CN112309577A - Multi-mode feature selection method for optimizing Parkinson voice data - Google Patents

Multi-mode feature selection method for optimizing Parkinson voice data Download PDF

Info

Publication number
CN112309577A
CN112309577A CN202011078465.0A CN202011078465A CN112309577A CN 112309577 A CN112309577 A CN 112309577A CN 202011078465 A CN202011078465 A CN 202011078465A CN 112309577 A CN112309577 A CN 112309577A
Authority
CN
China
Prior art keywords
individual
optimal
individuals
population
parkinson
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202011078465.0A
Other languages
Chinese (zh)
Other versions
CN112309577B (en
Inventor
胡晓敏
张首荣
李敏
陈伟能
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Guangdong University of Technology
Original Assignee
Guangdong University of Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Guangdong University of Technology filed Critical Guangdong University of Technology
Priority to CN202011078465.0A priority Critical patent/CN112309577B/en
Publication of CN112309577A publication Critical patent/CN112309577A/en
Application granted granted Critical
Publication of CN112309577B publication Critical patent/CN112309577B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/70—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for mining of medical data, e.g. analysing previous cases of other patients
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/60—Information retrieval; Database structures therefor; File system structures therefor of audio data
    • G06F16/61—Indexing; Data structures therefor; Storage structures
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/60—Information retrieval; Database structures therefor; File system structures therefor of audio data
    • G06F16/63—Querying
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/60—Information retrieval; Database structures therefor; File system structures therefor of audio data
    • G06F16/65—Clustering; Classification
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/004—Artificial life, i.e. computing arrangements simulating life
    • G06N3/006—Artificial life, i.e. computing arrangements simulating life based on simulated virtual individual or collective life forms, e.g. social simulations or particle swarm optimisation [PSO]
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/20—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for computer-aided diagnosis, e.g. based on medical expert systems

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Health & Medical Sciences (AREA)
  • Databases & Information Systems (AREA)
  • Physics & Mathematics (AREA)
  • Public Health (AREA)
  • General Engineering & Computer Science (AREA)
  • Biomedical Technology (AREA)
  • Medical Informatics (AREA)
  • General Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Multimedia (AREA)
  • Primary Health Care (AREA)
  • Epidemiology (AREA)
  • Pathology (AREA)
  • Software Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Evolutionary Computation (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Medical Treatment And Welfare Office Work (AREA)

Abstract

The invention discloses a multi-modal feature selection method for optimizing Parkinson voice data, which comprises the following steps: establishing a Parkinson voice data set, initializing a population based on a particle swarm algorithm, and determining a characteristic character string of an individual according to a real number coding scheme; dividing individuals in the population into niches according to the individual adaptation values; updating the historical optimal value and the historical optimal position of each individual, and updating the position and the adaptive value of the optimal individual in each niche; updating the position and the speed of each individual, and evaluating the adaptive value of each individual by combining the Parkinson voice data set according to the characteristic character string of each individual; taking the updated individuals as a new population, and comparing the new population with the initialized population to obtain a new generation of population; screening, reserving the optimal individuals of the two populations, and removing repeated individuals to obtain a new generation of population for evolution; and outputting all optimal individuals of each generation, wherein the characteristic combination of the optimal individuals is used for assisting in judging whether the Parkinson disease exists.

