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

Abstract

The application discloses a Parkinson frozen gait event detection method based on an LSTM and GRU fusion model and belongs to the intelligent medical field. The application provides a fusion model based on LSTM and GRU for the three-axis acceleration time sequence provided by a subject receiving a Parkinson frozen gait excitation experiment, pre-processes input data through filtering, normalization, feature extraction and label displacement, combines a multi-head attention mechanism encoder to assist coding, simultaneously uses pseudo-label expansion training data, realizes detection on a Parkinson frozen gait event, and has high accuracy and correctness on a large amount of data samples.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical technology, and in particular to a method for detecting Parkinson's disease frozen gait events based on a fusion model of LSTM and GRU. Background Technology

[0002] Parkinson's disease is a chronic neurodegenerative disease that predominantly affects middle-aged and elderly people, with over 6 million patients worldwide. Its prevalence increases with age. Patients may experience various motor dysfunctions, including resting tremor, bradykinesia, and frozen gait, often accompanied by multiple complications. Frozen gait is one of the most common motor disorders, characterized by a brief, sudden cessation of walking or a significant reduction in stride length during initiation, walking, or turning, despite the patient's intention to walk. It is most common in the middle and late stages of Parkinson's disease and severely impacts patients' quality of life. Treatment with dopamine-based medications can reduce the frequency of frozen gait while on medication, but like most Parkinson's gait disorders, frozen gait often develops resistance to drug treatment. Another common non-pharmacological treatment is external sensory cues called rhythmic auditory stimulation (RAS), which is generally considered effective in alleviating frozen gait; however, the effectiveness of RAS treatment gradually diminishes over time. Therefore, using machine learning or deep learning methods, which have seen significant development in recent years, to accurately detect or predict gait disturbances caused by Parkinson's disease can effectively provide patients with information such as the frequency of frozen gait episodes, thereby guiding them to reduce the frequency of frozen gait episodes and improving their quality of life.

[0003] Existing machine learning algorithms and deep learning research have achieved preliminary results on a small number of samples. However, due to the complexity of Parkinson's disease pathology and the high variability, sporadic and heterogeneous nature of frozen gait events, a single model or method is difficult to completely predict all frozen gait situations. Therefore, the accuracy of its detection, prediction and classification cannot be effectively guaranteed. Summary of the Invention

[0004] This invention provides a Parkinson's disease frozen gait event detection method based on an LSTM and GRU fusion model. This method detects the start and end of frozen gait events in Parkinson's disease and classifies the events. A fusion model based on LSTM (Long Short-Term Memory) and GRU (Gated Recurrent Unit) is constructed using triaxial acceleration time series provided by subjects who underwent Parkinson's frozen gait provocation experiments. The input data is preprocessed through filtering, normalization, feature extraction, and label shifting. A multi-head attention encoder is used for assisted encoding, and pseudo-labels are used to augment the training data. This method achieves high accuracy and correctness on a large number of data samples.

[0005] A first aspect of this invention provides a method for detecting Parkinson's disease frozen gait events based on a fusion model of LSTM and GRU, comprising the following steps:

[0006] Step 1: Obtain the triaxial acceleration time series data of Parkinson's frozen gait;

[0007] Step 2: Preprocess the triaxial acceleration time series data;

[0008] Step 3: Fine-tune and classify the preprocessed triaxial acceleration time series data;

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

[0010] Step 5: Weight the detection data and stitch the two types of detection data together to complete the frozen gait event detection.

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

[0012] Step 2-1: Use a low-pass filter with a cutoff frequency of 35Hz to denoise and scale the triaxial acceleration time series data;

[0013] Step 2-2: Normalize the triaxial acceleration time series data using mean standardization, limiting the values ​​to the interval [-1, 1]. The formulas for noise reduction and normalization are as follows:

[0014]

[0015] Where, N ACC For the normalized result, F ACC The result is the filtered result, where Mean and Std represent the mean and standard deviation, and frac is the scaling factor.

[0016] Steps 2-3 generate forward difference temporal features, using the following formula:

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

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

[0019] Steps 2-4 generate backward difference time series features, using the following formula:

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

[0021] Among them, Acc lag-diff For the generated forward interpolation time series features, Acc(-index) is the offset data, and index is the offset amount;

[0022] Steps 2-5: Concatenate the generated new features with the original triaxial acceleration time series data column by column;

[0023] Steps 2-6 involve performing mean standardization on the new spliced ​​data again.

