Classification of Deep Brain Stimulation states in Parkinson’s disease using functional MRI
Poster No:
1411
Submission Type:
Abstract Submission
Authors:
Dimitra Kiakou1,2, Karsten Mueller2,1, Pavel Filip1,3, Andrej Lasica1, Filip Růžička1, Dušan Urgošík4, Kristína Burdová1, Robert Jech1,4
Institutions:
1Department of Neurology, Charles University, 1st Faculty of Medicine and General University Hospital, Prague, Czech Republic, 2Max Planck Institute for Human Cognitive and Brain Sciences, Leipzig, Sachsen, Germany, 3Center for Magnetic Resonance Research (CMRR), University of Minnesota, Minneapolis, MN, United States, 4Na Homolce Hospital, Prague, Czech Republic
First Author:
Dimitra Kiakou
Department of Neurology, Charles University, 1st Faculty of Medicine and General University Hospital|Max Planck Institute for Human Cognitive and Brain Sciences
Prague, Czech Republic|Leipzig, Sachsen, Germany
Department of Neurology, Charles University, 1st Faculty of Medicine and General University Hospital|Max Planck Institute for Human Cognitive and Brain Sciences
Prague, Czech Republic|Leipzig, Sachsen, Germany
Co-Author(s):
Karsten Mueller
Max Planck Institute for Human Cognitive and Brain Sciences|Department of Neurology, Charles University, 1st Faculty of Medicine and General University Hospital
Leipzig, Sachsen, Germany|Prague, Czech Republic
Max Planck Institute for Human Cognitive and Brain Sciences|Department of Neurology, Charles University, 1st Faculty of Medicine and General University Hospital
Leipzig, Sachsen, Germany|Prague, Czech Republic
Pavel Filip
Department of Neurology, Charles University, 1st Faculty of Medicine and General University Hospital|Center for Magnetic Resonance Research (CMRR), University of Minnesota
Prague, Czech Republic|Minneapolis, MN, United States
Department of Neurology, Charles University, 1st Faculty of Medicine and General University Hospital|Center for Magnetic Resonance Research (CMRR), University of Minnesota
Prague, Czech Republic|Minneapolis, MN, United States
Andrej Lasica
Department of Neurology, Charles University, 1st Faculty of Medicine and General University Hospital
Prague, Czech Republic
Department of Neurology, Charles University, 1st Faculty of Medicine and General University Hospital
Prague, Czech Republic
Filip Růžička
Department of Neurology, Charles University, 1st Faculty of Medicine and General University Hospital
Prague, Czech Republic
Department of Neurology, Charles University, 1st Faculty of Medicine and General University Hospital
Prague, Czech Republic
Kristína Burdová
Department of Neurology, Charles University, 1st Faculty of Medicine and General University Hospital
Prague, Czech Republic
Department of Neurology, Charles University, 1st Faculty of Medicine and General University Hospital
Prague, Czech Republic
Robert Jech
Department of Neurology, Charles University, 1st Faculty of Medicine and General University Hospital|Na Homolce Hospital
Prague, Czech Republic|Prague, Czech Republic
Department of Neurology, Charles University, 1st Faculty of Medicine and General University Hospital|Na Homolce Hospital
Prague, Czech Republic|Prague, Czech Republic
Introduction:
Deep brain stimulation (DBS) in the sub-thalamic nucleus (STN-DBS) is an effective treatment for Parkinson's disease (PD), however, the precise mechanisms are not yet well established. Recent studies have investigated the effects of DBS in brain networks (Lamoš et al., 2023; Arévalo Sáenz et al., 2022; Accolla et al., 2016), but limited research involved machine learning with post-operative magnetic resonance imaging (MRI) (Peralta et al., 2021). Therefore, the characterization of STN-DBS connectivity alterations using functional MRI may reveal therapeutic markers and improve our understanding of DBS in PD. Thus, this study aimed to differentiate the STN-DBS ON and OFF states using machine learning and identify the brain patterns contributed to the classification in PD patients.
Methods:
A total of 104 patients with advanced type of PD (34 females, age 59.5±7.9 years, disease duration 15.1±6.3 years) were recruited. Resting state functional MRI (rs-fMRI) was acquired in two sessions in the STN-DBS ON and OFF state. Data pre-processing included realignment, unwarping, slice-time correction, normalization, and spatial filtering with Gaussian kernel of 10 mm. White matter, cerebrospinal fluid, and the translational and rotational parameters from head movements were regressed out and a high-pass filter with a cutoff frequency of 0.01 Hz was applied. Thereafter, global correlation (GCOR) and eigenvector centrality (EC) maps were estimated using rs-fMRI for each participant and each session resulting 208 network maps respectively. All above analyses were conducted using CONN toolbox, SPM12, and LIPSIA software package.