Description

Multi-mode feature selection method for optimizing Parkinson voice data
Technical Field
The invention relates to the technical field of medical technology and evolutionary computation, in particular to a multi-modal feature method for reducing the dimensionality of Parkinson voice data.
Background
At present, the cause of the Parkinson disease is not clear and can not be completely cured. The early detection of the disease in the sick period has great significance for improving the life experience of patients and treating the Parkinson disease. Researchers have proposed a variety of schemes to assist physicians in diagnosing parkinson's disease, including hand-drawn signal diagnosis and voice data prediction. In recent years, more and more researchers analyze voice data through a voice signal processing algorithm, a machine learning algorithm, a support vector machine and the like to judge the Parkinson's disease. The speech data collected by researchers also reveals UCI databases such as the parkinson speech data set of Max Little, oxford university and the parkinson multi-type speech data set of Olcay et al. However, the dimension of the voice data set collected in real life is often large, and the calculation cost and the time cost are greatly increased when a large number of example analysis training is faced.
The evolutionary algorithm has the characteristics of low calculation cost, high convergence speed and simple and understandable structure, and is widely applied to solving the problem of feature selection. However, the conventional evolutionary algorithm can only provide one solution if applied to feature selection, either multi-objective optimization or unimodal optimization.
Disclosure of Invention
The invention aims to provide a multi-modal feature selection method for optimizing Parkinson voice data, which is used for overcoming the defects that the time cost of the existing algorithm is high and only a single prediction scheme can be provided.
In order to realize the task, the invention adopts the following technical scheme:
a multi-modal feature selection method for optimizing Parkinson's speech data comprises the following steps:
extracting original Parkinson voice data, determining attributes and labels and establishing a Parkinson voice data set; wherein the attribute represents the collection standard of the voice data, and the label represents whether the person corresponding to the voice data is healthy or sick;
initializing a population based on a particle swarm algorithm, and initializing the position range of an individual according to the dimensionality of a Parkinson voice data set so as to determine a search space; determining a characteristic character string of an individual according to a real number coding scheme;
randomly dividing the whole population according to the individual adaptive value, and dividing the individuals in the population into niches;
updating the historical optimal value and the historical optimal position of each individual, and guiding the search direction of the individual; updating the position and the adaptive value of the optimal individual in each niche, taking the optimal individual of each niche as the global optimal individual of all the individuals of the niche, and further guiding the search direction of the individual;
updating the position and the speed of each individual, and evaluating the adaptive value of each individual by combining the Parkinson voice data set according to the characteristic character string of each individual; taking the updated individuals as a new population, and comparing the new population with the old population to obtain a new generation of population;
screening a new generation population and an old population, reserving optimal individuals of the two populations, and removing repeated individuals to obtain a new generation population for evolution;
outputting all optimal individuals of each generation; all the combinations of features of the optimal individuals will be applied for prediction of parkinson's disease.
Further, initializing the position range of the individual according to the dimensionality of the Parkinson voice data set so as to determine a search space; determining a characteristic string of an individual according to a real number encoding scheme, comprising:
setting the range of each dimension position of the individual in a [0,1] interval, and randomly initializing a real number for each dimension position in the interval by the individual; each dimension of an individual corresponds to each attribute of the data set, and each individual is a potential solution; converting the continuous real value of the individual position into a discrete 01 character string, and the coding scheme is as follows:
Figure BDA0002717885420000021
wherein
Figure BDA0002717885420000022
A binary value representing the d-th dimension of the ith individual,
Figure BDA0002717885420000023
representing the position of the ith individual in the d-th dimension; after conversion to a string, each individual represents a potential solution, each binary bit of an individual represents an attribute of the data set, 0 represents that the attribute is valid, and 1 represents that the attribute is invalid.
Further, the randomly dividing the whole population according to the individual adaptive value, and dividing the individuals in the population into niches includes:
firstly, setting the size N of each sub-population, then sequencing the individuals of the whole population according to adaptive values, selecting the optimal individual P in the current population, calculating the distance between all the individuals and the optimal individual P, and finding out N-1 individuals closest to the optimal individual P, wherein the N-1 individuals and the optimal individual P form a niche; and finally, removing the N individuals forming the niche from the population, and repeating the steps until all the individuals are classified into the corresponding niches.
Further, the updating the historical optimal value and the historical optimal position of each individual includes:
if the current adaptation value of the individual is better than the historical optimum value, then the position x of the current individual is usedi(t +1) and the adaptation value fit (x)i(t +1)) replacing the historical optimum adapted value and the historical optimum position;
if the adaptive value of the current individual is equal to the historical optimal adaptive value, selecting a random value to be compared with a threshold value of 0.5, and updating the historical optimal position by using the position of the current individual with half probability; and if the adaptive value of the current individual is smaller than the historical optimal adaptive value, keeping the historical optimal position and the optimal adaptive value.
Further, the updating of the position and velocity of each individual is performed according to the following formula:
Figure BDA0002717885420000031
Figure BDA0002717885420000032
wherein k denotes the kth niche, i denotes the ith individual, d denotes the historical optimum position of the d-th dimension,