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

[0025] Step 3-1: Process the triaxial acceleration time series data using the label displacement algorithm, the formula is as follows:

[0026]

[0027] Where i is the offset result of the 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 records the number of 1s in the logical AND result, σ is the offset marking function that records the current offset value, Max is the maximum / minimum value operation, and & is the logical AND operation.

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

[0029] Step 3-3: Based on the type of input data, split the input data into two groups by row: one group from laboratory data and the other group from daily records.

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

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

[0032]

[0033] The loss function used is a weighted sum of the binary cross-entropy function and the focus loss function, where f i·BCE f is the binary cross-entropy loss for the i-th feature channel. i·FL Let α be the focus loss for the i-th feature channel, α be the weight coefficient, and Norm be the final normalization process.

[0034] Step 4-1-1: Encode using a fully connected layer;

[0035] Step 4-1-2: Perform a cyclical random rolling process on the encoded data, using the following formula:

[0036] w r =σ(w,τ(size))

[0037] Among them, w r For 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: Overlay the enhanced data and the original data through an add layer and then randomly discard the data;

[0039] Step 4-1-4: Encode the data generated in step 4-1-3 using a multi-head attention mechanism, with the formula: MHA(Q,K,V)=Concat(head1,…,herad) n W o

[0040]

[0041] First, the self-attention mechanism is applied to multiple independent subspaces. Weights are calculated for values ​​based on the matching degree between the query and the key. Then, attention is calculated using the respective weight matrices. Finally, the attention weights and the input acceleration (ACC) time series are weighted and summed, and linearly projected to obtain the final attention representation, W. i Q W i K With W i V It is a linear projection matrix. Softmax is the scaling factor, and Softmax is the activation function;

[0042] Step 4-1-5: Process the data encoded in step 4-1-4 using a combination of add overlay layer, LayerNorm normalization layer, and Dropout layer;

[0043] Step 4-1-6: Use a two-layer bidirectional long short-term memory network to capture frozen gait event features;

[0044] Step 4-1-7: Perform fully connected processing on the results of step 4-1-6, and use the Sigmoid activation function to generate the probability of frozen gait event detection;

[0045] Step 4-2 processes the data from daily records, and the loss function used throughout 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 i-th 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 StratifiedGroupKFold grouped cross-validation strategy to divide the data into 5 groups;

[0051] Step 4-2-4: Use a gated recurrent unit (GRU) to train the grouped data to capture temporal features related to frozen gait events;

[0052] Step 4-2-5: Process the output of step 4-2-4 using a combination of fully connected layers, LayerNorm normalization layers, and Dropout layers, and use the 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 frozen gait event has occurred, but do not give the type of frozen gait event. Based on the detection results of the one-stage model and the fuzzy labels, replace the original fuzzy labels with pseudo-labels of three types of events.

[0054] Step 4-2-7: Concatenate the labeled data portion 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 phase of training.

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

[0057] 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;

[0058] Step 5-2: Collect the detection results of the daily record data generated in Step 4-2-8 and perform weighted processing according to the given weights;

[0059] Step 5-3: Combine the results of Step 5-1 and Step 5-2 line by line;

[0060] Step 5-4: Based on the detection probability, generate binary classification results of 0 and 1 for each row of three types of events.

[0061] The present invention provides a Parkinson's frozen gait event detection method based on a fusion model of LSTM and GRU, which has the following characteristics:

[0062] Beneficial effects:

[0063] (1) This invention provides a fusion model with stronger detection capabilities, which can effectively cope with the high variability, sporadic nature and heterogeneity of frozen gait.

[0064] (2) This invention uses specific filters and normalization methods to effectively remove noise from the input acceleration time series. At the same time, a label displacement algorithm is designed to fine-tune misaligned labels.

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

[0066] (4) This invention constructs a fusion network based on Long Short-Term Memory Network (LSTM) and Gated Recurrent Unit (GRU), and uses data after generating pseudo-labels to expand the training set, which can capture frozen gait events of multiple subjects under various conditions and classify them with high accuracy.

