Abnormal gait recognition method and system based on machine learning

Abstract

The invention relates to an abnormal gait recognition method and system based on machine learning, and belongs to the technical field of human body stepping tribology. According to the scheme, gait contact force data of subjects with normal gaits and abnormal gaits are collected, and foot sole contact force of eight phase points in each gait cycle of each subject is taken from the gait contact force data to form a data set; and training a machine learning model by using the data set to obtain an abnormal gait recognition model for judging whether the gait is abnormal or not. According to the method, eight most significant phase points are selected based on mechanical behaviors in a gait cycle, model training is performed based on sole contact force data of the eight phase points, and abnormal gaits are recognized as features, so that the recognition accuracy is improved.

Description

TECHNICAL FIELD

[0001] The application relates to an abnormal gait recognition method and system based on machine learning, and belongs to the technical field of human gait tribology. BACKGROUND

[0002] With the progress of science and technology, the life expectancy of human beings has been significantly prolonged; the economic society has developed rapidly, and the pace of life has also significantly accelerated. The incidence of diseases related to the nervous system and joint degenerative diseases, such as depression, Alzheimer's disease, Parkinson's disease, stroke, cerebral small vessel disease, lumbar disease, and osteoarthritis, has significantly increased. The diagnosis of these diseases depends on the assessment of the patient's motor symptoms by neurologists.

[0003] Taking Parkinson's disease, a typical nervous system disease, as an example, the symptoms of early Parkinson's patients are motor slowness and reduced amplitude of lower limb movement. When walking, the affected limb is dragged, and the arm swing amplitude is reduced. Abnormal gait, as an important indicator for screening early Parkinson's disease, can effectively help diagnosis. In clinical practice, experienced neurologists will quantitatively assess the patient's motor symptoms and make a diagnosis. However, this quantitative assessment relies on the experience of the doctor and may be subject to subjective misjudgment. Existing technologies have emerged to use artificial intelligence to exclude the influence of human subjective factors on abnormal gait analysis and recognition methods.

[0004] For example, the Chinese patent application with the application publication number CN117481607A discloses a Parkinson's disease and Parkinson's syndrome evaluation method and system. The method obtains the relevant data of speed, pressure, friction coefficient, single foot support, and double foot support during the walking process of normal people and Parkinson's patients, trains a machine learning network, and obtains a Parkinson's evaluation model for evaluating abnormal gait.

[0005] This method trains the evaluation model according to multiple gait data, which may introduce some irrelevant variables, and the amount of data within a gait cycle is large, which may affect the quality of modeling. SUMMARY

[0006] The purpose of the present application is to provide an abnormal gait recognition method and system based on machine learning to solve the problem of large amount of data affecting the quality of modeling in the existing evaluation of abnormal gait based on gait data.

[0007] To achieve the above-mentioned purpose, the scheme of the present application includes:

[0008] The abnormal gait recognition method based on machine learning of the present application comprises:

[0009] The gait contact force data of subjects including normal gait and abnormal gait is collected, the plantar contact force data of 8 phase points in each gait cycle of each subject is formed into a data set, and a machine learning model is trained by using the data set to obtain an abnormal gait recognition model; the input data of the same data structure as when the data set is trained is input into the abnormal gait recognition model, and the recognition result of whether the gait is abnormal is obtained according to the output of the model; the 8 phase points are: right foot force rising left foot force falling intersection, right foot landing phase force maximum point, right foot landing phase and take-off phase exchange point, right foot take-off phase force maximum point, right foot force falling left foot force rising intersection, left foot landing phase force maximum point, left foot landing phase and take-off phase exchange point, and left foot take-off phase force maximum point.

[0010] Further, the data set further includes a normalized walking time of each phase point associated with the plantar contact force of the phase point; the walking time is normalized to a unified set range by using the minimum value and the maximum value of the original time data in the corresponding gait cycle, to obtain the normalized walking time.

[0011] Further, after collecting the gait contact force data of the subjects, the gait contact force of different people's feet and the ground is compared with the body weight of the corresponding subjects, the gait contact force is normalized to remove the body weight influence, and the plantar contact force of different phase points is extracted from the normalized contact force data.

