1.

Introduction

Alzheimer’s disease (AD) can be characterized by dementia that usually begins with a slight failure of memory that slowly becomes more acute.¹ In fact, the International Working Group on AD’s proposal for early diagnosis of AD considers the evidence of significant and progressive episodic memory deficit as the core diagnostic criterion. Thus, the aim of researchers is to find efficient tests to detect early memory impairment to be used in a clinical environment. Among the variety of tests to define memory impairment, the free and cued selective reminding test (FCSRT)^2,3 is considered to be highly sensitive and specific in distinguishing AD patients from healthy controls⁴ and people affected by other forms of dementia.⁵ In addition, the test makes it possible to differentiate patients at an early stage of AD from mild cognitive impairment (MCI) patients’ nonconverters.⁶ During this test, 12 pictures of everyday life are shown to the patient, which they are asked to memorize. After this encoding phase, the participant has to recall all the figures previously shown. If he is not able to remember all the figures by himself, a cued recall phase follows, during which the doctor reminds the participant the semantic field of each figure not retrieved. The same procedure is repeated after 30 min. During this time, other tests are administered to the patient. Failures in recalling the pictures are considered highly significant for early memory impairment. However, despite the validity of the FCSRT, it could be very useful for the diagnosis of AD to overlap neuroimaging techniques with the administration of the test.

So far, working memory (WM) has been investigated using different neuroimaging techniques, such as functional magnetic resonance imaging (fMRI) and positron emission tomography (PET).^7,8 However, for the administration of the FCSRT, the doctor is required to sit next to the patient and interact with him. Therefore, it is not possible to perform the task in an fMRI scanner. Furthermore, performing this test lying down in a scanner would not be comfortable for the patient, and the result of the test could be invalidated. Moreover, this test is widely used for ambulatory patients, thus the employment of a portable technique is advantageous with respect to the fMRI.

Therefore, to preserve the interaction between the doctor and the patient, and the ecological conditions of the FCSRT, in this study we used functional near-infrared spectroscopy (fNIRS) as it is more suitable for this type of application.

fNIRS is a noninvasive hemodynamic-based neuroimaging technique that measures the relative changes in concentration of oxygenated (HbO) and deoxygenated hemoglobin (Hbb) secondary to neuronal activity. Given its better spatial resolution⁹ and lower sensitivity to motion artifacts compared to electroencephalography (EEG), and due to its portability and light weight,¹⁰ this technique is very suitable for ecological applications, such as a clinical interview.

fNIRS has already been proposed to study AD patients. For example, it was used to investigate visuospatial deficits¹¹ and different brain responses to a verbal fluency task (VFT) in healthy individuals and AD patients.^12–14 In particular, it was shown that a change in the cortex oxygenation could be due to an alteration of the regional brain functions or to deteriorated neurovascular coupling. In addition, fNIRS was also employed for monitoring the effect of cholinesterase inhibitors on AD patients during a VFT,¹⁵ demonstrating a variation in the oxygenation of the prefrontal cortex and the areas involved with speech processing. Furthermore, differences between healthy, MCI, and AD patients¹⁶ were investigated using fNIRS, making the potential relevance of this technique for the diagnosis of the AD undeniable.

Researchers argued that the complexity of a biological system output could change with age and disease.^17–20 Complexity can be defined as the difficulty of predicting a signal. Along the same line, we hypothesized a change in fNIRS signal complexity due to neurovascular coupling dysregulation related to the task and to the disease. Thus, we investigated how the complexity of an fNIRS signal changed between healthy adults and patients. In particular, we used the sample entropy (SampEn)²¹ and the multiscale entropy (MSE)²² as metrics for the complexity.

