Rehabilitation result prediction method based on chain type encoder rehabilitation prediction model

Abstract

The invention discloses a rehabilitation result prediction method based on a chain encoder rehabilitation prediction model, which comprises the following steps of: firstly collecting rehabilitation record data of a patient, then constructing an encoder and a decoder, and training by using electroencephalogram data of the patient, namely pre-training of resting electroencephalogram reduction for the encoder; the method comprises the following steps of: firstly, training different types of encoders by using electroencephalograms related to three different events, calculating the importance of different electroencephalogram channels by using an attention layer, constructing an encoder prediction block in combination with a state vector of a patient, and finally constructing a chain type encoder rehabilitation prediction model by using a cascade training method. Weight parameters in the attention layer of the model are extracted, the importance of different electroencephalogram channels on prediction is obtained, and rehabilitation prediction is completed. According to the chain type encoder structure constructed by the method, heterogeneous data are fused, the periodicity of the rehabilitation process and the influence of different electroencephalogram channels on the rehabilitation effect are fully considered, the accuracy of rehabilitation prediction is improved, and decision support can be provided for doctors.

Description

Technical Field

[0001] The invention belongs to the technical field of rehabilitation prediction, and in particular relates to a rehabilitation result prediction method based on a chain encoder rehabilitation prediction model. Background Art

[0002] In the field of rehabilitation prediction, existing methods are usually based on statistical models, such as linear regression, support vector machines, etc., to establish relationship models between diseases and related features. However, these methods have certain limitations when dealing with large-scale medical data and high-dimensional features, and cannot well capture the complexity of nonlinear relationships and high-dimensional features. With the development and application of machine learning and deep learning technologies, their powerful representation learning capabilities and parallel computing capabilities provide new methods for rehabilitation prediction.

[0003] CN107280666A provides a method and system for predicting postoperative rehabilitation of deaf patients after CI based on machine learning. First, the patient's EEG signal is obtained and preprocessed, and features are extracted from the preprocessed signal. A training set and a test set are selected, and a kernel function is selected to predict rehabilitation through a support vector machine. This method only considers a single EEG signal, and the manual feature extraction is limited by experience. The selected features have a direct impact on the final prediction result.

[0004] CN118285790A provides a motor function rehabilitation prediction method based on multimodal asynchronous data and related equipment, firstly obtaining the patient's bioelectric signal and electroencephalography data and preprocessing, then extracting features to obtain the first feature vector, the second feature vector, the third feature vector and the fourth feature vector, and obtaining the target feature vector through feature fusion, and finally inputting the target feature vector into the trained model to obtain the rehabilitation prediction result. This rehabilitation prediction method integrates different modal data, but ignores the process characteristics of patient rehabilitation.

[0005] In summary, existing technologies rarely consider heterogeneous data in rehabilitation prediction, and do not take into account the procedural nature of patient rehabilitation, as well as the different degrees of influence of different EEG channels on rehabilitation outcomes, which affects the accuracy of the final results. Summary of the invention

[0006] To solve the above technical problems, the present invention provides a rehabilitation outcome prediction method based on a chain encoder rehabilitation prediction model, which predicts the recovery of the next stage according to the EEG data and scale data of the patient's rehabilitation process, thereby improving the accuracy of rehabilitation prediction.

[0007] The technical solution adopted by the present invention is: a rehabilitation result prediction method based on a chain encoder rehabilitation prediction model, and the specific steps are as follows:

[0008] S1. Collect patient rehabilitation record data;

[0009] Among them, the patient's EEG data, scale data and intervention methods collected at each stage are called rehabilitation records; the patient's EEG data include: the patient's resting state EEG data and event-related potential EEG data.

[0010] S2, constructing an encoder and a decoder, and training the encoder and the decoder based on the patient's resting state EEG data collected in step S1;

[0011] S3, based on the encoder and decoder pre-trained in step S2, use event-related potential EEG data for secondary training;

[0012] S4, constructing an encoder prediction block based on the encoder and decoder trained in step S3;

[0013] S5. Based on step S4, a cascade training method is used to construct a rehabilitation prediction model based on a chain encoder, that is, prediction training is performed on the encoder prediction block;

[0014] S6. Based on the trained rehabilitation prediction model obtained in step S5, the rehabilitation record data of the patient to be predicted is input to obtain a rehabilitation prediction result.

