Medical data processing-based depression assessment system

Abstract

According to the depression assessment system based on medical data processing, a data fusion module extracts features of medical texts, voice data and physiological electric signal data through a multi-branch neural network, and dynamic interaction is carried out by adopting a bidirectional cross attention mechanism; the evaluation optimization module performs confidence scoring on the feature set and judges whether data is sensitive information, the sensitive data is processed by an authoritative knowledge base, the non-sensitive data is processed by a general knowledge base, and multi-branch neural network parameters are updated through an adaptive learning rate; the state evaluation module uses an MM-RNN model to calculate a depression state score and determine a depression level. According to the method, a sensitive information judgment mechanism based on confidence scoring is introduced, the data quality is intelligently evaluated, the data processing path is dynamically adjusted, it is ensured that sensitive data is effectively protected, the accuracy and reliability of the evaluation result are improved, and the adaptability and safety of the system for processing various medical data are enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of depression assessment, and specifically to a depression assessment system based on medical data processing. Background Art

[0002] Depression is a common mental illness that affects a large number of people globally. Its symptoms usually include low mood, loss of interest, sleep disorders, etc. In severe cases, it can affect the normal life and work of patients. Traditional depression assessment methods mainly rely on clinical diagnosis and standardized questionnaires, such as the Hamilton Depression Rating Scale and the Beck Depression Inventory, to evaluate the depressive symptoms of patients through self-report and doctor observation. With the progress of computer technology and big data analysis, depression assessment systems have been proposed to assist in the diagnosis and assessment of depression, so as to provide more accurate and objective assessment results.

[0003] Existing depression assessment systems mostly rely on static data analysis, such as single scale scoring or single-modal physiological signal data analysis. This method ignores the dynamic relationships and interactions between different data modalities, resulting in incomplete and inaccurate assessment results. At the same time, existing depression assessment systems handle data quality in a relatively single way, lacking in-depth exploration and weighing of the mutual influence between different types of data, resulting in insufficient accuracy and robustness of the assessment results when the system faces complex and variable clinical data. Summary of the Invention

[0004] In order to solve the technical problems mentioned in the current background art, the purpose of the present invention is to provide a depression assessment system based on medical data processing.

[0005] To achieve the above purpose, the present invention adopts the following technical solutions: A depression assessment system based on medical data processing, comprising: M1, a data fusion module, which extracts modal features from medical text data, voice data, and physiological electrical signal data through a multi-branch neural network, and performs dynamic interaction on the modal features by using a bidirectional cross-attention mechanism to generate a structured feature set; M2, an evaluation optimization module, which performs confidence scoring on the structured feature set. The confidence scoring is calculated by weighting the medical text feature stability index, voice clarity factor, and physiological electrical signal factor, compares the confidence scoring with a confidence scoring threshold to determine whether it is sensitive information. The retrieval path for sensitive information corresponds to an authoritative knowledge base, and the retrieval path for non-sensitive information corresponds to a general knowledge base, and outputs an optimal medical data reference set; Based on the optimal medical data reference set, an adaptive learning rate strategy is used to iteratively update the parameters of the multi-branch neural network; M3. The state evaluation module dynamically weights the reference feature vectors within the optimal medical data reference set, calculates the depression state score using the MM-RNN model based on the processed reference feature vectors, generates a dynamic depression threshold for depression state evaluation according to the dynamic threshold function; and determines the depression state level based on the dynamic threshold.

[0006] Further, the medical text data mainly comes from the medical records of patients and doctor's diagnosis reports; the voice data mainly comes from the voice symptom descriptions of patients and relevant voice records; the physiological electrical signal data mainly comes from the electrocardiograms and electroencephalograms of patients. The multi-branch neural network includes a BERT model, a self-supervised speech model, and a CNN-LSTM hybrid model. The extraction process of the modal features is as follows: The medical text data extracts the medical text modal feature set through the BERT model , the voice data extracts the voice modal feature set through the self-supervised speech model , and the physiological electrical signal data extracts the physiological electrical signal modal feature set through the CNN-LSTM hybrid model .

