Quantitative assessment of resting state for mild cognitive impairment detection: A functional near-infrared spectroscopy and deep learning approach

Background : Mild cognitive impairment (MCI) is considered a prodromal stage of Alzheimer’s disease, which is the sixth leading cause of death in the United State. Early diagnosis of MCI can allow for treatment to improve cognitive function and reduce modifiable risk factors. Currently, the combination of machine learning and neuroimaging plays a role in identifying and understanding neuropathological diseases. However, some challenges still remain, and these limitations need to be optimized for clinical MCI diagnosis. Methods: In this study, for stable identification with functional near-infrared spectroscopy (fNIRS) using the minimum resting-state time, nine different measurement durations (i.e., 30, 60, 90, 120, 150, 180, 210, 240, and 270 s) were evaluated based on 30 s intervals using a traditional machine learning approach and graph theory analysis. The machine learning methods were trained using temporal features of the resting-state fNIRS signal and included linear discriminant analysis (LDA), support vector machine, and K-nearest neighbor (KNN) algorithms. To enhance the diagnostic accuracy, feature representation-and classification-based transfer learning (TL) methods were used to detect MCI from the healthy controls through the input of connectivity maps with 30 and 90 s durations. Results: As in the results of the traditional machine learning and graph theory analysis, there was no significant difference among the different time


