Subtype Parkinson's disease type identification system based on electroencephalogram signals and deep learning

Abstract

The invention discloses a subtype Parkinson's disease type identification system based on electroencephalogram signals and deep learning. The system comprises the steps that basic information, motion symptom information, non-motion symptom information and objective indexes of a person to be identified are collected; assigning values to the basic information; performing noise reduction and filtering on tremor signals and corresponding electroencephalogram signals thereof, rigid signals and corresponding electroencephalogram signals thereof, posture signals and gait signals and corresponding electroencephalogram signals thereof, and motion delay signals and corresponding electroencephalogram signals thereof in the motion symptom information; scoring each piece of information in the non-exercise symptom information; gender and age assignment features are extracted from the basic information; extracting time domain features and frequency domain features of each signal from the exercise symptom information; respectively extracting score assignment features from each piece of information in the non-exercise symptom information; constructing a one-dimensional feature vector based on each feature; and inputting the one-dimensional feature vector into a target convolutional neural network type recognition model to perform type recognition of the subtype Parkinson's disease, and outputting the type of the subtype Parkinson's disease of the person to be recognized.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a subtype Parkinson's disease type recognition system based on electroencephalogram signals and deep learning. Background Art

[0002] Parkinson's disease (PD) is a common chronic, progressive, degenerative neurological disease. This has been proven by more and more evidence-based medicine. The manifestations of Parkinson's disease are complex and diverse, including tremor, rigidity, bradykinesia, postural instability, etc. These symptoms are important factors affecting the quality of life of Parkinson's patients and may even lead to progressive functional disorders. In addition, due to different factors, the constitution, clinical symptoms, age, etc. of Parkinson's patients show large individual differences and heterogeneity. The clinical characteristics of Parkinson's patients help to further understand the pathogenesis, clinical progression and individualized treatment of Parkinson's disease.

[0003] In order to objectively reflect the clinical heterogeneity of Parkinson's disease, scholars used data-driven classification technology to analyze Parkinson's disease patients. One of the most common classifications is the tremor-dominant (TD) subtype, the postural instability and gait difficulty (PIGD) subtype, and the mixed subtype. By comparing the clinical characteristics of TD and PIGD patients, it was found that most PIGD subtype Parkinson's disease patients had rapid disease progression and poor clinical prognosis.

[0004] In addition to typical motor symptoms, many Parkinson's disease patients also have some non-motor symptoms, such as sleep disorders, depression and anxiety, vision loss, gastrointestinal disorders, etc. These symptoms often exist before the diagnosis of Parkinson's disease and may be important precursor features of Parkinson's disease. Some clinical reports show that patients with the PIGD subtype have a higher risk of cognitive impairment, dementia, hallucinations, depression, apathy and other psychological and emotional problems, which are usually unresponsive or poorly responsive to dopaminergic treatment. The clinical symptoms of the TD subtype are milder, motor and cognitive symptoms progress more slowly, and the risk of dementia and mental illness is lower.

[0005] Tus believes that early detection of different subtypes based on the treatment of motor symptoms and individualized intervention of non-motor symptoms related to the subtype may be an important way to delay disease progression and improve patient prognosis. However, previous studies were mostly based on a single motor symptom questionnaire, and few studies reported on the differentiation of non-motor symptoms of different subtypes, let alone the technology to comprehensively consider motor symptoms and non-motor symptoms to identify the subtypes of Parkinson's disease. Summary of the invention

[0006] In view of the problems and shortcomings of the prior art, the present invention provides a subtype Parkinson's disease type identification system based on EEG signals and deep learning.

[0007] The present invention solves the above technical problems through the following technical solutions:

[0008] The present invention provides a subtype Parkinson's disease type identification system based on EEG signals and deep learning, which is characterized in that it includes an information acquisition module, an information preprocessing module, a feature extraction module, a feature vector construction module and a type identification module;

[0009] The information collection module is used to collect basic information, motor symptom information, non-motor symptom information and objective indicators of the person to be identified;

[0010] Among them, basic information includes gender and age; movement symptom information includes tremor data, stiffness data, postural instability and gait disorder data and bradykinesia data detected by multimodal sensors, tremor data includes tremor signals and their corresponding electroencephalogram signals, stiffness data includes stiffness signals and their corresponding electroencephalogram signals, postural instability and gait disorder data includes posture signals, gait signals and their corresponding electroencephalogram signals, and bradykinesia data includes bradykinesia signals and their corresponding electroencephalogram signals;

[0011] Non-motor symptom information includes gastrointestinal symptom information, urinary tract symptom information, cardiovascular symptom information, temperature regulation symptom information, pupil adjustment symptom information, sexual function symptom information, sleep disorder symptom information, rapid eye movement sleep behavior disorder symptom information (RBD symptom information), anxiety level symptom information and depression level symptom information; objective indicators include salivary cortisol content and salivary melatonin content;

[0012] The information preprocessing module is used to perform a value assignment preprocessing operation on the basic information; for the motor symptom information, perform a noise reduction and filtering preprocessing operation on the tremor signal and its corresponding EEG signal, the rigidity signal and its corresponding EEG signal, the posture signal and gait signal and its corresponding EEG signal, and the bradykinesia signal and its corresponding EEG signal; and perform a scoring preprocessing operation on each piece of information in the non-motor symptom information;

