Network Hyperexcitability in Early-Stage Alzheimer’s Disease: Evaluation of Functional Connectivity Biomarkers in a Computational Disease Model

Background: There is increasing evidence from animal and clinical studies that network hyperexcitability (NH) may be an important pathophysiological process and potential target for treatment in early Alzheimer’s disease (AD). Measures of functional connectivity (FC) have been proposed as promising biomarkers for NH, but it is unknown which measure has the highest sensitivity for early-stage changes in the excitation/inhibition balance. Objective: We aim to test the performance of different FC measures in detecting NH at the earliest stage using a computational approach. Methods: We use a whole brain computational model of activity dependent degeneration to simulate progressive AD pathology and NH. We investigate if and at what stage four measures of FC (amplitude envelope correlation corrected [AECc], phase lag index [PLI], joint permutation entropy [JPE] and a new measure: phase lag time [PLT]) can detect early-stage AD pathophysiology. Results: The activity dependent degeneration model replicates spectral changes in line with clinical data and demonstrates increasing NH. Compared to relative theta power as a gold standard the AECc and PLI are shown to be less sensitive in detecting early-stage NH and AD-related neurophysiological abnormalities, while the JPE and the PLT show more sensitivity with excellent test characteristics. Conclusions: Novel FC measures, which are better in detecting rapid fluctuations in neural activity and connectivity, may be superior to well-known measures such as the AECc and PLI in detecting early phase neurophysiological abnormalities and in particular NH in AD. These markers could improve early diagnosis and treatment target identification.

evidence for the validity of this model by comparing empirical resting-state recordings of MEG (magnetoencephalography) to the ADD model.For this purpose, we re-analyzed a dataset of 18 persons with subjective cognitive decline (SCD), 18 subjects with mild cognitive impairment (MCI), and 18 subjects with Alzheimer's disease (AD).Spectral features of these data were previously described in detail in an earlier paper in this journal [1].Results with the joint permutation entropy have also been described previously [2].Descriptive information on these groups can be found in Supplementary Table 1.Further background information can be found in [1,2].
For the present analysis, we first computed spectral features for 20 consecutive epochs (sample frequency: 1250 Hz; length 3.2768 s (4096 samples); 78 AAL ROIs, corresponding to the 78 ROIs used for the ADD model).Since relative power in the theta band, averaged over all 78 ROIs, is a promising biomarker in early AD we used this to match the empirical data to the model.Specifically, for each epoch of each subject we determined the model time step (range: 1-100) where the squared difference between model and empirical average relative theta power obtained the smallest value.With this approach we determined for each subject an average (over 20 epochs) best matching time to the ADD model.Since each model time also corresponds to specific values of underlying model parameters such as excitatory and inhibitory firing rates, E/I balance, coupling strengths between excitatory and inhibitory populations, and between thalamic input and excitatory populations, these we taken into account as well.Finally, we determined whether optimal matching times and internal model parameters were significantly different between the MCI group compared to the SCD group and the AD group compared to the SCD group.
Results of this analysis are shown in Supplementary Table 2.As can be seen, the SCD group showed an optimal match to the ADD model at an earlier time step (mean 20.01 SD 2.88) compared to the MCI group (mean 23.52; SD 4.25) and the AD group (mean 35.06;SD 16.48).
The matching times of the MCI and AD groups were both significantly later than the of the SCD group (p < 0.005).This shows that progressive pathology from SCD to MCI to AD corresponds to later matching times (later simulated stages of the degenerative process) in the ADD model.
Of interest, this also allows to infer from the model internal features such as firing rates and connection strengths which are not directly accessible in empirical data (columns 2-7 in Supplementary Table 2).These findings suggest significantly increased excitatory and decreased inhibitory firing rates and increased E/I balance in MCI and even more so in AD.This is accompanied by a loss of structural connectivity between the excitatory and inhibitory populations (parameters C1 and C2) and a loss of thalamic input (Pt).
Next, we evaluated whether and to what extent the patterns of chance in functional connectivity measures predicted by the ADD model could be found in empirical MEG recordings of the SCD, MCI, and AD groups.In particular, we were interested in the prediction by the model that new measures (JPE and PLT) would be more sensitive to early pathology (MCI phase as opposed to AD phase) compared to conventional measures (AECc and PLI), in particular in the theta band.Results of this analysis are shown in Supplementary Table 3 and Supplementary Figure 1.Note that the results for the JPE of this dataset have been described in [2].Significant changes in functional connectivity in the AD group compared to the SCD group could be demonstrated with the AECc in the alpha and beta band, with the PLI in the delta band, with the JPE in the theta band, and with the PLT in the theta and beta band.Significant changes in functional connectivity between the MCI group and the SCD group could only be demonstrated with the JPE and the PLT in the theta band.The empirical data therefore show a pattern of changes, in particular in the theta band, which is qualitatively similar to the predictions made by the model.In particular, the main prediction of the model-superiority of JPE and PLT in the theta band in detecting abnormalities-in the early phase along the AD spectrum is confirmed by the empirical findings.

Table 2 .
Fitting empirical MEG recordings to ADD model. of time steps in ADD model corresponding to optimal fit with empirical data; E rate, firing rate of excitatory population in spikes/second; E rate, firing rate of inhibitory population in spikes per second; E/(E+I), balance between excitatory and inhibitory firing rates; C1, coupling strength between excitatory and inhibitory populations; C2, coupling strength between inhibitory and excitatory populations; Pt, thalamic input to excitatory population in spikes per second; SCD, subjective cognitive decline; MCI, mild cognitive impairment; AD, Alzheimer's disease; SD, standard deviation.**p<0.005(comparison of MCI or AD to SCD; Permutation test using BrainWave).