F67. Distinguishing Parkinson’s Disease Dementia (PDD) patients from Parkinson’s Disease (PD) patients using EEG frequency and connectivity measures

Introduction The aims of this study are to investigate the usage of Phase Lag Index and frequency-band power measures as parameters for classification of PD and PDD patients, and dealing with the challenge of handling imbalanced data for classification. Methods EEG data for a group of 81 PD patients and 19 PDD patients were collected from three centres and analysed using automated segmentation and Inverse Solution post-processing. The PD group was a mix of MCI, Non MCI and unclassified early stage PD patients. 63 Frequency measures and 216 Phase Lag Index measures were obtained for all patients. To overcome the problem of imbalanced data, Random Forest algorithm was applied to the data and compared with Random Forest using cost-sensitive learning as well as Random Forest with stratified sampling. Classification models were built using frequency measures, PLI measures and frequency combined with PLI measures respectively. Results Applying cost-sensitive learning or stratified sampling to Random Forest increased the predictive performance of the model, in comparison to using Random Forest alone. In the case of stratified sampling, using 63 frequency measures for classification gave a ROC curve with average AUC value of 0.68. The AUC value increased to 0.75 when using PLI measures alone, which further increased to 0.8 when combining PLI and frequency measures. Further analysis revealed many more PLI measures than frequency measures to be amongst the top features distinguishing the two groups accurately. Conclusion Phase Lag Index measures may contain more information than EEG-band power measures and can be useful in distinguishing PD patients from PDD. Furthermore, band-power and PLI measures contain non-redundant information.