For instance, SampEn was used to study the nonlinear properties of heart rate (HR) time series^23,24 in neonates to evaluate the reduced variability of HR in the course of neonatal sepsis, and in adults, to investigate the obstructive sleep apnea syndrome.²⁵ In addition, it was used to evaluate the complexity of fMRI signal in patients affected by attention-deficit hyperactivity disorder²⁶ and in patients with schizophrenia during a social task.²⁷ With regards to AD, it was used to analyze background brain activity through magnetoencephalography (MEG)¹⁸ and EEG signals.^19,20

If we consider the fNIRS signal from an information viewpoint, it is a nonstationary and nonlinear time series and its entropy can thus be used to discriminate among different healthy brain states. In particular, in this study, a complexity analysis was carried out to distinguish early AD patients from healthy people. This problem has been tackled already by several researchers, but, to the best of our knowledge, this is the first time that this kind of analysis has been performed on fNIRS signals recorded during a WM test in completely ecological conditions. The possibility of using this complex analysis as an indicator of early AD could support the traditional clinical diagnosis of this dementia. Because of the dysregulation in neural responses in patients affected by AD, we expected the patients with AD to exhibit higher complexity than healthy controls.

2.

Materials and Methods

2.2.

Free and Cued Selective Reminding Test

This test was proposed by Buschke,³ and it was validated for AD diagnosis by Lemos et al.² This test is composed of an encoding phase, in which the participant has to memorize different figures, and an immediate and a delayed recall during which the participant has to name the figures previously seen. In the encoding phase, 12 selected stimulus figures are presented, 4 at a time on a sheet, 1 figure per quadrant. The subject is required to indicate and mention aloud each illustration after the experimenter verbally discloses the semantic cue of each of them. After this phase, three immediate recall trials follow. Every trial is preceded by 30 s of countdown to prevent recall from short-term memory. In this phase, the subject is asked to freely recall the items within 2 min [immediate free recall (IFR)]; next, the experimenter reminds the subject of the semantic cues of any items not freely retrieved [immediate cued recall (ICR)]. If the participant does not remember the figures with the cue, the experimenter reminds the subject of the images, providing him with both the items and the cues. The total recall is the sum of the free and cued recall. After 30 min, during which the subject has to perform nonverbal tasks, the same procedure of recall is carried out again [delayed free recall/delayed cued recall (DFR/DCR)]. The nonverbal tests²⁹ used to separate the immediate and the delayed recalls are: clock drawing test (CDT),³⁰ digit span test (DST),³¹ Corsi span test (CST),³² trial making test (TMT) (versions A and B),³³ and Benton test (BT).³⁴

During the administration of FCSRT, a rest period (RP) of 1 min between two different tasks was added to the original structure of the test to reset the brain activity. The experimental design is shown in Fig. 1.

Fig. 1

Experimental protocol: FCSRT with rest periods and filler tests added.

2.3.

Functional Near-Infrared Spectroscopy Measurement

To measure the cerebral hemodynamic, a commercial frequency-domain near-infrared spectroscopy instrument (Imagent, ISS Inc., Champaign, Illinois) was used. This instrument is made up of 32 laser diodes sources, 16 emitting light at 690 nm and 16 at 830 nm, and 4 photomultiplier tubes (PMT). The light is modulated at 110 MHz, while the PMTs are at 110.005 MHz. Time multiplexing was employed for the sources during the testing (switch mode: 16; sample frequency: 10 Hz). The light power emitted by the sources on the scalp of the participants is lower than $4\mathrm{mW}/{\mathrm{cm}}^2$, with respect to the American National Standards Institute standard limits and allowed for safe acquisition.

Nine sources for each wavelength and four detectors were used, resulting in 17 measurement channels in a multidistance configuration (3 to 4 cm). To ensure the comfort of the patients, a helmet with fixed holes for the fibers was used, positioned according to the international 10–20 electrodes placement system.³⁵ The brain areas involved in memory deficits, as well as in impaired abstract thinking and inhibition, are Brodmann areas (BA) 8, 9, and 47.³⁶ We tried to find the best compromise between covering the brain areas involved in this function and the fixed positions of the holes for the fibers. The best solution we found is presented in Fig. 2. The lowest detector is positioned in correspondence with the Fp point, according to the 10–20 system,^35,38 and the channels cover BAs 8, 9, 10, and 46.

