Depression emotion analysis method based on mobile phone sensor under complication background

Abstract

The invention relates to a method for analyzing depressive emotion based on mobile phone sensor data in the presence of complications, and belongs to the technical field of depression screening. The method comprises the following steps: firstly, establishing a database, dividing data in the established database into a training set and a test set according to a ratio of 7: 3, and constructing an uncertain reasoning neural network URNN to analyze and collect sensor data so as to evaluate the depressive emotion of a user; the training set is used for training the constructed URNN for uncertainty reasoning, and finally the test set is used for evaluating the trained URNN. According to the method, data from different sensors are integrated, machine learning and deep learning algorithms are combined, characteristics related to depression are extracted from the data, and the influence of complications is considered, so that more accurate depression emotion analysis is realized.

Description

Technical Field

[0001] The invention relates to a method for analyzing depression based on mobile phone sensor data in the presence of complications, and belongs to the technical field of depression screening. Background Art

[0002] Depression is a growing social problem with increasing incidence and costs. An estimated 280 million people worldwide suffer from depression. Depression can severely affect the patient's quality of life and has resulted in more than 50 million years of life lost (YLD) worldwide. At the same time, depression is also a common mental illness associated with some other diseases (such as Parkinson's disease), which can affect the incidence and treatment of other diseases. In addition, the presence of complications can lead to misdiagnosis of depressive symptoms, which in turn leads to ineffective treatment strategies. Therefore, it is very important to improve the ability to analyze depressive mood in the context of complications to determine the most effective and appropriate treatment options.

[0003] The use of sensors to assess patient behavior and symptom performance and then analyze depression is an emerging trend in intelligent analysis of depression. Today, smartphones and wearable devices are often equipped with various sensors for collecting different types of data. For example, built-in sensors on smartphones, such as accelerometers, gyroscopes, microphones, light sensors, and GPS, can collect motion, sound, optical, and location data from users. In addition, based on these sensors, how to obtain effective and objective depression symptom measurements from sensor data is a very important step in intelligent analysis of depression. Existing methods extract two feature categories from sensor data of mobile devices or fixed sensor data: motion features and behavioral features. Motion features include continuous physical activity or limb movement features measured by sensors. Behavioral features include features such as phone use, sleep, location information, social isolation, and food intake. At present, the methods for extracting sensor data features mainly include feature engineering and deep learning-based methods. Feature engineering extracts predictive variables of traditional machine learning models, such as the mean and standard deviation of acceleration values, but this method requires special solutions designed according to task requirements and is relatively labor-intensive. Deep learning-based methods can automatically extract features from sensor data and are now widely adopted.

[0004] Using features extracted from data collected by mobile sensors, machine learning and deep learning models can be trained to detect patients with depression. Machine learning models include models such as support vector machine (SVM), decision tree (DT), logistic regression (LR), random forest (RF), etc. These models are mainly classifiers that predict the presence or absence of depression. For deep learning models, convolutional neural network (CNN) is the most common model. For example, Kumar et al. and Yan et al. developed convolutional neural network (CNN) to predict whether a person has a mental illness. Sensor data is a series of high-speed data points that describe the activity status of an individual. It is first input into the deep learning model for training. The trained model can automatically learn the feature representation of sensor data. The learned feature vector captures the unique characteristics of the data collected by each sensor and is then used for the task of detecting depression.

[0005] Some patients with depression may have other comorbidities. For example, patients with depression may also suffer from Parkinson's disease or diabetes. These comorbidities may share some symptoms with depression (e.g., bradykinesia and anhedonia), which may lead to similar features extracted from sensor data, making depression detection difficult. So far, only a few detection methods have considered the comorbidity problem in depression detection. Since depression and arrhythmia are common complications in the elderly, Huang et al. used two different methods to classify patients into different types of arrhythmia and different degrees of depression, respectively. For arrhythmia, they used heart rate as a feature and estimated the intervals between different waveforms in electrocardiogram (ECG) data for classification, while for depression, they used questionnaires to examine mental health status. Sarda et al. compared five machine learning methods, including SVM, DT, RF, adaptive boosting (AdaBoost), and extreme gradient boosting (XGBoost), to detect depression in diabetic patients based on activity, mobility, sleep, and communication features extracted from smartphone data. The results showed that the XGBoost and RF methods performed better than the other methods. These studies have used different methods to detect diseases based on the unique diagnostic criteria of each disease, and still have not addressed the issues of symptom similarity and diagnostic uncertainty faced by depression detection based on sensor data in the context of comorbidities.

