Parkinson's syndrome behavior evaluation method and device based on smart watch

Abstract

The invention belongs to the field of artificial intelligence and mode recognition, and discloses a Parkinson's syndrome behavior evaluation method and device based on a smart watch. According to the method, motion data of a user in a natural state is collected through a built-in sensor of the smart watch, remote monitoring can be achieved without additional equipment, and six actions including sitting posture rest, two-arm horizontal lifting, wrist forward and backward rotation, palm slapping, two-arm forward stretching and linear walking are collected. Inertial measurement unit data contains abundant signal features, and the signal features are processed by adopting a deep learning technology, so that key feature information is effectively extracted. According to the classification model based on the data of the inertial measurement unit of the smart watch, the feature extraction efficiency and the pattern recognition precision are remarkably improved, and a reliable technical means is provided for objective evaluation of Parkinson's syndrome.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence and pattern recognition, and relates to a Parkinson's syndrome behavior assessment method and device based on a smart watch. Background Art

[0002] Amidst the global aging population, the incidence of Parkinson's disease continues to rise, posing a significant challenge to the healthcare system. Current clinical practice primarily relies on clinical examinations and rating scales performed by neurologists for diagnosis and assessment. This traditional model has significant limitations: The uneven geographical distribution of specialist resources hinders some patients from obtaining timely professional diagnosis and treatment. Furthermore, the highly subjective nature of assessment methods, combined with time constraints during outpatient examinations, often makes it difficult to accurately reflect a patient's true motor function. These factors highlight the urgency of developing new, objective assessment tools.

[0003] Wearable devices provide a new technical path to solving this problem. As a popular wearable terminal, smart watches are equipped with sensors such as accelerometers and gyroscopes, and have the unique advantage of continuously collecting motion physiological parameters. Studies have confirmed that the characteristic movement disorders of Parkinson's syndrome, such as bradykinesia and resting tremor, can be quantified through sensor data. Compared with traditional outpatient assessments, this technical solution has significant advantages: it can obtain natural state motion data over a longer time span, and its portability also supports long-term disease monitoring in a home environment.

[0004] The rapid development of artificial intelligence (AI) technology has provided new solutions for data processing. Deep learning algorithms can extract clinically relevant digital biomarkers from multi-dimensional motion data collected by smartwatches, enabling objective quantification of disease characteristics and prediction of disease progression. This intelligent assessment system has dual benefits: it not only provides an objective basis for clinical decision-making and reduces diagnostic bias, but also establishes a remote monitoring system to support dynamic adjustment of personalized treatment plans. Based on comprehensive technical feasibility and clinical needs analysis, the Parkinson's syndrome assessment system, which integrates smartwatches and AI technologies, will not only help improve disease screening efficiency but also optimize medical resource allocation, possessing significant clinical application value. Summary of the Invention

[0005] The present invention solves the problems of poor behavioral assessment of Parkinson's syndrome and time-consuming and labor-intensive traditional assessment methods, and proposes a Parkinson's syndrome behavioral assessment method and equipment based on smart watches, which provides an effective solution for realizing remote and large-scale screening of Parkinson's syndrome at home. The three-dimensional accelerometer and gyroscope data of six actions, namely, sitting and resting, arms raised, wrist pronation and supination, palm slapping, arms extended, and straight walking, are collected and sent to the deep learning model to extract features and classify them. The multi-scale perception module learns features of different scales, the frequency perception module focuses on the 3-7Hz tremor frequency band, the shared channel attention mechanism is combined with the pooling layer and the multi-layer perceptron to dynamically weight the importance of the sensor channel, and the prediction result network performs deep feature learning through the convolution layer, batch normalization, ReLU and pooling layer, and finally outputs the evaluation score through the fully connected layer. This method realizes the evaluation of Parkinson's syndrome behavior through the collaborative analysis of multi-scale spatiotemporal features and prior frequency domain knowledge.

[0006] The specific technical solutions of the present invention are:

[0007] A method for assessing Parkinson's disease behavior based on a smartwatch, comprising the following steps:

[0008] (1) The user wears a smartwatch and completes six actions: sitting and resting, raising arms horizontally, wrist pronation and supination, palm clapping, extending arms forward, and walking in a straight line;

[0009] (2) Using the accelerometer and gyroscope in the smartwatch to capture acceleration data and rotation rate data during the user's specific movements;

[0010] (3) The acceleration data and rotation rate data are filtered using an L1 trend filter to eliminate gravity offset and constant deviation to capture dynamic motion patterns; the first half second of each data is deleted to reduce artifacts caused by vibration; the preprocessed data is divided into a training set, a validation set, and a test set;

[0011] (4) building an evaluation model for Parkinson's syndrome behavior based on a smartwatch, including a multi-scale perception module, a frequency perception module, a shared channel attention mechanism, and a prediction result network; (5) training the evaluation model for Parkinson's syndrome behavior based on inertial measurement unit data built in step (4), adjusting hyperparameters to obtain optimal performance, and iterating the evaluation model for Parkinson's syndrome behavior based on inertial measurement unit data in step (4) so ​​that the validation set loss converges and the optimal model is saved;

[0012] (6) Use the optimal model to evaluate the individuals in the test set to verify the effectiveness of the method.

