2.1.

Participants

From May 2016 to August 2017, 23 AD dementia patients, 25 aMCI patients, and 30 sex-, age-, and education-matched HCs participated in this study. All participants were right-handed Han Chinese with normal or corrected-to-normal version, who provided written informed consent for inclusion prior to the study. All clinical tests were approved by the Ethics Committee of the Beijing Normal (Beijing, China) and were then carried out as a standardized clinical evaluation, including a medical history interview and a battery of neuropsychological tests. The neuropsychological tests consisted of the Chinese version of the mini-mental state examination (MMSE),²⁵ the Beijing version of Montreal cognitive assessment (MoCA),²⁶ clinical dementia rating (CDR), the auditory verbal learning test (AVLT),²⁷ Hachinski ischemic scale (HIS), Hamilton depression rating scale,²⁸ the center for Epidemiologic Studies depression scale,²⁹ and activities of daily living scale. Patients with aMCI and AD dementia were recruited from the memory clinic of the Neurology Department, Xuan Wu Hospital, Capital Medical University (Beijing, China), whereas the HCs were recruited from the local community by advertisements. The diagnoses of aMCI and AD dementia were determined by the consensus of two experienced neurologists according to the published criteria from Petersen et al.³⁰ and National Institute on Aging-Alzheimer’s Association,³¹ respectively. For details, the inclusion criteria for aMCI were as follows: (1) a memory complaint confirmed by an informant; (2) cognitive decline in a single domain or multiple domains; abnormal objective cognitive impairment documented by scores falling 1.5 SD below the age and education matched-specific norms on standard neuropsychological tests; (3) CDR score of 0.5; and (4) being free from dementia according to the Diagnostic and Statistical Manual of Mental Disorders, fourth edition, revised. The inclusion criteria for HCs were: (1) no complaint of memory or other cognitive impairment, (2) CDR score is 0, and (3) no severe visual or auditory impairment. All the participants were excluded if they demonstrated: (1) a clear history of stroke; (2) depression (Hamilton depression rating scale score $\ge 24$ points); (3) other nervous system diseases that can cause cognitive impairment (such as brain tumors, Parkinson’s disease, encephalitis, or epilepsy); (4) traumatic brain injury; (5) other systemic diseases that can cause cognitive impairment, such as thyroid dysfunction, severe anemia, syphilis, and HIV; and (6) histories of psychosis or congenital mental growth retardation.

2.2.

fNIRS Data Acquisition

The fNIRS scanning was conducted in a dimly lit room in Xuan Wu hospital. All participants were instructed to sit in a comfortable chair and remain motionless with their eyes closed [Fig. 1(a)], to avoid sleeping, and not to think about anything. The fNIRS data were acquired using a CW6 optical imaging system (CW6, Techen Co., Massachusetts) with a 46-channel array of optodes. This array consisted of 12 light emitters and 24 optical detectors [Figs. 1(b) and 1(c)], each of which detected the NIR light of its neighboring/surrounding emitters. The sources and detectors were alternately placed on each participant’s right and left hemispheres with an interoptode distance of 3.2 cm. The bottom column of the probe array was placed along the inion reference point according to the international 10/20 system^32,33 so that the midpoint of the lower edge of the array was placed directly above the inion. The probe positioning was examined and adjusted to ensure consistency of the positions among the participants. Resting state was defined as no specific cognitive task during the fMRI scanning. The resting-state fNIRS data acquisition lasted $\sim 11\min$ for each individual participant.

