Study Design and Participants
The research was approved by the Human Research Protection Office at Washington University School of Medicine. This study was a prospective observational cohort study using a convenience sample. Three different groups were recruited for the study: 1) stroke patients, 2) older healthy, neurologically normal subjects to be used as controls, and 3) young healthy subjects to be used as a reference population.
We enrolled ischemic stroke patients who met the following enrollment criteria. Inclusion criteria included: 1) age greater than 18 years, 2) anterior circulation ischemic stroke with onset within 72 hours, and 3) NIH Stroke Scale (NIHSS) greater than or equal to 1 at time of initial evaluation in ER. Exclusion criteria included: 1) symptoms suggestive of a small sub-cortical stroke, 2) bilateral strokes, 3) prior stroke, 4) pregnancy, and 5) enrollment in an experimental therapeutic trial. Patients who met the inclusion and exclusion criteria were approached for consent. If the patient had difficulty with language production or comprehension due to the stroke, consent was obtained from the patient’s legally authorized representative. Of those enrolled, 25 were excluded from analyses: one due to hair incompatibility, seven due to discomfort, five due to excessive movement, seven due to physician or family visits, two due to poor data quality caused by snoring, and three with less than 15 minutes of acquired data. We used a cutoff of 15 minutes of clean resting state data to optimize fidelity of the fc-maps at the level of the individual participant 30, 31. Of the remaining patients (n = 13 with four females, mean age = 73.4 (± 13.5) years), four had right-sided MCA strokes.
We recruited two control groups with a sole exclusion criterion of diagnosed neurological or psychological conditions. Nine young healthy volunteers were enrolled through an inclusion criteria of age between 18 and 30 years. Of these nine subjects, seven are used in this study to generate an average young healthy reference baseline with 2 being excluded due to motion artifacts limiting the total usable data to less than 15 minutes (n = 7, mean age = 25.9 (± 1.7) years. Similarly, 11 older normal control subjects were recruited with the inclusion of age greater than 55 years (n = 11, mean age = 61.1 (± 5.6) years).
To measure stroke severity, a National Institute of Health Stroke Scale (NIHSS) score evaluation was performed at the time of the HD-DOT scan. The NIHSS is a quantitative bedside assessment of neurological deficits, scoring disruptions in several different domains, including consciousness, vision, language, neglect, motor (extraocular, upper extremity, and lower extremity) and sensory 32. These domains have been correlated with fMRI fc within the default mode network, the visual network, the fronto-parietal network, the dorsal attention network, and the somatomotor network respectively 10, 13, 14, 33, 34. Some patients underwent a structural head CT or brain MRI as part of their standard clinical care (see below). The bedside HD-DOT assessment did not interrupt standard clinical care.
Clinical HD-DOT System Console
The clinical HD-DOT console was designed to support high-fidelity measurement of light levels at the bedside in a small form-factor that would present no potential disruption to standard clinical care in the intensive care unit. The entire console of the system fit on a custom-built hospital-compliant console (Minitec, New York) measuring 4 feet in height, 3 feet in length, and 2 feet in width. This small footprint allowed for transport through hospital rooms for data collection without disrupting clinical care. The clinical HD-DOT console supported beside measurements arising from 48 source and 34 detector positions placed bilaterally in two rectangular grids on the participant’s head (Fig. 1). The continuous wave HD-DOT system contained LED sources, illuminating with near-infrared (NIR) wavelengths 750 nm and 850 nm (750-03AU and OPE5T85, Roithner Lasertechnik), and avalanche photodiode (APD) detectors (Hamamatsu C5460-01). Source boxes (n = 6) regulated power and delivered source encoding flashing patterns to 32 LEDs each (three 750 nm and one 850 nm LED per source position). The NIR light from the LEDs was directly coupled into 4 − 1 optical fiber bundles (CeramOptec, silicone cladded, 2.5 mm diameter bundles of 50 µm fiber, numerical aperture 0.66) via SMA connectors. Light collected from the head was transmitted to detector boxes (n = 6, each housing up to six APD detectors) using similar 1–1 optical fibers and SMA connectors. The APDs were digitized by dedicated 15 24-bit analog-to-digital converters at 96 kHz (M32AD, RME). The system utilized temporal, frequency, and spatial encoding to sample the entire field of view at a 10 Hz framerate 35. Specifications of the light level of the sources and noise levels of the detection opto-electronics have been previously described 35.