[0015] Furthermore, the step S1 is specifically as follows:

[0016] First, obtain the scale score of the patient during the rehabilitation process, and record the scale score of the mth stage in the vector Sc m In the same time, the intervention methods implemented on the patient during the patient's rehabilitation process are recorded in the vector in m .

[0017] Among them, the scales include: Hamilton Depression Rating Scale HAMD, Stressful Event Questionnaire SRSS, Hamilton Anxiety Rating Scale HAMA, Pittsburgh Sleep Quality Index Score Change PSQI; the intervention methods include: acupuncture + medication, non-meridian and non-acupoint, symmetrical acupuncture, Tai Chi and placebo.

[0018] Two types of EEG data were collected using a multi-electrode brain cap:

[0019] 1) Patient's resting EEG data: The patient sits quietly and stays awake, and the patient's resting EEG with eyes closed is recorded for q minutes.

[0020] The body should not make any unnecessary movements, including shaking, blinking, or shaking the head, and the eyes should not make any excessive glances or movements. qSet according to the actual situation.

[0021] 2) Event-related potential EEG data: The patients watched three types of pictures, including three themes: happiness, calmness, and fear. Whenever a picture appeared on the screen, the patient had to quickly press a key with his left hand to insert a mark in the EEG signal at the time of the key press, and record the EEG signal 1 second before the mark and 2 seconds after the mark.

[0022] There are p pictures of each type, and p, s1, and s2 are set according to actual conditions.

[0023] After the resting state and event-related potential tests were completed, the patient's EEG data in the resting state and the event-related potential EEG data were obtained, and the EEG data were preprocessed using EEGLAB as follows:

[0024] (1) Downsampling: downsampling the sampling rate from 1000 Hz to 100 Hz;

[0025] (2) Filtering: Use a notch filter to remove mains artifacts between 49 Hz and 50 Hz, remove low-frequency noise below 1 Hz, and remove high-frequency noise above 100 Hz;

[0026] (3) ICA analysis: ICA algorithm runica in EEGLAB is used to perform ICA decomposition, decomposing the noise signals of electrooculography, electromyography, channel noise and electrocardiogram, and removing them;

[0027] (4) segment the resting-state EEG data, with each segment length being s2-(-s1) seconds;

[0028] (5) Using EEGLAB, extract the EEG data segments before and after observing the pictures from the three types of event-related potential EEG data;

[0029] The length of the event-related potential EEG data segment is s2-(-s1) seconds, and the range is [-s1*1000, s2*1000] (ms), with the unit being milliseconds.

[0030] Furthermore, the step S2 is specifically as follows:

[0031] First, build the encoder and decoder.

[0032] The encoder includes: 3 CNN layers and 1 LSTM layer; the CNN layer includes: convolution layer, ReLU layer and pooling layer.

[0033] Among them, the convolution layer captures EEG features from the resting state EEG data segment and inputs it into the Relu layer and the pooling layer to increase the depth, compress the number of parameters and reduce the feature dimension. The output of the CNN layer is input into the LSTM layer to extract sequential features and output the encoding vector corresponding to the resting state EEG data segment.

[0034] The decoder includes: 1 LSTM layer and 3 CNN layers; the CNN layer includes: upsampling layer, ReLU layer and convolution layer.

[0035] Among them, the LSTM layer restores the encoded vector into sequence data, and then the sequence data is upsampled to restore the data dimension, and finally the CNN layer outputs the restored resting state EEG data fragment.

[0036] Then, the resting-state EEG data in the rehabilitation record obtained in step S1 is used to perform end-to-end pre-training on the encoder and decoder to preliminarily learn the feature representation of EEG.

[0037] The j′th resting state EEG data segment after segmentation in step S1 is represented as e (-，j′) =[e (1，j′) , e (2，j′) , e (3，j′) , …e (i，j′) , …e (c，j′) ].

[0038] Among them, i∈[1,c] represents the channel number, c represents the number of EEG channels, j′∈[1,n′] represents the resting state EEG data segment number, and n′ represents the number of resting state EEG data segments.