Fig. 1

(a) Photograph of fNIRS data collection from a participant. (b) The arrangement of the 46 measurement channels across the whole head. The red and green solid circles represent sources and detectors, respectively. The digits represent the positions of the measurement channels. (c) Anatomical position of each measurement channel. (d) Illustration of dynamic FC analysis steps. The spectrum power of the dynamic FC time series for each channel was first extracted. The AUC of the power spectrum across the low frequency ($<0.1\mathrm{Hz}$) band was then used as a dynamic FC variability index to mark time-varying characteristics of dynamic FC. K-means clustering was utilized to identify brain FC states.

fNIRS data from individual channels were acquired at two different wavelengths (690 and 830 nm) with a sampling rate of 50 Hz. The signal quality was evaluated using signal-to-noise ratio, which was calculated as the ratio between mean signal and stand deviation of the raw signal on 690 and 830 nm, respectively. Generally, the signal was acceptable for the signal-to-noise ratio larger than 2 Hz. Changes in oxygenated (HbO), deoxygenated (HbR), and total hemoglobin (HbT) signals were generated by using the modified Beer–Lambert law.³⁴ The HbO signal was primarily analyzed for the present study due to its high signal-to-noise in the correlation with regional cerebral blood flow.³⁵ The HbR results were also analyzed and presented in the supplement as a complement.

For each individual’s HbO dataset, a temporal independent component analysis (ICA) was conducted to remove typical motion-induced artifacts and systematic physiological noise.^36–38 Specifically, these noise components were identified according to the components’ temporal profiles, spatial maps, and power spectra. A component would be considered noise if it met one of the following conditions:³⁷ (1) the corresponding temporal profile included sudden jumps, slowly varied U or inverted U-shaped spike, or numerous intercurrent quick spikes (e.g., motion artifacts); (2) the dominant frequency of power spectra of the component was outside the range of 0.01 to 0.1 Hz; and (3) the spatial map of the component presented a global and spatially dispersive pattern (e.g., physiological interference). Once the noise components were identified, the concentration signal was subsequently reconstructed with these particular components eliminated from the original hemoglobin time course by replacing zero in the corresponding column of mixing matrix.³⁸ After the ICA procedure, a bandpass filter from 0.01 to 0.1 Hz was implemented to the denoised hemoglobin signals and then 10-min recordings (i.e., 30,000 time points) from the continuous time courses were extracted for later analysis.

2.3.

Dynamic FC Computation

Figure 1(d) illustrates the dynamic FC analysis steps. Similar to the approach adopted by Allen et al.¹⁹ and in our previous study,³⁹ a sliding window-based correlation approach was first utilized to generate a series of dynamic FC maps. Specifically, for time series (i.e., 10 min) on all measurement channels, 60-s time windows were selected and then shifted in an increment of 1 s along the entire time length. The FC within each time window was quantitatively calculated between any pair of brain regions using the Pearson correlation strategy. The 10-min measurement duration and a 60-s time window brought about 541 sliding time windows and thus constructed 541 dynamic FC maps. To quantify the whole-cortex FC fluctuation across a series of sliding time windows, we also calculated the FC variability index ($Q$)^19,39 on each participant. Specifically, we first of all calculated the spectrum power of the dynamic FC time series (with Fourier analysis),¹⁹ and then the area under the curve (AUC) of the power spectrum across the low frequency ($<0.1\mathrm{Hz}$) band was quantified as the FC variability index, $Q$, to mark time-varying characteristics of dynamic FC. A larger $Q$ value denoted more variable FC, whereas a smaller $Q$ value represented less variable FC. Of note, the $Q$ index takes into account the information of power and amplitude in the frequency domain by filtering the potential fluctuating noises contained in the high frequencies ($>0.1\mathrm{Hz}$), which is slightly different from a simple measure of standard deviation.⁴⁰ We point out that the static FC was also calculated based on total signal length with Pearson correlation analysis.

2.4.

Dynamic FC Differences between the Various Groups

To localize the specific pairs of regions in which the dynamic FC was altered in patient groups, we introduced a network-based statistic (NBS) approach⁴¹ to compare group difference in FC variability $Q$. The identified alternation in the value of index $Q$ was shown in the brain cortex according to their different spatial connectivity patterns: homotopic connectivity, long intrahemispheric connectivity, short intrahemispheric connectivity, and heterotopic connectivity. The details for comparison can be found in Sec. 2.6. The receiver operating characteristic (ROC) curve approach was further used to evaluate whether the $Q$ metric might serve as a biomarker for diagnosing patients from HCs. ROC is a graphical plot that illustrates the performance of a binary classifier, which was created by plotting the sensitivity against the 1-specificity at various thresholds.