[0013] The feature extraction module is used to extract gender-assigned features and age-assigned features based on basic information; for motor symptom information, extract tremor time domain features and tremor frequency domain features in tremor signals, extract tremor EEG time domain features and tremor EEG frequency domain features in EEG signals corresponding to tremor signals, extract tonic time domain features and tonic frequency domain features in tonic signals, extract tonic EEG time domain features and tonic EEG frequency domain features in EEG signals corresponding to tonic signals, extract posture time domain features and posture frequency domain features in posture signals, extract gait time domain features and gait frequency domain features in gait signals, extract EEG signals corresponding to posture signals Extract the postural EEG time domain features and postural EEG frequency domain features in the signal, extract the bradykinesia time domain features and bradykinesia frequency domain features in the bradykinesia signal, and extract the bradykinesia time domain features and bradykinesia frequency domain features in the EEG signal corresponding to the bradykinesia signal; for each information in the non-motor symptom information, extract the gastrointestinal tract score assignment features, urinary tract score assignment features, cardiovascular score assignment features, body temperature regulation score assignment features, pupil adjustment score assignment features, sexual function score assignment features, sleep disorder score assignment features, RBD score assignment features, anxiety level score assignment features and depression level score assignment features respectively;

[0014] The feature vector construction module is used to construct a one-dimensional feature vector based on each feature=(gender assignment feature, age assignment feature, tremor time domain feature, tremor frequency domain feature, tremor EEG time domain feature, tremor EEG frequency domain feature, tonic time domain feature, tonic frequency domain feature, tonic EEG time domain feature, tonic EEG frequency domain feature, posture time domain feature, posture frequency domain feature, gait time domain feature, gait frequency domain feature, posture EEG time domain feature, posture EEG frequency domain feature, bradycy time domain feature, bradycy frequency domain feature, bradycy time domain feature, bradycy frequency domain feature, gastrointestinal tract score assignment feature, urinary tract score assignment feature, cardiovascular score assignment feature, body temperature regulation score assignment feature, pupil adjustment score assignment feature, sexual function score assignment feature, sleep disorder score assignment feature, RBD score assignment feature, anxiety level score assignment feature, depression level score assignment feature, salivary cortisol content feature, salivary melatonin content feature);

[0015] The type recognition module is used to input the one-dimensional feature vector of the person to be identified into the target convolutional neural network type recognition model constructed based on the deep learning algorithm to perform type recognition of subtype Parkinson's disease, thereby identifying and outputting the specific subtype Parkinson's disease type of the person to be identified. The subtype Parkinson's disease types include tremor-dominant subtype Parkinson's disease, namely TD subtype, postural instability and gait difficulty subtype Parkinson's disease, namely PIGD subtype, and mixed subtype Parkinson's disease.

[0016] The positive and progressive effects of the present invention are:

[0017] The present invention simultaneously takes into account the basic information, motor symptom information, non-motor symptom information and objective indicators of the person to be identified, and extracts multiple features based on these factors. These features and the target convolutional neural network type recognition model constructed based on the deep learning algorithm can accurately identify the type of subtype Parkinson's disease of the person to be identified. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 Schematic diagram of the structure of a subtype Parkinson's disease type identification system according to a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0020] like Figure 1 As shown, this embodiment provides a subtype Parkinson's disease type identification system based on EEG signals and deep learning, which includes an information acquisition module 1, an information preprocessing module 2, a feature extraction module 3, a feature vector construction module 4, a type identification module 5 and a model construction module 6.

[0021] The information collection module 1 is used to collect basic information, motor symptom information, non-motor symptom information and objective indicators of the person to be identified.

[0022] Among them, basic information includes gender and age.

[0023] The motor symptom information includes tremor data, rigidity data, postural instability and gait disorder data, and bradykinesia data. Tremor data includes tremor signals (obtained by tremor sensor detection) and their corresponding EEG signals (obtained by EEG sensor detection), rigidity data includes rigidity signals (obtained by flexible tactile sensor detection) and their corresponding EEG signals (obtained by EEG sensor detection), postural instability and gait disorder data includes posture signals (obtained by posture sensor detection), gait signals (obtained by gait sensor detection) and their corresponding EEG signals (obtained by EEG sensor detection), and bradykinesia data includes bradykinesia signals detected by inertial sensor and their corresponding EEG signals (obtained by EEG sensor detection).

[0024] Non-motor symptom information includes gastrointestinal symptom information, urinary tract symptom information, cardiovascular symptom information, temperature regulation symptom information, pupil adjustment symptom information, sexual function symptom information, sleep disorder symptom information, rapid eye movement sleep behavior disorder symptom information (RBD symptom information), anxiety level symptom information and depression level symptom information.

[0025] Objective indicators include salivary cortisol content and salivary melatonin content.

[0026] The information preprocessing module 2 is used to perform value assignment preprocessing operations on basic information (such as assigning 1 to males and 0 to females); for motor symptom information, perform noise reduction, filtering and artifact removal preprocessing operations on tremor signals and their corresponding EEG signals, rigidity signals and their corresponding EEG signals, posture signals and gait signals and their corresponding EEG signals, bradykinesia signals and their corresponding EEG signals; perform scoring preprocessing operations on each information in the non-motor symptom information (such as for gastrointestinal symptom information, give a scoring value based on the described gastrointestinal related information); no preprocessing operations are required for objective indicators.

