Parkinson freezing gait event detection method based on LSTM and GRU fusion model

Abstract

The invention discloses a Parkinson's freezing gait event detection method based on an LSTM and GRU fusion model, and belongs to the field of intelligent medical treatment. According to the method, a fusion model based on LSTM and GRU is constructed for a three-axis acceleration time sequence provided by a subject receiving a Parkinson's freezing gait excitation experiment, input data is preprocessed through filtering, normalization, feature extraction, label displacement and the like, multi-head attention mechanism encoders are combined to assist encoding, training data are expanded by using pseudo labels, and the accuracy of training data is improved. The Parkinson's freezing gait event is detected, and high accuracy and correctness are achieved on a large number of data samples.

Description

Technical Field

[0001] The present invention relates to the field of intelligent medical technology, and in particular to a Parkinson's freezing gait event detection method based on an LSTM and GRU fusion model. Background Art

[0002] Parkinson's disease is a chronic neurodegenerative disease that is common in the middle-aged and elderly population. There are more than 6 million patients worldwide. Its prevalence increases with age. Patients may experience a variety of motor dysfunctions such as resting tremor, bradykinesia and freezing of gait, which are often accompanied by multiple complications. Among them, freezing of gait is one of the most common movement disorders. It is manifested as "a short and sudden cessation of steps or a significant reduction in stride length when starting, walking or turning, despite the patient's willingness to walk." It often occurs in the middle and late stages of Parkinson's disease and seriously affects the patient's quality of life. Treatment with dopamine drugs can reduce the frequency of freezing gait in patients taking medication, but like most Parkinson's gait disorders, freezing gait is usually resistant to drug treatment. Another common non-drug treatment is external sensory cues called rhythmic auditory stimulation (RAS), which is generally considered to be effective in alleviating freezing of gait, but the therapeutic effect of RAS will gradually weaken over time. Therefore, using machine learning or deep learning methods that have been well developed in recent years to accurately detect or predict gait disorders caused by Parkinson's disease can effectively provide information such as the frequency of patients' freezing of gait, thereby guiding patients to reduce the frequency of freezing of gait and improve their quality of life.

[0003] Existing machine learning algorithms and deep learning research have achieved initial results on a small number of samples. However, due to the complex pathology of Parkinson's disease and the high variability, sporadicity and heterogeneity of freezing gait events, a single model or method is difficult to fully predict all freezing gait situations. Therefore, the accuracy of its detection, prediction and classification cannot be effectively guaranteed. Summary of the invention

[0004] The present invention provides a Parkinson's freezing gait event detection method based on an LSTM and GRU fusion model, which is used to detect the start and end of freezing gait events in Parkinson's disease and classify the events. A fusion model based on LSTM (Long Short-Term Memory) and GRU (Gated Recurrent Unit) is constructed for a three-axis acceleration time series provided by a subject who undergoes a Parkinson's freezing gait stimulation experiment. The input data is pre-processed by filtering, normalization, feature extraction, label shifting, etc., combined with a multi-head attention mechanism encoder to assist in encoding, and training data is expanded by using pseudo labels. The detection of Parkinson's freezing gait events is achieved, and the method has high accuracy and correctness on a large number of data samples.

[0005] The first aspect of the present invention provides a method for detecting Parkinson's freezing gait events based on an LSTM and GRU fusion model, comprising the following steps:

[0006] Step 1, obtaining the three-axis acceleration time series data of Parkinson's freezing gait;

[0007] Step 2, preprocessing the three-axis acceleration time series data;

[0008] Step 3, fine-tuning and classifying the preprocessed three-axis acceleration time series data;

[0009] Step 4: Feed the classified three-axis acceleration time series data into the target networks of LSTM and GRU respectively to obtain two types of detection data of each submodule;

[0010] Step 5: weight the detection data and concatenate the two types of detection data to complete the frozen gait event detection.

