these features using tensor concatenation, dynamically calculates the weights of each modality using a self-attention module, and generates hidden layer vectors through dimensionality reduction using a fully connected layer. The output layer uses a single-node fully connected layer with a sigmoid activation function to output the baseline risk score based on the hidden layer vectors. The model is trained based on the training set and the multimodal fusion neural network, and an individualized baseline risk model is obtained after training. The training method for the short-term early warning model is as follows: Acquire historical physiological and behavioral data for model training; historical physiological data includes user heart rate variability, skin conductance, and diurnal temperature variation; historical behavioral data includes duration of social activities and exercise intensity. The architecture of the short-term early warning model was determined to be an LSTM time series prediction model; The model is trained using a modified risk score, historical physiological data, historical behavioral data, and an LSTM time series prediction model. After training, a short-term early warning model is obtained.

2. The method according to claim 1, characterized in that, The self-assessment data includes symptom indicators, immediate status, behavioral data, and derived characteristics; Symptom indicators include the severity of depressive symptoms, sleep quality index, and suicidal ideation scale; Real-time status includes real-time emotion score and attention score; Behavioral data includes social time and moderate-to-vigorous intensity exercise time. The derived features include dynamic change features and risk-related features; among them, dynamic change features are used to calculate short-term volatility indicators, and risk-related features are used to measure the increase in suicidal ideation.

3. The method according to claim 2, characterized in that, The correction coefficient is generated by a gradient boosting tree regression model. The input features of the gradient boosting tree regression model include symptom indicators, immediate status, behavioral data, and derived features. The output is the correction coefficient λ∈[0.9,1.1]. The corrected risk score is calculated as the product of the baseline risk score and the correction coefficient λ.

4. The method according to claim 1, characterized in that, The method for generating immediate intervention strategies using the immediate intervention model is as follows: The system acquires real-time voice data and real-time text data from users, extracts voice keywords from the real-time voice data, extracts text keywords from the real-time text data, and generates an original keyword list based on the voice keywords and text keywords. The original keyword list is then input into a depression emotion knowledge graph for implicit emotion inference to obtain implicit emotion labels. Acoustic features are extracted from real-time speech data, and the acoustic features and implicit emotion labels are input into an emotion classification model for processing to obtain the user's real-time emotion category. Extract time-series features based on the user's historical sentiment categories; Real-time sentiment categories and time-series features are input into a dynamic risk monitoring model for processing to obtain the user's risk level label; Short-term warning labels, risk level labels, and implicit sentiment labels are input into the sentiment intensity quantification model for processing to obtain an immediate intervention intensity rating. Based on risk level labels, implicit sentiment labels, and immediate intervention intensity classifications, immediate intervention strategies are generated.

5. The method according to claim 4, characterized in that, The intensity grading of immediate intervention is obtained as follows: Based on the risk level labels output by the dynamic risk monitoring model, risks are divided into three levels: low risk, medium risk, and high risk; among which, the weight of low risk is 0.2, the weight of medium risk is 0.5, and the weight of high risk is 0.

8. The emotional intensity values ​​are divided into three levels based on the implicit emotional tags of the depression emotional knowledge graph: mild, moderate, and severe. The intensity coefficient for mild is 1, for moderate it is 1.5, and for severe it is 2. If the short-term warning label is high risk, the warning correction factor is 20%. Intervention intensity value = (weight × 0.6 + intensity coefficient × 0.4) × (1 + early warning correction factor); The calculated intervention intensity values ​​are then mapped to different levels to obtain an instantaneous intervention intensity level.

6. A device for intervening in the risk of relapse of depression, used to implement the method of claim 1, characterized in that, include: The first assessment module is used to acquire the user's genetic risk score data, clinical history data, and neuroimaging data, and input them into a pre-trained individualized baseline risk model, which outputs the patient's baseline risk score. The second assessment module is used to acquire the user's self-assessment data, determine the correction coefficient based on the self-assessment data, and correct the baseline risk score based on the correction coefficient to obtain the corrected risk score. The third assessment module is used to acquire the user's physiological and behavioral data, and input the corrected risk score, physiological data and behavioral data into a pre-trained short-term early warning model. The short-term early warning model outputs a short-term early warning label. The fourth evaluation module is used to acquire users' voice data and chat text data in real time, and input the short-term warning tags, voice data and chat text data into a pre-trained real-time intervention model, which generates a real-time intervention strategy.

7. An electronic device, characterized in that, It includes a processor and a memory, the processor being configured to execute a computer program stored in the memory to implement the method as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one instruction that, when executed by a processor, implements the method as described in any one of claims 1 to 5.

Images