Electrocorticography is superior to subthalamic local field potentials for movement decoding in Parkinson’s disease

  1. Timon Merk  Is a corresponding author
  2. Victoria Peterson
  3. Witold J Lipski
  4. Benjamin Blankertz
  5. Robert S Turner
  6. Ningfei Li
  7. Andreas Horn
  8. Robert Mark Richardson
  9. Wolf-Julian Neumann  Is a corresponding author
  1. Movement Disorder and Neuromodulation Unit, Department of Neurology, Charité - Universitätsmedizin Berlin, corporate member of Freie Universität Berlin and Humboldt Universität zu Berlin, Germany
  2. Brain Modulation Lab, Department of Neurosurgery, Massachusetts General Hospital, United States
  3. Harvard Medical School, United States
  4. Department of Neurobiology, University of Pittsburgh, United States
  5. Department of Computer Science, Technische Universität Berln, Germany
15 figures, 1 table and 2 additional files

Figures

Figure 1 with 1 supplement
Movement induced spectral changes are more dominant for ECoG than STN-LFP signals for a grip force task before and after a machine learning feature signal processing pipeline.

(A) ECoG, STN, and gripping force were recorded simultaneously during performance of a Go / No-Go task. (B) Individual ECoG and STN electrodes were localized and transformed into in Montreal …

Figure 1—figure supplement 1
Analyzed movements show variability in maximum amplitude and velocity.

(A) All used normalized and baseline corrected grip force traces. (B) Maximum peak amplitude histogram (C) All movement trace velocities.

Linear Models and Wiener Filters reveal temporally and spectrally specific coefficient distributions with grip-force decoding performance gain by including signals preceding the target sample by up to 500ms.

(A) Multivariable linear model coefficients trained only from the instantaneous sample (0 time lag with respect to decoded target sample) including all frequency bands from best channels per patient …

Figure 2—source data 1

Best channel Linear Model coefficients trained from instantaneous sample.

https://cdn.elifesciences.org/articles/75126/elife-75126-fig2-data1-v1.csv
Figure 2—source data 2

Univariate Linear Model coefficients of single frequency band and time lag.

https://cdn.elifesciences.org/articles/75126/elife-75126-fig2-data2-v1.zip
Figure 2—source data 3

Wiener Filter multiple time-step comparison.

https://cdn.elifesciences.org/articles/75126/elife-75126-fig2-data3-v1.csv
XGBOOST outperforms other machine learning methods for ECoG based grip-force decoding.

Based on the presented real-time compatible signal processing pipeline Neural Networks, Elastic - Net regularized Linear Models, Wiener Filters and extreme Gradient Boosting (XGBOOST) regression …

Figure 3—source data 1

Cross-validated ECoG and STN machine learning model performances for single channels.

https://cdn.elifesciences.org/articles/75126/elife-75126-fig3-data1-v1.csv
Figure 3—source data 2

Cross-validated combined and best-channel XGBOOST performances for best ECoG and STN channels.

https://cdn.elifesciences.org/articles/75126/elife-75126-fig3-data2-v1.csv
Figure 3—source data 3

Cross-validated XGBOOST performances for multichannel models based on ECoG, LFP and combined ECoG-LFP channels.

https://cdn.elifesciences.org/articles/75126/elife-75126-fig3-data3-v1.csv
Grand average grip-force decoding performances correlate inversely with preoperative PD motor sign severity.

UPDRS-III scores show significant negative correlations with patient-wise XGBOOST grip-force decoding performance averages for (A) ECoG (ρ=–0.55, p=0.039) and (B) STN-LFP signals (ρ=–0.55, p=0.042) …

Figure 4—source data 1

ECoG and STN single channel performances and UPDRS ratings.

https://cdn.elifesciences.org/articles/75126/elife-75126-fig4-data1-v1.csv
Figure 4—source data 2

ECoG and STN Force prediction UPDRS correlation.

https://cdn.elifesciences.org/articles/75126/elife-75126-fig4-data2-v1.csv
Subthalamic low-beta bursts relate to PD motor impairment and are associated with lower ECoG decoding performance.

UPDRS-III scores are significantly correlated with time spent in subthalamic low-beta bursts in the motor preparation period (A) and during movement (not shown). Average XGBOOST decoding performance …

Figure 5—source data 1

Time spend in low-beta burst performance and UPDRS correlation.

https://cdn.elifesciences.org/articles/75126/elife-75126-fig5-data1-v1.csv
Figure 5—source data 2

Movement onset aligned low-beta bursts for two subjects.

https://cdn.elifesciences.org/articles/75126/elife-75126-fig5-data2-v1.zip
Grip-force decoding performances spatially peak in sensorimotor cortex and the dorsolateral STN.

(A) Channels are color coded for individual XGBOOST grip-force regression performances per channel. Performance differences shown are in favor of ECoG over STN and contralateral over ipsilateral …

Figure 6—source data 1

Single channel XGBOOST coordinates and performances.

https://cdn.elifesciences.org/articles/75126/elife-75126-fig6-data1-v1.csv
Figure 7 with 1 supplement
Structural and functional movement decoding network analysis reveals cerebellar as well as sensorimotor cortical decoding capacity.

(A) Visualization of fibers originating from the ECoG recording locations of subject 1. (B) Decoding performance across all subjects and channels significant fiber tracts are displayed. All ECoG …

Figure 7—source data 1

Fiber Tracking network decoding performance prediction.

https://cdn.elifesciences.org/articles/75126/elife-75126-fig7-data1-v1.zip
Figure 7—source data 2

fMRI network decoding performance prediction.

https://cdn.elifesciences.org/articles/75126/elife-75126-fig7-data2-v1.zip
Figure 7—figure supplement 1
'Prediction Network Mapping’ allows for prediction of machine learning decoding performances using functional and structural connectivity.

(A) Functional connectivity ‘Fingerprints’ are estimated using fMRI resting state correlations of the Volume of Tissue Activated (VTA) voxels correlation to all other voxels. (B) The correlation of …

Author response image 1
Author response image 2
Author response image 3
Author response image 4
Author response image 5
Author response image 6
Author response image 7
Author response image 8

Tables

Table 1
Subject characteristics.
NGenderUPDRS totalHemisphereAgeMovementsDisease duration [years]ECoG Strip Contact Number LeftECoG Strip Contact Number Right
0Male28R60.312810.706
1Male27L+R51.2464142828
2Male33L+R53.82137.288
3Male31L+R44.228510.188
4Male322L+2 R63.638113.128+828+8
5Male52L59.6845.960
6Male55L71.61611.460
7Male50L52.51318.760
8Male62L+R66.85479.866
9Male48L67.98617.160
10Female31R6920510.406

Additional files

Supplementary file 1

Electrode details, Bayesian optmization hyperparemeters and best subject performances.

(a) Electrode Details (b) Bayesian Optimization Hyperparameters (c) Best channel R2 performances.

https://cdn.elifesciences.org/articles/75126/elife-75126-supp1-v1.docx
Transparent reporting form
https://cdn.elifesciences.org/articles/75126/elife-75126-transrepform1-v1.docx

Download links