[0011] Optionally, in one embodiment of the present invention, step 2 further comprises:

[0012] Step 2-1, use a low-pass filter with a cutoff frequency of 35 Hz to perform noise reduction and scaling on the three-axis acceleration time series data;

[0013] Step 2-2, use mean normalization to normalize the three-axis acceleration time series data, limit the value to the interval [-1,1], and the noise reduction and normalization formula is:

[0014]

[0015] Among them, N ACC is the normalized result, F ACC is the result after filtering, Mean and Std represent the mean and standard deviation, and frac is the scaling factor;

[0016] Step 2-3, generate the forward difference time series feature, the formula is:

[0017] Acc lead-diff =Acc(index)-Acc(0)

[0018] Among them, Acc lead-diff is the generated forward difference time series feature, Acc(index) is the data after offset, Acc(0) is the data before offset, and index is the offset;

[0019] Step 2-4, generate the backward difference time series feature, the formula is:

[0020] Acc lag-diff =Acc(-index)-Acc(0)

[0021] Among them, Acc lag-diff is the generated forward difference time series feature, Acc(-index) is the offset data, and index is the offset;

[0022] Step 2-5, concatenate the generated new features with the original three-axis acceleration time series data by column;

[0023] Step 2-6, the concatenated new data is normalized again by means of mean standardization.

[0024] Optionally, in one embodiment of the present invention, step 3 further comprises:

[0025] Step 3-1: Use label displacement algorithm to process the three-axis acceleration time series data. The formula is:

[0026]

[0027] Among them, i is the offset result of label displacement, label is the three types of event labels, Valid and Task are the Valid label and Task label, size is the sequence length, μ is the counting function, which records the number of 1s in the logical AND result, σ is the offset marking function, which records the current offset value, Max is the maximum value operation, and & is the logical AND operation;

[0028] Step 3-2, replace the original event label with the offset event label to complete label fine-tuning;

[0029] Step 3-3, according to the type of input data, the input data is split into two groups by row, one group is from laboratory data, and the other group is from daily record data.

[0030] Optionally, in one embodiment of the present invention, step 4 further comprises:

[0031] Step 4-1 processes the data from the laboratory, and the loss function used in the entire step 4-1 is expressed as:

[0032]

[0033] The loss function used is the weighted combination of the binary cross entropy function and the focal loss function, where f i·BCE is the binary cross entropy loss of the i-th feature channel, f i·FL is the focal loss of the i-th feature channel, α is the weight coefficient, and Norm is the final normalization process;

[0034] Step 4-1-1, use the fully connected layer for encoding;

[0035] Step 4-1-2: Perform random rolling of the encoded data in a loop. The formula is:

[0036] w r =σ(w,τ(size))

[0037] Among them, w r is the data after random rolling enhancement, w is the original data, size is the sequence length, τ is the random value function, and σ is the cyclic rolling function;

[0038] Step 4-1-3, superimpose the enhanced data with the original data through the add layer and randomly discard them;

[0039] Step 4-1-4, use the multi-head attention mechanism to encode the data generated in step 4-1-3. The formula is: MHA(Q,K,V)=Concat(head1,…,herad n )W o

[0040]

[0041] First, the self-attention mechanism is applied to multiple independent subspaces, and the value Value is weighted according to the match between the query Query and the key Key. Then, the attention is calculated using the respective weight matrices, and finally the attention weights and the input acceleration ACC time series are weighted summed. After linear projection, the final attention representation, W, is obtained. i Q , W i K With W i V is the linear projection matrix, is the scaling factor, Softmax is the activation function;

[0042] Step 4-1-5, use the combination of add overlay layer, LayerNorm normalization layer and Dropout layer to process the data encoded in step 4-1-4;

[0043] Step 4-1-6, using a two-layer bidirectional long short-term memory network to capture the freezing gait event characteristics;

[0044] Step 4-1-7, perform full connection processing on the result of step 4-1-6, and use the Sigmoid activation function to generate the probability of freezing gait event detection;

[0045] Step 4-2 processes the data from daily records, and the loss function used in the entire step 4-2 is:

[0046]

