2.1.

Array Design

Traditional NIRS devices have a sparse design in which a source and detector are placed $\sim 3\mathrm{cm}$ apart and the area underneath and between them is monitored. fNIRS assumes that the volume of optical sensitivity includes the brain and therefore that hemoglobin oxygenation changes measured are brain-activity related. For a typical, healthy adult brain, this is likely true. However, in dementia (particularly in AD), the brain can severely atrophy, manifesting as a loss of neurons and neuronal connections (see Fig. 2), meaning that the optical array may no longer be sensitive to absorption changes in the brain.

Fig. 2

Examples of cortical sensitivity of low-density [(a), (c)] and high-density [(b), (d)] array NIRS on a brain with severe atrophy due to AD [(a), (b)] and a healthy brain [(c), (d)].

To ascertain whether optodes placed on the scalp are sensitive to the cortex, we modeled the sensitivity of two different fNIRS arrays using two different head models: one for a healthy adult brain using the Colin27 MRI template²³ and one for an atrophied brain using MRI data collected from a patient with AD from the Multimodal Imaging in Lewy Body Disorders (MILOS) study (IRAS: 202332). Both MRI datasets were segmented using SPM12²⁴ to produce a five-layer head model, which was then converted to a tetrahedral volume mesh using iso2mesh.²⁵

To model the state-of-the-art in terms of cortical sampling, we modeled the sensitivity from an array based on the LUMO modular design (Gowerlabs Ltd, London, UK), which we term the high-density array. In this high-density array, 12 hexagonal modules (each module containing three sources and four detectors) were positioned on the anterior scalp overlying the frontal cortex/lobe, yielding a total of 36 sources and 48 detectors, with over 400 channels and a source-detector separation in the range 25 to 45 mm. For the low-density array, the center positions of each module were taken and assigned as either a source or a detector, forming an array consisting of six sources and six detectors, with the mean nearest neighbor source-detector separation being 30 mm for the dementia head model and 32 mm for the Colin27 head model—a sampling density typical of previous fNIRS research in dementia.

The high- and low-density arrays were registered to each head model. For each array-head model combination (four in total), TOAST++²⁶ was used to model near-infrared light propagation from sources to detectors, and the sensitivity to the cortex for each channel was summed to produce a sensitivity distribution.

The results of these simulations are shown in Fig. 2. In both the atrophied and healthy brains, there is a factor of 50 increase in cortical sensitivity using the high-density array compared with the low-density array. There is a similar spatial distribution of cortical sensitivity between the healthy and atrophied brain, particularly in the superior and middle frontal gyri, indicating that high-density arrays are capable of sampling the atrophied brain in this case. (Note that further work needs to be done to assess NIRS brain sensitivity with different types of atrophy in different dementia subtypes and stages and in comparison to age-matched healthy controls.)

However, in the atrophied brain, we see a mean decrease of 22.3% in the relative brain sensitivity of nearest-neighbor channels of the low-density array with respect to the healthy brain. This is comparable to the decrease in sensitivity in the high-density array (reduction in relative brain sensitivity for 25 to 40 mm channels in the high-density array is 20.4%). The decrease in relative brain sensitivity of the low-density array will lead to partial volume effects in which apparent differences in function may be due to changes in anatomy. This is particularly a problem for comparisons between subjects with dementia and healthy controls and for longitudinal studies in which progressive atrophy over time is expected to occur. Crucially, though, the aim of the high-density array is to provide a high level of sensitivity to the cortex, which is demonstrated in the atrophied brain in Fig. 2. This permits an image reconstruction approach to be taken to recover changes in cortical hemoglobin concentration while avoiding the partial volume issues that present substantial limitations for a channel-space analysis of activation in the atrophied brain.

These results highlight the importance of using high-density DOT (HD-DOT). HD-DOT is an imaging technique in which fNIRS data, collected using a high-density array, is combined with a model of light transport—produced using a structural prior of the subject’s head structure—to produce a three-dimensional image localizing hemoglobin concentration changes to the cortex. Overlapping channel measurements (i.e., channels that exhibit sensitivity profiles that partially sample the same volume) in HD-DOT arrays increase the spatial resolution,²⁷ improving the precision of functional mapping. A range of source–detector separations allows for depth discrimination, permitting a tomographic approach to study depth-dependent responses, which is important when there is an increased distance of the cortex from the scalp surface due to atrophy. Finally, the inclusion of short separation channels, which predominantly sample nonbrain tissue, enables contamination from scalp hemodynamics to be removed from the cortex-originating functional signal.

Another advantage of DOT is its use of anatomical head structures to model light transport. An anatomically-accurate structural prior increases the accuracy of resulting images,²⁸ producing images that are inherently registered to cortical anatomy and enabling improved image interpretation. There is a need for up-to-date, patient-specific data when studying patients whose brains are undergoing atrophy. As can be seen in Fig. 2, there are clear differences in gyrification between the atrophied brain and the brain of a healthy younger adult. Thus, the use of a healthy adult brain for patients with brain atrophy will lead to potential misinterpretations of where activation is localized, so patient-specific structural priors are imperative. Further, given the progressive nature of atrophy in dementia, brain structure will be altered over time, so structural priors derived from a recent MRI scan of a patient are needed. Though MRI scanning is needed to acquire structural data, it will only need to be performed once to allow for multiple HD-DOT scans, such as when performing continuous or longitudinal monitoring over a period (depending on the rate of atrophy).