[0039] The resting state EEG data segment (-，j′) Input the encoder and get e (-，j′) The corresponding encoding vector x (-，j′) =[x (1，j′) , x (2，j′) , x (3，j′) , …x (c，j′) ].

[0040] Then the encoder outputs the encoded vector x (-，j′) Input to the decoder, the decoder reconstructs the resting state EEG data segment, and the resting state EEG data segment output by the decoder is recorded as

[0041] The loss function expression of encoder pre-training is as follows:

[0042]

[0043] Where n′ represents the number of resting state EEG data segments.

[0044] Finally, the pre-trained encoder and decoder are retained, denoted as encoder E and decoder D.

[0045] Furthermore, the step S3 is specifically as follows:

[0046] The encoder E and decoder D are trained again using the three different types of event-related potential EEG data described in step S1. The encoder and decoder obtained are represented as E k and D k .

[0047] Among them, k∈{1, 2, 3} represents the event-related potential EEG data segment type number corresponding to happiness, calmness, and fear respectively.

[0048] The encoder E and decoder D are trained end-to-end. First, the event-related potential EEG data fragments in the event-related potential EEG dataset are used. As the input of encoder E, encoder E outputs event-related potential EEG data segments The corresponding encoding vector

[0049] Where j∈[1,n], n represents the number of event-related potential EEG data segments.

[0050] Then the encoding vector output by encoder E is As the input of decoder D, decoder D reconstructs the event-related potential EEG data fragment, and the output is expressed as

[0051] Among them, the loss function expression of the encoder E training is as follows:

[0052]

[0053] Where n represents the number of event-related potential EEG data segments.

[0054] Finally, keep the encoder E after training k , k∈{1, 2, 3}.

[0055] Furthermore, the step S4 is specifically as follows:

[0056] The encoder prediction block includes: encoder layer, attention layer, fully connected layer and activation function layer.

[0057] The encoder layer is the encoder E obtained after step S3 training. k .

[0058] The encoder prediction block adds an attention mechanism to calculate the contribution of different brain regions to the rehabilitation outcome. For channel i, the encoding vector of the event-related potential EEG data of type k is The expression for calculating attention is as follows:

[0059]

[0060] in, Represents the potential feature representation of the EEG encoding vector, which is used to calculate the corresponding channel attention value w i , b i Represent weight and bias values ​​respectively; θ represents the EEG channel number, and its value range is [1, c]; express The EEG encoding vector with attention weight output by the attention layer is used to obtain the EEG signals of different segments. Combined, expressed as For different event-related potential EEG data segment type numbers k, Combine and flatten into one-dimensional form, record the attention vector

[0061] In the encoder prediction block, in the patient’s mth stage rehabilitation record data, the attention vector and the patient state vector [SC m , in m ] The combined input is sent to the fully connected layer, and then through the activation function layer Relu to predict the scale vector in the patient's m+1 stage rehabilitation record

[0062] Furthermore, the step S5 is specifically as follows:

[0063] A cascade training method is used to construct a rehabilitation prediction model based on a chain encoder, that is, the encoder prediction block is trained for prediction. When a patient has rehabilitation record data for the mth stage, the encoder prediction block is used to predict the scale vector of the patient's rehabilitation record for the m+1th stage according to the patient's rehabilitation record data. When the patient has data on the rehabilitation record of the m+1th stage, the rehabilitation record scale vector Sc of the m+1th stage can be used. m+1 and the predicted scale vector The loss is constructed, the encoder prediction block is trained, and the encoder parameters are shared for each training process.

[0064] Among them, the encoder prediction block calculates the loss through the mean square error loss function, which is expressed as follows:

[0065]

[0066] Where N represents the scale vector length, Sc m+1 The value of the i-th element, express The value of the i-th element of .

[0067] After training the rehabilitation prediction model, output and retain the model, and output the weight parameters W in the attention layer = [w1, w2, ...wc-1 , w c ], we can get the importance of EEG channels on the rehabilitation outcomes during the patient's rehabilitation process.

