SP 1. Parkinsonian rest tremor can be detected based on frequency domain features of the cerebral tremor network

• Published in: Clinical neurophysiology Vol. 127; no. 9; p. e168
• Main Authors: Hirschmann, J., Schoffelen, J.-M., Schnitzler, A., van Gerven, M.A.J.
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.09.2016
• Subjects: Neurology
• ISSN: 1388-2457; 1872-8952
• DOI: 10.1016/j.clinph.2016.05.187

Recent research has revealed several features of electrophysiological signals, which change upon manifestation of Parkinsonian rest tremor. These observations spur hope that the onset and the presence of tremor can be inferred by online analysis of brain signals, enabling demand-driven deep brain stimulation for tremor. It is an open question, however, whether it is possible to detect the presence of tremor given only a short data epoch. Furthermore, it is unclear which signals need to be considered for this task. In this study, we re-analyze data from 18 subthalamic nuclei (STN) of 12 tremor-dominant Parkinson patients in whom magnetoencephalography (MEG) and STN local field potentials (LFPs) were recorded simultaneously. Patients rested while tremor waxed and waned spontaneously. Epochs were attributed either to the tremor condition or to the tremor-free condition based on inspection of forearm electromyograms. MEG source activity was extracted from the cerebral tremor network proposed by Timmermann et al. (2003). Data were segmented into 4s epochs and power and connectivity features were computed for each epoch. These features were used as input to elastic net logistic regression to decode the tremor condition. Feature importance was evaluated through analysis of activation patterns and re-classification with single features or a subset of features. When testing and training was performed in different subjects, the decoder accuracy was 70%. Accuracy could be substantially improved by performing cross-validated training within subjects (81±12%). An analysis of feature importance revealed that power features were much more informative than connectivity features. Among others, STN and primary motor cortex beta power as well as thalamic gamma power were found to be particularly discriminative. Considering only STN power features yielded an accuracy of 77±12% in within-subject classification, which was increased to 86±11% when adding MEG power features. A similar accuracy could be obtained when using MEG power features only (85±12%). We conclude that Parkinsonian rest tremor can be distinguished from tremor-free periods with low error rate, provided that training data are available for each subject. Furthermore, connectivity features contribute little to tremor detection on a timescale of a few seconds. Nevertheless, it is of advantage to record cortical, cerebellar and thalamic sources in addition to the STN, as their power features provide additional information. Future research will integrate the current and related findings to design an algorithm for online stimulation control, incorporating estimates of the tremor state and patient-specific parameters set by the clinician.