[0067] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

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

[0069] Figure 1 A flowchart of a Parkinson's frozen gait event detection method based on an LSTM and GRU fusion model provided according to an embodiment of the present invention;

[0070] Figure 2 This is a diagram illustrating the overall framework of the method of the present invention;

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

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

[0073] Figure 5 This is a diagram showing the overall structure of the LSTM-based model of the present invention.

[0074] Figure 6 This is a diagram showing the overall structure of the invention based on the GRU model. Detailed Implementation

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

[0076] This invention utilizes filters, normalization, feature extraction, and label displacement algorithms to preprocess acceleration input data, enhances the 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 frozen gait information in the acceleration data, while generating pseudo-labeled data to further enhance the training data. Finally, it completes the detection and classification of Parkinson's frozen gait events.

[0077] To address the aforementioned technical problems, this invention provides a method for detecting Parkinson's disease-related gait freezing events based on an LSTM and GRU fusion model, enabling the detection and classification of highly variable, sporadic, and heterogeneous gait freezing events. This invention utilizes acceleration data collected from 136 subjects of different ages under laboratory and daily recording conditions. This data was extracted by a triaxial accelerometer worn on the lower back of the subjects. All subjects underwent experimental tasks that induced gait freezing in the laboratory or at home, providing a series of detailed acceleration data in three directions: vertical acceleration (AccV), mediolateral acceleration (AccML), and anterior-posterior acceleration (AccAP). Based on the specific symptoms and manifestations observed by the subjects in different situations, gait freezing was classified into three types: start hesitation (SH), turn hesitation (TH), and walking hesitation (WH). This invention performs noise reduction, normalization, feature extraction, and fine-tuning of three types of event labels using a label displacement algorithm on the above data. It also uses cyclic random rolling to enhance training data and multi-head attention mechanism to assist encoding. A fusion model based on LSTM and GRU is used to capture the features of acceleration time series, detect the start and end of complex frozen gait events, and classify them.

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

[0079] Step 1: Obtain the triaxial acceleration time series data of Parkinson's frozen gait.

[0080] Step 2: Preprocess the triaxial acceleration time series data.

[0081] Step 3: Fine-tune and classify the preprocessed triaxial acceleration time series data.

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

[0083] Step 5: Weight the detection data and stitch the two types of detection data together to complete the frozen gait event detection.

[0084] Furthermore, 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 35Hz to perform noise reduction and scaling on the original input data.

[0086] Step 2-2: Normalize the data using mean standardization, limiting the values ​​to the interval [-1, 1]. The above two steps can be described as follows:

[0087]

[0088] Where N ACC For the normalized result, F ACC This is the filtered result, where Mean and Std represent the mean and standard deviation, and frac is the scaling factor.

[0089] Step 2-3, generating forward interpolation time series features, can be represented as:

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

[0091] Among them Acc lead-diff For the generated forward difference time series features, Acc(index) is the offset data, and index is the offset amount.

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

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

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

[0095] Steps 2-5 involve concatenating the generated 6 new features (2 features per direction) with the original data column by column.

[0096] Steps 2-6 involve normalizing the new data generated in Step 2-5 again using mean standardization.

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

[0098] Step 3-1: The data from daily records includes two types of manually labeled tags: Valid and Task. A frame is considered valid only if both tag values ​​are true. Due to subjective factors in manual labeling and sensor latency, the three types of labels (SH, TH, and WH) marking frozen gait events in the training data may be misaligned. Therefore, a label shifting algorithm is used to process this type of data. This step can be represented as:

[0099]

[0100] Where i is the offset result of the 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 / minimum value operation, and & is the logical AND operation.

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

[0102] Step 3-3: Based on the type of input data, split the input data into two groups by row: one group from laboratory data and the other group from daily records.

[0103] Furthermore, 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 a weighted sum of the binary cross-entropy (BCE) function and the focal loss (FL) function, where f i·BCE f is the binary cross-entropy loss for the i-th feature channel. i·FL Let α be the focus loss for the i-th feature channel, α be the weight coefficient, and Norm be the final normalization process.

[0107] Step 4-1-1: Encode using a fully connected layer.

