KEYWORDS
fNIRS, brain aging, aging, machine learning
ABBREVIATIONS AD, Alzheimer’s disease; bacc, balanced accuracy; CSF, cerebrospinal fluid; CV, cross-validation; DPF, differential path length factor; dPhaseTE, differential phase transfer entropy; EEG, electroencephalography; ESSC, envelope-signal spectra correlation; fMRI, functional magnetic resonance imaging; fAPF, fractional amplitude of physiological fluctuations; fNIRS, functional near-infrared spectroscopy; FC, functional connectivity; H2O, water, HbO, oxygenated-hemoglobin; HbR, deoxygenated-hemoglobin; HbT, total-hemoglobin; LED, light-emitting diodes; LOOCV, leave-one-out-cross-validation; MBBL, modified beer-lambert law; MCCV, Monte Carlo cross-validation; MREG, magnetic resonance encephalography; MRI, magnetic resonance imaging; MRMR, minimum redundancy maximum relevance; MCI, mild cognitive-impairment ; NDD, neurodegenerative disease; NIR, near-infrared; PFC, prefrontal cortex; PSD, power spectral density; SampEn, sample entropy; SD, source-detector; SVM, support vector machine; TBI: traumatic brain injury
1 | INTRODUCTION
The life expectancy of the global population is steadily continuing to increase [1]. As aging possess an increased risk of developing neurodegenerative diseases (NDD), such as Alzheimer’s disease (AD), the better understanding of healthy brain aging process is of importance. The assessment of the brain aging trajectory is of interest to detect of deviations from the healthy trajectory, and to monitor the efficacy of the treatments and interventions. The increased understanding of the brain aging process can lead to development of improved interventions, with an aim to delay age-related disease onset and to reduce their severity in the later years of life.
Brain aging has been widely studied, and multiple interconnected structural and functional changes have been reported. The changes can be observed in different scales of interest, such as micro or macro scale. Some notable changes are DNA damage, which is considered as one of the key hallmarks of aging, cerebrovascular changes observed as decreased vessel size, reduced number of capillaries, small infractions and microbleeds, causing overall decreased cerebral perfusion, changes in functional connectivity (FC), and regional brain atrophy, manifested in overall decrease of brain volume and weight. [2] In recent years, neuroimaging methods such as magnetic resonance imaging (MRI) and electroencephalography (EEG) have been utilized to measure the healthy brain aging trajectory. The methods commonly aim to quantify the biological age of the brain, which is assumed to be congruent with the chronological age of the healthy individuals, while being increased in the presence of aging related disease and decreased in successful aging [3]
Functional near-infrared spectroscopy (fNIRS) presents an intriguing method for the assessment of aging related-changes in the brain due to its good temporal resolution, affordability, and portable light-weight equipment. This enables development of brain health monitoring applications which can be used in a natural environment. The method is based on applying light source-detector (SD) pairs with minimum distance of 2.5 cm to the scalp [4], utilizing wavelengths in the optical window of approximately 650 nm – 1000 nm, where the near-infrared (NIR) light is able to propagate through the superficial biological layers of the skin, skull and cerebrospinal fluid (CSF), reaching the cortical brain layer [5]. Finally, the back scattered light is recorded by the detector pair. By applying modified beer-lambert law (MBBL), the method can be used to quantify chromophore relative concentrations [6]. The method requires selection of at least two wavelengths from the different sides of the isobestic point of the chromophores of interest. Commonly the concentration changes of oxygenated-hemoglobin (HbO) and deoxygenated-hemoglobin (HbR) are monitored, as the cortical changes are related to the brain metabolism caused by neural activity [5]. However, the changes in water (H2O) can be detected in NIR region as well by selecting wavelength accordingly. The measurement of neurohydrodynamics have been demonstrated recently in the assessment of glymphatic system by fNIRS [7].