[0007] Further, the dynamic interaction process is as follows: Input the medical text modal feature set, voice modal feature set, and physiological electrical signal modal feature set into the bidirectional cross-attention mechanism to obtain the attention matrix, and then use the attention matrix to dynamically weight the medical text modal feature set, voice modal feature set, and physiological electrical signal modal feature set. Dynamically interact the medical text modal feature set through the bidirectional cross-attention mechanism , voice modal feature set and physiological electrical signal modal feature set between features; After the bidirectional cross-attention mechanism finishes calculation, use the obtained attention matrix to dynamically weight and fuse the medical text modal feature set , voice modal feature set and physiological electrical signal modal feature set , and the calculation formula is: ; ; ; where is the medical text modal feature set after dynamic interaction , is the medical text modal feature set voice modal feature set The fused features; is the physiological electrical signal modal feature set after dynamic interaction , is the medical text modal feature set and the physiological electrical signal modal feature set The fused features; is the voice modal feature set after dynamic interaction , is the voice modal feature set and the physiological electrical signal modal feature set The fused features; is the medical text modal feature set and the voice modal feature set The attention matrix between them; is the medical text modal feature set and the physiological electrical signal modal feature set The attention matrix between them; is the voice modal feature set and the physiological electrical signal modal feature set The attention matrix between them The structured feature set is denoted as F, and .

[0008] Furthermore, the calculation steps of the confidence score include calculating the medical text feature stability index, calculating the voice clarity factor, and calculating the physiological electrical signal clarity factor; The medical text feature stability index The calculation process is to calculate the mean of the medical text modal feature set , calculate the standard deviation of the medical text modal feature set and finally calculate the medical text feature stability index ; The voice clarity factor The calculation process is to calculate the power of the voice signal, calculate the power of the noise in the voice signal, and finally calculate the voice clarity factor ; The physiological electrical signal clarity factor The calculation process is to calculate the power of the physiological electrical signal, calculate the power of the noise in the physiological electrical signal, and finally calculate the physiological electrical signal clarity factor , and finally obtain the physiological electrical signal clarity factor ; Based on the medical text feature stability index , voice clarity factor and physiological electrical signal clarity factor Calculate the confidence score , the calculation formula is: ; Among them, 、 and are weight coefficients, and ; corresponds to the weight coefficient of the medical text feature stability index ; corresponds to the weight coefficient of the voice clarity factor ; corresponds to the weight coefficient of the physiological electrical signal clarity factor ; Set the confidence score threshold to 0.8; when , it is judged as sensitive information, and the corresponding retrieval path is the authoritative knowledge base; when , it is judged as non-sensitive information, and the corresponding retrieval path is the general knowledge base; finally, the optimal medical data reference set is output according to the retrieval results of different paths.

[0009] Furthermore, denote the optimal medical data reference set as , and Among them, is the reference feature vector after linear fusion of the th medical text data, voice data, and physiological electrical signal data; Perform dynamic weighting on the reference feature vectors inside the optimal medical data reference set, and the calculation formula is: ; Among them, is the weight coefficient of the th reference feature vector at the th moment; is the number of reference feature vectors; is the time point most relevant to the th reference feature vector; is the smoothing parameter used to control the sensitivity of weighting; During the evaluation process, first, it is necessary to calculate the weighted dynamic feature set through dynamic weighting, and the calculation formula is: ; Based on the dynamic feature set , using the MM-RNN model, after dynamically learning the temporal dependence relationship of the reference feature vectors, calculate the depression state score: ; Among them, is the output depression state score, ranging from [0, 1], representing the patient's current depression state; is the weight matrix of the output layer; ; is the hidden state at time step ; is the activation function; The generation function of the dynamic depression threshold is as follows: ; B ; where and are both the adjusted dynamic depression thresholds; is the initially set depression threshold; is the change amplitude factor; is the decay factor for controlling the state change rate; is the Gaussian noise term, used to simulate the natural fluctuations of the mental state; is the offset constant; Based on the dynamic depression threshold, determine the depression state level: ; where is mild depression; is moderate depression; is severe depression.

[0010] Furthermore, use the features of the optimal medical data reference set as input data to calculate the loss function , the loss function depends on the parameter , in each iteration, calculate the gradient of the loss function with respect to the current network parameters , meanwhile, the learning rate is dynamically adjusted according to the historical information of the gradient, and finally update the multi-branch neural network parameters based on the gradient and the adaptive learning rate .