[0108] Step 4-1-2 involves performing a cyclical, random rolling process on the encoded data. This step can be represented as:

[0109] w r =σ(w,τ(size))

[0110] Among them, w rThe data is enhanced by random rolling, where w is the original data, size is the sequence length, τ is the random value function, and σ is the cyclic rolling function.

[0111] Step 4-1-3: Overlay the enhanced data and the original data through the add layer and then randomly discard the data.

[0112] Step 4-1-4, encoding the generated data using a multi-head attention mechanism, can be represented 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). The value (V) is weighted according to the matching degree between the query (Q) and the key (K). Then, the attention is calculated using their respective weight matrices. Finally, the attention weights and the input acceleration (ACC) time series are weighted and summed, and linearly projected to obtain the final attention representation. i Q W i K With W i V It is a linear projection matrix. is the scaling factor, and Softmax is the activation function. This operation is performed frame by frame.

[0116] Step 4-1-5: Process the data encoded in step 4-1-4 using a combination of add overlay layer, LayerNorm normalization layer, and Dropout layer.

[0117] Step 4-1-6: Use a two-layer bidirectional long short-term memory network to capture frozen gait event features.

[0118] Step 4-1-7: Perform fully connected processing on the results of step 4-1-6, and use the Sigmoid activation function to generate the probability of frozen gait event detection.

[0119] Step 4-2 processes the data from daily records, and the loss function used throughout 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] Step 4-2-3: Use the StratifiedGroupKFold grouped cross-validation strategy to divide the data into 5 groups.

[0125] Step 4-2-4: Use a gated recurrent unit (GRU) to train the grouped data to capture temporal features related to frozen gait events.

[0126] Step 4-2-5: Process the output of step 4-2-4 using a combination of fully connected layers, LayerNorm normalization layers, and Dropout layers, and use the Sigmoid activation function to output the detection probability, thus completing 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 for this portion of data only indicate whether a frozen gait event occurred, but do not specify the type of frozen gait event. Based on the detection results of the one-stage model and the fuzzy labels, replace the original fuzzy labels with pseudo-labels for the three event types.

[0128] Step 4-2-7: Combine the labeled data portion 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 phase of training.

[0130] Furthermore, 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 the daily record data generated in Step 4-2-8 and perform weighted processing according to the given weights.

[0133] Step 5-3: Combine the results of Step 5-1 and Step 5-2 row by row.

[0134] Step 5-4: Based on the detection probability, generate binary classification results of 0 and 1 for each row of three types of events.

[0135] The Parkinson's frozen gait event detection method proposed in this invention, based on an LSTM and GRU fusion model, preprocesses the input acceleration time series data using specific filters, normalization, feature extraction, and label shifting methods to remove noise from the input data, normalize the input data to a defined range, and construct effective features to assist event detection, while fine-tuning the position of artificial labels; it uses a multi-head attention mechanism to assist encoding, enabling the model to focus on multi-dimensional data features; it uses a complex model architecture based on a Long Short-Term Memory (LSTM) network and a gated recurrent unit (GRU) to address the highly sporadic, variable, and heterogeneous nature of Parkinson's frozen gait events; and it utilizes a pseudo-label generation algorithm to process the data portion with only fuzzy labels, thereby expanding the training data. This invention constructs a fusion model based on LSTM and GRU for the triaxial acceleration time series provided by subjects undergoing Parkinson's frozen gait provocation experiments. By preprocessing the input data through filtering, normalization, feature extraction, and label displacement, and combining it with a multi-head attention mechanism encoder for assisted encoding, while using pseudo-labels to expand the training data, the invention achieves the detection of Parkinson's frozen gait events, and has high accuracy and correctness on a large number of data samples.

[0136] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0137] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0138] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or N executable instructions for implementing custom logic functions or processes, and the scope of preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of the invention pertain.