[0047] The loss function used is the binary cross entropy loss function with logits, f i·BCEWL is the binary cross entropy loss with logits for the ith feature channel, and frac is the scaling factor;

[0048] Step 4-2-1, use fully connected layer encoding;

[0049] Step 4-2-2, perform LayerNorm normalization on the data encoded in step 4-2-1;

[0050] Step 4-2-3, use the group cross-validation strategy of StratifiedGroupKFold to divide the data into 5 groups;

[0051] Step 4-2-4, use the gated recurrent unit GRU to train the grouped data to capture the temporal characteristics of the freezing gait event;

[0052] Step 4-2-5, use the combination of fully connected layer, LayerNorm normalization layer and Dropout layer to process the output of step 4-2-4, and use Sigmoid activation function to output the detection probability to complete the first stage of training;

[0053] Step 4-2-6, using the one-stage model trained in step 4-2-5, detect daily data with fuzzy labels. The fuzzy labels only indicate whether a freezing gait event occurs. In order to give the type of freezing gait event, according to the one-stage model detection results and the fuzzy labels, the original fuzzy labels are replaced with pseudo labels of three types of events;

[0054] Step 4-2-7: concatenate the labeled data generated in step 4-2-6 with the original daily data row by row to form new data;

[0055] Step 4-2-8: Repeat steps 4-2-1 to 4-2-5 to complete the second stage of training.

[0056] Optionally, in one embodiment of the present invention, step 5 further comprises:

[0057] Step 5-1, collecting the test results of the laboratory data generated in step 4-1-7, and weighting them according to the given weights;

[0058] Step 5-2, collecting the detection results of the daily record data generated in step 4-2-8 and performing weighted processing according to the given weights;

[0059] Step 5-3, concatenate the results of step 5-1 and step 5-2 row by row;

[0060] In step 5-4, based on the detection probability, a binary classification result of 0 or 1 is generated for each row for the three types of events.

[0061] The Parkinson's freezing gait event detection method based on the LSTM and GRU fusion model in the embodiment of the present invention has the following features:

[0062] Beneficial effects:

[0063] (1) The present invention provides a fusion model with stronger detection capability, which can effectively deal with the high variability, sporadicity and heterogeneity of freezing gait.

[0064] (2) The present invention uses a specific filter and normalization method to effectively remove noise in the input acceleration time series, and designs a label displacement algorithm to fine-tune the misplaced labels.

[0065] (3) The present invention designs a cyclic random rolling algorithm to enhance the training data and improve the detection effect of the model. At the same time, it uses a multi-head attention mechanism to assist encoding so that the model can focus on the temporal features of multiple dimensions.

[0066] (4) The present invention constructs a fusion network based on the long short-term memory network LSTM and the gated recurrent unit GRU, and uses the data after generating pseudo-labels to expand the training set, which can capture the freezing gait events of multiple subjects under various conditions and classify them with high accuracy.

[0067] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0069] Figure 1 A flow chart of a method for detecting Parkinson's freezing gait events based on an LSTM and GRU fusion model provided according to an embodiment of the present invention;

[0070] Figure 2 It is the overall framework structure diagram of the method of the present invention;

[0071] Figure 3 For the present invention Figure 2 LSTM Group combination structure diagram in;

[0072] Figure 4 For the present invention Figure 2 GRU Group combination structure diagram in;

[0073] Figure 5 It is the overall structure diagram based on LSTM model of the present invention;

[0074] Figure 6 It is the overall structure diagram based on the GRU model of the present invention. DETAILED DESCRIPTION

[0075] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and should not be construed as limiting the present invention.

[0076] The present invention utilizes filters, normalization, feature extraction, and label shift algorithms to preprocess acceleration input data, enhances input data with a cyclic random rolling algorithm, and assists encoding with a multi-head attention mechanism. It uses a fusion model based on LSTM and GRU to capture the freezing gait information in acceleration data, and generates pseudo-labeled data to further enhance training data, ultimately completing the detection of Parkinson's freezing gait events and performing classification.