[0013] Furthermore, the frequency perception module focuses on the tremor signal in the frequency range of 3-7 Hz through fast Fourier transform and inverse fast Fourier transform; the frequency perception module converts the time domain information in the preprocessed data into frequency domain information through fast Fourier transform, generates a feature mask in the 3-7 Hz frequency domain; converts it back to time domain information through inverse fast Fourier transform, and then extracts new features through convolution layer, normalization and activation function.

[0014] Furthermore, the multi-scale perception module includes a convolution layer and an activation function layer; the multi-scale perception module extracts multi-scale features from the preprocessed data through convolution layers with different scale convolution kernel sizes, obtains features of different scales through the activation function layer, and fuses them with the features of the frequency perception module before inputting them into the shared channel attention mechanism.

[0015] Furthermore, the shared channel attention mechanism consists of a pooling layer and a multi-layer perceptron. The results of the frequency perception module and the multi-scale perception module serve as inputs to the shared channel attention mechanism, which is used to learn the importance weights of different channels. Specifically, the inputs of the shared channel attention mechanism are fed into the average pooling layer and the maximum pooling layer for feature extraction. The extracted features are then further learned through a multi-layer perceptron, with the activation function responsible for obtaining the feature learning results. The shared channel attention mechanism enables the model to learn the importance of sensor channels, allowing it to focus on the most information-rich channels while suppressing irrelevant or noisy channels. By sharing the multi-layer perceptron, computational complexity is reduced and the efficiency of learning attention weights is improved. Max pooling and average pooling accelerate computation to prevent overfitting, and the multi-layer perceptron can dynamically adjust the importance of each sensor channel.

[0016] The prediction network further includes convolutional layers, normalization layers, activation functions, and max pooling layers. The prediction network first uses convolutional layers to further extract features and capture tremor behavior patterns. Normalization layers stabilize the training distribution and eliminate individual and device differences. The activation function introduces nonlinearity to enhance the model's expressiveness. The max pooling layer focuses on key movement patterns. Finally, a fully connected layer predicts the score.

[0017] A Parkinson's syndrome behavior assessment device based on a smart watch, characterized by comprising: at least one processor, and a memory communicatively connected to the at least one processor;

[0018] The memory stores instructions that can be executed by the one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the above-mentioned Parkinson's syndrome behavior assessment method based on a smart watch.

[0019] The present invention has the following beneficial effects:

[0020] (1) The present invention utilizes the built-in sensor of the smart watch to collect motion data in daily life scenarios without the need for external devices, thereby realizing the remote home monitoring function of Parkinson's syndrome.

[0021] (2) The present invention proposes a classification model based on the data features of the inertial measurement unit of a smart watch. By optimizing the feature extraction and pattern recognition process of the inertial sensor data of the smart watch, the accuracy of the behavioral assessment of Parkinson's syndrome is significantly improved.

[0022] (3) The present invention proposes six movements that can reflect the movement characteristics of Parkinson's syndrome: sitting and resting, raising arms horizontally, wrist pronation and supination, palm slapping, extending arms forward, and walking in a straight line. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 This is a structural diagram of the evaluation model for smartwatch inertial measurement unit data according to the present invention;

[0024] Figure 2 This is a multi-scale feature extraction network structure diagram for three-dimensional inertial measurement unit data, corresponding to Figure 1 mid-scale multi-scale perception module;

[0025] Figure 3 This is a feature extraction network structure diagram for 3-7 Hz frequency tremor of the present invention, corresponding to Figure 1 and Figure 2 mid-frequency sensing module;

[0026] Figure 4 This is a structural diagram of the present invention for learning high-level features of three-dimensional inertial measurement unit data, corresponding to Figure 1 The shared channel attention mechanism in

[15] ;

[0027] Figure 5 It is the test result of the optimal model trained by the present invention. DETAILED DESCRIPTION

[0028] In order to make the above objects and advantages more understandable, the present invention is further described below with reference to the accompanying drawings and specific embodiments.