[0068] Furthermore, the step S6 is specifically as follows:

[0069] For the patient to be predicted, the patient's rehabilitation record data is input into the rehabilitation prediction model based on the chain encoder trained in step S5, and the HAMD, SRSS, HAMA, and PSQI score results of the next stage of the patient's rehabilitation process are output, that is, the rehabilitation prediction results.

[0070] Beneficial effects of the present invention: The method of the present invention first collects patient rehabilitation record data, then constructs an encoder and a decoder, and uses the patient's EEG data for training, that is, pre-training the encoder for resting EEG restoration, and then uses three different event-related EEGs to train different types of encoders, and then uses the attention layer to calculate the importance of different EEG channels, and combines the patient's state vector to construct an encoder prediction block, and finally uses a cascade training method to construct a chain encoder rehabilitation prediction model, extract the weight parameters in the model's attention layer, obtain the importance of different EEG channels to the prediction, and complete the rehabilitation prediction. The chain encoder structure constructed by the method of the present invention integrates heterogeneous data, fully considers the stages of the rehabilitation process, and the influence of different EEG channels on the rehabilitation effect, improves the accuracy of rehabilitation prediction, and can provide decision support for physicians. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 The present invention is a flowchart of a method for predicting rehabilitation results based on a chain encoder rehabilitation prediction model.

[0072] Figure 2 1 is a structural diagram of an encoder and a decoder in an embodiment of the present invention.

[0073] Figure 3 This is a diagram of the encoder prediction block structure in an embodiment of the present invention.

[0074] Figure 4 Schematic diagram of the structure of a rehabilitation prediction model based on a chain encoder in an embodiment of the present invention. DETAILED DESCRIPTION

[0075] The method of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0076] like Figure 1 As shown, a flowchart of a rehabilitation result prediction method based on a chain encoder rehabilitation prediction model of the present invention, the specific steps are as follows:

[0077] S1. Collect patient rehabilitation record data;

[0078] Among them, the patient's EEG data, scale data and intervention methods collected at each stage are called rehabilitation records; the patient's EEG data include: the patient's resting state EEG data and event-related potential EEG data.

[0079] S2, constructing an encoder and a decoder, and training the encoder and the decoder based on the patient's resting state EEG data collected in step S1;

[0080] S3, based on the encoder and decoder pre-trained in step S2, use event-related potential EEG data for secondary training;

[0081] S4, constructing an encoder prediction block based on the encoder and decoder trained in step S3;

[0082] S5. Based on step S4, a cascade training method is used to construct a rehabilitation prediction model based on a chain encoder, that is, prediction training is performed on the encoder prediction block;

[0083] S6. Based on the trained rehabilitation prediction model obtained in step S5, the rehabilitation record data of the patient to be predicted is input to obtain a rehabilitation prediction result.

[0084] In this embodiment, the step S1 is specifically as follows:

[0085] First, obtain the scale score of the patient during the rehabilitation process, and record the scale score of the mth stage in the vector Sc m In the same time, the intervention methods implemented on the patient during the patient's rehabilitation process are recorded in the vector in m .

[0086] Among them, the scales include: Hamilton Depression Rating Scale HAMD, Stressful Event Questionnaire SRSS, Hamilton Anxiety Rating Scale HAMA, Pittsburgh Sleep Quality Index Score Change PSQI; the intervention methods include: acupuncture + medication, non-meridian and non-acupoint, symmetrical acupuncture, Tai Chi and placebo.

[0087] Use a multi-electrode brain cap (a 64-electrode brain cap is used in this embodiment) to collect two types of EEG data:

[0088] 1) Patient's resting EEG data: The patient sits quietly and stays awake, and the patient's resting EEG with eyes closed is recorded for q = 2 minutes.

[0089] The body should not have any unnecessary movements, including shaking, blinking, or shaking the head, and the eyes should not have excessive scanning and movement.

[0090] 2) Event-related potential EEG data: The patients watched three types of pictures, including three themes: happiness, calmness, and fear. Whenever a picture appeared on the screen, the patient had to quickly press a key with his left hand to insert a mark in the EEG signal at the time of the key press, and record the EEG signal 1 second before the mark and 2 seconds after the mark.

[0091] There are p=60 pictures of each type, and the resolution of the machine playing the pictures is 1980*1024; in this embodiment, s1=1 and s2=2.