[0006] Multi-sensor data fusion can improve the reliability, credibility, and accuracy of patient health assessment and risk prediction, especially for the fusion of heterogeneous sensor data. Heterogeneous sensors can collect different types of data and measure different aspects of the patient's status, thereby achieving a comprehensive health assessment and reducing diagnostic uncertainty. The uncertainty of depression analysis in the context of complications increases the necessity of multi-sensor fusion. In order to monitor and manage patient health, many researchers have made increasing efforts to integrate data collected by heterogeneous sensors.

[0007] Disadvantages of existing technology:

[0008] First, existing technologies ignore the uncertainty of depression analysis in the context of complications. Existing work on depression analysis using mobile phone sensor data faces an important challenge, namely, significant uncertainty in the presence of complications, which is often ignored by the analysis method. This uncertainty mainly stems from overlapping symptoms between depression and other diseases and diagnostic uncertainty. Depression and its complications may share some common symptoms. For example, depression and Parkinson's disease have some common motor symptoms, such as slow movements, fatigue, and hunchback posture. Therefore, sensor data collected from the mobile phones of patients with depression and patients with other complications and features extracted from the data may have similarities. In addition, complications have been shown to be a factor leading to diagnostic uncertainty, and the presence of complications increases the difficulty for doctors to make a confident diagnosis of depression. For algorithms, previous machine learning models have not considered the diagnostic uncertainty that may be caused by complications, so existing depression detection methods based on sensor data may not be effective for patients with other complications. There is no corresponding treatment method for this uncertainty in the context of complications.

[0009] Second, the existing technology does not consider the importance of different features in feature learning. As mentioned above, due to the influence of complications and uncertainty, the existing depression analysis technology based on mobile phone sensors still has limitations in data processing and feature learning. Although useful features can be extracted from the collected sensor signals through deep learning models or feature engineering, these features may be associated with both depression and complications due to the existence of overlapping symptoms. The existing methods do not consider the differences in the importance of different features for depression analysis during feature learning, and cannot distinguish similar features between depression and complications. This may cause the existing feature learning methods to be ineffective in distinguishing depression from its complications.

[0010] Third, the fusion strategies involved in the prior art do not consider the symptom categories measured by the sensors. Most current studies tend to directly fuse all sensor data without considering the impact of different symptom categories measured by sensor data on depression detection. Especially in the presence of complications, the symptoms shared by depression and complications make it challenging to learn effective features to detect depression. Taking full advantage of the advantages of multiple sensors, we can deeply explore the symptom features most relevant to depression analysis from similar symptoms. This requires classification and hierarchical fusion based on the symptom categories measured by sensor data to learn effective symptom features. However, existing methods have failed to effectively learn the symptom representations of patients with depression and complications, and there is an urgent need for sensor data fusion strategies based on symptom categories. Summary of the invention

[0011] The present invention aims to solve the technical problems faced in disease identification and emotion analysis when a patient with depression also suffers from other complications (such as Parkinson's disease).

[0012] The technical solution of the present invention is:

[0013] A method for analyzing depression based on mobile phone sensor data in the presence of complications, the method comprising the steps of:

[0014] S001: Establish database The data in the established database is divided into training set and test set according to 7:3;

[0015] in, represents the q-time sensor features sampled by the m-th sensor for the n-th patient. (1) ,…,X (m) ,…,X (M)} represents a set of multi-sensor data, X (m) represents the data collected from the mth sensor, and M represents the total number of sensors; in represents the i-th sample data of the m-th sensor, and q is the total number of samples. Contains sensor readings at different time points, expressed as in represents the reading at a certain time point, and T represents the total number of time points. n represents the depression state of the nth patient. The depression state of each patient is represented by Y = {y1,…,y n ,…,y |N|} means, where y n =1 means depression, y n =0 means non-depression.

