Robust EEG analysis of Parkinson’s disease: A machine learning approach

Abstract

Parkinson's disease (PD) is one of the most common neurodegenerative disorders worldwide. An electroencephalogram (EEG) has the potential to detect PD. PD is diagnosed by clinically examining the patient's symptoms. As the need for automatic and more reliable procedures in detecting PD increases, different Machine Learning (ML) models have been developed for this purpose. These models have been trained indistinctively with subjects ON and OFF medication, in spite of the known effect that medications such as levodopa can have on the EEG signal.

In this thesis, we aim to explore whether there is a difference when using subjects' ON or OFF medication when training and testing the model in order to assess whether the subjects' medication status affects the model's performance. We also aim to try to improve an already trained ML model by applying a conformal prediction algorithm, making the model more "confident" in the answer it gives. We also evaluated the effect of conformal predictions for different sub-populations based on genders, severity stage and different centres.

The results of the tests, which are used to determine whether there is a difference between using subjects' ON or OFF medication when training and testing the model, show that training on ON medication recordings increases the model's performance compared to training the model on subjects OFF medication. In addition, we also observed that severe subjects ON medication have a big influence in the learning of the algorithm. Another test we used to analyse the effect of medication was to try and predict whether a subject was either ON or OFF medication. The results of the tests indicate that the EEG signals do not contain enough information about the medication status to distinguish between ON/OFF medication. By applying a conformal prediction we were able to improve the performance of our overall model, with an increase in balanced accuracy from 75.92\% $\pm$ 6.024 to 79.29\% $\pm$ 9.601. We also saw different increases within each specified sub-population, concluding that the use of a conformal prediction can enhance a model's performance.

Parkinson's disease (PD) is one of the most common neurodegenerative disorders worldwide. An electroencephalogram (EEG) has the potential to detect PD. PD is diagnosed by clinically examining the patient's symptoms. As the need for automatic and more reliable procedures in detecting PD increases, different Machine Learning (ML) models have been developed for this purpose. These models have been trained indistinctively with subjects ON and OFF medication, in spite of the known effect that medications such as levodopa can have on the EEG signal.

In this thesis, we aim to explore whether there is a difference when using subjects' ON or OFF medication when training and testing the model in order to assess whether the subjects' medication status affects the model's performance. We also aim to try to improve an already trained ML model by applying a conformal prediction algorithm, making the model more "confident" in the answer it gives. We also evaluated the effect of conformal predictions for different sub-populations based on genders, severity stage and different centres.

The results of the tests, which are used to determine whether there is a difference between using subjects' ON or OFF medication when training and testing the model, show that training on ON medication recordings increases the model's performance compared to training the model on subjects OFF medication. In addition, we also observed that severe subjects ON medication have a big influence in the learning of the algorithm. Another test we used to analyse the effect of medication was to try and predict whether a subject was either ON or OFF medication. The results of the tests indicate that the EEG signals do not contain enough information about the medication status to distinguish between ON/OFF medication. By applying a conformal prediction we were able to improve the performance of our overall model, with an increase in balanced accuracy from 75.92\% $\pm$ 6.024 to 79.29\% $\pm$ 9.601. We also saw different increases within each specified sub-population, concluding that the use of a conformal prediction can enhance a model's performance.