[0011] Compared with the prior art, the advantages of the present invention are: 1. The present invention combines medical texts, voice data, and physiological electrical signal data, and adopts a multi-branch neural network and a bidirectional cross-attention mechanism to fully explore the deep relationships between various data modalities, enabling data of different modalities to be weighted according to their own importance during the dynamic interaction process, avoiding the one-sided evaluation caused by a single data source in traditional methods. Especially in the early diagnosis of depression, the patient's language expression and physiological signals often show no obvious symptoms. By dynamically fusing various data through this system, a more comprehensive and accurate evaluation can be achieved, improving the diagnostic accuracy of depression.

[0012] 2. The present invention introduces a dynamic weighting mechanism based on confidence evaluation. By calculating the medical text stability index, voice clarity factor, and physiological electrical signal factor, it dynamically evaluates the reliability of various types of data. This innovative mechanism can adaptively adjust to different quality data, ensuring that data noise or missing data will not have too much impact on the evaluation results. In practical applications, the system can optimize the weight allocation of each data modality in real time according to the quality of each data input, thereby improving the stability and reliability of the evaluation results in different patients and different clinical environments.

[0013] 3. The present invention significantly improves the system's processing ability for medical data of different qualities through a sensitive information determination mechanism based on confidence scores. This mechanism can intelligently identify the sensitivity of data and automatically select appropriate data paths according to the confidence of the data, ensuring strict protection of sensitive data and improving efficiency in non-sensitive data processing. Through the dynamic evaluation of the quality of medical data, the system can effectively avoid the risk of incorrect evaluation or data leakage caused by data quality differences, thereby improving the accuracy and reliability of the evaluation results while ensuring privacy protection. In addition, this mechanism enhances the flexibility and adaptability of the system, enabling it to more efficiently process multiple data sources when facing complex clinical data, improving the overall performance and clinical application value of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0015] Figure 1 It is a schematic diagram of the system working process of the present invention; Figure 2 It is a schematic diagram of the structured feature set calculation process of the present invention; Figure 3 It is a schematic diagram of the optimal medical data reference set calculation process of the present invention. Specific Embodiments

[0016] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the following will clearly and completely describe the technical solutions in the embodiments of this application with reference to the accompanying drawings in the embodiments of this application. Obviously, the described embodiments are some, but not all, of the embodiments of this application. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in this application without creative efforts fall within the scope of protection of this application.

[0017] To achieve the above objectives, the present invention is implemented through the following technical solutions. The present invention provides a depression assessment system based on medical data processing, as Figures 1 - 3 shown, the system includes: M1. A data fusion module that extracts modal features from medical text data, voice data, and physiological electrical signal data through a multi-branch neural network, and uses a bidirectional cross-attention mechanism to dynamically interact with the modal features to generate a structured feature set.

[0018] The medical text data mainly comes from the patient's medical records and doctor's diagnosis reports; the voice data mainly comes from the patient's voice symptom descriptions and related voice records; the physiological electrical signal data mainly comes from the patient's electrocardiogram and electroencephalogram; The multi-branch neural network includes a BERT model, a self-supervised voice model, and a CNN-LSTM hybrid model; The process of extracting the modal features is as follows: the medical text data extracts a medical text modal feature set through the BERT model , the voice data extracts a voice modal feature set through the self-supervised voice model , and the physiological electrical signal data extracts a physiological electrical signal modal feature set through the CNN-LSTM hybrid model ; The dynamic interaction process is as follows: Input the medical text modal feature set, voice modal feature set, and physiological electrical signal modal feature set into the bidirectional cross-attention mechanism to obtain an attention matrix, and then use the attention matrix to perform dynamic weighted fusion on the medical text modal feature set, voice modal feature set, and physiological electrical signal modal feature set; Dynamically interact with the medical text modal feature set through the bidirectional cross-attention mechanism , voice modal feature set and physiological electrical signal modal feature set The features between them, and the calculation formula is: ; ; ; Among them, is the attention matrix between the medical text modality feature set and the speech modality feature set ; is the attention matrix between the medical text modality feature set and the physiological electrical signal modality feature set ; is the attention matrix between the speech modality feature set and the physiological electrical signal modality feature set ; , and are the weight matrices for calculating attention, which are used for calculating the similarity between different modality pairs respectively; is the dimension of the feature vector, which is used to normalize the attention scores.