In the future, it would be ideal to remove the requirement for subject-specific head models. This method is successful in neurodevelopmental NIRS studies of infants in which there are similar challenges in terms of a vulnerable population who are not easy to MRI scan, as well as neuroanatomical challenges that occur between longitudinal measurements. One approach to this would be to produce an atlas of the dementia brain at various stages of progression by averaging structural MRI data taken from many individuals with dementia. Alternatively, another approach is to have a database of head models of subjects with dementia at various stages of progression and with varying head shapes and sizes, and a best matching model can be found for a particular individual based on such characteristics. The challenge moving forward is to systematically determine how to employ a nonsubject-specific head model that minimizes the increase in localization error relative to using subject-specific anatomy.

2.2.

Headgear Design

Adaptations need to be made to NIRS device designs to improve accessibility for people with dementia. This is a population with reduced mobility, a high probability of having contraindications, and greater frailty. Traditional methods such as computerized tomography (CT) or MRI techniques only provide a snapshot of a patient’s status without accounting for the well-established fluctuations in symptomatology, particularly present in DLB.²⁹ As such, devices used to study dementia populations must be able to be worn for continuous monitoring to capture these fluctuations and provide dynamic and richer information of a patient’s vascular state. Devices therefore should also be wearable, robust to movement, portable for use in care homes or at the bedside, comfortable, and easy to use.

2.3.

Device Conclusion

The results of our optical modeling analysis highlight the importance of using HD-DOT for higher sensitivity, localization of anatomy, and high spatial resolution. Given that optical sensitivity decreases with depth, longitudinal comparisons of brain activity from the same patient with conventional fNIRS may suggest changes in function that are reflective of the patient’s changing anatomy rather than genuine functional changes. Further to this, if high-density or DOT systems are not used, it is imperative that the subject has a recent anatomical (CT or MRI) scan that can confirm that there is no significant atrophy in the region beneath the NIRS optodes. It is also important that a comfortable device is used to increase acceptance in a vulnerable population and improve tolerance for longer, more ecologically valid studies.

3.1.

Ecologically Valid Study Designs

As a major advantage of NIRS is its lower sensitivity to movement compared with other neuroimaging methods, dementia patients can be tested during more intensive or naturalistic tasks such as motor tasks; this includes any task that requires speech, as movement of the mouth causes issues for research with both MRI and electroencephalogram (EEG). This is particularly pertinent as many dementia subtypes present with motor deficits, such as DLB, Parkinsonian dementia,³⁵ and FTD.³⁶ NIRS systems can also be portable and therefore used to perform continuous monitoring in patients’ homes to assess the cognitive fluctuations associated with dementia.²⁹ Fiberless NIRS systems offer the ability to study cognitive activity during real-world, naturalistic tasks.³⁷ For example, dual-task walking paradigms are often used to investigate the effects of aging on prefrontal activity, e.g., Ref. 38. In the same vein, NIRS systems are highly compatible with virtual reality (VR) systems, enabling the exploration of more naturalistic environments and dynamic conditions.³⁹ NIRS studies integrated with VR have been performed to measure prospective memory, which has been shown to be impaired in mild AD patients who were asked to interact with a virtual, immersive town.⁴⁰ Immersive VR settings allow the user to manipulate the virtual environment freely, a feature that can be leveraged to design naturalistic tasks that would provide meaningful insights into memory loss.⁴¹ These integrated studies offer the potential to perform region-of-interest analyses on a broad range of multimodal data collected during multiple VR task designs, facilitating machine learning (ML) and holistic analysis methods on these augmented datasets.

3.2.

Functional Connectivity

Although a few studies have explored FC using NIRS (e.g., Refs. 42 and 43), these analyses were done with small channel numbers, low-density systems, or without subject-specific image reconstruction.⁴⁴ Cognitive decline in amnestic MCI (aMCI) and AD is typically addressed using static (spatiotemporally invariant) FC models, in which reduced connectivity is observed in aMCI/mild AD patients.⁴⁵ However, studies utilizing dynamic FC maps have found that the temporal variability of FC is discontinuous in aMCI and AD patients compared with healthy controls.⁴⁶ Subtler alterations in FC may be identified across symptomatology profiles and clinical subgroups, using seed-based approaches and highly detailed topographical maps of brain activation, enabled by HD-DOT.⁴⁷ Additionally, NIRS can easily be compared with other modalities and even used in conjunction with them, such as by combining EEG and NIRS¹⁹ or PET and NIRS.⁴⁸

3.3.

Study Design Conclusion

Current fNIRS studies have demonstrated functional dysfunction in dementia, but how this relates to clinically relevant outcomes is yet to be determined. By taking advantage of the wearability of NIRS, improvements in task design with more naturalistic or resting state experiments may allow for the recovery of clinically important biomarkers.