Figure BDA0002717885420000033
the speed of the t-th generation individual is shown,
Figure BDA0002717885420000034
represents the historical optimal location of the individual,
Figure BDA0002717885420000035
indicating the optimal individual position of the kth niche,
Figure BDA0002717885420000036
denotes the position of the t-th generation individuals, w is the internal weight, c1 and c2 are the two acceleration factors, r1 and r2 are at [0,1]The intervals are uniformly distributed random values.
Further, the step of comparing the updated individuals as a new population with the initialized population to obtain a new generation population includes:
firstly, storing the optimal individuals of two populations, and if the number of the optimal individuals is more than or equal to the size of the population, storing the individuals with the number equivalent to the size of the population; if the number of the optimal individuals is smaller than the size of the population, all the optimal individuals are stored, and the superior individuals are selected from the two populations to make up for the number of the missing populations, and finally, the population which evolves for the new generation is obtained.
Further, the outputting all optimal individuals of each generation, the feature combination of all optimal individuals is used for assisting in judging whether the Parkinson's disease exists, and the method comprises the following steps:
selecting the optimal individuals from the new generation of population and storing the optimal individuals in an external set, and if the optimal individuals in the current generation are better than the optimal individuals stored in the external set in the old population, emptying the external set and then storing the optimal individuals in the current generation; if the current generation optimal individuals are equivalent to the optimal individuals stored in the external set before in the adaptive value, judging whether the current generation optimal individuals are repeated, adding the external set without repeating the current generation optimal individuals, and abandoning the repetition; if the current generation optimal individuals are worse than the optimal individuals stored in the external set before, directly abandoning the current generation optimal individuals; thus, the outer set is continuously updated with the evolution algebra; when the search is finished, all optimal individuals of each generation of external set are output, and the characteristic combination of the optimal individuals represents each scheme which can correctly judge the Parkinson's disease.
A multi-modal feature selection apparatus for optimizing parkinson's speech data, comprising:
the input module is used for extracting original Parkinson voice data, determining attributes and labels and establishing a Parkinson voice data set; wherein the attribute represents the collection standard of the voice data, and the label represents whether the person corresponding to the voice data is healthy or sick;
the initialization module is used for initializing a population based on a particle swarm algorithm and initializing the position range of an individual according to the dimensionality of the Parkinson voice data set so as to determine a search space; determining a characteristic character string of an individual according to a real number coding scheme;
the dividing module is used for randomly dividing the whole population according to the individual adaptive value and dividing the individuals in the population into niches;
the individual updating module is used for updating the historical optimal value and the historical optimal position of each individual and guiding the searching direction of the individual; updating the position and the adaptive value of the optimal individual in each niche, taking the optimal individual of each niche as the global optimal individual of all the individuals of the niche, and further guiding the search direction of the individual;
the population updating module is used for updating the position and the speed of each individual and evaluating the adaptive value of each individual by combining the Parkinson voice data set according to the characteristic character string of each individual; taking the updated individuals as a new population, and comparing the new population with the old population to obtain a new generation of population;
the screening module is used for screening the new generation population and the old population, reserving the optimal individuals of the two populations, and eliminating repeated individuals to obtain the new generation population for evolution;
the output module is used for outputting all the optimal individuals of each generation; all the combinations of features of the optimal individuals will be applied for prediction of parkinson's disease.
A computer comprising a memory, a processor and a computer program stored in said memory and executable on said processor, the processor implementing the steps of the method for multimodal feature selection for optimized parkinson's speech data when executing the computer program.
A computer-readable storage medium, having stored thereon a computer program which, when being executed by a processor, carries out the steps of the method for multimodal feature selection for optimized parkinson's speech data.
Compared with the prior art, the invention has the following technical characteristics:
1. the feature selection problem of optimizing the Parkinson speech data is treated as a multi-mode discrete optimization problem, so that the data dimensionality can be effectively reduced, the time and the cost can be reduced, and a plurality of optional schemes can be found out to predict the Parkinson disease.
2. And (3) dividing the sub-populations by using a species cluster-based niche technology. When the traditional particle swarm optimization is used for researching the feature selection problem, the optimization tends to be unimodal, namely an optimal solution is found. According to the invention, the particle swarm optimization is expanded from single-peak optimization to multi-mode optimization by using the species cluster-based niche technology, and an individual evolves in each niche, so that more optimal solutions can be found.
3. A screening rule for generating a next generation population is designed, and selection is made in the optimal individuals of the new and old populations, so that the optimal non-repetitive individuals can be continued to the next generation population. The screening rule not only can enable different optimal alternatives to be stored to guide the searching direction of the population, but also is beneficial to maintaining the diversity of the population and avoids trapping in a certain local optimal area.
4. And designing the screening and storing rules of the optimal alternative schemes. The optimal individuals generated in each generation are stored in an external set, so that the searching capability of the algorithm can be fully embodied, and the loss of the optimal individuals in the exploration process is avoided. And at the end of the search, all the individuals remaining in the outer set need only be analyzed to obtain all alternatives for predicting parkinson's disease.