Clinical HD-DOT Imaging Cap
The HD-DOT imaging cap was designed with two central goals: (i) to provide stability for efficient and reliable optical coupling, and (ii) to maximize ergonomic comfort for prolonged scanning sessions at the bedside in hospital patients. The custom optical fibers were manufactured with right-angle tips to couple into a soft and flexible neoprene-based imaging cap (Fig. 1b) that allowed for the patient to recline comfortably in their bed during an imaging session. The neoprene provided both a locally rigid structure, which maintained the regular HD grid, and flexibility that promoted fiber-scalp conformity for a wide variety of head shapes and sizes. Each of the 82 optical fiber tips was guided by a ‘top-hat’ style spacer to provide modest translation perpendicular to the head surface to facilitate consistent direct fiber-scalp coupling. Elastic rubber strips on the outside of the cap, held in place with plastic rivets, maintained pressure on the right-angle tips to ensure adequate coupling of the optical fiber tip with the scalp (Fig. 1b) while also maintaining a ~ 3 mm penetration length of the optical fiber tip within the cap to ensure the tips comb through hair. The cap was attached to the patient using comfortable hook-and-loop straps across the forehead and over the top of the head. The imaging cap supported an HD imaging array by maintaining the 48 sources and 34 detectors in two interlaced rectangular arrays with first- through fourth-nearest neighbor separations as follows: 1.3, 3.0, 3.9, and 4.7 cm (Fig. 1c), with as many as 124, 170, 54, and 92 usable source-detector measurements per wavelength at the respective distances (up to 440 total measurements per wavelength). This arrangement produces a spatial resolution of approximately 13 mm within 1.0 cm below the cortical surface (Fig. 1c) 36. To maximize lateral coverage over the MCA watershed area, the grid was designed with left and right panels (each containing 24 sources and 17 detectors) that were symmetrically oriented to each other relative to midline. The field of view of this HD-DOT array encompassed bilateral aspects of temporal, occipital, parietal, and prefrontal cortices (Fig. 1c,d). This coverage resulted in sensitivity to brain function within multiple key functional networks and much of the infarcted cortical tissue of the patients within this study. An infarct incidence map of the 10 patients with identifiable infarcts on either clinical MRI or clinical CT imaging shows damaged cortical tissue co-localized within the HD-DOT clinical field of view (Fig. 1d).
Fitting the HD-DOT Imaging Cap
To reliably acquire adequate coupling across the entire imaging cap, a simple set of steps were followed. First, the center of the cap was placed against the back of the head with the bottom row of fibers on the inion and angled such that the sides of the cap were situated approximately 0.5 cm above the helix-scalp intersect. Second, the cap was gently rubbed back and forth to comb the optical fiber tips through the patient’s hair and to obtain stable coupling against the scalp for all 82 fiber tips. Third, two hook and loop straps were secured over the eyebrows of the patient to secure the cap to the head. Next, the fibers from each side of the imaging cap were gently guided over the head to help the cap match the curvature of the head. The weight of the fibers was supported by the bed around the patient and was neither disruptive nor uncomfortable. The position of the left and right pieces of the cap were checked for symmetric placement with the third source from the front placed just above the tragus at the dorsal-anterior ear-scalp connection.
Real-time HD-DOT Data Quality Assessment
Real-time metrics of data quality were presented on a computer monitor on the HD-DOT console to facilitate efficient at-the-bedside optimization of the cap coupling and position (Fig. 2). First, a schematic of the spatial layout of the optical coupling coefficients for each source and detector position helped localize specific optical elements that were not optimally coupled (Fig. 2a). These fibers could be directly accessed by the user to improve combing through the hair and strengthen the coupling at the scalp interface. Next, source-detector measurement pairs passing a temporal noise-to-signal ratio (NSR) threshold of 7.5% or less were displayed as green lines on a similar spatial layout as the coupling coefficients (Fig. 2b). Third, the mean light level for each source-detector pair was displayed as a function of the source-detector separation (Fig. 2c). The quality of source-detector coupling was evaluated through inspection of the log-linear fall-off of the light-level plots as a function of source-detector distance. Proper coupling produces narrow variance in the light levels for a given source-detector distance while poor coupling distorts the light-level to distance relationship typically by decreasing the light levels. Fourth, a histogram of the NSR for all measurement pairs was displayed (Fig. 2d). Optimal coupling of the HD-DOT array leads to values below 7.5% in the NSR histogram. A fifth real-time readout leveraged the fact that pulsatility of the arterial blood flow is a reliable signature of strong coupling in a CW optical system 37. The spatial distribution of the pulse signature was also displayed to aid in optimizing cap fit at the bedside (Fig. 2d). Displaying these metrics in real time helped the user optimize cap fit within 10 minutes, thereby maximizing data acquisition time for the study.
Clinical structural imaging
All of the evaluated stroke patients (n = 13) underwent a head CT scan as part of their standard clinical care. The CT scans were performed during the 48–72 hours following stroke onset using a Neurologica CereTom portable CT scanner (1.25 × 1.25 × 10 mm) or a Siemens Sensation Open scanner (0.5 × 0.5 × 3 mm) depending on patient stability and scanner availability. The CT scans were then evaluated by a board-certified neurologist to define a patient-specific ischemic infarct mask. Of the evaluated CT scans, masks were able to be generated for nine of thirteen patients. Of the remaining four patients without visible infarct on CT, one patient underwent a brain MRI as part of standard clinical care (Siemens Symphony TIM, TRA Diffusion, 0.9 × 0.9 × 6 mm). This MRI was evaluated by the same neurologist as the CT scans, and an infarct mask was generated. The surface projections of these 10 masks were combined to generate an infarct incidence map (Fig. 1d).