[0019] After the bidirectional cross-attention mechanism is calculated, the obtained attention matrix is used to perform dynamic weighted fusion on the medical text modality feature set , the speech modality feature set and the physiological electrical signal modality feature set , and the calculation formula is: ; ; ; Among them, is the medical text modality feature set after dynamic interaction , and is the feature after the fusion of the medical text modality feature set and the speech modality feature set ; is the physiological electrical signal modality feature set after dynamic interaction , and is the feature after the fusion of the medical text modality feature set and the physiological electrical signal modality feature set ; is the speech modality feature set after dynamic interaction , and is the feature after the fusion of the speech modality feature set and the physiological electrical signal modality feature set ; The structured feature set is denoted as F, and .

[0020] M2. Evaluation and Optimization Module: It performs confidence scoring on the structured feature set. The confidence scoring is dynamically weighted and calculated through the medical text feature stability index, speech clarity factor, and physiological electrical signal factor. It compares the confidence scoring with the confidence scoring threshold to determine whether it is sensitive information. The retrieval path for sensitive information corresponds to the authoritative knowledge base, and the retrieval path for non-sensitive information corresponds to the general knowledge base, and outputs the optimal medical data reference set; The calculation steps of the confidence scoring include calculating the medical text feature stability index, calculating the speech clarity factor, and calculating the physiological electrical signal clarity factor; The medical text feature stability index The calculation process is as follows: Calculate the mean of the medical text modality feature set Calculate the standard deviation of the medical text modality feature set And finally calculate the medical text feature stability index , and the calculation formula is: ; Among them, is the mean of the medical text modality feature set ; is the number of medical text modality features; is the th medical text modality feature; ; Among them, is the standard deviation of the medical text modality feature set ; ; Among them, is the medical text feature stability index.

[0021] The speech clarity factor The calculation process is as follows: Calculate the power of the speech signal, calculate the power of the noise in the speech signal, and finally calculate the speech clarity factor. The calculation formula is: ; Among them, is the power of the speech signal; is the duration of the speech signal; is the value of the speech signal at time , that is, the amplitude of the speech signal; ; Among them, is the power of the noise in the speech signal, indicating the intensity of the background noise; is the duration of the noise signal; The value of the noise signal at a certain time ; ; wherein, is the signal-to-noise ratio of the voice signal; ; wherein, is the voice clarity factor; is a coefficient related to the clarity of the voice signal, used to adjust the influence of the factor.

[0022] The physiological electrical signal clarity factor is calculated by calculating the power of the physiological electrical signal, calculating the power of the noise in the physiological electrical signal, and finally calculating the physiological electrical signal clarity factor , and the calculation process is the same as that of the voice clarity factor , and finally the physiological electrical signal clarity factor is obtained ; Based on the medical text feature stability index , voice clarity factor and physiological electrical signal clarity factor calculate the confidence score , and the calculation formula is: ; wherein, , and are weight coefficients, and ; corresponds to the weight coefficient of the medical text feature stability index ; corresponds to the weight coefficient of the voice clarity factor ; corresponds to the weight coefficient of the physiological electrical signal clarity factor ; In this embodiment, , , ;

[0023] Set the confidence score threshold to 0.8; when , it is determined as sensitive information, and the corresponding retrieval path is the authoritative knowledge base; when , it is determined as non-sensitive information, and the corresponding retrieval path is the general knowledge base; finally, output the optimal medical data reference set according to the retrieval results of different paths.

[0024] Use the features of the optimal medical data reference set as input data to calculate the loss function , the loss function depends on the parameter , and the calculation formula is: ; Among them, is the number of feature samples; is the true value, that is, the actually observed data; is the predicted value, that is, the predicted output of the multi-branch neural network under the current parameter; is the multi-branch neural network parameter of the th iteration; In each iteration, calculate the loss function for the current network parameter gradient , and the calculation formula is: ; Among them, is the gradient of the th iteration, indicating the partial derivative of the loss function with respect to the multi-branch neural network parameter ; is the partial derivative; In each iteration, the learning rate is dynamically adjusted according to the historical information of the gradient to adapt to the data and gradient changes in network training, and the calculation formula is: ; Among them, is the weighted average of the squares of the gradients in the th iteration; The weighted average of the squares of the gradients in the th iteration; is the decay rate, and ; ; Among them, is the adaptive learning rate; is the initial learning rate; is a constant, and ; In this embodiment, ; Based on the gradient and the adaptive learning rate update the multi-branch neural network parameter , and the calculation formula is: ; Among them, is the multi-branch neural network parameter after the th iteration; For the multi-branch neural network parameters at the th iteration.