Drawings
FIG. 1 is a schematic flow diagram of the process of the present invention;
FIG. 2 is a schematic diagram of a species-based cluster niche algorithm pseudocode;
FIG. 3 is a pseudo-code diagram of a screening process for generating a next generation population;
FIG. 4 is a schematic diagram of pseudocode for screening and saving optimal individuals using an outer set.
Detailed Description
Because the Parkinson's speech data who gathers in the real life have the characteristics that the data dimension is big, directly take training and can show time cost and computational cost increase, moreover because possess redundant and irrelevant characteristic between the attribute of speech data, do not directly use through the screening and can reduce even categorised correct rate, cause medical personnel to the erroneous judgement of Parkinson's disease.
In order to reduce the dimensionality of the voice data, reduce the detection cost and quickly judge the result, the parkinson voice data set needs to be subjected to feature selection preprocessing operation. However, the current evolutionary algorithms for studying the feature selection problem are biased towards unimodal optimization and only one solution is sought. The multi-modal particle swarm algorithm designed by the invention not only can effectively reduce the dimensionality of the Parkinson voice data, but also provides various reliable schemes for predicting the Parkinson diseases.
The idea of the invention is that firstly a group is initialized by combining a Parkinson voice data set, then a plurality of sub-groups are divided according to the niche rule, and each instance evolves in the sub-groups to find the optimal position. Each instance, after finding a new location, will be compared to the old population to determine the next generation population, and the best instance will be saved as a feature combination for judging parkinson's disease.
Referring to fig. 1, the multi-modal feature selection method for optimizing parkinson speech data of the present invention includes the following steps:
step 1, taking original Parkinson voice data, determining attributes and labels and establishing a Parkinson voice data set.
Extracting original Parkinson voice data, determining attributes and labels, establishing a data set, and storing the data in an external document in an LIBSVM format. The attributes represent the collection criteria of the voice data and the labels represent whether the person is healthy or diseased; an external document holding the parkinsonian speech data will be used to evaluate the performance of each individual in the population. The LIBSVM format is (dimension value: data value), each attribute of the data set corresponds to each dimension value, and the data under the attribute corresponds to the data value under the dimension. Each piece of data saved in LIBSVM format can clearly show how many dimensions each piece of data has and the specific data value of each dimension.
Since the external document stores the preprocessed parkinsonian speech data set, subsequent steps acquire the parkinsonian speech data by reading the external document.
And 2, initializing a population based on a particle swarm algorithm, setting the size of the population according to the user requirement, and setting the size of the population to be 30 in consideration of the calculation cost. And initializing the position range of the individual according to the dimensionality of the Parkinson voice data set so as to determine a search space.
Wherein, the range of each dimension position of the individual is set in the interval of [0,1], and a real number is randomly initialized for each dimension position in the interval by the individual; each dimension of an individual corresponds to each attribute of the data set, and each individual is a potential solution. Because the position of each individual in the traditional particle swarm algorithm is a continuous real value and is not suitable for processing the discrete characteristic selection problem, a real number coding scheme is adopted to convert the continuous real value of the position of the individual into a discrete 01 character string. The coding scheme is as follows:
Figure BDA0002717885420000061
wherein
Figure BDA0002717885420000062
A binary value representing the d-th dimension of the ith individual,
Figure BDA0002717885420000063
representing the position of the ith individual in the d-th dimension; after conversion to a string, each individual represents a potential solution, each binary bit of an individual represents an attribute of the data set, 0 represents that the attribute is valid, and 1 represents that the attribute is invalid.
And 3, randomly dividing the whole population according to the individual adaptive values, and dividing the individuals in the population into niches.
Dividing the whole population by adopting a species cluster-based niche rule, firstly setting the size N of each sub-population, then sequencing the individuals of the whole population according to adaptive values, selecting the optimal individual P in the current population, calculating the distance between all the individuals and the optimal individual P, and finding out N-1 individuals closest to the optimal individual P, wherein the N-1 individuals and the optimal individual P form a niche; and finally, removing the N individuals forming the niche from the population, and repeating the steps until all the individuals are classified into the corresponding niches.
It is noted here that the distance criterion requires the use of hamming distances, since the feature selection problem for parkinson's speech data is performed in discrete space. At this point, in order to compare with the new population, an extra space is needed to store all individuals of this population, denoted as oldSwarm, and fig. 2 shows the process of niche division.
In the step, a species cluster-based niche technology is adopted to divide the sub-populations, each sub-population represents a potential mountain peak, and operations conforming to multi-modal optimization are selected in the next series of steps, so that individuals evolve and search in respective niches. And several alternatives for predicting parkinson's disease at the end of the evolution are available.
Step 4, updating the historical optimal position pbest of each individuali(t) and historical optima fit (pbest)i(t)) for guiding a search direction of the individual; if the current adaptation value of the individual is better than the historical optimum value, then the position x of the current individual is usedi(t +1) and the adaptation value fit (x)i(t +1)) replacing the historical optimum adapted value and the historical optimum position;