[0016] S002: Constructing a neural network URNN for uncertainty reasoning;

[0017] S003: Use the training set in step S001 to train the neural network URNN for uncertainty reasoning constructed in step S002;

[0018] S004: Evaluate the trained uncertainty reasoning neural network URNN.

[0019] In step S002, the constructed uncertainty reasoning neural network URNN includes module 1 sensor data feature extraction, module 2 feature fusion based on symptom categories and module 3 depression prediction based on multiple symptoms. The specific method is:

[0020] The first step is to extract features from sensor data. CNN is used as the feature extraction layer to extract features from sensor data. Extract the features at time point j As shown below:

[0021]

[0022] The second step is to obtain the first step The weights are assigned to the features in to reflect their importance for depression detection. This design is implemented using an attention layer, where the attention weights represent the importance of the learned representations. The implementation of this layer is as follows:

[0023]

[0024] Among them, W i q , W i k and is the parameter to be trained by the model, and l is the total number of time points contained in the features extracted by the CNN model.

[0025] In the third step, the weighted features extracted in the second step are used based on module 2 to obtain the representation of the symptoms by fusing all the individual feature representations.

[0026]

[0027] Through the above fusion, URNN learns symptom-related representations.

[0028] In the fourth step, in order to capture time series information and further learn the change information of symptom representation in the time dimension, the representation of the same symptom category obtained in the third step is input into a BiLSTM layer. The detailed implementation process of the BiLSTM layer is as follows:

[0029]

[0030] c i =f i *c i-1 +v i *g i

[0031] h i =o i *tanh(c i )

[0032] Among them, W v ,U v ,W f ,U f ,W o ,U o ,W g is the parameter matrix to be trained; b is the bias term; v i ,,f i ,o i are the activation vectors of the input gate, forget gate, and output gate respectively; g i is the cell input activation vector; c i is the cell state vector; h i is the hidden state vector; σ is a nonlinear function.

[0033] After calculating all the hidden state vectors h in the sequence i After that, URNN updates the representation of the symptom S as follows

[0034]

[0035] The fifth step is to model the uncertainty of the model prediction. Let Z represent the set of prediction results, which is defined as:

[0036] Z={ω1,ω2,Ω}

[0037] ω1 and ω2 represent the prediction of non-depression and depression, respectively, and Ω indicates uncertainty as to whether a patient has depression.

[0038] Step 6: Calculate a prediction probability for each prediction result in Z in step 5. The calculation method is:

[0039]

[0040] Here W c and are the parameters to be trained.

[0041] In the seventh step, the probabilities from all symptom representations obtained in the sixth step are fused according to Dempster's rule based on module three to reduce the uncertainty of the prediction. This rule extracts shared probabilities from multiple states in Z and ignores any non-shared probabilities. This combination ensures that all probabilities related to depression are integrated into the final predicted probability, effectively reducing the probability of uncertainty. The final prediction results are as follows:

[0042]

[0043] Beneficial Effects

[0044] (1) The method of the present invention solves the problem of complications faced by depression analysis. Based on the theory of evidence, the uncertainty in the analysis of depression in the presence of complications is modeled, and uncertainty reasoning is performed by integrating symptom representations. The uncertain prediction probability is recalculated based on the manifestations of multiple symptoms, and the probability of depression prediction is updated while considering the uncertain prediction probability, thereby reducing the uncertainty in the model prediction and providing a novel and practical solution for the analysis and evaluation of depression in the context of complications.

[0045] (2) The method of the present invention introduces CNN and attention mechanism for the first time to learn features from sensor data. Due to the influence of comorbidities and similar symptoms of depression, not all sensor features are effective for the analysis and evaluation of depressive mood. The proposed sensor feature learning method selects effective features for depression detection by assigning different weights to the extracted features. This method effectively solves the feature learning challenges brought about by the similarity of depression and comorbid symptoms.