[0077] In order to solve the above technical problems, the present invention provides a Parkinson's freezing gait event detection method based on LSTM and GRU fusion model to detect and classify highly variable, sporadic and heterogeneous freezing gait events. The present invention uses acceleration data provided by 136 subjects of different ages collected under laboratory conditions and daily recording conditions, which are extracted by a triaxial accelerometer worn on the lower back of the subjects. All subjects received experimental tasks that can stimulate freezing gait in the laboratory or at home, and provided a series of detailed acceleration data. The acceleration data in three directions are acceleration in the vertical direction (acceleration-vertical, AccV), acceleration from the center to the sides (acceleration-medio-lateral, AccML) and acceleration in the front and back directions (acceleration-anterior-posterior, AccAP). According to the specific symptoms and manifestations of the subjects in different situations, the types of freezing gait are divided into three categories, namely, start difficulty (Start Hesitation, SH), turn difficulty (Turn Hesitation, TH) and walking difficulty (Walking Hesitation, WH). The present invention performs denoising, normalization, feature extraction on the above data, and uses a label shift algorithm to fine-tune the labels of three types of events. It also uses cyclic random rolling to enhance the training data and a multi-head attention mechanism to assist in encoding. It uses a method based on LSTM and GRU to build a fusion model to capture the characteristics of acceleration time series, detect the start and end of complex freezing gait events, and give their classification.

[0078] like Figure 1-Figure 6 As shown, the Parkinson's freezing gait event detection method based on the LSTM and GRU fusion model includes the following steps:

[0079] Step 1, obtain the three-axis acceleration time series data of Parkinson's freezing gait.

[0080] Step 2: preprocess the three-axis acceleration time series data.

[0081] Step 3: fine-tune and classify the preprocessed three-axis acceleration time series data.

[0082] Step 4: Feed the classified three-axis acceleration time series data into the target networks of LSTM and GRU respectively to obtain two types of detection data for each submodule.

[0083] Step 5: weight the detection data and concatenate the two types of detection data to complete the frozen gait event detection.

[0084] Further, in an embodiment of the present invention, step 2 includes the following steps:

[0085] Step 2-1, use a low-pass filter with a cutoff frequency of 35 Hz to reduce noise and scale the original input data.

[0086] Step 2-2, use mean standardization to normalize the data and limit the values ​​to the interval [-1,1]. The above two steps can be expressed as:

[0087]

[0088] Where N ACC is the normalized result, F ACC is the result after filtering, Mean and Std represent the mean and standard deviation, and frac is the scaling factor.

[0089] Step 2-3, generate the forward difference time series features, which can be expressed as:

[0090] Acc lead-diff =Acc(index)-Acc(0)

[0091] Acc lead-diff is the generated forward difference time series feature, Acc(index) is the offset data, and index is the offset.

[0092] Step 2-4, generating backward difference time series features, this step can be expressed as:

[0093] Acc lag-diff =Acc(-index)-Acc(0)

[0094] Among them, Acc lag-diff is the generated forward difference time series feature, Acc(-index) is the offset data, and index is the offset.

[0095] In steps 2-5, the generated 6 new features in 3 directions (2 in each direction) are concatenated with the original data by column.

[0096] Step 2-6, the new data generated in step 2-5 is normalized again by normalizing the mean.

[0097] Further, in an embodiment of the present invention, step 3 includes the following steps:

[0098] Step 3-1, the data from daily records contains two types of manually labeled labels Valid and Task. The frame data is considered valid only when the values ​​of the two columns of labels are both true. Due to subjective factors of manual labeling and sensor delays, the three types of labels SH, TH and WH marking freezing gait events in the training data may be misplaced, so the label displacement algorithm is used to process such data. This step can be expressed as:

[0099]

[0100] Among them, i is the offset result of label displacement, label is the three types of event labels, Valid and Task are the Valid label and Task label, size is the sequence length, μ is the counting function that can record the number of 1s in the logical AND result, σ is the offset marking function that can record the current offset value, Max is the maximum value operation, and & is the logical AND operation.