if the adaptive value of the current individual is equal to the historical optimal adaptive value, selecting a random value to be compared with a threshold value of 0.5, and updating the historical optimal position by using the position of the current individual with half probability; if the adaptive value of the current individual is smaller than the historical optimal adaptive value, the historical optimal position and the historical optimal adaptive value are reserved; the specific formula is as follows:
Figure BDA0002717885420000071
where i represents the ith individual of the population and t represents the current generation.
Step 5, updating the position lbest of the optimal individual of each nichekAnd an adaptation value fit (lbestk); the optimal individual of each niche can be used as the global optimal individual of all the individuals of the niche, and the search direction of the individuals is further guided.
Conventional globally optimal individuals are only suitable for solving unimodal optimization problems. After the sub-populations are divided, the traditional method is not suitable for multi-mode optimization, and therefore the optimal individual of each newborn mirror is selected as a substitute in the scheme.
And 6, updating the position and the speed of each individual, wherein the speed and position updating formula is as follows:
Figure BDA0002717885420000081
Figure BDA0002717885420000082
wherein k denotes the kth niche, i denotes the ith individual, d denotes the historical optimum position of the d-th dimension,
Figure BDA0002717885420000083
the speed of the t-th generation individual is shown,
Figure BDA0002717885420000084
represents the historical optimal location of the individual,
Figure BDA0002717885420000085
indicating the optimal individual position of the kth niche,
Figure BDA0002717885420000086
representing the position of the t-th generation of individuals, w is an internal weight and controls the influence of the previous speed on the current speed; c1 and c2 are two acceleration coefficients, and control the optimal value position of the particle history and the global optimal position and the influence on the search process; r1 and r2 are in [0,1]]The intervals are uniformly distributed random values.
In this step, the new and old populations generated will be combined and screened for the next generation population. In order to reduce the collision probability among the optimal individuals and slow down the loss of the optimal individuals, the optimal individuals are screened from the new and old populations, and the optimal number N of non-repetitive individuals which does not exceed the size M of the population is saved at most. If the optimal number of individuals does not reach the population size, then randomly selecting (M-N) individuals from the suboptimal individuals to make up for the remaining number. Such rules ensure that the best individuals found are extended as far as possible to the next generation.
And 7, after the state of the individual is updated, evaluating the adaptive value of the individual by adopting a 1NN classifier and combining the Parkinson voice data according to the characteristic character string of each individual, wherein the adaptive value is the classification accuracy. Then, the updated individual is used as a new population, and the new population is obtained by comparing the updated individual with the population oldSwarm in the step 3, and the specific process is as follows:
the optimal individuals of the two populations are stored firstly, and if the number of the optimal individuals is larger than or equal to the size of the population, the individuals with the number equivalent to the size of the population are stored. If the number of optimal individuals is less than the population size, all optimal individuals are saved and the superior individual is selected from the two populations to compensate for the missing population number. Finally, the population evolving for the new generation is obtained. Figure 3 visually illustrates the population screening process.
And 8, outputting all the optimal individuals of each generation, wherein the characteristic combination of all the optimal individuals is used for assisting in judging whether the Parkinson disease exists or not.
Selecting the optimal individuals from the new generation of population and storing the optimal individuals in an external set archive, and if the current generation of optimal individuals are better than the optimal individuals stored in the external set archive in the old population, clearing the external set and then storing the current generation of optimal individuals; if the current generation optimal individuals are equivalent to the optimal individuals stored in the external set archive before in the adaptive value, judging whether the current generation optimal individuals are repeated, adding the external set without repeating the current generation optimal individuals, and abandoning the repeated current generation optimal individuals; if the current generation optimal individuals are worse than the optimal individuals stored in the external set, directly discarding the current generation optimal individuals; thus, the outer set is continuously updated with the evolution algebra; when the search is finished, all optimal individuals of each generation of external set are output, and the characteristic combination of the optimal individuals represents each scheme which can correctly judge the Parkinson's disease. Figure 4 illustrates the screening process of the outer set of best individuals.
This step employs an external set of archives to preserve the optimal non-repetitive individuals of each generation. In view of the fact that the optimal individuals are lost in the whole searching process, the algorithm searching capacity can be fully embodied by adopting an external set to store each generation of optimal individuals; after a new population for evolution of the next generation is generated, all optimal non-repetitive individuals of the population are extracted and compared and screened with an external set, so that the external set is ensured to store the optimal non-repetitive individuals. And finding out real and effective attribute columns capable of accurately predicting the Parkinson disease according to the 01 characteristic character string of each optimal individual, and judging whether a person has the Parkinson disease or not according to the combination of the attribute columns.
And 9, applying all the optimal individuals of each generation in the external set to predict the Parkinson's disease. While the population evolves continuously in the search space, the parkinson speech data set evaluates the fitness value of each individual through a 1NN classifier. In the invention, the adaptive value is the classification accuracy, and the adaptive value can judge the quality of the individual so as to guide the search of the population. The external set archive stores all the optimal non-repetitive individuals found in the population evolution process, namely all the individuals with the highest classification accuracy and without repetition. The optimal individuals are tested by the Parkinson voice data set, and the reliability is also achieved.
