Ensemble classification algorithms on acoustic features to predict Parkinson’s disease

Background and aims: Parkinson’s disease is a progressive neurodegenerative disorder that affectsa person’s movement. Dopamine is responsible for controlling movement in the humanbody. Parkinson’s disease destroys brain cells that produce dopamine, causing the brain’s level ofdopamine to decrease. Parkinson’s disease causes neurological and movement disorders. One ofthe factors to predict this disease is measuring the person’s vocal features. These features can helpto determine whether a person has Parkinson’s disease or is healthy. This article aims to developan intelligent algorithm using artificial intelligence, to analyze the voice features of people whohave Parkinson’s disease. According to the recorded features of the patient’s voice, this algorithmchecks whether the person has Parkinson’s disease.Method: Classification algorithms classify samples based on their input features. These algorithmslook for similar patterns between samples and put those that have the most similarity intoa category (training phase). Then, when a new sample is given to the algorithm, the algorithmchecks the sample’s features, and based on the classification done in the previous stage, it isplaced in the collection that is closest to it (test stage). Simple classification algorithms examineall samples once and then categorize them, while ensemble classification algorithms perform theclassification in several stages and based on voting, which increases the accuracy of the final algorithm.In this study, we used three simple classification algorithms (SVM, Decision Tree, andK-Nearest Neighbors) and five ensemble classification algorithms (Ensemble AdaBoost, EnsembleBagging, Ensemble Gradient Boosting, Ensemble Extra Trees Classifier, and Ensemble RandomForest Classifier) to develop a high-accuracy detection algorithm for detecting Parkinson’spatients among the dataset samples. The dataset used in this study was from the University of California,Irvine (UCI), which included ۲۴۰ samples of Parkinson’s patients and healthy subjects.At first, we used simple classification algorithms to build the detection model, and after checkingtheir accuracy, we used ensemble algorithms to create the final algorithm.Results: Firstly, the detection algorithm was implemented using three simple classification algorithms,which included SVM, Decision Tree, and K-Nearest Neighbors. Algorithms achievedfinal accuracy of ۸۷, ۹۰, and ۸۵, respectively. The implementation of ensemble classificationalgorithms showed that these algorithms achieved higher accuracy in making the final algorithmthan simple classification algorithms (all ensemble algorithms had a final accuracy rate of morethan ۹۰%). Secondly, experiments showed that the size of the dataset (number of records) is relatedto the number of features examined in the dataset. The accuracy rate in large datasets canimprove by limiting the features (using highly correlated features) in most algorithms. EnsembleAdaBoost, Ensemble Bagging and Ensemble Gradient Boosting, and Extra Trees Classifier algorithmsachieved ۱۰۰% accuracy in both large datasets and small datasets with limited features.The Random Forest Classifier algorithm obtained the worst accuracy results in all the examinedmodes. Ensemble classification algorithms obtain better results than simple classification algorithmsdue to the construction of several models of data.Conclusion: In this article, we tried to design an artificial intelligence algorithm with high accuracyto diagnose Parkinson’s disease. This algorithm can determine whether a person has Parkinson’sor not by receiving the voice features of each person. The dataset used in the training andtesting stages of the algorithm included the voice features of people with Parkinson’s disease andhealthy people. Classification algorithms are popular in the design of artificial intelligence algo rithms. We examined several types of simple and ensemble classification algorithms, and resultsshowed that the ensemble algorithms ultimately have higher accuracy than the simple classificationalgorithms. Also, the number of dataset samples and the features of people’s voices can havea high impact on improving the performance of the final algorithm.

کلیدواژه ها:

Artificial Intelligence ، Parkinson ، Ensemble classification ، Acoustic data ، Machine Learning