[0101] Step 3-2: Use the offset event label to replace the original event label to complete the label fine-tuning.

[0102] Step 3-3, according to the type of input data, the input data is split into two groups by row, one group is from laboratory data, and the other group is from daily record data.

[0103] Further, in an embodiment of the present invention, step 4 includes the following steps:

[0104] Step 4-1 processes the data from the laboratory, and the loss function used in the entire step 4-1 can be expressed as:

[0105]

[0106] The loss function used is the weighted combination of the binary cross entropy function (BCE) and the focal loss function (FL), where f i·BCE is the binary cross entropy loss of the i-th feature channel, f i·FL is the focal loss of the i-th feature channel, α is the weight coefficient, and Norm is the final normalization process.

[0107] Step 4-1-1, use the fully connected layer for encoding.

[0108] Step 4-1-2, cyclically and randomly rolling the encoded data, this step can be expressed as:

[0109] w r =σ(w,τ(size))

[0110] Among them, w ris the data after random rolling enhancement, w is the original data, size is the sequence length, τ is the random value function, and σ is the cyclic rolling function.

[0111] In step 4-1-3, the enhanced data is superimposed on the original data through the add layer and randomly discarded.

[0112] Step 4-1-4, use the multi-head attention mechanism to encode the generated data. This step can be expressed as:

[0113] MHA(Q,K,V)=Concat(head1,…,head n )W o

[0114]

[0115] First, the self-attention mechanism is applied to multiple independent subspaces (n=6), and the value Value(V) is weighted according to the match between the query Query(Q) and the key Key(K), and then the attention is calculated using the respective weight matrices. Finally, the attention weights and the input acceleration ACC time series are weighted summed, and the final attention representation is obtained after linear projection. i Q , W i K With W i V is the linear projection matrix, is the scaling factor, Softmax is the activation function, and the operation is performed frame by frame.

[0116] Step 4-1-5, use the combination of add overlay layer, LayerNorm normalization layer and Dropout layer to process the data encoded in step 4-1-4).

[0117] In step 4-1-6, a two-layer bidirectional long short-term memory network is used to capture the freezing gait event characteristics.

[0118] Step 4-1-7, perform full connection processing on the result of step 4-1-6, and use the Sigmoid activation function to generate the probability of freezing gait event detection.

[0119] Step 4-2, processing data from daily records, and the loss function used in the entire step 4-2) can be expressed as:

[0120]

[0121] The loss function used is the binary cross entropy loss function with logits (BCEWithLogitsLoss, BCEWL), where fi·BCEWL is the binary cross entropy loss with logits for the i-th feature channel, and frac is the scaling factor.

[0122] Step 4-2-1, use fully connected layer encoding.

[0123] Step 4-2-2, perform LayerNorm normalization on the data encoded in step 4-2-1).

[0124] In step 4-2-3, use the group cross-validation strategy of StratifiedGroupKFold to divide the data into 5 groups.

[0125] In step 4-2-4, the gated recurrent unit (GRU) is used to train the grouped data to capture the temporal features of the freezing gait events.

[0126] Step 4-2-5, use the combination of fully connected layer, LayerNorm normalization layer and Dropout layer to process the output of step 4-2-4, and use Sigmoid activation function to output the detection probability to complete the first stage of training.

[0127] Step 4-2-6, using the one-stage model trained in step 4-2-5, detect daily data with fuzzy labels. The fuzzy labels of this part of the data only indicate whether a freezing gait event occurs, but do not give the type of freezing gait event. According to the detection results of the one-stage model and the fuzzy labels, the original fuzzy labels are replaced with pseudo labels of three types of events.

[0128] Step 4-2-7, concatenate the labeled data generated in step 4-2-6 with the original daily data row by row to form new data.

[0129] Step 4-2-8, repeat steps 4-2-1 to 4-2-5 to complete the second stage of training.

[0130] Further, in an embodiment of the present invention, step 5 includes the following steps:

[0131] Step 5-1, collect the test results of the laboratory data generated in step 4-1-7, and perform weighted processing according to the given weights.