[0046] (3) The method of the present invention proposes a hierarchical fusion strategy. When faced with multiple sensors measuring physical activity and various symptoms, this innovative fusion strategy first fuses sensor features belonging to the same symptom category to generate symptom representations, and then fuses the representations of all symptom categories for depression detection. This sensor data fusion based on symptom categories can generate effective symptom representations by integrating the features of multiple similar sensors, thereby reducing the uncertainty caused by similar symptoms.

[0047] (4) The method of the present invention analyzes sensor technology, fully utilizes the advantages of multiple sensors, aggregates similar features in sensor data to learn the representation of each symptom. This fusion strategy can effectively identify symptoms related to depression and avoid the negative impact of complications on depression analysis.

[0048] (5) The method of the present invention adds an uncertain prediction probability. In view of the problem that the existing technology directly predicts whether a patient has depression through a model and obtains a probability, which may lead to misdiagnosis in the context of complications, an uncertainty modeling and reasoning method is proposed based on evidence theory. In this way, the uncertainty of diagnosis can be modeled, which is also conducive to further reasoning to reduce this uncertainty. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is the URNN model architecture;

[0050] Figure 2 It is an example of multi-sensor fusion method;

[0051] Figure 3 Evaluate the case for fusion strategies. DETAILED DESCRIPTION

[0052] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0053] Example

[0054] The purpose of this technical solution is to Train a deep learning model that can detect depression patients based on data collected by multiple mobile phone sensors. In dataset D, and n represent the q sensor features sampled by the mth sensor of the nth patient and the depression state of the nth patient, respectively. (1) ,…,X (m) ,…,X (M)} represents a set of multi-sensor data, where X (m) represents the data collected from the mth sensor, and M represents the total number of sensors; in represents the i-th sample data of the m-th sensor, and q is the total number of samples. Contains sensor readings at different time points, expressed as in represents the reading at a certain time point, and T represents the total number of time points. N is used to represent a group of patients, and |N| represents the number of patients. The depression status of each patient is represented by Y={y1,…,y n ,…,y |N|} means, where y n =1 means depression, y n =0 means non-depression.

[0055] The following summarizes the important symbols and their descriptions involved in the overall technical solution:

[0056] is the sampling data of the mth sensor for the i-th time;

[0057] is the feature matrix of dimension j extracted from the sensor data;

[0058] is the depression-specific representation learned from the i-th sampling data of the m-th sensor;

[0059] It is a representation of the aggregation of symptoms S;

[0060] is the representation of the symptom S learned by the model;

[0061] To test the trust level of depression based on the representation of symptom S;

[0062] Predict the degree of belief in each depressive state category based on the representation of symptom S;

[0063] y n is the depression state of the nth patient.

[0064] In order to solve the technical problems raised previously, an uncertainty reasoning neural network (URNN) was proposed to detect depression patients based on mobile phone sensor data and perform uncertainty reasoning during the detection process to avoid the negative impact of other complications of patients on sentiment analysis.

[0065] Figure 1 The proposed model architecture is presented. The entire model architecture consists of three modules: a sensor data feature extraction module, a symptom category-based feature fusion module, and a multi-symptom-based depression prediction module. Module one aims to extract features from the data of each M sensor and learn the sensor's depression-specific representation from the extracted features. A hierarchical fusion strategy jointly performed by modules two and three is then proposed to achieve symptom representation learning and symptom representation-based depression detection. Module two is designed to perform the first step of fusion, while module three is responsible for the second step. Module three also performs uncertainty reasoning by combining the predicted probabilities based on each symptom representation to eliminate the uncertainty of depression sentiment analysis in the context of complications.

[0066] In module 1, features are extracted from the raw sensor data to characterize the behavior of patients with depression. Given the high efficiency of convolutional neural networks (CNNs) in automatically extracting features from local signals, CNN is used as a feature extraction layer to extract features from sensor data. In particular, Extract the features at time point j As shown below:

[0067]

[0068] The details of the CNN-based feature extraction layers are shown in Table 1. The CNN layers consist of a series of 2D convolution (Conv2d) layers, batch normalization (BatchNorm2d) layers, ReLU layers, and maximum pooling (MaxPool2d) layers. Since the BatchNorm2d layer and the ReLU layer do not change the input shape, they are not listed in Table 1. To learn the features of the local signal, two consecutive convolution layers are used, and the two-dimensional three-axis sensor data is converted into a one-dimensional feature vector using a combination of Conv2d layers and MaxPool2d layers.