[0027] The feature extraction module 3 is used to extract gender-assigned features and age-assigned features based on basic information; extract tremor time domain features and tremor frequency domain features in the tremor signal based on motor symptom information, extract tremor EEG time domain features and tremor EEG frequency domain features in the EEG signal corresponding to the tremor signal, extract tonic time domain features and tonic frequency domain features in the tonic signal, extract tonic EEG time domain features and tonic EEG frequency domain features in the EEG signal corresponding to the tonic signal, extract posture time domain features and posture frequency domain features in the posture signal, extract gait time domain features and gait frequency domain features in the gait signal, and extract the EEG corresponding to the posture signal. The postural EEG time domain features and postural EEG frequency domain features in the signal are extracted, the bradykinesia time domain features and bradykinesia frequency domain features in the bradykinesia signal are extracted, and the bradykinesiology time domain features and bradykinesiology frequency domain features in the EEG signal corresponding to the bradykinesiology signal are extracted; for each information in the non-motor symptom information, the gastrointestinal tract scoring assignment features, urinary tract scoring assignment features, cardiovascular scoring assignment features, body temperature regulation scoring assignment features, pupil adjustment scoring assignment features, sexual function scoring assignment features, sleep disorder scoring assignment features, RBD scoring assignment features, anxiety level scoring assignment features and depression level scoring assignment features are extracted respectively.

[0028] Among them, for the time domain features of tremor, time domain features of rigidity, time domain features of posture, time domain features of gait and time domain features of bradycy, the extracted time domain features include amplitude, standard deviation and similarity entropy; for the frequency domain features of tremor, frequency domain features of rigidity, frequency domain features of posture, frequency domain features of gait and frequency domain features of bradycy, the extracted frequency domain features include frequency, maximum peak value of spectrum and spectrum energy value; for the time domain features of tremor, time domain features of rigidity, time domain features of posture, time domain features of gait and time domain features of bradycy, the extracted frequency domain features include frequency, maximum peak value of spectrum and spectrum energy value; Time domain features, the extracted time domain features include amplitude, standard deviation and similarity entropy; for tremor EEG frequency domain features, tonic EEG frequency domain features, postural EEG frequency domain features and bradyencephalogram frequency domain features, the extracted frequency domain features include the power spectral density of delta wave, theta wave, alpha wave, beta wave and gamma wave bands, the delta wave frequency band is 1-4Hz, the theta wave frequency band is 4-8Hz, the alpha wave frequency band is 8-12Hz, the beta wave frequency band is 13-30Hz and the gamma wave frequency band is 31-50Hz.

[0029] For each information in the non-motor symptom information, the extracted gastrointestinal tract score assignment feature = gastrointestinal tract score value * gastrointestinal tract score weight, the extracted urinary tract score assignment feature = urinary tract score value * urinary tract score weight, the extracted cardiovascular score assignment feature = cardiovascular score value * cardiovascular score weight, the extracted temperature regulation score assignment feature = temperature regulation score value * temperature regulation score weight, the extracted pupil accommodation score assignment feature = pupil accommodation score value * pupil accommodation score weight, the extracted sexual function score assignment feature = sexual function score value * sexual function score weight, the extracted sleep disorder score assignment feature = sleep disorder score value * sleep disorder score weight, the extracted RBD score assignment feature = RBD score value * RBD score weight, the extracted anxiety level score assignment feature = anxiety level score value * anxiety level score weight, the extracted depression level score assignment feature = depression level score value * depression level score weight.

[0030] Among these weights, the gastrointestinal tract score weight, body temperature regulation score weight, RBD score weight and anxiety level score weight are relatively large; the urinary tract score weight, cardiovascular score weight, pupil accommodation score weight, sexual function score weight, sleep disorder score weight and depression level score weight are relatively small.

[0031] The feature vector construction module 4 is used to construct a one-dimensional feature vector of the person to be identified based on various features = (gender assignment feature, age assignment feature, tremor time domain feature, tremor frequency domain feature, tremor EEG time domain feature, tremor EEG frequency domain feature, tonic time domain feature, tonic frequency domain feature, tonic EEG time domain feature, tonic EEG frequency domain feature, posture time domain feature, posture frequency domain feature, gait time domain feature, gait frequency domain feature, posture EEG time domain feature, posture EEG frequency domain feature, bradycardia time domain feature, bradycardia frequency domain feature, bradycardia time domain feature, bradycardia frequency domain feature, gastrointestinal tract score assignment feature, urinary tract score assignment feature, cardiovascular score assignment feature, body temperature regulation score assignment feature, pupil adjustment score assignment feature, sexual function score assignment feature, sleep disorder score assignment feature, RBD score assignment feature, anxiety level score assignment feature, depression level score assignment feature, salivary cortisol content feature, salivary melatonin content feature).