Claims

1. A method for detecting Parkinson's frozen gait events based on a fusion model of LSTM and GRU, characterized in that, Includes the following steps: Step 1: Obtain the triaxial acceleration time series data of Parkinson's frozen gait; Step 2: Preprocess the triaxial acceleration time series data; Step 3: Fine-tune the labels and classify the preprocessed triaxial acceleration time series data; Step 4: Feed the classified triaxial acceleration time series data into the target networks of LSTM and GRU respectively to obtain two types of detection data for each sub-module; Step 5: Weight the detection data and concatenate the two types of detection data to complete the frozen gait event detection; Step 2 further includes: Step 2-1: Use a low-pass filter with a cutoff frequency of 35Hz to denoise and scale the triaxial acceleration time series data; Step 2-2: Normalize the triaxial acceleration time series data using mean standardization, limiting the values ​​to the interval [-1, 1]. The formulas for noise reduction and normalization are as follows: in, The result after normalization. This is the filtered result. and Represents the mean and standard deviation. This is the scaling factor; Steps 2-3 generate forward difference temporal features, using the following formula: in, For the generated forward interpolation time series features, The data after offset. The data before the offset. This is the offset; Steps 2-4 generate backward difference time series features, using the following formula: in, For the generated backward interpolation time series features, The data after offset. This is the offset; Steps 2-5 involve concatenating the generated six new features (two features per direction) with the original triaxial acceleration time series data column by column. Steps 2-6: Perform mean standardization on the new spliced ​​data again. Step 3 further includes: Step 3-1: Process the triaxial acceleration time series data using the label displacement algorithm, the formula is as follows: in, This is the offset result for the label displacement. There are three types of event tags. and For the Valid tag and the Task tag, For sequence length, This is a counting function that records the number of 1s in the result of a logical AND operation. An offset marker function records the current offset value. For extreme value calculation, For logical AND operation; Step 3-2: Replace the original event label with the offset event label to complete the label fine-tuning; Step 3-3: Based on the type of input data, split the input data into two groups by row: one group from laboratory data and the other group from daily records.

2. 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 throughout Step 4-1 is expressed as follows: The loss function used is a weighted sum of the binary cross-entropy function and the focus loss function, where The binary cross-entropy loss for the i-th feature channel is... The focus loss for the i-th feature channel, These are the weighting coefficients. For the final normalization process; Step 4-1-1: Encode using a fully connected layer; Step 4-1-2: Perform a cyclical random rolling process on the encoded data, using the following formula: in, For data enhanced by random rolling, The original data, For sequence length, It is a function for randomly selecting values. This is a loop scrolling function; Step 4-1-3: Overlay the enhanced data and the original data through an add layer and then randomly discard the data; Step 4-1-4: Encode the data generated in step 4-1-3 using a multi-head attention mechanism, with the following formula: First, the self-attention mechanism is applied to multiple independent subspaces. Weights are calculated for values ​​based on the matching degree between the query and the key. Then, attention is calculated using the respective weight matrices. Finally, the attention weights and the input acceleration (ACC) time series are weighted and summed, and linearly projected to obtain the final attention representation. , and It is a linear projection matrix. As a scaling factor, For activation functions; Step 4-1-5: Process the data encoded in step 4-1-4 using a combination of add overlay layer, LayerNorm normalization layer, and Dropout layer; Step 4-1-6: Use a two-layer bidirectional long short-term memory network to capture frozen gait event features; Step 4-1-7: Perform fully connected processing on the results of step 4-1-6, and use the Sigmoid activation function to generate the probability of frozen gait event detection; Step 4-2 processes the data from daily records, and the loss function used throughout step 4-2 is: The loss function used is the binary cross-entropy loss function with logits. For the i-th feature channel, use the binary cross-entropy loss with logits. This 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 StratifiedGroupKFold grouped cross-validation strategy to divide the data into 5 groups; Step 4-2-4: Use a gated recurrent unit (GRU) to train the grouped data to capture temporal features related to frozen gait events; Step 4-2-5: Process the output of step 4-2-4 using a combination of fully connected layers, LayerNorm normalization layers, and Dropout layers, and use the 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 frozen gait event has occurred, but do not give the type of frozen gait event. Based on the detection results of the one-stage model and the fuzzy labels, replace the original fuzzy labels with pseudo-labels of three types of events. Step 4-2-7: Concatenate the labeled data portion 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 phase of training.

3. The method according to claim 2, characterized in that, Step 5 further includes: 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; Step 5-2: Collect the detection results of the daily record data generated in Step 4-2-8 and perform weighted processing according to the given weights; Step 5-3: Combine the results of Step 5-1 and Step 5-2 line by line; Step 5-4: Based on the detection probability, generate binary classification results of 0 and 1 for each row of three types of events.

Images