[0069] Table 1 Detailed description of the feature extraction layer based on CNN

[0070]

[0071] To learn the feature vector specific to depression identification from the features extracted from sensor data, URNN is The weights are assigned to the features in to reflect their importance for depression detection. This design is implemented using an attention layer, where the attention weights represent the importance of the learned representation. The implementation process of this layer is as follows:

[0072]

[0073] Among them, W i q , W i k and is the parameter to be trained by the model, and l is the total number of time points contained in the features extracted by the CNN model.

[0074] Figure 2 Three examples of multi-sensor fusion methods are presented, which realize the fusion of multiple sensor data from the data level, model level and result level. This scheme proposes a novel two-step, symptom-based hierarchical fusion strategy for fusing multiple sensor representations. Figure 1As shown in Figure 2, module 2 implements the first step of this hierarchical fusion strategy, which fuses the sensor representations learned in the previous part according to the symptom category to learn symptom representations, thereby capturing fine-grained information at the symptom level from multiple sensor data. Sensors (e.g., accelerometers, gyroscopes) that measure the same physiological state of patients (e.g., gait) are classified into a symptom category (e.g., bradykinesia). Let S represent the set of symptom categories to which the sensor belongs. Each symptom category in S corresponds to the set of feature representations learned by k sensors in module 1, expressed as The symptom representation is formed by fusing all individual feature representations

[0075]

[0076] Through the above fusion, URNN learns symptom-related representations. Since the original sensor data is collected through a series of time points, the feature representation learned by CNN is still a time series. In order to capture time series information and further learn the change information in the time dimension of symptom representation, the representation of the same symptom category is input into a BiLSTM layer, which is very effective for modeling time data. The detailed implementation process of the BiLSTM layer is as follows:

[0077]

[0078] c i =f i *c i-1 +v i *g i

[0079] h i =o i *tanh(c i )

[0080] Among them, W v ,U v ,W f ,U f ,W o ,U o ,W g is the parameter matrix to be trained; b is the bias term; v i, ,f i ,o i are the activation vectors of the input gate, forget gate, and output gate respectively; g i is the cell input activation vector; c i is the cell state vector; h i is the hidden state vector; σ is a nonlinear function. i After that, URNN updates the representation of the symptom S as follows

[0081]

[0082] Module 3 models and infers the uncertainty caused by similar symptoms of depression and complications based on evidence theory to integrate multiple symptom representations learned in module 2 and reduce uncertainty. First, the uncertainty of model prediction is modeled, and Z is set to represent the set of prediction results, which is defined as:

[0083] Z={ω1,ω2,Ω}

[0084] ω1 and ω2 represent the prediction of non-depression and depression respectively, and Ω represents uncertainty as to whether a patient has depression. Next, module three needs to calculate a prediction probability for each prediction result in Z, and the calculation method is:

[0085]

[0086] Here W c and is the parameter to be trained. After obtaining the probability of each prediction result based on each symptom representation according to the above calculation, module three fuses the probabilities from all symptom representations according to Dempster's rule to reduce the uncertainty of the prediction. This rule extracts shared probabilities from multiple states in Z and ignores any non-shared probabilities. For example, the predicted probability of depression after fusion is a combination of the probabilities from depression and depression uncertainty. This combination ensures that all probabilities related to depression are integrated into the final predicted probability, effectively reducing the probability of uncertainty. The final prediction results are as follows:

[0087]

[0088] Evaluation of the detection performance of the technical solution:

[0089] To verify the proposed URNN detection performance advantage, the public dataset mPower was used for evaluation, which contains sensor data collected from Parkinson's patients and their depression status. Parkinson's patients have a high prevalence of depression and share motor and non-motor symptoms with depression, such as attention difficulties, reduced sleep, and slow movements. Therefore, this dataset was selected to evaluate the model's ability to detect depression in the presence of Parkinson's complications. The mPower dataset collects smartphone-based sensor data from patients through walking and mobile screen tapping experiments to measure patients' gait, body balance, flexibility, and reaction speed. Each patient provided information on whether they were diagnosed with depression before the start of the experiment. In the walking experiment, five active sensors (i.e., accelerometer, posture sensor, gyroscope, linear accelerometer, and gravimeter) sampled 20 steps of data at a frequency of 100Hz to measure gait and body balance. In the mobile screen tapping experiment, one sensor (i.e., accelerometer) measured hand flexibility and reaction speed for 20 seconds at a frequency of 100Hz. The data from these sensors contains a series of readings along three axes (x, y, z). Any records with missing data were excluded from the original dataset. The final dataset included 7,721 records involving 1,013 patients, of which 512 had depression and 501 did not. 70% of the dataset was used as a training set and 30% as a test set.

[0090] Table 2 shows the performance results of the URNN model and other baseline models in depression detection. The evaluation results show that the proposed URNN model is significantly higher than all baseline models in terms of recall, F1 value, and precision, while XGBoost achieves the best precision among all models. The best performing baseline deep learning model is CNN, and the best machine learning model is XGBoost, but URNN significantly outperforms both in terms of accuracy, recall, and F1 value. Although URNN's precision is not optimal, there is considerable improvement in other indicators, indicating that URNN can screen out patients with depression more accurately than other baseline models. This shows that URNN reduces the possibility of misdiagnosis and is superior to existing depression detection methods based on sensor data in detecting depression in the presence of complications.

[0091] Table 2 Depression test results

[0092]

[0093] In order to evaluate the impact of the proposed components on the overall detection performance of the model, ablation experiments were conducted. The proposed components include sensor data feature extraction (A1), symptom category-based feature fusion (A2), and multi-symptom-based depression detection (A3). The results of the ablation experiments are shown in Table 3. The experimental results indicate that removing any single component will lead to a significant decrease in the detection accuracy, recall, and F1 value of the model. Although the accuracy did not decrease significantly, other indicators showed a considerable decrease, indicating that many patients with depression were misidentified as Parkinson's patients, proving that removing a single component within URNN increases the risk of misdiagnosis. In addition, when all the proposed components are combined together, URNN achieves the best performance, highlighting the importance of integrating all three components to achieve the best results.

[0094] Table 3 Ablation experiment

[0095]

[0096] The proposed hierarchical fusion strategy was compared with other fusion strategies commonly used in existing methods, including data-level fusion, representation-level fusion, and result-level fusion. At the same time, in order to further evaluate the effectiveness of the hierarchical fusion strategy in two steps of learning and fusing symptom representations, the three existing fusion strategies and the proposed two-step hierarchical fusion were combined, and six combinations were obtained: data level-data level, representation level-representation level, result level-result level, data level-representation level, data level-result level, and representation level-result level. Among these combinations, URNN adopted the representation level-representation level strategy. The evaluation results are shown in Table 4. In order to compare the two-step hierarchical fusion strategy with the existing single-step fusion method, the average values ​​of the evaluation indicators of the single-step fusion strategy and the two-step fusion strategy were compared, and the results are shown in Table 4. Figure 3 shown.

[0097] Table 4 Evaluation of hierarchical fusion strategies

[0098]

[0099] The results show that, overall, the proposed model URNN outperforms three existing fusion strategies. Figure 3 It is shown that the two-step hierarchical fusion strategy exhibits superior performance compared to the existing single-step fusion strategy, highlighting the effectiveness of the hierarchical fusion strategy for detecting depression through symptom representation learning in the presence of comorbidities.

Claims

1. A method for analyzing depression based on mobile phone sensor data in the presence of complications, characterized in that The steps of the method include: S001, establish a database and divide the data in the established database into a training set and a test set according to a ratio of 7:3; S002: Constructing a neural network URNN for uncertainty reasoning; S003, using the training set in step S001 to train the neural network URNN for uncertainty reasoning constructed in step S002; S004, using the test set in step S001 to evaluate the uncertainty reasoning neural network URNN trained in S003, to achieve analysis of depressive emotions.