[0025] M3, the state evaluation module, performs weighted processing on the reference feature vectors inside the optimal medical data reference set through dynamic weighting, calculates the depression state score using the MM-RNN model based on the processed reference feature vectors, generates a dynamic depression threshold for depression state evaluation according to the dynamic threshold function; and determines the depression state level based on the dynamic threshold; Denote the optimal medical data reference set as , and , where is the weighted fusion reference feature vector after linear fusion of the th medical text data, voice data, and physiological electrical signal data; Performs weighted processing on the reference feature vectors inside the optimal medical data reference set through feature weighting, and the calculation formula is: ; where is the weight coefficient of the th reference feature vector at time ; is the number of reference feature vectors; is the time point most relevant to the th reference feature vector; is the smoothing parameter, used to control the sensitivity of weighting; In this embodiment, ; During the evaluation process, first, it is necessary to calculate the weighted dynamic feature set through dynamic weighting, and the calculation formula is: ; Based on the dynamic feature set , using the MM-RNN model, dynamically learn the temporal dependence relationship of the reference feature vectors, and the calculation formula is: ; where is the hidden state at time step , representing the reference feature vector representation at the current time step; is the activation function, used to introduce a non-linear relationship; is the weight matrix of the hidden layer, used to connect the hidden state of the previous moment and the input of the current moment; is the hidden state of the previous time step; is the bias term of the hidden layer.

[0026] Finally, calculate the depression state score: ; Wherein, is the output depression state score, ranging from [0, 1], indicating the patient's current depression state; is the weight matrix of the output layer; is the bias term of the output layer; The generation function of the dynamic depression threshold is as follows: ; B ; Wherein, and are both the adjusted dynamic depression thresholds; is the initially set depression threshold, and ; is the change amplitude factor, and The value range of is [0.1, 0.3]; The decay factor that controls the state change rate; is the Gaussian noise term, used to simulate the natural fluctuations of the mental state; is the offset constant, and ; Based on the dynamic depression threshold, determine the depression state level: ; Wherein, is mild depression; is moderate depression; is severe depression.

[0027] The above is only the specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present application can easily think of changes or substitutions, which should all be covered by the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claimed rights.

[0028] Finally: The above is only the preferred embodiment of the present invention and is not used to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should all be included in the protection scope of the present invention.

Claims

1. A depression assessment system based on medical data processing, characterized in that, Including: M1, a data fusion module, which extracts modal features from medical text data, voice data, and physiological electrical signal data through a multi-branch neural network, and uses a bidirectional cross-attention mechanism to dynamically interact with the modal features to generate a structured feature set; M2, an evaluation and optimization module, which performs a confidence score on the structured feature set. The confidence score is weighted and calculated through a medical text feature stability index, a voice clarity factor, and a physiological electrical signal clarity factor. The confidence score is compared with a confidence score threshold to determine whether it is sensitive information. The retrieval path for sensitive information corresponds to an authoritative knowledge base, and the retrieval path for non-sensitive information corresponds to a general knowledge base, and an optimal medical data reference set is output; Based on the optimal medical data reference set, an adaptive learning rate strategy is used to iteratively update the parameters of the multi-branch neural network; M3, a state evaluation module, which dynamically weights the reference feature vectors within the optimal medical data reference set, calculates a depression state score using an MM-RNN model based on the processed reference feature vectors, and generates a dynamic depression threshold for depression state evaluation according to a dynamic threshold function; and determines the depression state level based on the dynamic threshold.

2. The depression assessment system based on medical data processing according to claim 1, characterized in that The medical text data mainly comes from the medical records of patients and doctor's diagnosis reports; the voice data mainly comes from the voice symptom descriptions of patients and related voice records; the physiological electrical signal data mainly comes from the electrocardiograms and electroencephalograms of patients; The multi-branch neural network includes a BERT model, a self-supervised voice model, and a CNN-LSTM hybrid model; The extraction process of the modal features is as follows: the medical text data is used to extract a medical text modal feature set through the BERT model , and the speech data is used to extract a speech modal feature set through a self-supervised speech model , and the physiological electrical signal data is used to extract a physiological electrical signal modal feature set through a CNN-LSTM hybrid model .