The feature combination of each optimal individual represents the selected attribute columns of the data set, which means that a medical staff or a researcher only needs to select or match these attribute columns and then judge whether a person has Parkinson's disease or not from the perspective of speech according to the attribute column combination. In practical application, medical staff or researchers need to collect voice data of a plurality of individuals, label the corresponding voice data with health or Parkinson's disease according to the health state of the individual, then train by adopting the method provided by the invention, find all optimal feature combinations, and finally judge whether the individual is suffered from the Parkinson's disease according to the feature combinations, namely label the individual with unknown label.
Based on the process, after the Parkinson voice data are obtained, the multi-mode particle swarm algorithm designed by the invention is used for continuously selecting the attribute which can accurately judge the Parkinson voice data, redundant attributes are removed, the calculation cost and the time cost are reduced, and the judgment accuracy is even further improved. Besides, under the condition of ensuring the accuracy, the invention can find a plurality of optimal characteristic solutions, and the 01 characteristic character strings represent a plurality of schemes which can judge the Parkinson's disease.
Example (b):
in order to embody the function and significance of the multi-modal feature method provided by the invention, the invention takes a Parkinson disease detection data set of Oxford university as an example, the data set comprises 195 pieces of voice data, and each piece of voice data has 22 dimensions. The results of the experiment are shown in the table:
Figure BDA0002717885420000101
in the table, the characteristic string 1 indicates that an attribute of the voice data is selected, and 0 indicates that the attribute is not selected. When feature selection is not performed, that is, judged according to all attribute columns, the classification accuracy is only 96.41%. But the classification accuracy of the feature selection method designed by the invention is improved to 99.49 percent, the required attributes are reduced, and at least three prediction schemes are found. The characteristic selection method not only effectively reduces data dimension and improves the classification accuracy of the Parkinson disease prediction, but also can provide a plurality of prediction schemes.
According to another aspect of the present application, there is provided a multimodal feature selection apparatus for optimizing parkinson's speech data, comprising:
the input module is used for extracting original Parkinson voice data, determining attributes and labels and establishing a Parkinson voice data set; wherein the attribute represents the collection standard of the voice data, and the label represents whether the person corresponding to the voice data is healthy or sick;
the initialization module is used for initializing a population based on a particle swarm algorithm and initializing the position range of an individual according to the dimensionality of the Parkinson voice data set so as to determine a search space; determining a characteristic character string of an individual according to a real number coding scheme;
the dividing module is used for randomly dividing the whole population according to the individual adaptive value and dividing the individuals in the population into niches;
the individual updating module is used for updating the historical optimal value and the historical optimal position of each individual and guiding the searching direction of the individual; updating the position and the adaptive value of the optimal individual in each niche, taking the optimal individual of each niche as the global optimal individual of all the individuals of the niche, and further guiding the search direction of the individual;
the population updating module is used for updating the position and the speed of each individual and evaluating the adaptive value of each individual by combining the Parkinson voice data set according to the characteristic character string of each individual; taking the updated individuals as a new population, and comparing the new population with the old population to obtain a new generation of population;
the screening module is used for screening the new generation population and the old population, reserving the optimal individuals of the two populations, and eliminating repeated individuals to obtain the new generation population for evolution;
the output module is used for outputting all the optimal individuals of each generation; all the combinations of features of the optimal individuals will be applied for prediction of parkinson's disease.
It should be noted that, for specific functions and related explanations of the above modules, refer to corresponding steps 1 to 9 in the foregoing method embodiment, which are not described herein again.
The embodiment of the application further provides a computer, and the calculator can be a terminal device, a controller, a server and the like; comprising a memory, a processor and a computer program stored in the memory and executable on the processor, the processor implementing the above-described method for multimodal feature selection for optimized parkinson's speech data when executing the computer program, e.g. steps 1 to 9 as described above.
The computer program may also be partitioned into one or more modules/units, which are stored in the memory and executed by the processor to accomplish the present application. One or more modules/units may be a series of computer program instruction segments capable of performing specific functions, where the instruction segments are used to describe an execution process of a computer program in a terminal device, for example, the computer program may be divided into an input module, an initialization module, a dividing module, an individual updating module, a population updating module, a screening module, and an output module, and functions of each module are described in the foregoing description of the apparatus and are not described in detail again.
Implementations of the present application provide a computer-readable storage medium having stored thereon a computer program that, when executed by a processor, performs the above-described steps of the multimodal feature selection method for optimizing parkinson's speech data, e.g., the above-described steps 1 to 9.
The above embodiments are only used to illustrate the technical solutions of the present application, and not to limit the same; although the present application has been described in detail with reference to the foregoing embodiments, it should be understood by those of ordinary skill in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some technical features may be equivalently replaced; such modifications and substitutions do not substantially depart from the spirit and scope of the embodiments of the present application and are intended to be included within the scope of the present application.