Background
Alzheimer's disease (AD) is the most common type of dementia, and it gradually but certainly influences the patient's memory and cognitive, mental, and language abilities [1]. In the late stage of AD (i.e., the severe stage), the symptoms have an increasing impact on the patient's motor and physical abilities, requiring around-the-clock care [2].
Typically, people with AD survive an average of four to eight years after diagnosis; however, some AD patients can live as long as 20 years. AD and related neurodegenerative diseases are arguably considered the most dreaded maladies of the aged [3]. None of the available pharmacologic or non-pharmacologic therapies are able to slow or stop the destruction of neurons caused by AD symptoms [2]. Mild cognitive impairment (MCI) is a transitional state between healthy aging and AD [4]. Researchers believe that treatment in the early stages (i.e., MCI) of the AD continuum may be effective for preventing the progression of AD and sustaining brain function [2]. Therefore, timely diagnosis of MCI from the healthy control (HC) presents an opportunity for interventions to improve cognitive function and reduce the modifiable risk factors implicated in AD progression [5].
Currently, the principal MCI diagnosis tools available to clinical doctors for making professional judgments rely on information obtained from the patient's medical history, mental status examination, imaging studies, and blood tests [6]. In comparison to subjective characteristics (i.e., the medical history and mental status examination), biomarker tests such as brain imaging have the potential to provide a quantitative clinical determination for suspected MCI. In particular, when assisted by computer science and mathematics, the use of brain imaging has provided the ability to understand the neural destruction caused by MCI in vivo [7].
There are two types of brain imaging techniques: (1) hemodynamic-metabolic types such as single-photon emission tomography (SPECT), positron emission tomography (PET), functional magnetic resonance imaging (fMRI), and functional near-infrared spectroscopy (fNIRS); and (2) electric-magnetic types, which include electroencephalography (EEG) and magnetoencephalography (MEG) [8]. PET and SPECT are based on the principle of radioactive isotopes. This characteristic has restricted their application in children and pregnant women [9]. Generally, brain imaging using electric-magnetic techniques provides high temporal resolution, but these methods are susceptible to motor artifacts and environmental interference [10]. Moreover, some of these techniques (e.g., EEG) also lack sufficient spatial resolution (>1 cm) [11]. In recent years, the use of fMRI has achieved valuable advancements in understanding neurological diseases [12]. However, the physically constrained participants must also be exposed to a loud noise environment [13]. Additionally, owing to the effects of the electric and magnetic fields, these brain imaging modalities cannot be applied simultaneously with brain stimulation (e.g., transcranial electrical stimulation) for further rehabilitation treatment [14]. In essence, fNIRS is an alternative to conventional hemodynamic response-based neuroimaging techniques that promises to shed additional light on functional brain activity in environments previously inaccessible to fMRI [15].
Owing to the promising advancement of fNIRS, several studies have investigated the competence of fNIRS for the detection of hemodynamic changes in AD and MCI.
Through studies over the past decade, researchers have demonstrated the feasibility of using fNIRS to identify reduced cerebral oxygenation either in the resting state or the task period, such as during word retrieval, memory tasks, motor tasks, and visuospatial perception [28]. In the resting-state cases, altered connectivity [29][30][31] and fluctuations [32][33][34][35][36] of cerebral oxygenation have been reported in MCI and AD groups. This difference is associated with the dysregulation of information integration in the patient's brain. In the task period cases, oxygenation hypoactivation has been found in the frontotemporal area of the MCI group during verbal fluency tasks [37][38][39]. Additionally, a reduced hemodynamic response has been reported in the prefrontal cortex of the MCI group when performing memory tasks [40][41][42][43][44]. Similarly, examination results during motor [45,46] and visuospatial [47,48] tasks also showed a difference between the dementia and healthy groups. This reveals that executive dysfunction and abnormal visuospatial perception may be underlying in the patient's brain. These findings, as mentioned above, are consistent with fMRI results indicating a disrupted functional network and decreased hemodynamic response in MCI pathology. This empirical evidence also indicates that fNIRS-based indicators could play a role in the identification of MCI from the HC. Therefore, it is essential to develop and examine the practicability of fNIRS-based biomarkers for the diagnosis of MCI, as these could assist the clinical doctor in making confident decisions.
In our initial studies [49,50], we assessed the use of fNIRS-based biomarkers to detect MCI from the HC when participants were performing different mental tasks, such as the N-back task, Stroop task, and verbal fluency task. Seven digital biomarkers were extracted from the fluctuating time series of ΔHbO and ΔHbR: the mean, slope, peak value, kurtosis, and skewness. The highest accuracy was 76.67%, classified by a linear discriminant analysis (LDA) based on the N-back task and Stroop task. The imaging biomarkers were evaluated using a convolutional neural network (CNN), which included a t-map, mean map, slope map, kurtosis map, skewness map, HbO map, and connectivity map. The highest accuracy was 98.61% for the slope map case during the N-back task.
However, after our initial investigations, some challenges and limitations still needed to be addressed: (1) small datasets may lead to overfitting and local minima during the training of the deep learning model, (2) some of the patients may be limited or dislike performing several mental tasks for the examination, and (3) the lengthy procedure of the cognitive task may cause fatigue problems for the patients.
Therefore, based on the limitations mentioned above, we further quantitatively investigated the possibility of using the features extracted from resting-state fNIRS data over a short period to detect MCI from the HC with a transfer learning method. Five minutes of resting-state data were acquired from 24 subjects. The digital biomarkers were extracted to examine the performance of the different resting-state periods. LDA, support vector machine (SVM), and k-nearest neighbor (KNN) algorithms were used to classify the digital biomarkers. The connectivity map was analyzed as the input to the transfer learning method. We hypothesized that: (1) the transfer learning method could fine tune the model trained with the small MCI dataset and obtain excellent performance, (2) the connectivity map of the resting state would be a useful biomarker for MCI detection, and (3) using a shorter time window of the resting state could achieve good classification accuracy. To the best of the authors' knowledge, this is the first quantitative fNIRS study to examine the resting state for MCI detection using deep learning methods. Before the fNIRS measurements, the mental health of the subjects was evaluated using the Korean mini-mental state examination (K-MMSE) [51], Seoul Neuropsychological Screening Battery [52], and magnetic resonance imaging (MRI). The Pusan National University Institutional Review Board approved the experimental protocol. Before the experiment, all subjects had the entire content of the experiment explained to them and provided informed consent. Figure 1 shows the system flowchart of the experimental process.

Participants and system protocol
[ Figure 1]

fNIRS measurements
A multi-channel NIRSIT continuous wave system (OBELAB Inc., Rep. of Korea) with 24 emitters and 32 detectors (Figure 2a) was employed to measure the fNIRS signals with a sampling rate of 8.138 Hz. The wavelengths used for the detection of ΔHbO and ΔHbR were 780 and 850 nm, respectively. The inter-optodes distance was set at 15 mm. In total, 48 channels were distributed equidistantly on the prefrontal cortex. The channel configuration is illustrated in Figure 2b, which was set to be consistent with the international 10-20 EEG system with reference electrode location FpZ. Each channel was defined as an emitter-detector pair with a distance of 30 mm. The experiment was conducted in a confined room to avoid disturbance. The participants were seated on a comfortable chair and asked to rest without unnecessary movement. The duration of the fNIRS resting-state measurement was approximately 5 min for each participant.