[0092] After the resting state and event-related potential tests were completed, the patient's EEG data in the resting state and the event-related potential EEG data were obtained, and the EEG data were preprocessed using EEGLAB as follows:

[0093] (1) Downsampling: downsampling the sampling rate from 1000 Hz to 100 Hz;

[0094] (2) Filtering: Use a notch filter to remove mains artifacts between 49 Hz and 50 Hz, remove low-frequency noise below 1 Hz, and remove high-frequency noise above 100 Hz;

[0095] (3) ICA analysis: ICA algorithm runica in EEGLAB is used to perform ICA decomposition, decomposing the noise signals of electrooculography, electromyography, channel noise and electrocardiogram, and removing them;

[0096] (4) segment the resting-state EEG data, with each segment length being s2-(-s1) seconds;

[0097] (5) Using EEGLAB, extract the EEG data segments before and after observing the pictures from the three types of event-related potential EEG data;

[0098] The length of the event-related potential EEG data segment is 3 seconds, and the range is [-1000, 2000] (ms), with the unit being milliseconds.

[0099] In this embodiment, step S2 is specifically as follows:

[0100] First, construct the encoder and decoder, the structure is as follows Figure 2 shown.

[0101] The encoder includes: 3 CNN layers and 1 LSTM layer; the CNN layer includes: convolution layer, ReLU layer and pooling layer.

[0102] Among them, the convolution layer captures EEG features from the resting state EEG data segment and inputs it into the Relu layer and the pooling layer to increase the depth, compress the number of parameters and reduce the feature dimension. The output of the CNN layer is input into the LSTM layer to extract sequential features and output the encoding vector corresponding to the resting state EEG data segment.

[0103] The decoder includes: 1 LSTM layer and 3 CNN layers; the CNN layer includes: upsampling layer, ReLU layer and convolution layer.

[0104] Among them, the LSTM layer restores the encoded vector into sequence data, and then the sequence data is upsampled to restore the data dimension, and finally the CNN layer outputs the restored resting state EEG data fragment.

[0105] Then, the resting-state EEG data in the rehabilitation record obtained in step S1 is used to perform end-to-end pre-training on the encoder and decoder to preliminarily learn the feature representation of EEG.

[0106] The j′th resting state EEG data segment after segmentation in step S1 is represented as e (-，j′) =[e (1，j′) , e (2，j′) , e (3，j′) , …e (i，j′) , …e (c，j′) ].

[0107] Among them, i∈[1,c] represents the channel number, c represents the number of EEG channels (64 in this embodiment), j′∈[1,n′] represents the resting state EEG data segment number, and n′ represents the number of resting state EEG data segments.

[0108] The resting state EEG data segment (-，j′) Input the encoder and get e (-，j′) The corresponding encoding vector x (-，j′ )=[x (1，j′) , x (2，j′) , x (3，j′) , …x (c，j′) ].

[0109] Then the encoder outputs the encoded vector x (-，j′) Input to the decoder, the decoder reconstructs the resting state EEG data segment, and the resting state EEG data segment output by the decoder is recorded as

[0110] The loss function expression of encoder pre-training is as follows:

[0111]

[0112] Where n′ represents the number of resting state EEG data segments.

[0113] Finally, the pre-trained encoder and decoder are retained, denoted as encoder E and decoder D.

[0114] In this embodiment, step S3 is specifically as follows:

[0115] In order to learn the EEG data features of three different types of event-related potentials, the encoder E and decoder D are trained again using the three different types of event-related potential EEG data described in step S1. The encoder and decoder obtained are represented as E k and D k .

[0116] Among them, k∈{1, 2, 3} represents the event-related potential EEG data segment type number corresponding to happiness, calmness, and fear respectively.

[0117] The encoder E and decoder D are trained end-to-end. First, the event-related potential EEG data fragments in the event-related potential EEG dataset are used. As the input of encoder E, encoder E outputs event-related potential EEG data segments The corresponding encoding vector

[0118] Where j∈[1,n], n represents the number of event-related potential EEG data segments.

[0119] Then the encoding vector output by encoder E is As the input of decoder D, decoder D reconstructs the event-related potential EEG data fragment, and the output is expressed as

[0120] Among them, the loss function expression of the encoder E training is as follows:

[0121]

[0122] Where n represents the number of event-related potential EEG data segments.