2.5.

Detection and Characterization of FC States

To detect brief FC patterns (i.e., FC states), we applied a k-means clustering algorithm to the windowed correlation matrices concatenated across all participants. The clustering procedure was the same as the one used by Allen et al.¹⁹ The number of clusters ($k$) was determined by using the elbow criterion of the cluster validity index that was denoted as the ratio between between-cluster distances to the within-cluster distance. For each cluster, we separately calculated the averaged FC map to represent the identified brain FC state. We also calculated the mean of the standard deviation among these FC maps, which characterized the FC variance within each brain state (i.e., cluster).

To demonstrate the difference in FC states between the patients and HC groups, we further introduced the metric of occurrence frequency, $F$, to characterize the occurrence of each FC state during dynamic brain activity. The metric $F$ was defined as: $F=N_{\mathrm{i}}/N_{\mathrm{t}}$, in which $N_i$ is the number of windows classified into a particular state and $N_t$ is the total number of the sliding time windows for a given time series. As such, for each participant, the occurrence frequency ($F$) in each FC state could be quantified. For each FC state, the mean $F$ value was used as a statistical index for the comparison between the patients and heathy controls.

2.6.

Statistical Analysis

3.1.

Demographic and Clinical Characteristics

The demographic data are shown in Table 1. There were no significant differences in age [$F(\mathrm{2,75})=1.891$; $p=0.16$], gender ($p=0.40$), or years of education [$F(\mathrm{2,75})=1.07$; $p=0.35$] among the three groups. However, the patient group (aMCI and AD) had significantly lower scores on the MMSE [$F(\mathrm{2,75})=51.51$, $p<0.001$], MoCA [$F(\mathrm{2,75})=53.80$, $p<0.001$], AVLT-immediate recall [$F(\mathrm{2,75})=58.34$, $p<0.001$], AVLT delayed recall [$F(\mathrm{2,75})=118.52$, $p<0.001$], and AVLT-recognition [$F(\mathrm{2,75})=65.26$, $p<0.001$] than the HC group (Table 1).

Table 1

Demographics and clinical characters of the participants.

3.2.

Visualization of Dynamic FC Variability

Using sliding window correlation analysis and Fourier analysis, we obtained quantitative FC variability $Q$ at each connection for the HCs, aMCI, and AD groups, respectively [Fig. 2(a)]. Visually, the FC variability was obviously increased in participants during the progression of HCs to aMCI to AD, with most of the connections showing larger $Q$ in AD and smaller $Q$ in HCs [Fig. 2(b)]. Quantitatively, the results also support the observation that averaged $Q$ values across all 1035 (i.e., $46\times 45/2$) connections were $4.2\pm 0.16$ for HCs, $4.5\pm 0.18$ for aMCI, and $4.7\pm 0.17$ for AD. One possible explanation for the larger $Q$ in AD is that AD patients could primarily stay in an aberrant brain state while that brain state could show larger brain FC fluctuations. The static FC was also calculated for the HCs, aMCI, and AD groups [Fig. 2(c)]. Visually, the static FC maps showed similar spatial patterns across the HCs, aMCI, and AD groups.

Fig. 2

(a) Maps of FC variability for HC, aMCI, and AD groups, respectively. The $Q$ values represent the variability of dynamic FC over time. (b) One-dimensional representation of the FC variability ($Q$) for these three groups. (c) Maps of static FC for HC, aMCI, and AD groups, respectively. The values represent the static functional connectivity strength.

3.3.

Disrupted Regional FC Dynamics in aMCI and AD

We used the NBS method to identify the disrupted connected components in patients (Fig. 3). Under the cluster defining threshold of $p<0.05$, networks with 307 and 548 connections were revealed to experience significant changes in dynamic FC variability in the aMCI and AD groups, respectively. Using a more rigorous threshold of $p<0.005$, the networks were reduced to 34 and 178 significant connections for aMCI and AD, respectively.

