Linear predictive coding electroencephalography algorithms predict mortality in Parkinson's disease.

Background: Mortality is increased in Parkinson's disease (PD) and is difficult to predict because of its heterogeneity and the availability of few reliable prognostic markers. Objectives: We used electroencephalography (EEG) recordings and the Linear Predictive Coding EEG Algorithm for PD (LEAPD) to classify 3-year mortality status and correlate LEAPD indices with time to death. Methods: 2-minutes resting-state EEG from 94 PD patients was used for binary classification of 3-year mortality status (22 deceased). Single-channel classification using a balanced dataset of 44 was performed using leave-one-out cross-validation (LOOCV). Robustness was evaluated by truncating the recordings. LOOCV Spearman's correlation coefficient (ρ) was obtained between LEAPD indices and time to death. Optimum hyperparameters obtained using a balanced training dataset of 30 were tested on the remaining 64 by 10,000 randomized comparisons of 7 vs 7, using 5 channel combinations. Separate hyperparameters for the best ρ, obtained using the same training dataset, were used for out-of-sample correlation for the remaining 7 deceased. Results: The deceased participants were older and had more severe disease and worse cognition at baseline. Several EEG channels yielded 100 % LOOCV accuracy. Five had robust performance under data truncation. The correlations ranged between ρ = -0.59 to -0.86 and were significant after adjusting for age, cognitive and motor impairment. Out-of-sample testing using the best-performing 5-channel combination yielded a mean accuracy of 83 %. Out-of-sample Spearman's ρ was -0.82. Conclusions: Short resting EEG using machine learning algorithms such as linear predictive coding can predict mortality in PD.