[0132] Step 5-2, collect the detection results of daily record data generated in step 4-2-8 and perform weighted processing according to given weights.

[0133] Step 5-3, concatenate the results of step 5-1 and step 5-2 row by row.

[0134] In step 5-4, based on the detection probability, a binary classification result of 0 or 1 is generated for each row for the three types of events.

[0135] According to the Parkinson's freezing gait event detection method based on the LSTM and GRU fusion model proposed in an embodiment of the present invention, specific filters, normalization, feature extraction and label displacement methods are used to preprocess the input acceleration time series data, remove the noise of the input data, normalize the input data to a limited range, and construct effective features to assist event detection, while fine-tuning the position of artificial labels; a multi-head attention mechanism is used to assist encoding so that the model pays attention to multi-dimensional data features; a complex model architecture is constructed based on a long short-term memory network LSTM and a gated recurrent unit GRU to target the high sporadic nature, variability and heterogeneity of Parkinson's freezing gait events; a pseudo-label generation algorithm is used to process the data portion that only has fuzzy labels to expand the training data. The present invention constructs a fusion model based on LSTM and GRU for the three-axis acceleration time series provided by subjects who undergo a Parkinson's freezing gait stimulation experiment, preprocesses the input data through filtering, normalization, feature extraction, label displacement, etc., combines multi-head attention mechanism encoder to assist encoding, and uses pseudo-labels to expand training data to achieve detection of Parkinson's freezing gait events, with high accuracy and correctness on a large number of data samples.

[0136] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or N embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0137] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In the description of the present invention, the meaning of "N" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0138] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, fragment or portion of code comprising one or N executable instructions for implementing the steps of a custom logical function or process, and the scope of the preferred embodiments of the present invention includes alternative implementations in which functions may not be performed in the order shown or discussed, including performing functions in a substantially simultaneous manner or in reverse order depending on the functions involved, which should be understood by technicians in the technical field to which the embodiments of the present invention belong.

Claims

1. A Parkinson's freezing gait event detection method based on LSTM and GRU fusion model, characterized in that: The following steps are involved: Step 1, obtaining the three-axis acceleration time series data of Parkinson's freezing gait; Step 2, preprocessing the three-axis acceleration time series data; Step 3, fine-tuning and classifying the preprocessed three-axis acceleration time series data; Step 4: Feed the classified three-axis acceleration time series data into the target networks of LSTM and GRU respectively to obtain two types of detection data of each submodule; Step 5: weight the detection data and concatenate the two types of detection data to complete the frozen gait event detection.

2. The method according to claim 1, characterized in that Step 2 further includes: Step 2-1, use a low-pass filter with a cutoff frequency of 35 Hz to perform noise reduction and scaling on the three-axis acceleration time series data; Step 2-2, use mean normalization to normalize the three-axis acceleration time series data, limit the value to the interval [-1,1], and the noise reduction and normalization formula is: Among them, N ACC is the normalized result, F ACC is the result after filtering, Mean and Std represent the mean and standard deviation, and frac is the scaling factor; Step 2-3, generate the forward difference time series feature, the formula is: Acc lead-diff =Acc(index)-Acc(0) Among them, Acc lead-diff is the generated forward difference time series feature, Acc(index) is the data after offset, Acc(0) is the data before offset, and index is the offset; Step 2-4, generate the backward difference time series feature, the formula is: Acc lag-diff =Acc(-index)-Acc(0) Among them, Acc lag-diff is the generated forward difference time series feature, Acc(-index) is the offset data, and index is the offset; Step 2-5, concatenate the generated new features with the original three-axis acceleration time series data by column; Step 2-6, the concatenated new data is normalized again by means of mean standardization.