Fig. 3

The brain regions with significantly increased $Q$ values in patients with (a) aMCI and (b) AD. The significant connections were categorized into four connectivity subgroups, i.e., homotopic, intrahemispheric long and short, and heterotopic connections. The fNIRS measurement channels with different colors showed different regions of functional networks. (c) The relationship between FC variability $Q$ and MMSE scores in the significant intrahemispheric long connections located in (a). (d) The relationship between FC variability $Q$ and MMSE scores in the significant intrahemispheric long connections located in panel (b).

For the aMCI group, the significantly changed connections (using the threshold of $p<0.005$) tended to be relatively long distance and were concentrated in regions including the prefrontal and parietal cortexes [Fig. 3(a)]. We further considered the roles of these regional structures in the context of the functional network organization derived from the HC group. We found that the targeted connections in the aMCI group belong mainly to the default mode network, and the internetwork connections were mainly between the default mode network and the frontal–parietal network. For the AD group, a single connected network consisting of 178 edges linking widely distributed brain regions was altered [Fig. 3(b)]. The involved connections also belonged mainly to long connections and were concentrated in regions including the prefrontal, parietal, and visual cortexes. We further found that the targeted connections in the AD group belong mainly to the within default mode network, the frontal–parietal network, and the internetwork connections that linked the functional regions in the somatosensory, visual, and frontal–parietal networks. Notably, all the significant connections found in the aMCI group were also found in the AD group at network level.

All of the connections exhibited increased dynamic FC variability in the patients compared with the controls. No significant differences were found with respect to connected components between the patient groups of aMCI and AD patients. For the static FC, no significant differences were found between the patients (i.e., aMCI or AD) and HCs. For dynamic FC, the $Q$ were significantly negatively related to the MMSE scores in the intrahemispheric long connections [Figs. 3(c) and 3(d)], irrespective of the long connections from aMCI [Fig. 3(a)] or AD [Fig. 3(b)]. More severely, impaired patients tended to have increased $Q$ in the long intrahemispheric connections that were abnormal in patients.

3.4.

Sensitivity and Specificity of Dynamic FC Variability in Differentiating Patients from HCs

Figure 4 showed the sensitivity and specificity of the index $Q$ in differentiating patients from HCs. The measurement was the mean $Q$ across significant connections between aMCI and HCs [Figs. 4(a) and 4(c)] and between AD and HCs [Figs. 4(b) and 4(d)], respectively. Overall, the classification exhibited a satisfactory power in differentiating patients from HCs, no matter which measurement was used. Specifically, to differentiate aMCI patients from HCs, the classification performance was better as using the measurement from aMCI group [Fig. 4(a): $\text{sensitivity}=84\%$, $\text{specificity}=70\%$, and $\mathrm{AUC}=82.5\%$] than that from AD group [Fig. 4(b): $\text{sensitivity}=84\%$, $\text{specificity}=50\%$, and $\mathrm{AUC}=71.2\%$]. Likewise, to differentiate AD patients from HCs, the classification performance was better as using the measurement from AD group [Fig. 4(d): $\text{sensitivity}=82.6\%$, $\text{specificity}=76.7\%$, and $\mathrm{AUC}=86.4\%$] than that from aMCI group [Fig. 4(c): $\text{sensitivity}=60.9\%$, $\text{specificity}=90\%$, and $\mathrm{AUC}=78.8\%$]. These results demonstrate that the index $Q$ could be used as a biomarker for the diagnosis of patients with aMCI and AD.

Fig. 4

The patients-controls classification. (a) and (b) The ROC curves of average $Q$ for aMCI and HCs, in which the average $Q$ was calculated from the significant connections between aMCI and HCs (a), and between AD and HCs (b), respectively. (c) and (d) The ROC curves of average $Q$ for AD and HCs, in which the average $Q$ was calculated from those significant connections between aMCI and HCs (a), and between AD and HCs (d), respectively.