[0012] Further, the gait contact force data of the subjects or the to-be-identified object is collected by using a three-dimensional force platform, and the gait contact force is orthogonally decomposed to obtain the plantar contact force in the X, Y and Z directions corresponding to the walking time.

[0013] Further, the data set is formed by the following method:

[0014] The plantar contact force in the X, Y and Z directions of each phase point and the corresponding walking time data are data tiled to obtain a phase point data matrix, and the 8 phase point data matrices of each gait cycle of each subject form the data set.

[0015] Further, after obtaining the phase point data matrix, the plantar contact force in the corresponding direction of the phase point data matrix is normalized by using the maximum value and the minimum value of the plantar contact force data in the corresponding direction of all phase points of all subjects.

[0016] Further, the machine learning model is a random forest algorithm model.

[0017] Further, 10-fold cross-validation is used to divide the training set and the test set of the data set to train the random forest algorithm model.

[0018] Further, the integer label of the gait abnormality judgment category is converted into a one-hot encoding, and the random forest algorithm model output variable is converted back into an integer form of the category label.

[0019] The machine learning-based abnormal gait recognition method of the application has the following advantages:

[0020] The application is based on the mechanical behavior within a gait cycle, and selects the most significant 8 phase points, and based on the plantar contact force data of the 8 phase points, the model is trained and used as a feature for abnormal gait recognition, which improves the recognition accuracy.

[0021] The data is normalized to remove the influence of different body weights on the contact force, further improving the accuracy of model recognition.

[0022] Further, the data is simplified, irrelevant variables are removed, and only the plantar force in the X, Y and Z directions is retained, avoiding the influence of irrelevant variables on the recognition accuracy.

[0023] The abnormal gait recognition system of the application comprises an abnormal gait recognition model, which is trained by the following method:

[0024] The gait contact force data of subjects including normal gait and abnormal gait is collected, the plantar contact force of 8 phase points in each gait cycle of each subject is taken from the gait contact force data to form a data set, and a machine learning model is trained using the data set to obtain the abnormal gait recognition model; the 8 phase points are: right foot force rising left foot force falling intersection, right foot landing phase force maximum point, right foot landing phase and take-off phase exchange point, right foot take-off phase force maximum point, right foot force falling left foot force rising intersection, left foot landing phase force maximum point, left foot landing phase and take-off phase exchange point, left foot take-off phase force maximum point.

[0025] Further, the data set further comprises a normalized walking time of each phase point associated with the plantar contact force of the phase point; the walking time is normalized to a unified set range using the minimum and maximum values of the original time data in the corresponding gait cycle to obtain the normalized walking time.

[0026] Further, after collecting the gait contact force data of the subjects, the gait contact force of different feet and the ground is compared with the body weight of the corresponding subject, and the gait contact force is normalized to remove the influence of body weight, and the plantar contact force of different phase points is extracted from the normalized contact force data.

[0027] Further, the gait contact force data of the subject or the to-be-identified object is collected by using a three-dimensional force platform, and the gait contact force is orthogonally decomposed to obtain the plantar contact force in the X, Y and Z directions corresponding to the walking time.

[0028] Further, the data set is formed by the following method:

[0029] The plantar contact force in the X, Y and Z directions of each phase point and the corresponding walking time data are tiled to obtain a phase point data matrix, and the eight phase point data matrices of each gait cycle of each subject form the data set.

[0030] Further, after obtaining the phase point data matrix, the plantar contact force in the corresponding direction of the phase point data matrix is normalized by using the maximum value and the minimum value in the plantar contact force data in the corresponding direction of all phase points of all subjects.

[0031] Further, the machine learning model is a random forest algorithm model.

[0032] Further, 10-fold cross-validation is used to divide the data set into a training set and a test set to train the random forest algorithm model.

[0033] Further, the judgment category of whether the gait is abnormal is converted from an integer label to a one-hot encoding, and the output variable of the random forest algorithm model is converted back to an integer form of the category label.

