Reorganization of Functional Networks in Mild Cognitive Impairment

Whether the balance between integration and segregation of information in the brain is damaged in Mild Cognitive Impairment (MCI) subjects is still a matter of debate. Here we characterize the functional network architecture of MCI subjects by means of complex networks analysis. Magnetoencephalograms (MEG) time series obtained during a memory task were evaluated by synchronization likelihood (SL), to quantify the statistical dependence between MEG signals and to obtain the functional networks. Graphs from MCI subjects show an enhancement of the strength of connections, together with an increase in the outreach parameter, suggesting that memory processing in MCI subjects is associated with higher energy expenditure and a tendency toward random structure, which breaks the balance between integration and segregation. All features are reproduced by an evolutionary network model that simulates the degenerative process of a healthy functional network to that associated with MCI. Due to the high rate of conversion from MCI to Alzheimer Disease (AD), these results show that the analysis of functional networks could be an appropriate tool for the early detection of both MCI and AD.


S2 Statistical analysis
In Figure S1 we have plot the percentage of variation of the topological parameters for the full bandwidth (8-45 Hz), together with its statistical signicance. Circles correspond to a p < 0.03 and stars to p < 0.01. Note that outreach is the parameter which is modied the most (and in addition p outreach < 0.01), which indicates that it is the most suitable parameter in order to dierentiate between healthy individuals and patients with MCI.
S3 Community structure and roles Figure S2 complements Fig. 4 of the main document, which shows changes in the roles played by the main nodes of the network in the intra and inter lobular connections. We use the node characterization proposed by Guimerà et al. [1] that classies the role of the nodes with regard to the function that they are playing inside and outside of their communities. Two parameters are calculated, the within-module degree z i (also known as z-score) and the participation coecient p i . The rst parameter computes the importance of the node inside its community and it is dened as: where k i and c i are, respectively, the degree and the community c i of the node i,k ci is the mean degree of the community c i and σ kci is the standard deviation of k in c i . On the other hand, the participation coecient p i indicates how connections of node i are distributed among the existing communities: where k ci is the number of connections between node i and community c i and N c is the total number of communities.
The participation coecient is zero when all links of a node are inside its own community and close to one when they are distributed among all modules of the network. In Fig. S2(A) we have plotted the [p i , z i ] phase space of all nodes in the network for the control group. We can observe that, during a memory task, most participating nodes (high p i ) are located over the two frontal lobes (blue and black circles), while nodes with higher relevance (i.e., those with higher z i ) are located over the occipital lobe. Figure S2(B) shows the variation of both parameters in the MCI group (MCI minus control). In accordance with previous results, the participation coecient increases in the majority of nodes, since connections between lobes are strengthen over the whole network. Regarding the within-module degree, we observe both positive and negative changes, which indicates a generalized reorganization inside each lobe.  [3], and the Global Deterioration Scale/Functional Assessment Staging GDS/FAST [4].

S4 Supplementary materials and methods
In order to avoid possible dierences due to the years of education, patients and controls were chosen so that the resulting average number of years of education was similar: 10 years for patients and 11 years for controls.
A modied version of the Sternberg's letter-probe task [5] was used as the memory test. A set of ve letters was presented and participants were asked to keep the letters in mind. After the presentation of the ve letter set, a series of single letters (500 ms in duration with a random ISI between 23 s) was presented one at a time, and the participants were asked to press a button with their right hand when a member of the previous set was detected. All participants completed a training session before the actual test, which did not start until participants demonstrated that they The MEG signal was recorded with a 254 Hz sampling frequency and a band pass of 0.5 to 50 Hz, using a 148-channel whole-head magnetometer (MAGNES 2500 WH, 4-D Neuroimaging) conned in a magnetically shielded room. An environmental noise reduction algorithm using reference channels at a distance from the MEG sensors was applied to the data. Thereafter, single trial epochs were visually inspected by an experienced investigator, and epochs containing visible blinks, eye movements or muscular artifacts were excluded from further analysis. Artifact-free epochs from each channel were then classied into four dierent categories, according to the subject's performance in the experiment: hits, false alarms, correct rejections and omissions. Only hits were considered for further analysis because we were interested in evaluating the functional connectivity patterns which support recognition success. Thirty-ve epochs were used to calculate Synchronization Likelihood (SL) between all pairs of nodes (electrodes) of each individual.
This lower bound was determined by the participant with least epochs. To have an equal number of epochs across participants, thirty-ve epochs were randomly chosen from each of the other participants.
In-house Fortran code was used to implement the SL algorithm as described by Stam et al. [6]. The SL algorithm was applied to the thirty-ve extracted (artifact-free) one-second epochs of each subject. For each frequency band optimal SL parameter values were chosen according to Montez et al. [7]: lag L = fs , theiler window W 1 = 2L(M − 1) and window length W 2 > 10 pref + (W 1 − 1), with parameter pref below 0.05. Finally, f s is the sampling rate, and H f and L f are the high and low frequency bounds, respectively.
As mentioned in the main document, the following frequency bands were considered: α 1 : (8 − 11)Hz, α 2 : (11 − 14)Hz, β 1 : (14 − 25)Hz, β 2 : (25 − 35)Hz, γ : (35 − 45)Hz. The SL index was not computed for bands under 8 Hz as the epoch length and sampling rate do not allow an accurate enough estimation [7]. All epochs were digitally ltered o-line at the above frequency bands. Subsequently, SL was calculated for each of the thirty-ve one-second epochs of the (148 × 147)/2 channel pairs, for each frequency band and the full-band signal, and for each subject (nineteen controls and nineteen patients).