3.5.

FC States across Patients and HCs

Figure 5 shows the plot of cluster number versus ratio of between-cluster and within-cluster distance. Ten FC states were observed across multiple participants. We extracted four FC states that dynamically reoccurred over time and across participants in both patients and HCs [Fig. 6(a)]. The participants expressed an average of 2.13 (SD, 0.97), 2.08 (SD, 0.86), and 2.26 (SD, 0.69) states for HCs, aMCI, and AD, respectively. The number of states expressed had no significant difference between participant groups.

Fig. 5

The plot of cluster number versus ratio of between-cluster and within-cluster distance.

Fig. 6

(a) These four FC states were characterized over all the subjects. (b) The FC pattern in state 1 closely resembled the static FC with a correlation coefficient of $r=0.96$. (c) FC variance (standard deviation) within FC state 1 and state 2.

Of the four FC states that were observable across participant groups, state 1 accounts for $\sim 61\%$ of all windows, followed by state 2 for $\sim 23\%$, state 3 for 13%, and state 4 for 3% of all windows. The FC pattern in state 1 closely resembled the static FC [Fig. 6(b), correlation coefficient $r=0.96$], with strong positive correlations along the diagonal (i.e., within functional networks) [Fig. 6(a)], consistent with previous dynamic findings,⁴⁶ which represented a steady FC. For state 2, we found that the FC in the default mode network, frontal–parietal network, and visual network was increased, although it was decreased in the attention and somatosensory networks compared to state 1. The patterns in both the whole-brain FC and the within subnetwork connectivity for state 2 were very similar to those in state 1, but the variance of FC patterns within state 2 (which was quantified as the standard deviation across all FC patterns within each state) was larger than that within state 1 [Fig. 6(c)]. For states 3 and 4, the FC patterns were less observed and also differed markedly from these patterns of states 1 and 2. Notably, the patterns showed quite uniform connectivity between different measurement channels, which might reflect signals from global and noncortical components. As such, the states 3 and 4 are not analyzed or discussed in the following sections.

3.6.

Group Differences in Occurrence Frequency (F) of FC States

Analysis of group differences in $F$ values exhibited significant group effects (all $p<0.05$) for state 1 [$F(\mathrm{2,72})=3.581$, $p=0.033$] and state 2 [$F(\mathrm{2,72})=8.532$, $p<0.001$], respectively. Posthoc comparisons revealed significantly reduced $F$ values in state 1 and significantly increased $F$ values in state 2 for AD patients relative to both the controls and the aMCI patients (all $p<0.05$) (Fig. 7). However, no differences were found in the $F$ values between the aMCI and controls (all $p>0.05$) for these two states.

Fig. 7

Group differences of occurrence frequency $F$ in FC states 1 and 2 among HCs, aMCI, and AD.

3.7.

Relationship between State Occurrence Frequency ($F$) and Behavioral Performance

The occurrence frequencies $F$ for state 1 and state 2 showed different linear relationships with the clinical variables of the patients (Fig. 8). For state 1, the index $F$ showed a significant positive correlation with the MMSE ($r=0.40$, $p=0.0062$) as well as the AVLT_recognition ($r=0.37$, $p=0.015$) [Fig. 8(a)]. For state 2, the index $F$ showed a significant negative correlation with the MMSE ($r=0.39$, $p=0.0087$), MoCA ($r=0.3$, $p=0.05$), AVLT_delayed recall ($r=0.35$, $p=0.025$), and the AVLT_recognition ($r=0.31$, $p=0.046$) [Fig. 8(b)], respectively.

Fig. 8

The correlations between the occurrence frequency $F$ and clinical variables. (a) Plots showing the significant increases of the $F$ in state 1 with MMSE and AVLT_R scores. (b) Plots showing the significant decreases of the $F$ in state 2 with MMSE, MoCA, AVLT_D, and AVLT_R.