Fading of brain network fingerprint in Parkinson's disease predicts motor clinical impairment

Abstract

The clinical connectome fingerprint (CCF) was recently introduced as a way to assess brain dynamics. It is an approach able to recognize individuals, based on the brain network. It showed its applicability providing network features used to predict the cognitive decline in preclinical Alzheimer's disease. In this article, we explore the performance of CCF in 47 Parkinson's disease (PD) patients and 47 healthy controls, under the hypothesis that patients would show reduced identifiability as compared to controls, and that such reduction could be used to predict motor impairment. We used source-reconstructed magnetoencephalography signals to build two functional connectomes for 47 patients with PD and 47 healthy controls. Then, exploiting the two connectomes per individual, we investigated the identifiability characteristics of each subject in each group. We observed reduced identifiability in patients compared to healthy individuals in the beta band. Furthermore, we found that the reduction in identifiability was proportional to the motor impairment, assessed through the Unified Parkinson's Disease Rating Scale, and, interestingly, able to predict it (at the subject level), through a cross-validated regression model. Along with previous evidence, this article shows that CCF captures disrupted dynamics in neurodegenerative diseases and is particularly effective in predicting motor clinical impairment in PD.

Keywords: Parkinson's disease; brain fingerprint; brain network; clinical connectome fingerprint; magnetoencephalography; motor impairment; neurodegenerative disease.

FIGURE 1

Processing of the functional connectomes and their application for fingerprint analysis. (a) Visual representation of the data analysis pipeline. Through a magnetoencephalography (MEG) system composed of 154 sensors, we recorded the magnetic field emitted by neural activity. Noisy MEG signals were cleaned and artifacts were removed. Source reconstruction (beamforming) was achieved according to the Automated Anatomical Labeling atlas. Connectivity estimation was performed through phase linearity measurement (PLM) algorithm. (b) Fingerprint analysis scheme. Different identifiability matrices were built in order to investigate the functional connectomes (FCs) identifiability in healthy subjects (HS) and patients with Parkinson's disease (PD). Correlating test–retest individuals' FCs we obtained the blue and the red boxes, that represents the identifiability characteristics of HS and PD, respectively. Cross correlating test and retest FCs of subjects of different groups we obtained hybrid identifiability matrices. From these matrices, we were able to calculate the similarity of each patient's FC with respect to the ones belonging to the healthy group (Iclinical score).

FIGURE 2

Brain identification in healthy and Parkinson's disease (PD). (a) Identifiability matrices of healthy subjects (HS) and patients with PD. The main diagonal is representative of the self‐identifiability (I‐self), while off‐diagonal elements are representative of the similarity among different individuals (I‐others). The difference between those values is described as differential identification (I‐diff) and gives an estimation of the fingerprinting level of a group. These matrices are based on the functional connectomes computed in beta band. Note that the more the main diagonal is visible, the more the subjects turn out to be identifiable. Success rate (SR) is reported too, as a percentage of the number of times an individual is recognizable with respect to other individuals within the same group. (b) Statistical comparison between fingerprint parameters calculated on the identifiability matrices of HS and PD. HS shows higher identifiability with respect to PD. Significance p‐value: *p < .05, **p < .01, ***p < .001

FIGURE 3

Edge contribution to connectome fingerprint. Intraclass correlation (ICC) for the beta band connectivity, assessing the brain regions contribution to identifiability. Higher ICC values of an edge means major contribution of that edge to the identifiability. The same results are shown as brain renders displaying the nodal strength of most reliable edges (above the 75 percentile of the distribution; colorbar borders represent the 5 and 95 percentiles).

FIGURE 4

Identifiability based on the edge contribution. Success rate (SR) distribution in identifying individuals when performing fingerprint analysis including 100 edges at a time. SR distributions of healthy subjects (HS, blue line) and Parkinson's disease patients (PD, red line), were obtained adding the edges from the most contributing to the least contributing to identifiability, relying on the intraclass correlation (ICC) values. Actual distributions were compared to their respective null distribution (light blue for HS, and light red for PD) obtained repeating the same analysis 100 times, including the edges in a random order. The left panel shows the analysis performed using the ICC matrices belonging to each group. The right panel shows the analysis performed considering the ICC matrix of the healthy individuals for both HS and PD group.

FIGURE 5

Clinical variables contribution to patients' heterogeneity. (a) The panel displays the variance explained by the additive model including nine variables (age, education, gender, disease duration, clinical subtype, depression level [BDI], cognitive assessment [MoCA], levodopa equivalent daily dose [LEDD], and Unified Parkinson's Disease Rating Scale part III [UPDRS‐III]). Significant predictors in bold; positive/negative coefficients indicated with β+/β−. (b) The panel shows the correspondence between the actual I‐others values and the ones predicted by the model with k‐fold cross validation (k = 5). (c) The panel displays the distribution of the standardized residuals with k‐fold cross validation (k = 5).

FIGURE 6

Motor impairment prediction based on “clinical fingerprint.” The analysis aims to predict the motor impairment of the patients assessed through Unified Parkinson's Disease Rating Scale part III (UPDRS‐III), relying on the clinical identifiability (Iclinical) score. (a) Edges are added iteratively (100 per time up to whole‐brain) based on the Parkinson's disease (PD) patients' intraclass correlation (ICC) values, from the most to the least contributing to identifiability (x axis). Hence, the prediction performance (k‐fold cross validation with k = 5) of each multilinear model based on the Iclinical is evaluated as the Spearman correlation coefficient (Spearman's ρ, on y axis) between actual and predicted UPDRS‐III values (blue line). For comparison, we built a null model obtained by repeating the same analysis 100 times, but selecting the edges randomly (the red line represents the mean prediction of the null model; the shaded red area represents the standard deviation of the null model predictions). The following panels show the results of multilinear model with the highest performance, that is, when the Iclinical is calculated considering the 500 most reliable edges for PD patients identification (ICC score). (b) The panel shows the variance explained by the additive model including nine variables (age, education, gender, disease duration, clinical subtype, depression level [BDI], cognitive assessment [MoCA], levodopa equivalent daily dose [LEDD], and Iclinical in beta band). Significant predictors in bold; positive/negative coefficients indicated with β+/β−. (c) The panel shows the correspondence between the actual UPDRS‐III values and the ones predicted by the model. (d) The panel displays the distribution of the standardized residuals.

FIGURE 7

Relationship between motor impairment and clinical identifiability. Pearson's correlation between motor impairment assessed through Unified Parkinson's Disease Rating Scale part III (UPDRS‐III) and clinical identifiability expressed by Iclinical score in beta band. The Iclinical was computed including the 500 most reliable edges according to the intraclass correlation (ICC) scores. The negative significant coefficient indicates better motor condition (low UPDRS‐III values) when patients connectivity is more similar to the healthy individuals (high Iclinical), and vice versa.