[0034] The abnormal gait recognition system of the present application has the same beneficial effects as the abnormal gait recognition method based on machine learning of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 is a gait cycle division schematic diagram of the abnormal gait recognition method based on machine learning of the present application;

[0036] Figure 2 is an eight-phase point schematic diagram in a gait cycle of the abnormal gait recognition method based on machine learning of the present application;

[0037] Figure 3 is a random forest algorithm flowchart of the abnormal gait recognition method based on machine learning of the present application;

[0038] Figure 4 is a phase point data matrix schematic diagram constructed in the abnormal gait recognition method based on machine learning of the present application;

[0039] Figure 5 is a random forest model topological structure schematic diagram of the abnormal gait recognition method based on machine learning of the present application;

[0040] Figure 6 Figure 1 is a schematic diagram of 10-fold cross-validation of a random forest algorithm of the abnormal gait recognition method based on machine learning of the present application;

[0041] Figure 7 Figure 2 is a bar chart of accuracy test results of a random forest model of the abnormal gait recognition method based on machine learning of the present application. DETAILED DESCRIPTION

[0042] In order to make the purpose, technical solutions and advantages of the present application clearer and more apparent, the present application will be clearly and completely described in detail below in combination with the drawings and examples.

[0043] Implementation of the abnormal gait recognition method based on machine learning:

[0044] In the present application, the plantar mechanics indicators at different phase points in the walking process of a human body are used as relevant features for abnormal gait recognition of the human body.

[0045] 1. Division of gait cycle and normalization of gait parameters.

[0046] Since the walking postures of different people are inconsistent, in order to remove the influence of some irrelevant variables on the human body step friction, the present application divides the gait cycle according to the mechanical behavior in the walking process of different people, and the definition of one gait cycle is from the right heel landing to the next right heel landing again. A schematic diagram of one gait cycle T is shown in Figure Figure 1 As other embodiments, it can also be from the left heel landing to the next left heel landing again.

[0047] In order to remove the influence of different walking times on the data samples, and to normalize the walking time of one gait cycle of different people.

[0048] In the present embodiment, the normalization is performed by using the measurement method of the internal angle of a circle with certain generality and 360 as one gait cycle, i.e. the walking time of one gait cycle of different people is normalized to [0, 360]. The normalization formula is shown in (1).

[0049]

[0050] Wherein, x is the original time data in the walking process, min(x) is the minimum value of the original time data in one gait cycle (i.e. the time data at the starting point of the gait cycle, which should be 0 under normal circumstances, but in order to make the model have certain applicability, the minimum value of the original time data is used here), max(x) is the maximum value of the original time data in one gait cycle (i.e. the time data at the end point of the gait cycle), and y is the normalized data.

[0051] During walking, there is a contact point between the foot and the ground, which generates a gait contact force. This gait contact force is a vector force, and its direction and magnitude change with the gait cycle. To remove the influence of different body weights on the contact force, the contact force between the foot and the ground is normalized for different individuals. The normalization formula is shown in (2):

[0052]

[0053] Where N represents the original gait contact force data, W represents the weight of different individuals, and F represents the normalized gait contact force data.

[0054] 2. Selection of feature points within the gait cycle.

[0055] Since the amount of mechanical data in a gait cycle is huge and will affect the quality of modeling, this invention uses 8 of the most significant phase points for modeling and uses the normalized gait contact force F (hereinafter also called plantar force) during human walking as the benchmark for feature site marking.

[0056] like Figure 2 As shown, the eight selected characteristic phase points are: Phase point 1: the intersection of the right foot force rising and the left foot force falling; Phase point 2: the highest point of the right foot landing phase force; Phase point 3: the point where the right foot landing phase and takeoff phase exchange; Phase point 4: the highest point of the right foot takeoff phase force; Phase point 5: the intersection of the right foot force falling and the left foot force rising; Phase point 6: the highest point of the left foot landing phase force; Phase point 7: the point where the left foot landing phase and takeoff phase exchange; Phase point 8: the highest point of the left foot takeoff phase force. Here, "force" refers to the plantar force, for example, "right foot force" refers to the "right foot plantar force," and "the highest point of the left foot landing phase force" refers to the "highest point of the left foot landing phase plantar force."

