The prediction of Parkinson's disease based on Pearson coefficient feature screening and machine learning

Abstract

Parkinson's disease is a common and severe type of brain disease. Its incidence rate is relatively high among brain diseases. At present, there is no very effective treatment for Parkinson's disease. So researchers have focused on diagnosing Parkinson's disease. At present, machine learning methods have been applied in the medical field and have played a very positive role in the diagnosis of diseases. It has been proven that analyzing the patient's voice and trembling condition through machine learning can accurately diagnose Parkinson's disease. In this study, before training the model, we used the Pearson coefficient feature screening method to improve the accuracy of diagnosis. Then, we conducted training on six major models (Random Forest, GBDT, Adaboost, Logistic Regression, Decision Tree, XGboost) in order to find the model with the best performance. In this study, we found that the performance of Random Forest is the best in these models (Accuracy: 91.53%, recall: 100%), then is the GBDT model (Accuracy: 91.53%, recall: 97.78%). The other four models all have a great disparity on accuracy and recall, which are the two most important metrics on the detection of diseases. The research results have demonstrated that the feature selection method based on Pearson's coefficient indeed comprehensively improves the accuracy of diagnosis for Parkinson's disease. And we also found that in the process of diagnosing Parkinson's disease, the performance of the Random Forest and GBDT models is the best.