3. The depression assessment system based on medical data processing according to claim 2, wherein The dynamic interaction process is as follows: Input the medical text modal feature set, voice modal feature set, and physiological electrical signal modal feature set into the bidirectional cross-attention mechanism to obtain an attention matrix, and then use the attention matrix to dynamically weight the medical text modal feature set, voice modal feature set, and physiological electrical signal modal feature set; Dynamically interact the feature sets of medical text modalities, speech modalities, and physiological electrical signal modalities through the bidirectional cross-attention mechanism , speech modality feature sets and physiological electrical signal modality feature sets ; After the calculation of the bidirectional cross-attention mechanism, the obtained attention matrix is used to perform dynamic weighted fusion on the medical text modality feature set , the speech modality feature set , and the physiological electrical signal modality feature set . The calculation formula is as follows: ； ； ； Among them, is the medical text modal feature set after dynamic interaction , and is the medical text modal feature set is the voice modal feature set is the fused feature; is the physiological electrical signal modal feature set after dynamic interaction , and is the medical text modal feature set and the physiological electrical signal modal feature set is the fused feature; is the voice modal feature set after dynamic interaction , and is the voice modal feature set and the physiological electrical signal modal feature set is the fused feature; is the medical text modal feature set and the voice modal feature set is the attention matrix between them; is the medical text modal feature set and the physiological electrical signal modal feature set is the attention matrix between them; is the voice modal feature set and the physiological electrical signal modal feature set is the attention matrix between them; The structured feature set is denoted as , and .

4. The depression assessment system based on medical data processing according to claim 2, wherein The calculation steps of the confidence score include calculating the medical text feature stability index, calculating the voice clarity factor, and calculating the physiological electrical signal clarity factor; The stability index of the medical text features The calculation process is as follows: calculate the mean of the medical text modality feature set , calculate the standard deviation of the medical text modality feature set and finally calculate the stability index of the medical text features ; The speech intelligibility factor is calculated by calculating the power of the speech signal, calculating the power of the noise in the speech signal, and finally calculating the speech intelligibility factor ; The physiological electrical signal clarity factor The calculation process is as follows: calculate the power of the physiological electrical signal, calculate the power of the noise in the physiological electrical signal, and finally calculate the physiological electrical signal clarity factor , and finally obtain the physiological electrical signal clarity factor ; Based on the medical text feature stability index , speech clarity factor and physiological electrical signal clarity factor calculate the confidence score , and the calculation formula is: ； Among them, , and are weight coefficients, and ; is the weight coefficient corresponding to the stability index of medical text features ; is the weight coefficient corresponding to the voice clarity factor ; is the weight coefficient corresponding to the physiological electrical signal clarity factor ; Set the confidence score threshold to 0.8; when it is judged as sensitive information, and the corresponding retrieval path is the authoritative knowledge base; when it is judged as non-sensitive information, and the corresponding retrieval path is the general knowledge base; finally, the optimal medical data reference set is output according to the retrieval results of different paths.

5. The depression assessment system based on medical data processing according to claim 1, characterized in that Denote the optimal medical data reference set as , and , where is the reference feature vector after linear fusion of the th medical text data, voice data, and physiological electrical signal data; Dynamically weight the reference feature vectors within the optimal medical data reference set, and the calculation formula is: ； Among them, is at the weight coefficient of the -th reference feature vector at the moment; is the number of reference feature vectors; is the time point most relevant to the -th reference feature vector; is a smoothing parameter used to control the sensitivity of weighting; During the evaluation process, it is first necessary to calculate the weighted dynamic feature set through dynamic weighting , and the calculation formula is: ； Based on the dynamic feature set , after dynamically learning the temporal dependence relationship of the reference feature vector by using the MM-RNN model, calculate the depression state score: ； Among them, is the output depression state score, ranging from [0, 1], indicating the patient's current depression state; is the weight matrix of the output layer; is the bias term of the output layer; is at time step is the hidden state; is the activation function; The generation function of the dynamic depression threshold is as follows: ； ； Among them, and are both adjusted dynamic depression thresholds; is the initially set depression threshold; is the change amplitude factor; is the decay factor for controlling the state change rate; is the Gaussian noise term used to simulate the natural fluctuations of the mental state; is the offset constant; Based on the dynamic depression threshold, determine the depression state level: ； Among them, is mild depression; is moderate depression; is severe depression.

6. The depression assessment system based on medical data processing according to claim 1, wherein Calculate the loss function using the features of the optimal medical data reference set as input data , the loss function depends on parameters , in each iteration, calculate the loss function for the current network parameters gradient , while the learning rate is dynamically adjusted according to the historical information of the gradient, and finally update the multi-branch neural network parameters based on the gradient and the adaptive learning rate .

Images