[0057] The landing phase and takeoff phase will be explained using the right foot as an example.

[0058] Right foot landing phase:

[0059] Starting point: Right heel touches the ground (at this time, the left foot is at the end of the takeoff phase and is about to leave the ground); End point: Right foot fully bears weight, and the body's center of gravity shifts to directly above the right foot (at this time, the left foot is completely off the ground and enters the swing phase; during the swing phase, there is no plantar force as the left foot is not in contact with the ground). During this process, the change in the height of the body's center of gravity will generate the highest point of plantar force in the right foot landing phase (i.e., phase point 2). The dynamic significance of the right foot landing phase is that the right foot absorbs the forward momentum of the body, achieving braking.

[0060] Takeoff phase with right foot:

[0061] Starting point: the right heel is lifted, and the body weight is moved forward beyond the support point (at this time, the left foot is in the initial stage of landing, and the heel just touches the ground). End point: the right toe is completely off the ground (the left foot enters the fully loaded state), and during this process, the body is accelerated forward by the right foot, thus generating a maximum point of the right foot take-off phase foot force (i.e., phase point 4). The dynamics of the right foot take-off phase is that the right foot accelerates the body forward.

[0062] The direction of the normalized foot force F has a certain relationship with the body center of gravity at the corresponding time during walking, so the directions of this vector force F at the same phase point are different for different people, especially for patients with abnormal gait, there will be obvious differences. The present application captures this difference through the model to identify abnormal gait.

[0063] 3. Reference Figure 3 The establishment of the abnormal gait recognition model based on machine learning is as follows.

[0064] The three-dimensional force platform system is used to collect gait contact force data of normal people and people with abnormal gait (such as Parkinson's patients), and further data processing (such as data cleaning, denoising and normalization) and orthogonal decomposition are performed to obtain foot contact force in each direction. The foot contact force of the above-mentioned 8 phase points of each subject is extracted as the data set for model training, the data set is divided into training set and test set, and the model is trained.

[0065] The three-dimensional force platform will normalize the original contact force N to remove the influence of different body weights on the contact force, and obtain the normalized foot force F; then the normalized foot force F is orthogonally decomposed into X, Y, Z direction foot contact force (F X , F Y , F Z ), wherein the X axis is parallel to the horizontal plane and consistent with the walking direction, the Z axis is perpendicular to the horizontal plane upward, and the Y axis is parallel to the horizontal plane and perpendicular to the walking direction. At the same time, according to the time data recorded during the collection of the subject's foot contact force by the three-dimensional force platform, the gait cycle is counted.

[0066] Specifically, in the present embodiment, the three-dimensional force platform system adopts Kistler three-dimensional force platform, which is a multi-component aluminum plate force platform that can measure ground reaction force, moment and pressure center in gait and balance analysis. The measurement range of Fx and Fy is between-2.5KN and 2.5KN, and the measurement range of Fz is between 0 and 5KN. The natural frequency: X axis and Y axis direction: about 400Hz, Z axis direction: about 200Hz.

[0067] The embodiment takes Parkinson's patients as an example to explore the internal correlation between the plantar mechanical indicators of normal people and Parkinson's patients, realizes the recognition of abnormal gait of Parkinson's patients under data driving, and sorts out 400 volunteers of different genders as subjects for collecting data and training models, including 200 normal people and 200 Parkinson's patients.

[0068] The embodiment first extracts the features of each gait cycle at 8 phase points for each subject, and establishes a gait feature database based on this. Then, 10-fold cross-validation is used to divide the training set and the test set, and different training sets are modeled based on the random forest algorithm. Finally, the performance of the model on the whole data set is evaluated.

[0069] 3.1, data preprocessing.