Fig. 2

(a) Channels configuration plot using NIRS-SPM software³⁷ and (b) schematic optodes placement and channels distances.

3.

Results

During the RP and IFR, the values of MSE ($\tau =\mathrm{1,2},3$) do not show significant differences between healthy controls and patients for all of the channels. MSE ($\tau =1$) and MSE ($\tau =2$) during DFR show significant differences between the two groups for the channels reported in Table 1. After the FDR correction, only HbO Ch 15 and Hbb Ch 8 confirm statistical differences. Patients exhibit higher values of entropy than healthy controls, as shown in Fig. 3. MSE with scale $\tau =3$ does not show significant differences between patients and healthy groups for DFR. Using Brodmann’s Atlas, we found that channels that showed a significant difference between the two groups cover Broadmann areas 9 and 46. The areas of the brain for which there are distinctions between healthy controls and patients are shown in Fig. 4.

Table 1

Area under the ROC curves for the channels that show significant difference between healthy controls and AD patients.

Fig. 3

Boxplot of MSE ($\tau =2$) for (a) Ch 15 oxyhemoglobin during DFR and (b) Ch 8 deoxyhemoglobin during DFR.

Fig. 4

Channels that show a significant difference between the two groups through Wilcoxon–Mann–Whitney test ($p<0.05$) survived to FDR correction for (a) DFR, MSE ($\tau =2$), oxyhemoglobin and (b) DFR, MSE ($\tau =2$), deoxyhemoglobin.

Table 1 reports the area under the ROC curve for the channels that exhibit significant differences between the two groups, and Fig. 5 shows the ROC curves underlying the higher area under the curve (AUC). They were obtained for Ch 10 during the DFR phase for MSE ($\tau =2$) on the Hbb signal ($\text{cutoff}=0.3843$, $\text{sensitivity}=0.727$, $1\text{-}\text{specificity}=0.091$) and for Ch 18 during the DFR phase for MSE ($\tau =2$) on HbO signal ($\text{cutoff}=0.3159$, $\text{sensitivity}=0.909$, $1\text{-}\text{specificity}=0.182$).

Fig. 5

(a) ROC curve for MSE ($\tau =2$) of Ch 15 oxyhemoglobin during DFR and (b) ROC curve for MSE ($\tau =2$) of Ch 8 deoxyhemoglobin during DFR.

4.

Discussion

The aim of this study is to demonstrate the feasibility of a complexity analysis approach for fNIRS signals collected during ecological conditions to discriminate healthy controls and AD patients. In particular, according to the hypothesis that aging and disease lead to different complexity outputs of a system, we analyzed the complexity of HbO and Hbb signals using SampEn and MSE algorithms during the IFR and DFR phases of the FCSRT for both healthy controls and patients. This was done to discriminate between the two groups.

Typically, block or repeated event paradigms are used for fNIRS studies and general linear model or block averaging-based analyses are carried out. Unfortunately, because of the ecological paradigm, this kind of analysis was not feasible in this study. In fact, the timing of the test is not fixed, but is subject dependent. In the FCSRT, only the maximum time of the free recalling phase is fixed, and its duration depends on the performance of the subjects. Moreover, the subject can name the pictures whenever he feels ready during the available time, without a predetermined timing. It is thus impossible to create a proper design matrix and properly model the expected functional response. Therefore, it is necessary to investigate the possibility of using a different approach for signal analysis, such as a complexity investigation.

The main finding of the present study is that significant differences for MSE ($\tau =1$) and MSE ($\tau =2$) between AD patients and healthy controls performing the FCSRT, during the DFR phase, were observed in the salient region involved in this kind of task, as shown in Table 1. After the FDR correction, only MSE ($\tau =2$) gives some significant results. More precisely, patients with AD exhibit higher values of entropy than controls (Fig. 4). This analysis can discriminate BAs 9 and 46 (Fig. 4), regions involved in WM tasks.³⁶ This result confirms the hypothesis that brain signals from AD constitute a dysregulation of neurological patterns and that there is an increase in complexity, as demonstrated by Vaillancourt and Newell.¹⁷ MSE with $\tau =3$ does not seem to be significant for this kind of analysis. This is probably caused by the limited number of samples or by the specific dysregulation of these brain areas, which have a temporal structure that requires shorter temporal scales.

