Deep learning-based EEG analysis for Parkinson’s disease

Abstract

Parkinson’s Disease (PD) is the fastest-growing neurological condition, with over 10 million cases worldwide. The most common way of diagnosing PD has been based on the observation of symptoms of the subject. Electroencephalography (EEG) is a non-invasive technique that allows clinicians to analyse brain patterns and has the potential to detect PD. Manually analysing EEG signals requires a skilled and trained expert to interpret the results accurately. For this reason, Deep Learning (DL) has the potential to bypass this step and detect PD automatically based on the EEG signals. Such models have been developed, with most reaching very promising results. However, some of these models do not practice a strict division of subjects on the subject level, resulting in potential cross-contamination of data between the testing and training set, not representing a real test scenario.

In our work, we aim to compare the performance of DL models in detecting PD using EEG signals represented as time-frequency series and spectrograms as inputs. Additionally, we aim to assess the difference in model performance between 2D and 3D DL architectures. Two implementations were tested for the 3D models: subject and segment level. For that purpose, we used the datasets from Turku, California, and New Mexico, which contained in total 50 PD and 60 non-PD subjects.

Our results show that incorporating the splitting on the subject level reduces the model’s performance compared to other approaches that do not perform that splitting, with the best Area Under the Receiver Operating Characteristic Curve (AUC) and subject level accuracy, respectively, being 0.69±0.07 and 60.00%±4.97% achieved by EEGNet run on 320x320 spectrograms. In addition, combining the 244x244 spectrograms on a subject level using the ResNet18 3D model showed a better result compared to the standard 2D ResNet18 model without ImageNet, with an AUC score of 0.68±0.09 vs. 0.64±0.09, respectively. These results indicate that EEG has the potential for detection and the performance benefits of representing EEG signals as spectrograms.

Parkinson’s Disease (PD) is the fastest-growing neurological condition, with over 10 million cases worldwide. The most common way of diagnosing PD has been based on the observation of symptoms of the subject. Electroencephalography (EEG) is a non-invasive technique that allows clinicians to analyse brain patterns and has the potential to detect PD. Manually analysing EEG signals requires a skilled and trained expert to interpret the results accurately. For this reason, Deep Learning (DL) has the potential to bypass this step and detect PD automatically based on the EEG signals. Such models have been developed, with most reaching very promising results. However, some of these models do not practice a strict division of subjects on the subject level, resulting in potential cross-contamination of data between the testing and training set, not representing a real test scenario.

In our work, we aim to compare the performance of DL models in detecting PD using EEG signals represented as time-frequency series and spectrograms as inputs. Additionally, we aim to assess the difference in model performance between 2D and 3D DL architectures. Two implementations were tested for the 3D models: subject and segment level. For that purpose, we used the datasets from Turku, California, and New Mexico, which contained in total 50 PD and 60 non-PD subjects.

Our results show that incorporating the splitting on the subject level reduces the model’s performance compared to other approaches that do not perform that splitting, with the best Area Under the Receiver Operating Characteristic Curve (AUC) and subject level accuracy, respectively, being 0.69±0.07 and 60.00%±4.97% achieved by EEGNet run on 320x320 spectrograms. In addition, combining the 244x244 spectrograms on a subject level using the ResNet18 3D model showed a better result compared to the standard 2D ResNet18 model without ImageNet, with an AUC score of 0.68±0.09 vs. 0.64±0.09, respectively. These results indicate that EEG has the potential for detection and the performance benefits of representing EEG signals as spectrograms.