[0123] Finally, keep the encoder E after training k , k∈{1, 2, 3}.

[0124] In this embodiment, step S4 is specifically as follows:

[0125] like Figure 3 As shown, the encoder prediction block includes: encoder layer, attention layer, fully connected layer and activation function layer.

[0126] The encoder layer is the encoder E obtained after step S3 training. k .

[0127] In order to better explore the role of different EEG channels, an attention mechanism is added to calculate the contribution of different brain regions to the rehabilitation results. For channel i, the encoding vector of event-related potential EEG data of type k is The expression for calculating attention is as follows:

[0128]

[0129] in, Represents the potential feature representation of the EEG encoding vector, which is used to calculate the corresponding channel attention value w i 、b i Represent weight and bias values ​​respectively; θ represents the EEG channel, and its value range is [1, c]; express The EEG encoding vector with attention weight output by the attention layer is used to obtain the EEG signals of different segments. Combined, expressed as For different event-related potential EEG data segment type numbers k, Combine and flatten into one-dimensional form, record the attention vector

[0130] In the encoder prediction block, in the patient’s mth stage rehabilitation record data, the attention vector and the patient state vector [Sc m , in m ] The combined input is sent to the fully connected layer, and then through the activation function layer Relu to predict the scale vector in the patient's m+1 stage rehabilitation record

[0131] In this embodiment, step S5 is specifically as follows:

[0132] like Figure 4 As shown, a cascade training method is used to construct a rehabilitation prediction model based on a chain encoder, that is, the encoder prediction block is trained for prediction. When the patient has the mth stage rehabilitation record data (i.e. Figure 4 When the mth rehabilitation record data is obtained, the encoder prediction block is used to predict the scale vector of the patient's m+1th stage rehabilitation record according to the patient's rehabilitation record data. When the patient has data on the rehabilitation record of the m+1th stage, the rehabilitation record scale vector Sc of the m+1th stage can be used. m+1 and the predicted scale vector The loss is constructed, the encoder prediction block is trained, and the encoder parameters are shared for each training process.

[0133] Among them, the encoder prediction block calculates the loss through the mean square error loss function, which is expressed as follows:

[0134]

[0135] Where N represents the scale vector length, Sc m+1 The value of the i-th element, express The value of the i-th element of .

[0136] After training the rehabilitation prediction model, output and retain the model, and output the weight parameters W in the attention layer = [w1, w2, ...w c-1 , w c ], the importance of EEG channels on the rehabilitation outcomes during the patient’s rehabilitation process (channel importance distribution) can be obtained.

[0137] In this embodiment, step S6 is specifically as follows:

[0138] For the patient to be predicted, the patient's rehabilitation record data is input into the rehabilitation prediction model based on the chain encoder trained in step S5, and the HAMD, SRSS, HAMA, and PSQI score results of the next stage of the patient's rehabilitation process are output, that is, the rehabilitation prediction results.

[0139] In summary, the chain encoder structure constructed by the method of the present invention integrates heterogeneous data, fully considers the stages of the rehabilitation process and the influence of different EEG channels on the rehabilitation effect, improves the accuracy of rehabilitation prediction, and can provide decision support for doctors.

[0140] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A rehabilitation outcome prediction method based on a chain encoder rehabilitation prediction model, the specific steps are as follows: S1. Collect patient rehabilitation record data; in, The EEG data, scale data and intervention methods collected by patients at each stage are called rehabilitation records; the EEG data of patients include: resting state EEG data of patients and event-related potential EEG data; S2, constructing an encoder and a decoder, and training the encoder and the decoder based on the patient's resting state EEG data collected in step S1; S3, based on the encoder and decoder pre-trained in step S2, use event-related potential EEG data for secondary training; S4, constructing an encoder prediction block based on the encoder and decoder trained in step S3; S5. Based on step S4, a cascade training method is used to construct a rehabilitation prediction model based on a chain encoder, that is, prediction training is performed on the encoder prediction block; S6. Based on the trained rehabilitation prediction model obtained in step S5, the rehabilitation record data of the patient to be predicted is input to obtain a rehabilitation prediction result.