[0032] The type recognition module 5 is used to input the one-dimensional feature vector of the person to be identified into the target convolutional neural network type recognition model constructed based on the deep learning algorithm to perform type recognition of subtype Parkinson's disease, thereby identifying and outputting the specific subtype Parkinson's disease type of the person to be identified. The subtypes of Parkinson's disease include tremor-dominant subtype Parkinson's disease (TD subtype), postural instability and gait difficulty subtype Parkinson's disease (PIGD subtype), and mixed subtype Parkinson's disease.

[0033] With respect to the target convolutional neural network type recognition model applied in the type recognition module 5, the model construction module 6 is used to construct this target convolutional neural network type recognition model.

[0034] Specifically, the information collection module 1 is called to collect basic information, motor symptom information, non-motor symptom information and objective indicators of multiple past patients, and the type of subtype Parkinson's disease of each past patient is marked. These past patients include male and female patients of different age groups with TD subtype, male and female patients of different age groups with PIGD subtype, and male and female patients of different age groups with mixed subtype. This makes the training samples diverse, and thus makes the constructed convolutional neural network type recognition model more accurate.

[0035] Among them, the basic information of each past patient includes gender and age; the motor symptom information of each past patient includes tremor data, rigidity data, postural instability and gait disorder data and bradykinesia data, each tremor data includes a tremor signal and its corresponding EEG signal, each rigidity data includes a rigidity signal and its corresponding EEG signal, each postural instability and gait disorder data includes a posture signal, a gait signal and its corresponding EEG signal, and each bradykinesia data includes a bradykinesia signal and its corresponding EEG signal; the non-motor symptom information of each past patient includes gastrointestinal symptom information, urinary tract symptom information, cardiovascular symptom information, temperature regulation symptom information, pupil adjustment symptom information, sexual function symptom information, sleep disorder symptom information, rapid eye movement sleep behavior disorder symptom information, i.e., RBD symptom information, anxiety level symptom information and depression level symptom information; the objective indicators of each past patient include salivary cortisol content and salivary melatonin content.

[0036] Information preprocessing module 2 is called to perform value preprocessing operations on the basic information of each past patient; for the motor symptom information of each past patient, tremor signals and their corresponding EEG signals, rigidity signals and their corresponding EEG signals, posture signals and gait signals and their corresponding EEG signals, bradykinesia signals and their corresponding EEG signals are preprocessed by noise reduction, filtering and artifact removal; and scoring preprocessing operations are performed on each information in the non-motor symptom information of each past patient.

[0037] Call feature extraction module 3 to extract gender-assigned features and age-assigned features for each past patient's basic information; extract tremor time domain features and tremor frequency domain features from the tremor signal for each past patient's motor symptom information, extract tremor EEG time domain features and tremor EEG frequency domain features from the EEG signal corresponding to the tremor signal, extract tonic time domain features and tonic frequency domain features from the tonic signal, extract tonic EEG time domain features and tonic EEG frequency domain features from the EEG signal corresponding to the tonic signal, extract posture time domain features and posture frequency domain features from the posture signal, extract gait time domain features and gait frequency domain features from the gait signal, extract posture signal corresponding to the posture signal. The postural EEG time domain features and postural EEG frequency domain features in the EEG signals are extracted, the bradykinesia time domain features and bradykinesia frequency domain features in the bradykinesia signals are extracted, and the bradykinesiia time domain features and bradykinesiia frequency domain features in the EEG signals corresponding to the bradykinesiia signals are extracted; for each past patient's non-motor symptom information, the gastrointestinal tract score assignment features, urinary tract score assignment features, cardiovascular score assignment features, body temperature regulation score assignment features, pupil adjustment score assignment features, sexual function score assignment features, sleep disorder score assignment features, RBD score assignment features, anxiety level score assignment features and depression level score assignment features are extracted respectively.

[0038] Among these weights, the gastrointestinal tract score weight, body temperature regulation score weight, RBD score weight and anxiety level score weight are relatively large; the urinary tract score weight, cardiovascular score weight, pupil accommodation score weight, sexual function score weight, sleep disorder score weight and depression level score weight are relatively small.

[0039] Call the feature vector construction module 4, and construct the corresponding one-dimensional feature vector based on each feature of each past patient, respectively = (gender assignment feature, age assignment feature, tremor time domain feature, tremor frequency domain feature, tremor EEG time domain feature, tremor EEG frequency domain feature, rigidity time domain feature, rigidity frequency domain feature, rigidity EEG time domain feature, rigidity EEG frequency domain feature, posture time domain feature, posture frequency domain feature, gait time domain feature, gait frequency domain feature, posture EEG time domain feature, posture EEG frequency domain feature, bradykines time domain feature , bradyencephalogram frequency domain features, bradyencephalogram time domain features, bradyencephalogram frequency domain features, gastrointestinal tract score assignment features, urinary tract score assignment features, cardiovascular score assignment features, temperature regulation score assignment features, pupil adjustment score assignment features, sexual function score assignment features, sleep disorder score assignment features, RBD score assignment features, anxiety level score assignment features, depression level score assignment features, salivary cortisol content features, salivary melatonin content features), thereby obtaining the one-dimensional feature vector of each past patient.

[0040] A sample set is constructed, which includes the one-dimensional feature vectors of each past patient and the corresponding subtype of Parkinson's disease.