[0070] Each phase point corresponds to the plantar contact force in X, Y and Z directions. In the embodiment, the data tiling method is used for modeling, and the data of one phase point of a person is constructed into a 1x32 phase point data matrix as shown in Figure 4 , which includes 8 time phase point data representing the occurrence time of the corresponding phase point, 8 X direction contact force phase point data representing the X direction contact force F X , 8 Y direction contact force phase point data representing the Y direction contact force F Y and 8 Z direction contact force phase point data representing the Z direction contact force F Z . Among them, the time phase point data is the walking time after normalization in the corresponding gait cycle.

[0071] After constructing the phase point data matrix, in order to eliminate the differences between the data, the normalized processing is performed on the plantar contact force in each direction in the sorted input and output data before machine learning modeling, and the formula is: where x is the plantar contact force data in a certain direction to be normalized, x norm is the data after normalization, x max , x min are the maximum and minimum values of the plantar contact force data in the corresponding direction before normalization.

[0072] 3.2, model establishment.

[0073] After many experiments, it is found that the judgment accuracy of the model is the highest when using the random forest machine learning algorithm, so the machine learning algorithm of random forest is adopted in this embodiment. The random forest model is a machine learning algorithm containing multiple decision trees. Among them, the input of the model is 32 input variables, which are 8 time phase point data, 8 time phase point corresponding F Z directional contact force phase point data, F X directional contact force phase point data and F Y directional contact force phase point data. The output of the model is two output variables. The structure of the model adopts a random forest model with 50 decision trees and a minimum leaf number of 2 for modeling.

[0074] The topology structure of the random forest model is shown in Figure 5

[0075] In order to increase the accuracy of judging patients with gait abnormalities and normal people, we convert the judgment category from integer label to one-hot encoding, and then convert the output variable from one-hot encoding back to integer form category label. Among them, the training samples are encoded, and the normal people are 01 and the patients are 10. Finally, the output variable is converted to integer form.

[0076] The trained optimal model can be used as an abnormal gait recognition model. The three-dimensional force platform system is used to collect the gait contact force of the diagnosis object to be recognized for abnormal gait recognition. The same way as in the model training is used for gait cycle division and data normalization processing, and the data of 8 feature phase points in each gait cycle are selected to form a phase point data matrix. The to-be-recognized data composed of the phase point data matrix is input (gait feature point input) into the trained abnormal gait recognition model, and according to the output category label (gait feature point output), the recognition result of whether the gait of the diagnosis object is abnormal is obtained.

[0077] 3.3, Model accuracy test.

[0078] For the test of the accuracy of the model, the 10-fold cross-validation is used to test the accuracy of the embodiment. The 10-fold cross-validation is a commonly used model evaluation method in statistics, which is used to estimate the performance of a prediction model on independent data sets. First, as Figure 6 ​As shown, data partitioning is performed on the dataset, and the dataset is randomly divided into 10 equal-sized folds. Then, model training and verification are performed, and in each round, one fold is selected as the verification set, and the remaining 9 folds are selected as the training set. The model is trained using the training set, and at the same time, the model is tested using the verification set, and the performance indicators are calculated. Finally, repeated training is performed, and the above steps are repeated 10 times, and each time a different fold is selected as the verification set. The average of the performance indicators obtained by the 10 verifications can obtain the performance of the overall dataset.

[0079] wherein the test accuracy formula is:

[0080]

[0081] In the formula, TP represents the number of correctly predicted patient samples; TN represents the number of correctly predicted normal samples; FP represents the number of incorrectly predicted patient samples; and FN represents the number of incorrectly predicted normal samples.

[0082] 3.4, the performance verification and evaluation of the abnormal gait recognition model of the application.

[0083] Figure 7 For the accuracy test results of the abnormal gait recognition model of the application, it can be seen that the accuracy of the 10 training sets is above 99, and the average accuracy is 99.92%; at the same time, it can also be seen that the accuracy of the 10 test sets is above 80%, and the average accuracy is 90.2%, the average accuracy of the training set and the test set is shown in Table 1 below, the abnormal gait recognition model of the application shows good performance. It is proved that the random forest algorithm in the application has good recognition effect on 8 phase points in a gait cycle, and has good significance for recognizing abnormal gait of Parkinson's disease.

