The progressive loss of brain network fingerprints in Amyotrophic Lateral Sclerosis predicts clinical impairment

Abstract

Amyotrophic lateral sclerosis (ALS) is a neurodegenerative disease characterised by functional connectivity alterations in both motor and extra-motor brain regions. Within the framework of network analysis, fingerprinting represents a reliable approach to assess subject-specific connectivity features within a given population (healthy or diseased). Here, we applied the Clinical Connectome Fingerprint (CCF) analysis to source-reconstructed magnetoencephalography (MEG) signals in a cohort of seventy-eight subjects: thirty-nine ALS patients and thirty-nine healthy controls. We set out to develop an identifiability matrix to assess the extent to which each patient was recognisable based on his/her connectome, as compared to healthy controls. The analysis was performed in the five canonical frequency bands. Then, we built a multilinear regression model to test the ability of the "clinical fingerprint" to predict the clinical evolution of the disease, as assessed by the Amyotrophic Lateral Sclerosis Functional Rating Scale-Revised (ALSFRS-r), the King's disease staging system, and the Milano-Torino Staging (MiToS) disease staging system. We found a drop in the identifiability of patients in the alpha band compared to the healthy controls. Furthermore, the "clinical fingerprint" was predictive of the ALSFRS-r (p = 0.0397; β = 32.8), the King's (p = 0.0001; β = -7.40), and the MiToS (p = 0.0025; β = -4.9) scores. Accordingly, it negatively correlated with the King's (Spearman's rho = -0.6041, p = 0.0003) and MiToS scales (Spearman's rho = -0.4953, p = 0.0040). Our results demonstrated the ability of the CCF approach to predict the individual motor impairment in patients affected by ALS. Given the subject-specificity of our approach, we hope to further exploit it to improve disease management.

Keywords: Brain network identifiability; Clinical connectome fingerprint; Functional connectome; Magnetoencephalography, Phase Linearity Measurement; Motor neurons disease; Neurodegenerative diseases.

Graphical abstract

Fig. 1

Pipeline analysis and clinical connectome fingerprint application. (A) a: the neuronal activity was recorded using a magnetoencephalography (MEG) composed by 154 sensors; b: raw MEG signals including noise, and cardiac and blinking artefacts; c: MEG signals after removing noise and physiological artefacts; d: magnetic resonance image (MRI) of a subject; e: coregistration of MEG and MRI signals to obtain the source reconstruction (beamforming); f: functional connectivity matrix (one for each frequency band) estimated using the phase linearity measurement (PLM). Each matrix displays on both rows and columns the 90 ROIs (B) The blue and the green blocks represents the two identifiability matrices of healthy controls (HC) and Amyotrophic Lateral Sclerosis (ALS) patients, respectively obtained by correlating the test and re-test individual functional connectomes, in each group separately. Crossing the FCs test of the HC with the FCs retest of the ALS and vice-versa, we obtained two hybrid identifiability matrices, that allowed us to calculate the I-clinical value of each patient (expressing how much an ALS patient is similar to the HC group). (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 2

Identifiability matrices and fingerprint features comparison. (A). Identifiability matrices comparison in the alpha band. The main diagonal is representative of the I-self while the I-others is represented by all the elements outside the main diagonal. The main diagonal of the healthy controls’ (HC) identifiability matrix (IM) is more visible as compared to the amyotrophic lateral sclerosis patients’s (ALS) one, revealing a reduction in self similarity (i.e lower I-self values) of the ALS compared to the HC. (B) Fingerprint features comparison in the alpha band. HC displays higher values of both I-self and I-others as compared to ALS patients. The right panel shows the comparison between the HC I-others and the I-clinical of the patients. Even in this case, HC displays higher values of I-others compared to the I-clinical of the patients confirming the drop of identifiability. FCs = functional connectomes. significant p value is indicated with **(p < 0.01) and *** (p < 0.001).

Fig. 3

Edge contribution to connectomes’ identifiability. (A) The intra-class-correlation (ICC) matrices show the edges contribution to the identifiability in the alpha band. We can observe a drop in ICC in the amyotrophic lateral sclerosis (ALS) as compared to the healthy controls (HC). This means that in the patients there are less reliable brain regions that contribute to self-identification. The rank of the 90 ROIs is displayed in the left panel. (B) The same results are represented as brain renders, showing the nodal strength of the most reliable edges. (from 5th to 95th percentile). The nodal strength values are obtained by summing the elements of both rows and columns of the lower part of the ICC matrices. (C) Difference between the HC ad ALS ICC matrices, with the consequent brain renders in which we highlighted the brain regions whose nodal strength significantly differed between the two groups. The healthy controls showed higher nodal strength values in all four brain regions.

Fig. 4

Iterative model of edgewise subjects’ identification. The success rate (SR) of healthy controls (HC, blue line) and Amyotrophic Lateral Sclerosis patients (ALS, red line) were obtained performing the fingerprint analysis by adding 50 edges at a time from those contributing the least to identifiability to those which contributed the most, according to the intraclass correlation (ICC) values. The light blue line and the light red line were representative of the null distributions (obtained by adding the edges in a random order, one hundred times at each step) for the HC and ALS patients respectively. The observed SR values are higher values as compared to the null model distributions for both patients and HC. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 5

Motor impairment prediction in the alpha band. A multilinear regression model with leave-one-out cross validation (LOOCV) was performed to test the capacity of the “clinical fingerprint” (i.e., the I-clinical score) to predict the motor impairment in amyotrophic lateral sclerosis (ALS) patients. The predictive models of the ALS functional rating scale revised (ALSFRS-r) (A), the King’s disease staging system (B) and the Milano Torino staging system (MiTos) are represented on the rows. The left column reports the explained variance obtained by adding the five predictors (age, education, gender, duration of disease and I-clinical in alpha band). The significant predictors are highlighted in bold while the positive and negative coefficients are indicated with β+/β-, respectively. The significant p value is indicated with * (p < 0.05) ** (p < 0.01) and *** (p < 0.001). The central column shows the comparison between the observed and the predicted values of the response variable, validated through LOOCV for A, B and C, respectively. The right column represents the distribution of the standardised residuals (i.e., standardisation of the difference between observed and predicted values).

Fig. 6

Correlation between motor impairment and I-clinical. (A) Spearman’s correlation between the King’s disease staging system and the clinical identifiability (i.e., the I-clinical). The negative correlation coefficient indicates that as the I-clinical scores increase, the King’s scores decrease and, thus, the ALS patients show better motor functions. (B) Negative correlation between the Milano-Torino staging system (MiToS) and the I-clinical. Higher MiToS scores (i.e., worse motor condition) correspond to lower I-clinical scores and vice versa.