[0041] Convolutional neural network is used to construct the convolutional neural network type recognition model, and the particle swarm improvement algorithm is used to optimize the hyperparameters in the training process of the convolutional neural network type recognition model. The optimal hyperparameters are obtained by continuous updating and optimization, and the target convolutional neural network type recognition model is established using the optimized hyperparameters. Among them, the convolutional neural network type recognition model consists of multiple convolutional layers, each convolutional layer is followed by a pooling layer, the last pooling layer is followed by a fully connected layer, and the fully connected layer is followed by a Softmax classifier, in which the number of convolution kernels of each convolutional layer is set to be a multiple of 2, and the number of convolution kernels of the previous convolutional layer is greater than the number of convolution kernels of the next convolutional layer.

[0042] The specific process of optimization using particle swarm improved algorithm is as follows:

[0043] Initialization: Initialize the hyperparameters of the model, including the number of convolutional layers, the number of convolution kernels of each convolutional layer, the number of neurons in the fully connected layer, and the learning rate. Set the number of particles in the particle swarm to N, initialize the speed and position of each particle in the particle swarm, set the maximum number of iterations T and the particle space dimension D that corresponds to the hyperparameters one by one. The position of each particle represents a set of hyperparameters. Set the number of convolutional layers, the number of convolution kernels of each convolutional layer, the number of neurons in the fully connected layer, and the range of learning rates.

[0044] Based on the initialized position, the individual optimal position and the global optimal position are obtained: the hyperparameters corresponding to the spatial position of any particle are substituted into the convolutional neural network type recognition model and trained using the sample set. The fitness of the particle is calculated using the positively correlated fitness function established by the model training accuracy. In this way, the fitness corresponding to each particle can be obtained as the individual optimal fitness and the maximum value can be taken as the global optimal fitness. The current position of each particle is taken as the corresponding individual optimal position, and the current position of the particle corresponding to the global optimal fitness is taken as the global optimal position.

[0045] Update the speed and position of each particle:

[0046]

[0047] Among them, W t represents the inertia weight of the t-th iteration, and represents the velocity and position of the tth iteration, and represents the speed and position of the t+1th iteration, C1 and C2 represent learning factors, rand1 and rand2 represent random numbers (0,1), represents the optimal position of the individual after the tth iteration, represents the global optimal position after the tth iteration, W min represents the minimum inertia weight, W max represents the maximum inertia weight, and η represents the inertia adjustment factor.

[0048] Update the individual optimal position and the global optimal position: use the fitness function to calculate the fitness of each particle after the tth iteration. Only when the fitness of the nth particle after the update is greater than the individual optimal fitness of the nth particle, update the individual optimal fitness of the nth particle = the fitness of the nth particle after the update, and update the individual optimal position of the nth particle = the updated position of the nth particle. n is 1-N in turn, so that the individual optimal fitness and individual optimal position of each particle can be obtained. Then, the maximum value of each individual optimal fitness is taken as the current optimal fitness. Only when the current optimal fitness is greater than the current global optimal fitness, update the global optimal fitness = the current optimal fitness, and update the global optimal position = the current position of the particle to which the current optimal fitness belongs.

[0049] Cut-off condition: When the global optimal fitness reaches the preset value or the number of iterations reaches the maximum number of iterations T, the iteration ends. At this time, the optimal hyperparameters are obtained based on the global optimal position, and the target convolutional neural network type recognition model is established using the optimal hyperparameters. Otherwise, it is further determined whether t≤T / 2. When t≤T / 2, the current individual optimal fitness of each particle is arranged in descending order, and the last n1 particles are discarded. The particle with the global optimal fitness is used as the main elite particle, and n1 particles are randomly selected from the remaining particles that have not been discarded as the slave elite particles. The main elite particles are divided into Genetic hybridization is performed with each selected slave elite particle to obtain n1 particles. The current position of each particle in the n1 particles obtained by hybridization is the position of the particle after the t-th iteration. The current speed of each particle in the n1 particles obtained by hybridization is the speed of the particle after the t-th iteration = the average speed of the main elite particle and the corresponding slave elite particle participating in the hybridization. In this way, a new particle population can be obtained, and the speed and position of each particle in the new particle population are continuously updated. n1 is a positive integer and 1≤n1≤L / 3, t=t+1, and when t>T / 2, the speed and position of each particle are directly updated.

[0050] In this embodiment, t is divided into two cases: t≤T / 2 and t>T / 2. When t≤T / 2, bad particles are discarded and good particles are hybridized to replace the bad particles to achieve rapid convergence. When t>T / 2, most of the particles are good particles and no particle replacement is required.

[0051] Although the specific embodiments of the present invention are described above, those skilled in the art should understand that these are only examples, and the protection scope of the present invention is defined by the appended claims. Those skilled in the art may make various changes or modifications to these embodiments without departing from the principles and essence of the present invention, but these changes and modifications all fall within the protection scope of the present invention.