[0084] Abnormal gait recognition system implementation:

[0085] The abnormal gait recognition system of the application is a computer system, wherein the abnormal gait recognition model in the abnormal gait recognition method implementation based on machine learning is stored, the gait contact force of the diagnostic object to be subjected to abnormal gait recognition is collected by using a three-dimensional force platform system, the same way as in the model training is used for gait cycle division and data normalization processing, the data of 8 characteristic phase points in each gait cycle are selected to form a phase point data matrix, and the to-be-recognized data composed of the phase point data matrix is input (gait feature point input) into the trained abnormal gait recognition model in the abnormal gait recognition system of the application. According to the output class label (gait feature point output), the recognition result of whether the gait of the diagnostic object is abnormal is obtained.

[0086] The training of the abnormal gait recognition system and the acquisition, data processing, etc. of the data set for training have been described in the machine learning-based abnormal gait recognition method embodiments clearly enough, and will not be described here again.

Claims

1. A method for identifying abnormal gait based on machine learning, characterized in that, include: Gait contact force data were collected from subjects with normal and abnormal gait. From this data, eight phase points of plantar contact force were extracted for each subject within each gait cycle to form a dataset. This dataset was used to train a machine learning model to obtain an abnormal gait recognition model. The same data structure used during the model training was applied to the abnormal gait recognition model, representing the plantar contact force at eight phase points within each gait cycle of the subject to be identified. The model's output determined whether the gait was abnormal. The eight phase points were: the intersection of right foot force rising and left foot force falling; the highest point of right foot landing phase force; the point where right foot landing and takeoff phases interchange; the highest point of right foot takeoff phase force; the intersection of right foot force falling and left foot force rising; the highest point of left foot landing phase force; the point where left foot landing and takeoff phases interchange; and the highest point of left foot takeoff phase force.

2. The abnormal gait recognition method based on machine learning according to claim 1, characterized in that, The dataset also includes the normalized walking time for each phase point, which is associated with the plantar contact force at that phase point; The walking time is normalized to a uniform set range by using the minimum and maximum values ​​of the original time data in the corresponding gait cycle, thus obtaining the normalized walking time.

3. The abnormal gait recognition method based on machine learning according to claim 1, characterized in that, After collecting the gait contact force data of the subjects, the gait contact force of different people's feet to the ground was compared with the weight of the corresponding subjects. The gait contact force was normalized to remove the influence of weight, and the plantar contact force at different phase points was extracted from the normalized contact force data.

4. The abnormal gait recognition method based on machine learning according to claim 1, characterized in that, The gait contact force data of the subject or the object to be identified is collected using a three-dimensional force measuring platform. The gait contact force is orthogonally decomposed to obtain the plantar contact force in the X, Y and Z directions corresponding to the walking time.

5. The abnormal gait recognition method based on machine learning according to claim 4, characterized in that, The dataset was formed using the following method: The foot contact force and corresponding walking time data in the X, Y, and Z directions of each phase point are tiled to obtain a phase point data matrix. The dataset consists of 8 phase point data matrices for each subject in each gait cycle.

6. The abnormal gait recognition method based on machine learning according to claim 5, characterized in that, After obtaining the phase point data matrix, the maximum and minimum values ​​of the plantar contact force data in the corresponding directions of all phase points of all subjects are used to normalize the plantar contact force in the corresponding directions of the phase point data matrix.

7. The abnormal gait recognition method based on machine learning according to claim 1, characterized in that, The machine learning model is the random forest algorithm model.

8. The abnormal gait recognition method based on machine learning according to claim 7, characterized in that, The dataset was divided into training and test sets using 10-fold cross-validation to train the random forest algorithm model.

9. The abnormal gait recognition method based on machine learning according to claim 7, characterized in that, The gait abnormality judgment category is converted from integer label to one-hot encoding, and then the output variable of the random forest algorithm model is converted back into integer category label by one-hot encoding.

10. An abnormal gait recognition system, characterized in that, This includes the abnormal gait recognition model in the abnormal gait recognition method based on machine learning as described in any one of claims 1 to 9.

Images