Claims (10)

1. A multi-modal feature selection method for optimizing Parkinson's speech data, comprising the steps of:
extracting original Parkinson voice data, determining attributes and labels and establishing a Parkinson voice data set; wherein the attribute represents the collection standard of the voice data, and the label represents whether the person corresponding to the voice data is healthy or sick;
initializing a population based on a particle swarm algorithm, and initializing the position range of an individual according to the dimensionality of a Parkinson voice data set so as to determine a search space; determining a characteristic character string of an individual according to a real number coding scheme;
randomly dividing the whole population according to the individual adaptive value, and dividing the individuals in the population into niches;
updating the historical optimal value and the historical optimal position of each individual, and guiding the search direction of the individual; updating the position and the adaptive value of the optimal individual in each niche, taking the optimal individual of each niche as the global optimal individual of all the individuals of the niche, and further guiding the search direction of the individual;
updating the position and the speed of each individual, and evaluating the adaptive value of each individual by combining the Parkinson voice data set according to the characteristic character string of each individual; taking the updated individuals as a new population, and comparing the new population with the old population to obtain a new generation of population;
screening a new generation population and an old population, reserving optimal individuals of the two populations, and removing repeated individuals to obtain a new generation population for evolution;
outputting all optimal individuals of each generation; all the combinations of features of the optimal individuals will be applied for prediction of parkinson's disease.
2. The method for multi-modal feature selection for optimizing parkinson's speech data of claim 1, wherein the initializing of the location ranges of the individuals according to the dimensionality of the parkinson's speech data set determines the search space; determining a characteristic string of an individual according to a real number encoding scheme, comprising:
setting the range of each dimension position of the individual in a [0,1] interval, and randomly initializing a real number for each dimension position in the interval by the individual; each dimension of an individual corresponds to each attribute of the data set, and each individual is a potential solution; converting the continuous real value of the individual position into a discrete 01 character string, and the coding scheme is as follows:
Figure FDA0002717885410000011
wherein
Figure FDA0002717885410000021
A binary value representing the d-th dimension of the ith individual,
Figure FDA0002717885410000022
representing the position of the ith individual in the d-th dimension; after conversion to a string, each individual represents a potential solution, each binary bit of an individual represents an attribute of the data set, 0 represents that the attribute is valid, and 1 represents that the attribute is invalid.
3. The method for multi-modal feature selection for optimizing parkinson's speech data of claim 1, wherein the randomly classifying the entire population according to the individual fitness value, and classifying the individuals in the population into niches comprises:
firstly, setting the size N of each sub-population, then sequencing the individuals of the whole population according to adaptive values, selecting the optimal individual P in the current population, calculating the distance between all the individuals and the optimal individual P, and finding out N-1 individuals closest to the optimal individual P, wherein the N-1 individuals and the optimal individual P form a niche; and finally, removing the N individuals forming the niche from the population, and repeating the steps until all the individuals are classified into the corresponding niches.
4. The method of multimodal feature selection for optimizing parkinson's speech data of claim 1, wherein the updating the historical optimal value and the historical optimal location for each individual comprises:
if the current adaptation value of the individual is better than the historical optimum value, then the position x of the current individual is usedi(t +1) and the adaptation value fit (x)i(t +1)) replacing the historical optimum adapted value and the historical optimum position;
if the adaptive value of the current individual is equal to the historical optimal adaptive value, selecting a random value to be compared with a threshold value of 0.5, and updating the historical optimal position by using the position of the current individual with half probability; and if the adaptive value of the current individual is smaller than the historical optimal adaptive value, keeping the historical optimal position and the optimal adaptive value.
5. The method of multi-modal feature selection for optimizing parkinson's speech data of claim 1, wherein the updating of the location and velocity of each individual is according to the following:
Figure FDA0002717885410000023
Figure FDA0002717885410000024
wherein k denotes the kth niche, i denotes the ith individual, d denotes the historical optimum position of the d-th dimension,
Figure FDA0002717885410000025
the speed of the t-th generation individual is shown,
Figure FDA0002717885410000026
represents the historical optimal location of the individual,
Figure FDA0002717885410000027
indicating the optimal individual position of the kth niche,
Figure FDA0002717885410000028
denotes the position of the t-th generation individuals, w is the internal weight, c1 and c2 are the two acceleration factors, r1 and r2 are at [0,1]The intervals are uniformly distributed random values.
6. The method of claim 1, wherein comparing the updated individuals as a new population with the initialized population to obtain a new generation population comprises:
firstly, storing the optimal individuals of two populations, and if the number of the optimal individuals is more than or equal to the size of the population, storing the individuals with the number equivalent to the size of the population; if the number of the optimal individuals is smaller than the size of the population, all the optimal individuals are stored, and the superior individuals are selected from the two populations to make up for the number of the missing populations, and finally, the population which evolves for the new generation is obtained.
7. The method of claim 1, wherein outputting all optimal individuals for each generation, the combination of features of all optimal individuals being used to assist in determining whether parkinson's disease is present comprises:
selecting the optimal individuals from the new generation of population and storing the optimal individuals in an external set, and if the optimal individuals in the current generation are better than the optimal individuals stored in the external set in the old population, emptying the external set and then storing the optimal individuals in the current generation; if the current generation optimal individuals are equivalent to the optimal individuals stored in the external set before in the adaptive value, judging whether the current generation optimal individuals are repeated, adding the external set without repeating the current generation optimal individuals, and abandoning the repetition; if the current generation optimal individuals are worse than the optimal individuals stored in the external set before, directly abandoning the current generation optimal individuals; thus, the outer set is continuously updated with the evolution algebra; when the search is finished, all optimal individuals of each generation of external set are output, and the characteristic combination of the optimal individuals represents each scheme which can correctly judge the Parkinson's disease.
8. A multimodal feature selection apparatus for optimizing parkinson's speech data, comprising:
the input module is used for extracting original Parkinson voice data, determining attributes and labels and establishing a Parkinson voice data set; wherein the attribute represents the collection standard of the voice data, and the label represents whether the person corresponding to the voice data is healthy or sick;
the initialization module is used for initializing a population based on a particle swarm algorithm and initializing the position range of an individual according to the dimensionality of the Parkinson voice data set so as to determine a search space; determining a characteristic character string of an individual according to a real number coding scheme;
the dividing module is used for randomly dividing the whole population according to the individual adaptive value and dividing the individuals in the population into niches;
the individual updating module is used for updating the historical optimal value and the historical optimal position of each individual and guiding the searching direction of the individual; updating the position and the adaptive value of the optimal individual in each niche, taking the optimal individual of each niche as the global optimal individual of all the individuals of the niche, and further guiding the search direction of the individual;
the population updating module is used for updating the position and the speed of each individual and evaluating the adaptive value of each individual by combining the Parkinson voice data set according to the characteristic character string of each individual; taking the updated individuals as a new population, and comparing the new population with the old population to obtain a new generation of population;
the screening module is used for screening the new generation population and the old population, reserving the optimal individuals of the two populations, and eliminating repeated individuals to obtain the new generation population for evolution;
the output module is used for outputting all the optimal individuals of each generation; all the combinations of features of the optimal individuals will be applied for prediction of parkinson's disease.
9. A computer comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor when executing the computer program performs the steps of the method for multimodal feature selection for optimized parkinson's speech data according to any of claims 1 to 7.
10. A computer-readable storage medium, in which a computer program is stored which, when being executed by a processor, carries out the steps of the method for multi-modal feature selection for optimizing parkinson's speech data according to any of claims 1 to 7.
CN202011078465.0A 2020-10-10 2020-10-10 Multi-mode feature selection method for optimizing parkinsonism voice data Active CN112309577B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202011078465.0A CN112309577B (en) 2020-10-10 2020-10-10 Multi-mode feature selection method for optimizing parkinsonism voice data