[0029] A smartwatch-based behavioral assessment method for Parkinson's disease Figure 1 The specific implementation process is as follows:

[0030] Step 1: Create a dataset;

[0031] A data acquisition application was developed using Android Studio to collect inertial measurement unit data for six actions: sitting and resting, arms raised, wrist pronation and supination, palm slapping, arms extended, and straight-line walking. The specific acquisition process is as follows:

[0032] Seated rest data collection: 1) Wear the smartwatch on both hands and click the start button through the central control; 2) Wait for the announcement of "test start"; 3) The participant places both hands on the armrests and both feet on the ground naturally for 10 seconds; 4) "test end" is announced;

[0033] Arms raised data collection: 1) Wear the smartwatch on both hands and click the start button on the central control; 2) Wait for the announcement of "Test Start"; 3) The participant places both hands palms down, with five fingers naturally extended and middle fingers facing each other, in front of the chest, with both arms parallel to the ground, for 10 seconds; 4) The announcement of "Test End" is made;

[0034] Wrist pronation and supination data collection: 1) Wear the smartwatch on both hands and click the start button through the central control; 2) Wait for the announcement of "test start"; 3) The participant extends their arm forward with their palm facing down and alternately flips their palm up and down at the fastest speed and maximum amplitude for 10 seconds; 4) The announcement of "test end" is made;

[0035] Palm tapping data collection: 1) Wear the smartwatch on both hands and click the start button on the central control; 2) Wait for the announcement of "Test Start"; 3) The participant bends their arm, fixes their elbow on the table, places their palm flat on the table, and slowly taps the table with their palm 10 times, testing both hands separately;

[0036] Data collection for arm extension: 1) Wear the smartwatch on both hands and click the start button on the central control panel; 2) Wait for the announcement of "Test Start"; 3) The participant extends their arm forward with their palm facing downward, wrists straight, and fingers spread apart without touching, for 10 seconds; 4) "Test End" is announced.

[0037] Straight-line walking data collection: 1) Wear the smartwatch on both hands and click the start button on the central control; 2) Wait for the announcement of "test start"; 3) Participants walk in a straight line for 10 meters, turn around, and return to the starting point;

[0038] Each subject repeated the activity more than five times, that is, five pieces of data of each type were collected; a total of 500 individuals with Parkinson's syndrome and 500 healthy individuals were recruited.

[0039] Step 2: Preprocess the original data and divide it into training set, validation set and test set;

[0040] For three-dimensional acceleration and rotation speed, an L1 trend filter is used to eliminate gravity offset and constant deviation to capture dynamic motion patterns. The first half second of each data is deleted to reduce artifacts caused by vibration to obtain pre-processed inertial measurement unit data.

[0041] A nested five-fold cross-validation strategy was used, with 80% of the data in each outer fold used for model training and 20% for model testing. Each inner fold further used 80% of the training data for training and 20% for validation, that is, the training set of each inner fold accounted for 64% and the validation set accounted for 16%. The outer folds were only used to evaluate the generalization performance of the model, which could avoid data from the same individual appearing in the training set and test set of the same fold, thereby overestimating the model performance.

[0042] Step 3: Build a smartwatch-based Parkinson's behavior assessment model, such as Figure 1 As shown;

[0043] The model mainly includes multi-scale perception module, frequency perception module, shared channel attention mechanism, and prediction result network;

[0044] The pre-processed data are used as the input of the multi-scale perception module and the frequency perception module for feature extraction and learning; the multi-scale perception module is as follows: Figure 2 As shown in the figure, the input data is subjected to multi-scale feature extraction through convolution layers with convolution kernels of 3, 5, and 7, and features of different scales are obtained through ReLU activation function, which are then fused with the features of the frequency perception module to obtain the extracted features. The frequency perception module is as follows: Figure 3 As shown in the figure, the input data is converted from time domain information to frequency domain information through fast Fourier transform, generating a feature mask in the 3-7 Hz frequency domain, which is then converted back to time domain information through inverse fast Fourier transform, and then new features are extracted through convolutional layers, normalization, and ReLU activation functions;

[0045] The results of the frequency domain perception module and the multi-scale perception module are used as the input of the shared channel attention mechanism to learn the importance weights of different channels; the shared channel attention mechanism is as follows Figure 4 As shown in the figure, the average pooling layer and the maximum pooling layer are input respectively for feature extraction. The extracted features are further learned by a multi-layer perceptron, and the Sigmoid activation function is responsible for obtaining the feature learning results.