Resting-state preprocessing
In this study, ΔHbO and ΔHbR throughout the whole experiment were obtained from the raw optical density using the modified Bear-Lambert Law [53,54]. A fourth-order Butterworth low-and high-pass filter with a cut-off frequency of 0.0018-0.15 Hz was applied to remove cardiac noise (~1.1 Hz), respiration (0.25 Hz), and other physiological noise [55][56][57]. In addition, a detrending algorithm was used to remove shifts in the baseline. However, the slowly varying U-shaped noise (< 0.1 Hz) associated with deep breath or slow motion and other global unknown noise (< 0.1 Hz) could not be removed by the bandpass filer or low-/high-pass filter [58]. Therefore, after passing the signal through the filter, we also manually checked and removed the noise channel with the obvious abnormal fluctuation by the experience. All of the preprocessing procedures were analyzed offline using MATLAB TM software (MathWorks, Natick, Massachusetts).

Temporal features and functional connectivity
To identify MCI and examine the minimum required duration of the resting-state data, we selected six temporal features to conduct the classification: the mean of ΔHbO, mean of ΔHbR, standard deviation of ΔHbO, standard deviation ΔHbR, variance of ΔHbO, and variance of ΔHbR. Each feature was calculated for nine different time intervals (i.e., 0-30 s, 0-60 s, 0-90 s, 0-120 s, 0-150 s, 0-180 s, 0-210 s, 0-240 s, and 0-270 s).
The functional connectivity of the brain signals reveals the interaction among different brain regions, which shows the temporal correlations between spatially distant neural activity [59]. Pearson's correlation coefficients (r) were used to analyze the correlation between the temporal signals of each channel. The connectivity matrix consists of the underlying interhemispheric relationships in the prefrontal cortex. The connectivity matrices used in this study were divided into two types: (1) calculations for graph theory analysis, and (2) imaging biomarkers for the transfer learning method. In the first case

Graph theory
Generally, graph theory has been employed to characterize the network communication ability. The nodes of the network refer to the channels, and the correlation coefficient is defined as the edge for each channel pair. Typically, the calculated network parameters (i.e., global efficiency, local efficiency, and small-worldness) have been used to describe changes in brain network architecture, disease development, and brain evolution [60][61][62].
To quantify the network characteristics and assess the minimum necessary measurement duration, the global efficiency, local efficiency, and small-worldness were analyzed with an increased sparsity threshold range (0.5-0.9) with an interval of 0.1 to evaluate the different periods in the resting-state functional connectivity. The threshold was applied to remove spurious connections, defined as when the edge is lower than the threshold. In essence, the global efficiency (Eglob) measures the integration of the network, which indicates the ability to transfer information over the entire measured brain network (G).
The global efficiency is defined in Equation 1. The local efficiency (Eloc) of the network represents the communication ability in a local region. As given in Equation 2, it can be calculated based on the global efficiency. The small-worldness (Sw) is an ensemble of the measured networks that refers to the high-transferability network using the shortest path distance. Practically, it can be measured based on a comparison (shown in Equation 3) between the average cluster coefficient (K) and path length (Li) for a random network.
where N refers to the nodes of the network, and d (i, j) represents the length of the shortest path between two random nodes, i and j. The direct neighbor of the i th node produces the local subgraph network (Gi). The path length and clustering efficiency of the random network are given by Lrand and Crand, respectively.