Claims

1. A subtype Parkinson's disease type identification system based on EEG signals and deep learning, characterized in that: It includes an information collection module, an information preprocessing module, a feature extraction module, a feature vector construction module and a type recognition module; The information collection module is used to collect basic information, motor symptom information, non-motor symptom information and objective indicators of the person to be identified; Among them, basic information includes gender and age; movement symptom information includes tremor data, stiffness data, postural instability and gait disorder data and bradykinesia data detected by multimodal sensors, tremor data includes tremor signals and their corresponding electroencephalogram signals, stiffness data includes stiffness signals and their corresponding electroencephalogram signals, postural instability and gait disorder data includes posture signals, gait signals and their corresponding electroencephalogram signals, and bradykinesia data includes bradykinesia signals and their corresponding electroencephalogram signals; Non-motor symptom information includes gastrointestinal symptom information, urinary tract symptom information, cardiovascular symptom information, temperature regulation symptom information, pupil adjustment symptom information, sexual function symptom information, sleep disorder symptom information, rapid eye movement sleep behavior disorder symptom information (RBD symptom information), anxiety level symptom information and depression level symptom information; objective indicators include salivary cortisol content and salivary melatonin content; The information preprocessing module is used to perform a value assignment preprocessing operation on the basic information; for the motor symptom information, perform a noise reduction and filtering preprocessing operation on the tremor signal and its corresponding EEG signal, the rigidity signal and its corresponding EEG signal, the posture signal and gait signal and its corresponding EEG signal, and the bradykinesia signal and its corresponding EEG signal; and perform a scoring preprocessing operation on each piece of information in the non-motor symptom information; The feature extraction module is used to extract gender-assigned features and age-assigned features based on basic information; for motor symptom information, extract tremor time domain features and tremor frequency domain features in tremor signals, extract tremor EEG time domain features and tremor EEG frequency domain features in EEG signals corresponding to tremor signals, extract tonic time domain features and tonic frequency domain features in tonic signals, extract tonic EEG time domain features and tonic EEG frequency domain features in EEG signals corresponding to tonic signals, extract posture time domain features and posture frequency domain features in posture signals, extract gait time domain features and gait frequency domain features in gait signals, extract EEG signals corresponding to posture signals Extract the postural EEG time domain features and postural EEG frequency domain features in the signal, extract the bradykinesia time domain features and bradykinesia frequency domain features in the bradykinesia signal, and extract the bradykinesia time domain features and bradykinesia frequency domain features in the EEG signal corresponding to the bradykinesia signal; for each information in the non-motor symptom information, extract the gastrointestinal tract score assignment features, urinary tract score assignment features, cardiovascular score assignment features, body temperature regulation score assignment features, pupil adjustment score assignment features, sexual function score assignment features, sleep disorder score assignment features, RBD score assignment features, anxiety level score assignment features and depression level score assignment features respectively; The feature vector construction module is used to construct a one-dimensional feature vector based on each feature=(gender assignment feature, age assignment feature, tremor time domain feature, tremor frequency domain feature, tremor EEG time domain feature, tremor EEG frequency domain feature, tonic time domain feature, tonic frequency domain feature, tonic EEG time domain feature, tonic EEG frequency domain feature, posture time domain feature, posture frequency domain feature, gait time domain feature, gait frequency domain feature, posture EEG time domain feature, posture EEG frequency domain feature, bradycy time domain feature, bradycy frequency domain feature, bradycy time domain feature, bradycy frequency domain feature, gastrointestinal tract score assignment feature, urinary tract score assignment feature, cardiovascular score assignment feature, body temperature regulation score assignment feature, pupil adjustment score assignment feature, sexual function score assignment feature, sleep disorder score assignment feature, RBD score assignment feature, anxiety level score assignment feature, depression level score assignment feature, salivary cortisol content feature, salivary melatonin content feature); The type recognition module is used to input the one-dimensional feature vector of the person to be identified into the target convolutional neural network type recognition model constructed based on the deep learning algorithm to perform type recognition of subtype Parkinson's disease, thereby identifying and outputting the specific subtype Parkinson's disease type of the person to be identified. The subtype Parkinson's disease types include tremor-dominant subtype Parkinson's disease, namely TD subtype, postural instability and gait difficulty subtype Parkinson's disease, namely PIGD subtype, and mixed subtype Parkinson's disease.

2. The subtype Parkinson's disease type identification system based on EEG signals and deep learning as claimed in claim 1, characterized in that: The feature extraction module is used for extracting time domain features including amplitude, standard deviation and similar entropy for tremor time domain features, rigidity time domain features, posture time domain features, gait time domain features and bradycy time domain features; for extracting frequency domain features including frequency, maximum spectrum peak and spectrum energy value for tremor frequency domain features, rigidity frequency domain features, posture frequency domain features, gait frequency domain features and bradycy frequency domain features; for extracting time domain features including amplitude, standard deviation and similar entropy for tremor EEG time domain features, rigidity frequency domain features, posture EEG time domain features and bradycy EEG time domain features; for extracting frequency domain features including power spectral density of δ wave, θ wave, α wave, β wave and γ wave frequency bands for tremor EEG frequency domain features, rigidity frequency domain features, posture EEG frequency domain features and bradycy EEG frequency domain features.