fNIRS has been used in multiple aging studies showing evidence for the use of measurement method in the brain aging assessment. In FC studies fNIRS has been used successfully to confirm the finding of increased brain region co-operation, assumed to be due to brain’s adaptation to the structural changes. [8] Furthermore, Nguyen et al. found evidence of age-related decreased FC detected during verbal fluency task, although evidence of age-related changes during oddball and resting state was not found. [9]
Arterial stiffness has been associated with aging and is one of the leading risk factors for hypertension [10]. As younger subject’s arteries are more compliant, the mechanical stress to the brain caused by pulsatility is reduced in comparison to older adults with stiffer arteries, causing damage to the brain’s microvasculature. The arterial stiffness can be detected from the changed pulse shape form, and the distinguishing of younger and older adults has been shown to be successful by fNIRS by using pulse shape parameters. [11] Additionally, fNIRS has been used with MRI to show association between aging related cortical thinning and regional pulsatility. [12]
fNIRS has been used widely to study aging related changes in prefrontal cortex (PFC) activation, which is involved in the executive function. The results have shown age related differences in the PFC activation [13–16]. Furthermore, fNIRS has been used to characterize differences between younger and older adults, and mild cognitive-impairment (MCI) and AD groups [17,18], and to show age-related differences in within-session trainability of hemodynamic response. [19]
The results of the previous findings suggest potential for the use of fNIRS in the study of aging. To investigate this further, the use of single channel resting state fNIRS to detect aging caused differences in healthy adults is studied by utilizing machine learning (ML) methods in an age group classification task. In addition to traditional analysis of HbO, HbR and total hemoglobin (HbT) changes, the effects of relative H2O concentration changes and raw fNIRS signals are used. Multiple features based on the latest fNIRS, EEG and functional magnetic resonance imaging (fMRI) studies are applied. The results presented are based on the research conducted for the master’s thesis of Martti Ilvesmäki [20].
2 | EXPERIMENTAL
2.1 | fNIRS measurements and data acquisition
The data collection followed the guidelines established by the Declaration of Helsinki and the study was approved by the regional Ethical Committee of Northern Ostrobothnia Hospital District in Oulu University Hospital. All participants of the study signed informed consent letters before the measurements. The data was collected with hospital multi-modal MRI compatible frequency coded fNIRS utilizing high power light-emitting diodes (LED) coupled to optical fibres, with modifiable wavelength selection [21,22].
Total of 56 healthy controls participated in the study. The participants restrained from the use of alcohol 12 hours prior the measurements. With the exception of one individual, the participants were non-smoking. Data from single optode with 3 cm source detector (SD) separation distance, placed on the left side of the forehead corresponding to the Fp1 of the 10-20 system of the international federation for EEG electrode placement, was used with sampling frequency of 800 Hz. The subjects were in a supine position in resting state during the measurement, which lasted for approximately 5.5 – 10 minutes in total. After visual inspection, the data of six subjects was discarded, as the data of four subjects had gross movement artefacts and for two subjects the length of the measurement was insufficient. The final dataset consisted of 50 subjects with mean age of 42.3 ± 15.2 and age range of 23 - 67 (22F, 28M). The signal lengths were unified to the length 5.5 minutes by selecting signal of corresponding length from the start of the measurement. It is noteworthy, that the chronological and biological age of the brain may differ as pathologies, such as schizophrenia, traumatic brain injury (TBI) and AD, have been shown to affect the aging trajectory of the brain [23–25]. Thus, although the subjects in the study were considered as healthy controls and thus the chronological age is assumed to be similar to the subject’s brain age, the true brain age is not known. Therefore, a potential undiagnosed pathology could affect the results of the study.
The used wavelengths of 690 nm, 830 nm and 980 nm were selected due to the absorption being dominated by HbR, HbO and H2O respectively in the corresponding wavelength. Additionally, wavelength of 810 nm was used, which is close to the isobestic point of the hemoglobin.