Complexity analysis of brain signals for patients with AD has been already carried out by several authors. Gómez et al.¹⁸ found that MEG signals from patients are less complex than controls during resting state. Deng et al.⁴⁶ have the same findings on EEG signals. According to the literature, an increase in the complexity of fNIRS signal is observed during the performance of a mental arithmetic task⁴⁷ and during a WM task⁴⁸ with respect to the RP for healthy people. Conversely, we did not find significant differences between the two groups for SampEn and MSE during the rest phase. This might be due to the limited length of this phase and, therefore, the limited number of samples. On the contrary, the higher complexity of patients with respect to healthy people observed during DFR could be a result of a dysfunction in the neurovascular coupling in the frontal area due to the task and to the activation of these brain regions. In fact, cerebral blood flow (CBF) depends on the brain activity, and it is regulated by the coordinate action of neurons, glia, and vascular cells. When neurons are activated, various vascular changes are elicited and lead to CBF increases in the activated area.⁴⁹ In AD patients, there is deposition of amyloid $\beta$-peptide in neuropil and vessels, so neurovascular coupling is disrupted and CBF is not appropriate to support the metabolic demand of the tissue.⁴⁹

It is also possible that the increased complexity observed in our study is caused by a degradation of functional neural connectivity. For example, Pijnenburg et al.⁵⁰ found a loss of connectivity during a WM task using MEG signals on AD patients. Stam et al.⁵¹ observed a loss of small-world network characteristics in AD patients using EEG, and Supekar et al.⁵² had the same findings in fMRI.

We found similar results for both HbO and Hbb. This finding could be considered relevant because these two signals are physiologically anticorrelated, hence, their predictabilities should be almost the same. In fact, we would suppose to find similar values of entropy for both the time series.

The ROC analyses show that the MSE ($\tau =2$) evaluated on Ch 8 and 15 can discriminate AD patients from healthy controls because of its higher area under ROC curve (Table 1). They are placed over BA 46 and BA 9, respectively.

Even though our results are preliminary, this study demonstrates that it is possible to use the SampEn and MSE to discriminate efficaciously healthy people from AD patients during ecological protocols. In particular, to the best of our knowledge, it is the first time that the fNIRS signal was collected during the administration of the FCSRT in a clinical condition, giving a physiological correlation to the diagnosis of AD obtained by this test. Due to fNIRS’ portability and it being noninvasive, this technique could become an important support for the diagnosis of AD.

One limitation of our study is the limited sample size and the limited length of the fNIRS signal. However, it is due to the test’s structure and to its ecological features. In fact, probably because of the restricted number of samples, we were not able to find differences during the RP.⁵³ In our future studies, we will enlarge the sample size of the rest phase to obtain more useful information to better discriminate AD patients from healthy controls.

We will also combine fNIRS with other neuroimaging techniques to investigate which one would be more effective in assessing regional differences in the brain in terms of sensitivity and specificity. Furthermore, we plan to investigate the complexity of the HbO and Hbb signals by testing different parameters, in particular regarding the similarity factor, for the calculation of the entropy, in order to find the best parameters settings for HbO and Hbb. In addition, it could be of great interest to study the systemic contribution to the complexity of the fNIRS signal, measuring some physiological signals (i.e., breathing rate and HR) and placing short separation channels.

5.

Conclusions

To the best of our knowledge, this is the first study investigating the complexity of brain fNIRS signals in patients with AD during the administration of FCSRT, in completely ecological conditions.

The results show that MSE with scale $\tau =2$ is appropriate to investigate brain signal complexity for this kind of task. AD patients show higher values of entropy than healthy controls during DFR phase in BAs 9 and 46. These results are consistent with the hypothesis that AD may be characterized by a dysregulation of complex functional networks and support the hypothesis of a complexity variation with age or disease of biological systems.