[0046] The prediction result network first further extracts features through the convolutional layer to capture the behavioral patterns of tremors. The normalization layer stabilizes the training distribution and eliminates individual and device differences. The ReLU activation function introduces nonlinearity to enhance the model's expressiveness. The maximum pooling layer focuses on key movement patterns. This process is performed three times to ensure that the prediction result network can fully learn the behavioral patterns of Parkinson's syndrome. The fully connected layer outputs the evaluation results.

[0047] Step 4: Set hyperparameters and train the model;

[0048] During model training, the model's hyperparameters were set to 50 epochs, the learning rate was set to 0.0005, the Adam optimizer was used, and the batch size was set to 16. In addition, a dynamic learning rate reduction strategy was used. When the loss did not decrease for 10 consecutive epochs, the learning rate was reduced to half of its original value.

[0049] Step 5: Test the model. Use the trained optimal model to predict the individuals in the test set and calculate the accuracy, precision, recall, F1 score and AUC. The results are as follows: Figure 5 shown.

Claims

1. A method for Parkinson's disease behavior assessment based on a smartwatch, characterized in that: The steps include: (1) The user wears a smartwatch and completes six actions: sitting and resting, raising arms horizontally, wrist pronation and supination, palm clapping, extending arms forward, and walking in a straight line; (2) Using the accelerometer and gyroscope in the smartwatch to capture acceleration data and rotation rate data during the user's specific movements; (3) The acceleration data and rotation rate data are filtered using an L1 trend filter to eliminate gravity offset and constant deviation to capture dynamic motion patterns; the first half second of each data is deleted to reduce artifacts caused by vibration; Divide the preprocessed data into training set, validation set and test set; (4) Build a smartwatch-based Parkinson's disease behavior evaluation model, including a multi-scale perception module, a frequency perception module, a shared channel attention mechanism, and a prediction result network; (5) training the evaluation model of Parkinson's syndrome behavior based on inertial measurement unit data constructed in step (4), adjusting hyperparameters to obtain optimal performance, iterating the evaluation model of Parkinson's syndrome behavior based on inertial measurement unit data in step (4) so ​​that the validation set loss converges and saving the optimal model; (6) Use the optimal model to evaluate the individuals in the test set to verify the effectiveness of the method.

2. The method for Parkinson's disease behavior assessment based on a smartwatch according to claim 1, characterized in that: The frequency perception module focuses on the tremor signal in the frequency range of 3-7 Hz through fast Fourier transform and inverse fast Fourier transform; the frequency perception module converts the time domain information in the preprocessed data into frequency domain information through fast Fourier transform, generates a feature mask in the 3-7 Hz frequency domain; converts it back into time domain information through inverse fast Fourier transform, and then extracts new features through convolution layer, normalization and activation function.

3. The method for Parkinson's disease behavior assessment based on a smartwatch according to claim 1, characterized in that: The multi-scale perception module includes a convolution layer and an activation function layer. The multi-scale perception module extracts multi-scale features from the preprocessed data through convolution layers with different scale convolution kernel sizes, obtains features of different scales through the activation function layer, and fuses them with the features of the frequency perception module before inputting them into the shared channel attention mechanism.

4. The method for Parkinson's disease behavior assessment based on a smartwatch according to claim 1, characterized in that: The shared channel attention mechanism consists of a pooling layer and a multi-layer perceptron. The results of the frequency perception module and the multi-scale perception module are used as inputs to the shared channel attention mechanism, which is used to learn the importance weights of different channels. Specifically, the input of the shared channel attention mechanism is input into the average pooling layer and the maximum pooling layer for feature extraction. The extracted features are further learned by a multi-layer perceptron, and the activation function is responsible for obtaining the feature learning results.

5. The method for Parkinson's disease behavior assessment based on a smartwatch according to claim 1, characterized in that: The prediction result network includes convolutional layers, normalization layers, activation functions, and maximum pooling layers. The prediction result network first uses convolutional layers to further extract features and capture the behavioral patterns of tremors. The normalization layer stabilizes the training distribution and eliminates individual and device differences. The activation function introduces nonlinearity to enhance the model's expressiveness. The maximum pooling layer focuses on key movement patterns. Finally, the fully connected layer predicts the score.

6. A Parkinson's syndrome behavior assessment device based on a smart watch, characterized in that: include: at least one processor, and a memory communicatively coupled to the at least one processor; The memory stores instructions that can be executed by the one processor, and the instructions are executed by the at least one processor so that the at least one processor executes the Parkinson's syndrome behavior assessment method based on a smart watch as described in any one of claims 1-5.

Images