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202011078465.0A CN112309577B (en) 2020-10-10 2020-10-10 Multi-mode feature selection method for optimizing parkinsonism voice data

Publications (2)

Publication Number Publication Date
CN112309577A true CN112309577A (en) 2021-02-02
CN112309577B CN112309577B (en) 2023-10-13

Family

ID=74489493

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202011078465.0A Active CN112309577B (en) 2020-10-10 2020-10-10 Multi-mode feature selection method for optimizing parkinsonism voice data

Country Status (1)

Country Link
CN (1) CN112309577B (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113361563A (en) * 2021-04-22 2021-09-07 重庆大学 Parkinson's disease voice data classification system based on sample and feature double transformation

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102663100A (en) * 2012-04-13 2012-09-12 西安电子科技大学 Two-stage hybrid particle swarm optimization clustering method
CN108595499A (en) * 2018-03-18 2018-09-28 西安财经学院 A kind of population cluster High dimensional data analysis method of clone's optimization

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102663100A (en) * 2012-04-13 2012-09-12 西安电子科技大学 Two-stage hybrid particle swarm optimization clustering method
CN108595499A (en) * 2018-03-18 2018-09-28 西安财经学院 A kind of population cluster High dimensional data analysis method of clone's optimization

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113361563A (en) * 2021-04-22 2021-09-07 重庆大学 Parkinson's disease voice data classification system based on sample and feature double transformation

Also Published As

Publication number Publication date
CN112309577B (en) 2023-10-13

Similar Documents

Publication Publication Date Title
CN112712118B (en) Medical text data-oriented filtering method and system
CN108664589B (en) Text information extraction method, device, system and medium based on domain self-adaptation
CN112257449B (en) Named entity recognition method and device, computer equipment and storage medium
CN110097096B (en) Text classification method based on TF-IDF matrix and capsule network
CN113705396B (en) Motor fault diagnosis method, system and equipment
EP3912042A1 (en) A deep learning model for learning program embeddings
CN114742211B (en) Convolutional neural network deployment and optimization method facing microcontroller
CN113836896A (en) Patent text abstract generation method and device based on deep learning
US20230029947A1 (en) Medical disease feature selection method based on improved salp swarm algorithm
CN112309577B (en) Multi-mode feature selection method for optimizing parkinsonism voice data
CN116842460A (en) Cough-related disease identification method and system based on attention mechanism and residual neural network
CN113191133B (en) Audio text alignment method and system based on Doc2Vec
CN111816306B (en) Medical data processing method, and prediction model training method and device
CN111582287B (en) Image description method based on sufficient visual information and text information
CN112244863A (en) Signal identification method, signal identification device, electronic device and readable storage medium
CN113590867B (en) Cross-modal information retrieval method based on hierarchical measurement learning
CN111796173B (en) Partial discharge pattern recognition method, computer device, and storage medium
CN112465054B (en) FCN-based multivariate time series data classification method
CN116520150A (en) Anomaly detection system construction system and method based on dynamic strategy and active learning
CN112200224B (en) Medical image feature processing method and device
CN115116619A (en) Intelligent analysis method and system for stroke data distribution rule
CN115206539A (en) Multi-label integrated classification method based on perioperative patient risk event data
CN114118226A (en) ECG data classification method based on time convolution network model
CN113792879A (en) Case reasoning attribute weight adjusting method based on introspection learning
Vanhoucke et al. Interpretability in multidimensional classification

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant