A Deep Learning Approach to Determine Age-related EEG Features in Parkinson's Disease

• Published in: 2021 28th National and 6th International Iranian Conference on Biomedical Engineering (ICBME) pp. 322 - 328
• Main Authors: Mirian, AmirAli, Shirshekar, Hossna, Mirian, Maryam S., Hussain, Ramy, Lee, Soojin, McKeown, Martin J.
• Format: Conference Proceeding
• Language: English
• Published: IEEE 25.11.2021
• Subjects: Age prediction; Biomarkers; Brain modeling; Data models; EEG; EEG sub-bands; Electroencephalography; LSTM; Parkinson's disease; Predictive models; RNN; Training
• DOI: 10.1109/ICBME54433.2021.9750334

Oscillatory biomarkers are useful for development of Brain-computer interface (BCI) and EEG-based neuro-feedback systems, which may have therapeutic implications for seniors and those with the disease. Although many biomarkers for age and disease exist, the EEG has the benefit that it is widely available, inexpensive, and potentially may act as a biomarker over rapid time scales, which might be beneficial during, e.g., the performance of a specific task such as neurofeedback games. Parkinson's disease (PD) is ideally suited for the exploration of oscillatory biomarkers since abnormal oscillations have been widely implicated in the pathophysiology of PD. Specifically, beta-band oscillations may be broadly considered "anti-kinetic" and seen as inhibiting movement, while gamma-band oscillations are considered "pro-kinetic" and appear to facilitate movement. However, many domain-based EEG features in people with PD overlap considerably with those just seen in normal aging. Here, we contrast the age-related EEG features in PD subjects and age-matched healthy controls (HC). We employed an end-to-end training strategy and built deep recurrent neural network models with Long Short-Term Memory (LSTM) cells to predict age from 60-s of rest EEG recorded from PD and HC. When reliable models with reasonable errors were found for both groups ( MAE=1.897 for HC and MAE=2.172 for PD), we investigated their deterioration of predictive power when fed frequency band-limited data. In PD subjects, beta and gamma bands in channels T7, FP2, and F7 were significantly more important for predicting age in PD than in HC. After medication, differences in the frequency bands predicting age between PD and controls become more prominent when PD subjects were on medication. Our results suggest that after the development of PD, beta and gamma become more strongly associated with age, implying that future studies examining beta and gamma changes in PD will need to take particular care in controlling for the age of subjects.