3. The subtype Parkinson's disease type identification system based on EEG signals and deep learning as claimed in claim 1, characterized in that: The feature extraction module is used for each information in the non-motor symptom information, the extracted gastrointestinal tract score assignment feature = gastrointestinal tract score value * gastrointestinal tract score weight, the extracted urinary tract score assignment feature = urinary tract score value * urinary tract score weight, the extracted cardiovascular score assignment feature = cardiovascular score value * cardiovascular score weight, the extracted temperature regulation score assignment feature = temperature regulation score value * temperature regulation score weight, the extracted pupil adjustment score assignment feature = pupil adjustment score value * pupil adjustment score weight, the extracted sexual function score assignment feature = sexual function score value * sexual function score weight, the extracted sleep disorder score assignment feature = sleep disorder score value * sleep disorder score weight, the extracted RBD score assignment feature = RBD score value * RBD score weight, the extracted anxiety level score assignment feature = anxiety level score value * anxiety level score weight, the extracted depression level score assignment feature = depression level score value * depression level score weight.

4. The subtype Parkinson's disease type identification system based on EEG signals and deep learning as claimed in claim 3, characterized in that: The weights of gastrointestinal tract score weight, temperature regulation score weight, RBD score weight and anxiety level score weight are larger, while the weights of urinary tract score weight, cardiovascular score weight, pupil accommodation score weight, sexual function score weight, sleep disorder score weight and depression level score weight are smaller.

5. The subtype Parkinson's disease type identification system based on EEG signals and deep learning as claimed in claim 1, characterized in that: The system also includes a model building module, which is used to build a target convolutional neural network type recognition model: Collect basic information, motor symptom information, non-motor symptom information and objective indicators of multiple past patients, and mark the type of subtype of Parkinson's disease of each past patient, including male and female patients of different ages with TD subtype, male and female patients of different ages with PIGD subtype, and male and female patients of different ages with mixed subtype; Among them, the basic information of each past patient includes gender and age; the motor symptom information of each past patient includes tremor data, stiffness data, postural instability and gait disorder data and bradykinesia data detected by multimodal sensors, the tremor data includes tremor signals and their corresponding electroencephalogram signals, the stiffness data includes stiffness signals and their corresponding electroencephalogram signals, the postural instability and gait disorder data includes posture signals, gait signals and their corresponding electroencephalogram signals, and the bradykinesia data includes bradykinesia signals and their corresponding electroencephalogram signals; The non-motor symptom information of each past patient includes gastrointestinal symptom information, urinary tract symptom information, cardiovascular symptom information, temperature regulation symptom information, pupil adjustment symptom information, sexual function symptom information, sleep disorder symptom information, rapid eye movement sleep behavior disorder symptom information (RBD symptom information), anxiety level symptom information and depression level symptom information; the objective indicators of each past patient include salivary cortisol content and salivary melatonin content; The basic information of each past patient is assigned a value preprocessing operation; for the motor symptom information of each past patient, the tremor signal and its corresponding EEG signal, the rigidity signal and its corresponding EEG signal, the posture signal and gait signal and its corresponding EEG signal, the bradykinesia signal and its corresponding EEG signal are subjected to a noise reduction filtering preprocessing operation; the various information in the non-motor symptom information of each past patient is subjected to a scoring preprocessing operation; Gender-assigned features and age-assigned features are extracted based on the basic information of each past patient; based on the motor symptom information of each past patient, tremor time domain features and tremor frequency domain features are extracted from the tremor signal, tremor EEG time domain features and tremor EEG frequency domain features are extracted from the EEG signal corresponding to the tremor signal, tonic time domain features and tonic frequency domain features are extracted from the tonic signal, tonic EEG time domain features and tonic EEG frequency domain features are extracted from the EEG signal corresponding to the tonic signal, posture time domain features and posture frequency domain features are extracted from the posture signal, gait time domain features and gait frequency domain features are extracted from the gait signal, and the EEG signal corresponding to the posture signal is extracted. Extract the bradykinesiology time domain features and bradykinesiology frequency domain features in the bradykinesiology signal, and extract the bradykinesiology time domain features and bradykinesiology frequency domain features in the EEG signal corresponding to the bradykinesiology signal; extract the gastrointestinal tract score assignment features, urinary tract score assignment features, cardiovascular score assignment features, body temperature regulation score assignment features, pupil adjustment score assignment features, sexual function score assignment features, sleep disorder score assignment features, RBD score assignment features, anxiety level score assignment features, and depression level score assignment features for each piece of non-motor symptom information of each past patient; Based on each feature of each past patient, a corresponding one-dimensional feature vector is constructed respectively = (gender assignment feature, age assignment feature, tremor time domain feature, tremor frequency domain feature, tremor EEG time domain feature, tremor EEG frequency domain feature, rigidity time domain feature, rigidity frequency domain feature, rigidity EEG time domain feature, rigidity EEG frequency domain feature, posture time domain feature, posture frequency domain feature, gait time domain feature, gait frequency domain feature, posture EEG time domain feature, posture EEG frequency domain feature, bradycardia time domain feature, bradycardia frequency domain feature, bradycardia time domain feature, bradycardia frequency domain feature, gastrointestinal tract score assignment feature, urinary tract score assignment feature, cardiovascular score assignment feature, body temperature regulation score assignment feature, pupil adjustment score assignment feature, sexual function score assignment feature, sleep disorder score assignment feature, RBD score assignment feature, anxiety level score assignment feature, depression level score assignment feature, salivary cortisol content feature, salivary melatonin content feature), so as to obtain a one-dimensional feature vector of each past patient; Constructing a sample set, the sample set includes a one-dimensional feature vector of each past patient and the type of the corresponding subtype of Parkinson's disease; A convolutional neural network type recognition model is constructed based on the deep learning algorithm. The particle swarm improvement algorithm is used to optimize the hyperparameters in the training process of the convolutional neural network type recognition model. The optimal hyperparameters are obtained through continuous updating and optimization, and the target convolutional neural network type recognition model is established using the optimized hyperparameters.

