Classification of pallidal oscillations with increasing parkinsonian severity

Abstract

The firing patterns of neurons in the basal ganglia are known to become more oscillatory and synchronized from healthy to parkinsonian conditions. Similar changes have been observed with local field potentials (LFPs). In this study, we used an unbiased machine learning approach to investigate the utility of pallidal LFPs for discriminating the stages of a progressive parkinsonian model. A feature selection algorithm was used to identify subsets of LFP features that provided the most discriminatory information for severity of parkinsonian motor signs. Prediction errors <20% were achievable using 28 of the possible 206 features tested. For all subjects, a spectral feature within the beta band was chosen through the feature selection algorithm, but a combination of features, including alpha-band power and phase-amplitude coupling, was necessary to achieve minimal prediction errors. There was large variability between the discriminatory features for individual subjects, and testing of classifiers between subjects yielded prediction errors >50%. These results suggest that pallidal oscillations can be predictive biomarkers of parkinsonian severity, but the features are more complex than spectral power in individual frequency bands, such as the beta band. Additionally, the best feature set was subject specific, which highlights the pathophysiological heterogeneity of parkinsonism and the importance of subject specificity when designing closed-loop system controllers dependent on such features.

Keywords: Parkinson's disease; machine learning; phase-amplitude coupling; support vector machine.

Fig. 1.

A: power spectral densities (PSDs) for each subject across the naive, mild, moderate, and severe parkinsonian states. Solid lines show population means, and shaded areas represent ±1 SD. Arrows indicate median frequency of oscillations in the beta band for the naive, mild, moderate, and severe states. B: boxplots of average power in the low (11–22 Hz), high (22–32 Hz), and overall (11–32 Hz) beta bands across observations for a given subject and parkinsonian state. Black dots represent values for individual observations and are displayed with a jitter on the x-axis for improved visualization.

Fig. 2.

Phase-amplitude coupling (PAC) between 12 phase bands and 6 frequency bands. A: z scores were averaged across observations for individual subjects separated by parkinsonian state. z Scores above 3.197 were considered significant at α = 0.05 with Bonferroni correction for the 6 × 12 frequency band pairs. B: scatterplots show the PAC z scores for individual recordings between all phase bands and averaged between 2 high-frequency oscillation (HFO) bands (181–256 Hz and 256–362 Hz). Two phase-amplitude pairs of interest are highlighted, beta phase (11–16 Hz) to HFO amplitude (+) and gamma phase (48–64 Hz) to HFO amplitude (*). Horizontal lines represent a confidence threshold of α < 0.05 with Bonferroni correction.

Fig. 3.

Top: modified Unified Parkinson's Disease Rating Scale (mUPDRS) for subjects P1 (A), P2 (B), and P3 (C) associated with each observation. Bottom: each observation was given an in-versus-out score for the naive (red dot), mild (cyan dot), moderate (blue dot), and severe (black dot) support vector machine (SVM) classifiers. The predicted class was assigned based on the highest score. The lines show smoothed trends for classifier scores. Dashed black lines show the change from one parkinsonian state to another.

Fig. 4.

Lasso regularization was used to select the smallest subset of features yielding a mean-square error (MSE) within 1 SE of the minimum possible MSE. A: regularization coefficients for the PSD feature set for individual subjects (P1, blue; P2, gray; P3, red) and pooled data (All, black). B: red squares show the PAC features with nonzero regularization coefficients for individual subjects and pooled data.

Fig. 5.

Confusion matrices for radial basis function (RBF)-SVMs using the 28 lasso-selected features based on pooled data. A: classifiers were trained and tested on data from individual subjects (P1, P2, P3) or on the pooled data (All). B: the confusion matrix results from classification of pooled data were split up by individual subject. x- and y-axis labels indicate naive (Na), mild (Mi), moderate (Mo), and severe (Se) states.