2. A rehabilitation outcome prediction method based on a chain encoder rehabilitation prediction model according to claim 1, characterized in that: The step S1 is specifically as follows: First, obtain the scale score of the patient during the rehabilitation process, and record the scale score of the mth stage in the vector Sc m At the same time, the intervention methods implemented on the patient during the patient's rehabilitation process at this stage are recorded in vector in m ; The scales include: Hamilton Depression Rating Scale (HAMD), Stressful Events Questionnaire (SRSS), Hamilton Anxiety Rating Scale (HAMA), Pittsburgh Sleep Quality Index (PSQI); the intervention methods include: acupuncture + medication, non-meridian and non-acupoint, symmetrical acupuncture, Tai Chi and placebo; Two types of EEG data were collected using a multi-electrode brain cap: 1) Patient's resting EEG data: The patient sits quietly and stays awake, and the patient's resting EEG with eyes closed is recorded for q minutes; The body cannot make unnecessary movements, including shaking, blinking, and shaking the head, and the eyes cannot make excessive scans and movements; q should be set according to the actual situation; 2) Event-related potential EEG data: The patient watched three types of pictures, including three themes: happy, calm, and frightened. Whenever a picture appeared on the screen, the patient was required to quickly press a key with his left hand to insert a mark in the EEG signal at the time of the key press, and record the EEG signal 1 second before the mark and 2 seconds after the mark; There are p pictures of each type, and p, s1, and s2 are set according to the actual situation; After the resting state and event-related potential tests were completed, the patient's EEG data in the resting state and the event-related potential EEG data were obtained, and the EEG data were preprocessed using EEGLAB as follows: (1) Downsampling: downsampling the sampling rate from 1000 Hz to 100 Hz; (2) Filtering: Use a notch filter to remove mains artifacts between 49 Hz and 50 Hz, remove low-frequency noise below 1 Hz, and remove high-frequency noise above 100 Hz; (3) ICA analysis: ICA algorithm runica in EEGLAB is used to perform ICA decomposition, decomposing the noise signals of electrooculography, electromyography, channel noise and electrocardiogram, and removing them; (4) segment the resting-state EEG data, with each segment length being s2-(-s1) seconds; (5) Using EEGLAB, extract the EEG data segments before and after observing the pictures from the three types of event-related potential EEG data; The length of the event-related potential EEG data segment is s2-(-s1) seconds, and the range is [-s1*1000, s2*1000] (ms).

3. The rehabilitation result prediction method based on the chain encoder rehabilitation prediction model according to claim 1 is characterized in that: The step S2 is specifically as follows: First, build the encoder and decoder; The encoder includes: 3 CNN layers and 1 LSTM layer; the CNN layer includes: convolution layer, Relu layer and pooling layer; Among them, the convolution layer captures EEG features from the resting state EEG data fragments, and inputs them into the Relu layer and the pooling layer to increase the depth, compress the number of parameters and reduce the feature dimension; the output of the CNN layer is input into the LSTM layer to extract sequential features and output the encoding vector corresponding to the resting state EEG data fragment; The decoder includes: 1 LSTM layer and 3 CNN layers; the CNN layer includes: upsampling layer, ReLU layer and convolution layer; Among them, the LSTM layer restores the encoded vector into sequence data, and then the sequence data is upsampled to restore the data dimension, and finally the CNN layer outputs the restored resting state EEG data fragment; Then, the encoder and decoder are end-to-end pre-trained using the resting-state EEG data in the rehabilitation record obtained in step S1 to preliminarily learn the feature representation of EEG; The j′th resting state EEG data segment after segmentation in step S1 is represented as e (-，j′) =[e (1，j′) , e (2，j′) , e (3，j′) , …e (i，j′) , …e (c，j′) ]; Among them, i∈[1,c] represents the channel number, c represents the number of EEG channels, j′∈[1,n′] represents the resting state EEG data segment number, and n′ represents the number of resting state EEG data segments; The resting state EEG data segment (-，j′) Input the encoder and get e (-，j′) The corresponding encoding vector x (-，j′) =[x (1，j′) , x (2，j′) , x (3，j′) , …x (c，j′) ]; Then the encoder outputs the encoded vector x (-，j′) Input to the decoder, the decoder reconstructs the resting state EEG data segment, and the resting state EEG data segment output by the decoder is recorded as The loss function expression of encoder pre-training is as follows: Where n′ represents the number of resting-state EEG data segments; Finally, the pre-trained encoder and decoder are retained, denoted as encoder E and decoder D.