Classification algorithm
The vital objective of the classification algorithm is to describe and predict unknown terms through training with the existing dataset. LDA, SVM, and KNN are the most widely used machine learning methods for brain disease identification [63]. The principle of LDA is the generation of hyper-planes to discriminate data from different classes.
Similarly, the main function of the SVM is to construct a margin (i.e., support vectors) with a maximum boundary to generate a linear extrication hyper-plane. In contrast, KNN is a type of instance-based learning, in which the vote of its neighbors categorizes an object. In this study, we selected six temporal features (mean of ΔHbO, mean of ΔHbR, [ Figure 3] With the development of deep learning and neural imaging techniques, computer-assisted clinical diagnosis methods have achieved massive progress. In particular, CNNs have realized huge successes in brain disease identification. However, many challenges remain, such as the cost of neural imaging examinations, limited datasets for rare diseases, and long procedures for clinical diagnosis [64]. It is difficult to accumulate a sufficient dataset to train CNN models. To avoid overfitting and the lower generalization power of a small dataset, transfer learning is a good solution [65]. In concept, transfer learning is used as a pre-trained model for fine tuning the new model to extract features or classify unknown objects [66], as shown in Figure 3. In this study, we employed two different transfer learning strategies: feature representation-based transfer learning (FRTL) and classification-based transfer learning (CTL). In the FRTL case (i.e., Figure 3a), the pre-   a maximum of 30). The mean of the K-MMSE scores in the MCI group is higher than 24. Interestingly, the statistical analysis results also revealed a non-significant difference (p-value = 0.49 > 0.05) between the two groups.

Demographics and clinical score
[ Table 1]

Temporal feature classification results
Rather than using a statistical analysis to identify the differences between the two groups, individual identification using the machine learning method may offer promising advantages for clinical diagnosis. Figure 4 shows [ Figure 4]

Functional connectivity map
The connectivity maps are generated by the functional connectivity matrix, which demonstrates the temporal relationship among the 48 channels. Figure 5  [ Figure 5]

Graph theory analysis
In this study, graph theory was employed to further evaluate the functional connectivity network and quantify the difference between different measurement durations. Figure 6 shows the global network efficiency, local network efficiency, and small-worldness for  Table 3, the mean accuracy of the classification based on ΔHbR (i.e., 78.38%) was lower than that based on ΔHbO, and it ranged from 69.65% (with Alexnet) to 85.92% (with Densenet 201). For the 90 s time window, the accuracies of the classifications using both ΔHbO (i.e., 74.36%) and ΔHbR (i.e., 73.75%) were lower than those for the 30 s time window, as listed in Tables 4 and 5. [