To determine characteristic features of the STN-DBS ON and OFF states, we implemented Leave One Group Out Cross Validation (LOGOCV) classification (each group corresponds to one subject), to predict the ON and OFF connectivity of an individual subject. Support Vector Machines (SVM) classification with Radial Basis Function kernel was performed on the vectorized, masked network maps in Python (v. 3.10). In each iteration, the classifier was trained in 206 samples and tested in two images acquired from the same subject. Furthermore, we applied Random Forest (RF) and Linear Discriminant Analysis (LDA) classification as alternative approaches for the categorization of the two states. Since the input data were high dimensional, we utilized the Python feature selection approach, SelectKBest with the score function f_classif, to extract the best K features based on univariate statistical tests. We explored the values of K in the range of 1000 till maximum number of voxels in intervals of 2000. Finally, the models performance was evaluated on the average classification accuracies across iterations.
To determine characteristic features of the STN-DBS ON and OFF states, we implemented Leave One Group Out Cross Validation (LOGOCV) classification (each group corresponds to one subject), to predict the ON and OFF connectivity of an individual subject. Support Vector Machines (SVM) classification with Radial Basis Function kernel was performed on the vectorized, masked network maps in Python (v. 3.10). In each iteration, the classifier was trained in 206 samples and tested in two images acquired from the same subject. Furthermore, we applied Random Forest (RF) and Linear Discriminant Analysis (LDA) classification as alternative approaches for the categorization of the two states. Since the input data were high dimensional, we utilized the Python feature selection approach, SelectKBest with the score function f_classif, to extract the best K features based on univariate statistical tests. We explored the values of K in the range of 1000 till maximum number of voxels in intervals of 2000. Finally, the models performance was evaluated on the average classification accuracies across iterations.
Results:
SVM outperformed RF and LDA reaching a high accuracy of 82.69% with GCOR robustly across several K values (Figure 1). LDA achieved slightly lower, while RF had the poorest performance with prominent fluctuation (Figure 1). Classification based on EC maps presented analogous results with maximum accuracy 80.29% using SVM. All algorithms were improved with dimensionality reduction independently of the input maps. To identify and visualize the contributing features to the SVM classification the minimum K voxels required were retrieved and reshaped back to the three-dimensional brain for both GCOR (K = 33000) and EC (K = 27000) (Figure 2). The mask is estimated as the intersection of features (voxels) selected at each iteration of LOGOCV SVM classification with SelectKBest (Figure 2).
Conclusions:
SVM classification accurately differentiated the STN-DBS ON and OFF states and revealed specific brain patterns with predominant area the temporal lobe. GCOR and EC maps exhibited equivalent classification accuracy with significantly overlapping selected regions. The selected attributes that detected brain connectivity modifications in STN-DBS may present an essential signature for refining DBS in PD.
Brain Stimulation:
Deep Brain Stimulation 2
Disorders of the Nervous System:
Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s)
Modeling and Analysis Methods:
Classification and Predictive Modeling 1
Connectivity (eg. functional, effective, structural)
Keywords:
Degenerative Disease
FUNCTIONAL MRI
Machine Learning
Neurological
Other - Deep Brain Stimulation, Parkinson's disease
1|2Indicates the priority used for review
·Figure 1 : Classification performance of SVM, LDA, and RF using SelectKBest across different values of K.
·Figure 2: Brain mask illustrating the voxels used for classification with the highest averaged accuracy and minimum required number of features (GCOR (upper): K = 33000; EC (lower): K = 27000).
Provide references using author date format
Arévalo Sáenz, A. et al. (2022) 'Estimulación cerebral profunda en la enfermedad de parkinson: Análisis de la anisotropía fraccional cerebral en Pacientes Intervenidos Mediante estimulación cerebral profunda', Revista de Neurología, 74(04), p. 125. doi:10.33588/rn.7404.2021196.
Lamoš, M. et al. (2023) 'The effect of deep brain stimulation in parkinson’s disease reflected in EEG microstates', npj Parkinson’s Disease, 9(1). doi:10.1038/s41531-023-00508-x.
Peralta, M., Jannin, P. and Baxter, J.S.H. (2021) 'Machine learning in Deep brain stimulation: A systematic review', Artificial Intelligence in Medicine, 122, p. 102198. doi:10.1016/j.artmed.2021.102198.
Acknowledgements: Supported by a grant of the National Institute for Neurological Research, Czech Republic, Programme EXCELES (ID project No. LX22NPO5107) and the Charles University: Cooperatio Program in Neuroscience.