Whole-brain dynamical modeling for classification of Parkinson's disease

Introduction:Simulated whole-brain connectomes demonstrate disparate inter-individual variability depending on dataprocessing and modeling approach (Domhof et al., 2021; Jung et al., 2021; Popovych et al., 2021). Byconsidering the human brain connectome as an individualized attribute, we investigate how empirical andsimulated whole-brain connectome-derived features can be utilized to classify patients with Parkinson'sdisease against healthy controls in light of varying data processing and model validation.Methods:This study included 51 (30 males) healthy controls and 65 (45 males) patients with Parkinson's disease.Diffusion-weighted MRI (dMRI), T1-weighted MRI and resting-state functional MRI (rsfMRI) were acquired inthe subjects. Empirical functional connectivity (eFC) was calculated by Pearson correlation between regionsof a whole-brain parcellation using blood oxygenation level-dependent (BOLD) signals extracted from thersfMRI. Empirical structural connectivity (eSC) was reconstructed using extracted streamlines connectingregions of the parcellation from whole-brain tractography that was calculated using dMRI. We used twodifferent parcellation schemes based on functional and structural brain properties for calculation of the eFCand eSC. We also simulated BOLD signals by a whole-brain dynamical model of Jansen-Rit type (Jansen etal., 1995) derived from the eSC serving as a network backbone. The simulated data was used to calculatethe simulated FC that was employed together with eFC and eSC to reveal differences between patients andcontrols. In this study, we applied four temporal filtering conditions in four frequency bands to empirical andsimulated BOLD signals. In addition to the widely used model validation via fitting the dynamical model toempirical neuroimaging data (Deco et al., 2015; Honey et al., 2009; Naskar et al., 2021), we invented amodel validation against behavioral data, such as subject classes, which we refer to as behavioral modelfitting and applied it to a machine-learning (ML) classification of Parkinsonian patients.Results:The results of ML investigation demonstrated that performance of the patient classification can besignificantly improved when the empirical data are complemented by the simulation results. This issupported by both integrative performance measures (Fig. 1) and predicted probabilities for individualsubjects (Fig. 2). We also showed that temporal filtering of empirical and simulated BOLD signals influencesthe prediction results, where the filtering in the low-frequency band is advisable for Parkinsonian patientclassification. In addition, composing the feature space of empirical and simulated data from multiple brainparcellation schemes provided complementary features that further improved the prediction performance.The best performance (median of balanced accuracies) was 0.65 for unseen subjects (cf. 0.61 using theempirical data)