4.1.

Preprocessing

Signal preprocessing is a crucial step in removing noise and extracting useful hemodynamic information from the NIRS data. The vast majority of NIRS studies, including those on dementia, involve similar preprocessing steps. First, raw light intensity signals are converted into changes in HbO and HbR concentrations using the modified Beer–Lambert Law. Physiological sources of noise are commonly removed using a Butterworth bandpass filter with zero-phase filtering to account for phase distortion⁴⁹ and baseline shifts are often eliminated via detrending algorithms (e.g., Ref. 50). Motion correction methodologies vary greatly across NIRS studies in dementia. Most dementia studies exclude trials involving motion artifacts (e.g., Ref. 51), whereas others attempt to perform motion correction using wavelet-based motion artifact removal by decomposing the NIRS signals in the wavelet domain and extracting wavelet detail coefficients (e.g., Refs. 44 and 52).

4.2.

Traditional Statistical Analysis

Almost all NIRS studies in dementia employ traditional statistical analysis methods. Most commonly, tests of significance using simple statistics, such as $t$-tests, are used to identify differences in signal metrics across conditions. Activation refers to the increases in relative HbO concentration, and the significance of activation is often determined via per-channel $t$-tests across patient groups (e.g., Ref. 53). Paired $t$-tests have been used to compare group mean brain activation levels at different time steps, typically before and after a treatment course or intervention (e.g., Refs. 54 and 55). Analysis of variance (ANOVA) tests, both one-way (e.g., Ref. 56) and two-way (e.g., Refs. 57 and 58), are also often used for group-level comparisons of mean activation or tissue oxygenation index. Correlation analyses between behavioral data, such as the correct answer ratio within tasks, and the degree of brain activation are typically performed using Pearson’s correlation coefficient (e.g., Refs. 59 and 60). When multiple statistical tests are performed, a Bonferroni correction is typically applied to prevent family-wise errors (e.g., Ref. 61).

4.3.

Machine Learning and Multivariate Analysis

The wide range of neuroimaging modalities used to characterize dementia in the past decades, coupled with nonimaging clinical data from electronic medical records, has led to the generation of large-scale patient datasets.⁶² These large volumes of data can augment traditional methods used to characterize dementia progression through the introduction of ML analysis techniques. ML has a range of applications in dementia research, including predictive modeling of the relationship between input variables and clinical diagnoses, and pattern recognition within the data to study disease progression from MCI to AD.^63,64 A large portion of neuroimaging studies use ML to train a model that performs discriminative classification between different patient groups, the most common being classification between AD and healthy controls, but also including classification of MCI from AD and healthy controls, though with generally lower classification accuracies.⁶⁵

Although no dementia research has been performed using HD-DOT and ML, there is a clear opportunity for the high-density recording and localization afforded by HD-DOT to allow for multivariate analysis looking at spatiotemporal dynamics of cortical representations. HD-DOT offers higher spatial resolutions than traditional multichannel NIRS, which lends itself to increased image quality. This offers the possibility of applying convolutional neural networks (CNNs) directly on reconstructed brain images, rather than on statistical representations derived from NIRS channel data (e.g., $t$-maps).

Both traditional ML analyses and, to a lesser degree, deep learning analyses have been applied to NIRS imaging data to classify between different dementia stages. Finding discriminative NIRS features to classify different stages of AD (mild AD, moderate-severe AD) from healthy controls has been a recent problem, with accuracies around 60% using only NIRS features in multiclass classification, which is lower than with features extracted from other imaging modalities, such as EEG.¹⁹ In contrast, classification accuracies above 70% have been achieved for binary classification tasks using traditional linear discriminant analysis (LDA) and support vector machine (SVM) classifiers to distinguish between MCI and healthy controls, for example.⁶⁶

CNNs have been trained on spatial and temporal feature maps as input images from both MCI and healthy controls, yielding average classification accuracies of 80% using temporal features and higher classification accuracies using spatial and spatiotemporal features overall.⁶⁷ CNNs have also been trained using both t-maps and correlation maps, achieving classification accuracies of over 90% on binary tasks to classify MCI from healthy controls.⁶⁸ Recently, long short-term memory (LSTM) networks and combined CNN-LSTM models have demonstrated very high accuracies of around 85% on multiclass AD NIRS datasets, compared with a wide variety of traditional ML models, exemplifying the value of applying complex, multilayer models on NIRS data.⁶⁷ The predictive success of the models described above demonstrates the effectiveness of both signal and image biomarkers in the early detection of AD, though further work is required to identify discriminative features to perform diagnosis between less clearly delineated stages of both AD and other forms of dementia.

4.4.

Data Analysis Conclusion

As is the case for the wider NIRS field, preprocessing (and importantly, reporting of methods) needs to be standardized to ensure consistency across studies and pave the way for multicenter studies with large numbers of patients to produce impactful results. Automation and the use of ML tools, even from the pre-processing stage, may improve the interpretation and understanding of the increasingly complex data captured by NIRS devices.