Discussion
In this study, our goal was to evaluate the minimum resting-state fNIRS signal and investigate the possibility of using transfer learning to detect MCI from the HC based on the resting-state fNIRS. To the best of the authors' knowledge, this is the first study to examine the minimum resting-state duration required for MCI detection using fNIRS signals. In addition, we believe it is the first study to apply transfer learning for individual classification of MCI based on the resting-state fNIRS signal. The cost of the neural imaging technique, difficulty of the signal acquisition for patients, and limited data available for rare diseases have always been major challenges for the application of computer-assisted brain disease diagnosis. Our system provides a solution for the comfortable and accurate detection of individual MCI from the HC, which is obtained using the transfer learning classification method with an input of the minimum restingstate fNIRS data.
MCI and dementia are multifaceted diseases, which cause accumulated pathological brain injury leading to decline in progressive motor, cognitive, and language abilities. The wellestablished biomarkers are neurodegeneration, amyloid plaques, and neurofibrillary tangles [67]. With the development of non-invasive neuroimaging techniques, the use of imaging biomarkers has emerged owing to their promising advantages of ease of signal measurement, safety, non-invasive nature, and low cost. In comparison to fMRI, EEG, PET, and SPECT, fNIRS as a novel non-invasive neural imaging technique is more environmentally unconstrained and has moderate temporal and spatial resolution. Thus, it is a promising tool for assessing biomarkers for brain disease diagnosis.
Empirical fNIRS studies have investigated the differences between MCI and the HC [28].
Some of the literature [41,44] has stated that less hemodynamic response appears in the frontal and parietal regions of the MCI and AD groups in comparison to the HC group when performing working memory tasks. Similarly, in the resting state [30,61], and during other cognitive tasks [38], the hypoactivation measured by fNIRS has also been reported in statistical analyses at the group level. As such, the reduced hemodynamic response and functional connectivity were also obtained in our study, and the MCI group exhibited lower functional connectivity than the HC group ( Figure 5). Similarly, this finding is consistent with the literature on fMRI [67] and EEG [68]. These experiential studies provide interpretability of the dysregulation and neurodegeneration of MCI, and also verify the feasibility of using fNIRS as a biomarker for MCI detection.
Mass-univariate analysis (i.e., statistical parametric mapping) plays a vital role in the determination of abnormal hemodynamic responses and the neurofunctional difference between the patient and control groups [69]. However, for individual MCI diagnosis, the statistical analysis cannot be easily applied [70]. Similarly, in our initial evaluation study [50], the statistical analysis outcomes showed inconsistencies with the results of the individual classifications. Therefore, it is essential to validate the biomarker results based on individual classifications before the established biomarker can be applied as a clinical assistance tool.
In our previous study [49,50], digital biomarkers (i.e., mean, slope, peak, kurtosis, and skewness) and imaging biomarkers (i.e., mean map, slope map, HbO map, skewness map, kurtosis map, t-map, and connectivity map) during mental tasks (i.e., the N-back task, Stroop task, and VFT) were evaluated for MCI detection using LDA and CNN. The highest accuracies for the digital and imaging biomarkers were 76.67% (with LDA) and 98.61% (with CNN), respectively. The biomarkers in our initial work are comparable to the established biomarkers, such as PET, SPECT, and cerebrospinal fluid [71]. In addition, a novel study [31] reported that using a K-mean clustering analysis of the functional connectivity index could detect MCI with a sensitivity of 84% and specificity of 70%. In comparison to those studies, the current study used a resting-state signal with the minimum time window, in which the most salient feature is the convenience of data acquisition for MCI patients without performing any tasks. In addition, the transfer learning method offers the opportunity for using small datasets, which may occur owing to the limitation of cost or patient numbers, to make decisions based on deep learning methods.
In fNIRS, which is a novel and promising technique, the necessary resting-state acquisition duration for accurate and stable MCI detection remains largely unknown. In the fMRI case, the minimum scanning time should be longer than 5 min to achieve stable functional connectivity. One study on a child via the fNIRS resting state [72] demonstrated that 2.5 min may be the minimum time for accurate measurement. In our study, this is consistent with the results of the functional connectivity observations (i.e.,  Figures 4 and 6). In particular, the classification results demonstrate that a measurement duration of 0.5 min could also achieve satisfactory results. This is why 30 s and 90 s time windows were selected to divide the resting-state data for the transfer learning classification.
To avoid overfitting and local minima during the classification due to the small dataset, transfer learning was used to solve this problem for the traditional deep learning. In the FRTL case, we employed seven pre-trained CNN models to extract the features. Each CNN models exhibited no significant difference in this study. In addition, the performance of the 80% training set was generally higher than that of the 70% training set. One possible reason is that some of the datasets were not sufficient to fine tune the pre-trained CNN model. However, the 90 s time window could obtain a more stable performance than the 30 s window during the analysis using traditional machine learning methods and the graph theory analysis. Both transfer learning results classified using the 90 s duration datasets showed a lower accuracy than those classified using the 30 s duration datasets. This may be due to the imbalance in the input data.

Limitation
Although the CTL provided satisfactory diagnosis accuracy (maximum accuracy: 97.01%), there are some limitations to this technique, and future directions must be discussed. First, owing to the insufficient resting-state data, the clipped sections with durations of 30 s and 90 s were unbalanced. Increasing and balancing the input numbers may further improve the classification accuracy. Second, owing to the long time required for the CTL procedure and the similar performances obtained with the FRTL, we only selected three pre-trained CNN models to differentiate the MCI from the HC group using CTL. Therefore, a future study could examine the rest of the models with the assistance of a supercomputer to check whether the higher performance may achieve or not. Finally, the prefrontal cortex has the benefit of no overlying hair, which can reduce the scattering and attenuation effects. This study only examines the resting state in the prefrontal cortex.
However, the default mode network may have a high possibility of identifying biomarkers useful for MCI detection.

Conclusion
In this study, we assessed the minimum resting-state duration required to obtain an accurate functional connectivity map. Moreover, a novel method to identify MCI from the HC using transfer learning methods was proposed.

Acknowledgement
We would like to special thank all participants for taking part in this study and the doctors of Pusan National University Hospital for helping us with recruitment of participants.

Competing interests
The authors declare that they have no competing interests.

Consent for publication
Not applicable

Ethics approval and consent to participate
This study was reviewed and approved by the Ethics Committee of the Pusan National University Hospital (H-1809-0190971) and all procedures involved were conducted in accordance with the latest version of the Declaration of Helsinki. Prior to experiments, written informed consent after receiving detailed information about the experiment procedure was obtained from the subjects.

Data availability statement
The datasets generated for this study are available from the corresponding author on reasonable request.