3. The method according to claim 1, characterized in that Step 3 further includes: Step 3-1: Use label displacement algorithm to process the three-axis acceleration time series data. The formula is: Among them, i is the offset result of label displacement, label is the three types of event labels, Valid and Task are the Valid label and Task label, size is the sequence length, μ is the counting function, which records the number of 1s in the logical AND result, σ is the offset marking function, which records the current offset value, Max is the maximum value operation, and & is the logical AND operation; Step 3-2, replace the original event label with the offset event label to complete label fine-tuning; Step 3-3, according to the type of input data, the input data is split into two groups by row, one group is from laboratory data, and the other group is from daily record data.

4. The method according to claim 1, characterized in that: Step 4 further includes: Step 4-1 processes the data from the laboratory, and the loss function used in the entire step 4-1 is expressed as: The loss function used is the weighted combination of the binary cross entropy function and the focal loss function, where f i·BCE is the binary cross entropy loss of the i-th feature channel, f i·FL is the focal loss of the i-th feature channel, α is the weight coefficient, and Norm is the final normalization process; Step 4-1-1, use the fully connected layer for encoding; Step 4-1-2: Perform random rolling of the encoded data in a loop. The formula is: w r =σ(w,τ(size)) Among them, w r is the data after random rolling enhancement, w is the original data, size is the sequence length, τ is the random value function, and σ is the cyclic rolling function; Step 4-1-3, superimpose the enhanced data with the original data through the add layer and randomly discard them; Step 4-1-4, use the multi-head attention mechanism to encode the data generated in step 4-1-3. The formula is: MHA(Q,K,V)=Concat(head1,…,head n )W o First, the self-attention mechanism is applied to multiple independent subspaces, and the value Value is weighted according to the match between the query Query and the key Key. Then, the attention is calculated using the respective weight matrices, and finally the attention weights and the input acceleration ACC time series are weighted summed. After linear projection, the final attention representation, W, is obtained. i Q , W i K With W i V is the linear projection matrix, is the scaling factor, Softmax is the activation function; Step 4-1-5, use the combination of add overlay layer, LayerNorm normalization layer and Dropout layer to process the data encoded in step 4-1-4; Step 4-1-6, using a two-layer bidirectional long short-term memory network to capture the freezing gait event characteristics; Step 4-1-7, perform full connection processing on the result of step 4-1-6, and use the Sigmoid activation function to generate the probability of freezing gait event detection; Step 4-2 processes the data from daily records, and the loss function used in the entire step 4-2 is: The loss function used is the binary cross entropy loss function with logits, f i·BCEWL is the binary cross entropy loss with logits for the ith feature channel, and frac is the scaling factor; Step 4-2-1, use fully connected layer encoding; Step 4-2-2, perform LayerNorm normalization on the data encoded in step 4-2-1; Step 4-2-3, use the group cross-validation strategy of StratifiedGroupKFold to divide the data into 5 groups; Step 4-2-4, use the gated recurrent unit GRU to train the grouped data to capture the temporal characteristics of the freezing gait event; Step 4-2-5, use the combination of fully connected layer, LayerNorm normalization layer and Dropout layer to process the output of step 4-2-4, and use Sigmoid activation function to output the detection probability to complete the first stage of training; Step 4-2-6, using the one-stage model trained in step 4-2-5, detect daily data with fuzzy labels. The fuzzy labels only indicate whether a freezing gait event occurs. In order to give the type of freezing gait event, according to the one-stage model detection results and the fuzzy labels, the original fuzzy labels are replaced with pseudo labels of three types of events; Step 4-2-7: concatenate the labeled data generated in step 4-2-6 with the original daily data row by row to form new data; Step 4-2-8: Repeat steps 4-2-1 to 4-2-5 to complete the second stage of training.

5. The method according to claim 4, characterized in that Step 5 further includes: Step 5-1, collecting the test results of the laboratory data generated in step 4-1-7, and weighting them according to the given weights; Step 5-2, collecting the detection results of the daily record data generated in step 4-2-8 and performing weighted processing according to the given weights; Step 5-3, concatenate the results of step 5-1 and step 5-2 row by row; In step 5-4, based on the detection probability, a binary classification result of 0 or 1 is generated for each row for the three types of events.

Images