2. The method for analyzing depression based on mobile phone sensor data in the presence of complications according to claim 1, characterized in that: In step S001, the database established is in, represents the q-time sensor features sampled by the m-th sensor for the n-th patient; N is the number of patients; X={X (1) ,…,X (m) ,…,X (M) } represents a set of multi-sensor data; X (m) represents the data collected from the mth sensor, and M represents the total number of sensors; in, represents the i-th sample data of the m-th sensor, and q is the total number of samples; Each Contains sensor readings at different time points, expressed as in, represents the reading at a certain time point, and T represents the total number of time points; y n represents the depression state of the nth patient. The depression state of each patient is represented by Y = {y1,…,y n ,…,y |N| } means, where y n =1 means depression, y n =0 means non-depression.

3. A method for analyzing depression based on mobile phone sensor data in the presence of complications according to claim 1 or 2, characterized in that: In the step S002, the constructed uncertainty reasoning neural network URNN includes a sensor data feature extraction module, a symptom category-based feature fusion module and a multi-symptom-based depression prediction module.

4. The method for analyzing depression based on mobile phone sensor data in the presence of complications according to claim 3, characterized in that: The sensor data feature extraction module is used to extract features from the data of each sensor and learn the corresponding depression representation of the sensor data from the extracted features.

5. The method for analyzing depression based on mobile phone sensor data in the presence of complications according to claim 3, characterized in that: The symptom category-based feature fusion module implements the first step of the two-step, symptom category-based hierarchical fusion strategy, which fuses the sensor representations learned in the previous part according to the symptom category to learn the symptom representation, thereby capturing fine-grained information at the symptom level from multiple sensor data.

6. The method for analyzing depression based on mobile phone sensor data in the presence of complications according to claim 3, characterized in that: The multi-symptom based depression prediction module models and reasons about the uncertainty caused by similar symptoms of depression and complications based on evidence theory to fuse multiple symptom representations learned in the symptom category based feature fusion module and reduce uncertainty.

7. The method for analyzing depression based on mobile phone sensor data in the presence of complications according to claim 2, characterized in that: In step S002, the specific method of constructing the neural network URNN for uncertainty reasoning is: The first step is to extract features from sensor data. CNN is used as the feature extraction layer to extract features from sensor data. Extract the features at time point j As shown below: The second step is to obtain the first step The weights are assigned to the features in to reflect their importance for depression detection. This design is implemented using an attention layer, where the attention weights represent the importance of the learned representation. The implementation process of this layer is as follows: Among them, W i q , and is the parameter to be trained by the model, and l is the total number of time points contained in the features extracted by the CNN model; In the third step, the weighted features extracted in the second step are combined with the feature fusion module based on symptom categories to obtain the representation of the symptoms by fusing all the individual feature representations. Through the above fusion, URNN learns symptom-related representations; In the fourth step, in order to capture the time series information and further learn the change information of the symptom representation in the time dimension, the representation of the same symptom category obtained in the third step is input into a BiLSTM layer; After calculating all the hidden state vectors h in the sequence i After that, URNN updates the representation of the symptom S as follows The fifth step is to model the uncertainty of the model prediction. Let Z represent the set of prediction results, which is defined as: Z={ω1,ω2,Ω} ω1 and ω2 represent predictions of non-depression and depression, respectively, and Ω indicates uncertainty as to whether a patient has depression; Step 6: Calculate a prediction probability for each prediction result in Z in step 5. The calculation method is: Here W c and are the parameters to be trained; In the seventh step, the multi-symptom depression prediction module combines the probabilities from all symptom representations obtained in the sixth step according to Dempster's rule to reduce the uncertainty of the prediction. This rule extracts shared probabilities from multiple states in Z and ignores any non-shared probabilities. This combination ensures that all probabilities related to depression are integrated into the final prediction probability, effectively reducing the probability of uncertainty. The final prediction results are as follows:

Images