4. The rehabilitation result prediction method based on the chain encoder rehabilitation prediction model according to claim 1, characterized in that: The step S3 is specifically as follows: The encoder E and decoder D are trained again using the three different types of event-related potential EEG data described in step S1. The encoder and decoder obtained are represented as E k and D k ; Wherein, k∈{1, 2, 3} represents the event-related potential EEG data segment type number corresponding to happiness, calmness, and fear respectively; The encoder E and decoder D are trained end-to-end. First, the event-related potential EEG data fragments in the event-related potential EEG dataset are used. As the input of encoder E, encoder E outputs event-related potential EEG data segments The corresponding encoding vector Where, j∈[1,n], n represents the number of event-related potential EEG data segments; Then the encoding vector output by encoder E is As the input of decoder D, decoder D reconstructs the event-related potential EEG data fragment, and the output is expressed as Among them, the loss function expression of the encoder E training is as follows: Where n represents the number of event-related potential EEG data segments; Finally, keep the encoder E after training k , k∈{1, 2, 3}.

5. The rehabilitation result prediction method based on the chain encoder rehabilitation prediction model according to claim 1, characterized in that: The step S4 is specifically as follows: The encoder prediction block includes: encoder layer, attention layer, fully connected layer and activation function layer; The encoder layer is the encoder E obtained after step S3 training. k ; The encoder prediction block adds an attention mechanism to calculate the contribution of different brain regions to the rehabilitation outcome. For channel i, the encoding vector of the event-related potential EEG data of type k is The expression for calculating attention is as follows: in, Represents the potential feature representation of the EEG encoding vector, which is used to calculate the corresponding channel attention value w i , b i Represent weight and bias values ​​respectively; θ represents the EEG channel number, and its value range is [1, c]; express The EEG encoding vector with attention weight output by the attention layer is used to obtain the EEG signals of different segments. Combined, expressed as For different event-related potential EEG data segment type numbers k, Combine and flatten into a one-dimensional form, record the attention vector In the encoder prediction block, in the patient’s mth stage rehabilitation record data, the attention vector and the patient state vector [Sc m , in m ] The combined input is sent to the fully connected layer, and then through the activation function layer Relu, to predict the scale vector in the patient's m+1 stage rehabilitation record 6. The rehabilitation result prediction method based on the chain encoder rehabilitation prediction model according to claim 1, characterized in that: The step S5 is specifically as follows: A cascade training method is used to construct a rehabilitation prediction model based on a chain encoder, that is, the encoder prediction block is trained for prediction. When a patient has rehabilitation record data for the mth stage, the encoder prediction block is used to predict the scale vector of the patient's rehabilitation record for the m+1th stage according to the patient's rehabilitation record data. When the patient has data on the rehabilitation record of the m+1th stage, the rehabilitation record scale vector Sc of the m+1th stage can be used. m+1 and the predicted scale vector Construct the loss and train the encoder prediction block, and for each training process, the encoder parameters are shared; Among them, the encoder prediction block calculates the loss through the mean square error loss function, which is expressed as follows: Where N represents the scale vector length, Sc m+1 The value of the i-th element, express The value of the i-th element of ; After training the rehabilitation prediction model, output and retain the model, and output the weight parameters W in the attention layer = [w1, w2, ...w c-1 , w c ], we can get the importance of EEG channels on the rehabilitation outcomes during the patient's rehabilitation process.

7. The rehabilitation result prediction method based on the chain encoder rehabilitation prediction model according to claim 1 is characterized in that: The step S6 is specifically as follows: For the patient to be predicted, the patient's rehabilitation record data is input into the rehabilitation prediction model based on the chain encoder trained in step S5, and the HAMD, SRSS, HAMA, and PSQI score results of the next stage of the patient's rehabilitation process are output, that is, the rehabilitation prediction results.

Images