6. The subtype Parkinson's disease type identification system based on EEG signals and deep learning as claimed in claim 5, characterized in that: The convolutional neural network type recognition model consists of multiple convolutional layers, each convolutional layer is followed by a pooling layer, the last pooling layer is followed by a fully connected layer, and the fully connected layer is followed by a Softmax classifier, wherein the number of convolution kernels of each convolutional layer is set to be a multiple of 2, and the number of convolution kernels of the previous convolutional layer is greater than the number of convolution kernels of the next convolutional layer.

7. The subtype Parkinson's disease type identification system based on EEG signals and deep learning as claimed in claim 6, characterized in that: The model building module is used to: Construct a convolutional neural network type recognition model and initialize the model's hyperparameters, including the number of convolutional layers, the number of convolutional kernels in each convolutional layer, the number of neurons in the fully connected layer, and the learning rate. Set the number of particles in the particle swarm to N, initialize the speed and position of each particle in the particle swarm, set the maximum number of iterations T and the particle space dimension D that corresponds to the hyperparameters one by one, and the position of each particle represents a set of hyperparameters. Set the number of convolutional layers, the number of convolutional kernels in each convolutional layer, the number of neurons in the fully connected layer, and the range of learning rates. Substitute the hyperparameters corresponding to the spatial position of any particle into the convolutional neural network type recognition model and train it using the sample set. Calculate the fitness of the particle using the positive correlation fitness function established by the model training accuracy. Thus, the fitness corresponding to each particle can be obtained as the individual optimal fitness and the maximum value can be taken as the global optimal fitness. The current position of each particle is taken as the corresponding individual optimal position, and the current position of the particle corresponding to the global optimal fitness is taken as the global optimal position. Update the speed and position of each particle: Among them, W t represents the inertia weight of the t-th iteration, and represents the velocity and position of the tth iteration, and represents the speed and position of the t+1th iteration, C1 and C2 represent learning factors, rand1 and rand2 represent random numbers (0,1), represents the optimal position of the individual after the tth iteration, represents the global optimal position after the tth iteration, W min represents the minimum inertia weight, W max represents the maximum inertia weight, η represents the inertia adjustment factor; Update the individual optimal position and the global optimal position: use the fitness function to calculate the fitness of each particle after the tth iteration. Only when the updated fitness of the nth particle is greater than the individual optimal fitness of the nth particle, update the individual optimal fitness of the nth particle = the updated fitness of the nth particle, and update the individual optimal position of the nth particle = the updated position of the nth particle. n is 1-N in turn. Then, take the maximum value of each individual optimal fitness as the current optimal fitness. Only when the current optimal fitness is greater than the current global optimal fitness, update the global optimal fitness = the current optimal fitness, and update the global optimal position = the current position of the particle to which the current optimal fitness belongs. When the global optimal fitness reaches the preset value or the number of iterations reaches the maximum number of iterations T, the iteration ends. At this time, the optimal hyperparameters are obtained based on the global optimal position, and the target convolutional neural network type recognition model is established using the optimal hyperparameters. Otherwise, the speed and position of each particle continue to be updated.

8. The subtype Parkinson's disease type identification system based on EEG signals and deep learning as claimed in claim 7, characterized in that: When the global optimal fitness reaches the preset value or the number of iterations reaches the maximum number of iterations T, the iteration ends. At this time, the optimal hyperparameters are obtained based on the global optimal position, and the target convolutional neural network type recognition model is established using the optimal hyperparameters. Otherwise, it is determined whether t≤T / 2. If so, the current individual optimal fitness of each particle is arranged in descending order, the last n1 particles are discarded, and the particles with the global optimal fitness are used as the main elite particles. The remaining particles that are not discarded are randomly selected as the slave elite particles. The main elite particles are genetically hybridized with each selected slave elite particle to obtain n1 particles. The current position of each particle in the n1 particles obtained by hybridization is the position of the particle after the tth iteration. The current speed of each particle in the n1 particles obtained by hybridization is the speed of the particle after the tth iteration = the average speed of the speed of the main elite particle and the corresponding slave elite particle participating in the hybridization. In this way, a new particle population can be obtained, and the speed and position of each particle in the new particle population continue to be updated. n1 is a positive integer and 1≤n1≤L / 3, t=t+1, otherwise the speed and position of each particle are directly updated.

9. The subtype Parkinson's disease type identification system based on EEG signals and deep learning as claimed in claim 1, characterized in that: The tremor signal is obtained by detecting a tremor sensor, the electroencephalogram signal is obtained by detecting an electroencephalogram sensor, the rigidity signal is obtained by detecting a flexible tactile sensor, the posture signal is obtained by detecting a posture sensor, the gait signal is obtained by detecting a gait sensor, and the bradykinesia signal is obtained by detecting an inertial sensor.

Images