HD-DOT Data Preprocessing and Reconstruction
Source-detector measurements that exhibited a standard deviation < 7.5% of their mean signal were considered low-noise measurements and retained for further processing. The 10 Hz log-ratio source-detector measurements were band-pass filtered to 0.009–0.08 Hz. Superficial and systemic hemodynamic variance was removed through regression of the averaged low-noise first nearest-neighbor measurements 38. First, second, and third nearest-neighbor measurement data were down-sampled from 10 Hz to 1 Hz and used for the image reconstruction. Volumetric reconstructions of absorption coefficients at both 750 and 850 nm were obtained using an inverted sensitivity matrix generated with an atlas representation of the adult head (Montreal Neurological Institute (MNI) 152-subject ICBM non-linear registration atlas 39, 40, 41). The adult atlas-based model of sensitivity was calculated using the NIRFAST light modeling toolbox 42, utilizing an FEM mesh based on the previously described head tissue segmentation and optical properties for scalp, skull, CSF, gray and white matter 36, 43. Three-dimensional maps of relative concentration changes of oxyhemoglobin (ΔHbO2) and deoxyhemoglobin (ΔHbR) were computed using the extinction coefficients of each hemoglobin species 44, 45. Finally, the data were smoothed with a 12 mm FWHM Gaussian kernel and transformed into MNI atlas space. Because the sensitivity profile decays exponentially as a function of depth, the effective penetration is limited in this system to approximately 2 cm below the head surface, corresponding to a threshold in the top 90% of the inverted sensitivity profile 35.
Automated Motion Detection and Censoring
Ideal cap fits do not guarantee high quality data. Noise in the data, such as system fluctuations or patient motion, can be misinterpreted as brain fluctuations and in particular, can cause erroneous functional connectivity results, as in fMRI 46. Noisy measurements from system issues were identified by having a temporal standard deviation greater than 7.5% (Fig. 2c) and were removed before image reconstruction as detailed above. Additionally, we utilized a time-point specific motion detection algorithm, termed global variance of the temporal derivatives (GVTD) 47 that is analogous to the DVARS metric of fMRI 48, 49). The GVTD metric is calculated as the root mean square across measurements of the temporal derivative in light levels. Temporal sections of data were classified as clean when they exhibited a GVTD value below an empirically defined threshold of 0.0015 for at least 60 contiguous seconds. This fast and automated cropping of motion-contaminated data was performed using the optical measurements before reconstruction into the voxelated space to maximize sensitivity of the metric as described by Sherafati et al., 2020. Each epoch of measurement data that remained after the automated motion detection was then individually pre-processed and reconstructed as previously published 35.
Functional Connectivity Analyses
Individual participant’s epochs of clean and filtered volumetric ΔHbO2 were concatenated together. The global signal, averaged over the entire HD-DOT field of view, was regressed from every voxel35, 38. Region of interest time traces, corresponding to a given seed location, such as the motor cortex, were computed by averaging the voxels within a 5 mm radius sphere centered at a given seed location (Fig. 3). Seed-based zero-lag fc-maps were created by calculating the Pearson correlation coefficient between the seed time trace and time trace of every other voxel within the field of view. Correlation coefficients were Fisher-Z transformed before further analysis. Given the high dimensionality of the data (10,241 fc-maps per subject with 10,241 voxels each), a similarity metric was calculated to reduce dimensionality. The similarity metric for a given seed location was generated by calculating the spatial Pearson correlation coefficient between the fc-map of interest (e.g. stroke patient or older normal subject) and the averaged young healthy fc-map (Fig. 4). Similarity metric maps were created by calculating similarity metrics for every seed in the field of view (n = 10,241).
Statistical analyses
To assess potential stroke-induced disruption in functional connectivity, we first compared individual fc-maps to a reference healthy young adult dataset. The young healthy mean reference fc-map was generated by averaging the Fisher Z-transformed fc-maps of the seven young healthy participants (Fig. 4a – lower panel). Then, for each voxel in the FOV of a particular subject (older control or stroke patients), a Pearson correlation was calculated between the respective fc-maps of the participant and the young healthy mean reference. This spatial correlation produces a similarity metric for the subject’s connectivity patterns at a given seed compared to the patterns of a young healthy normal average reference. For stroke patients, datasets with infarcts in the right hemisphere were flipped across the sagittal axis in order to have all infarcts located in the left hemisphere. The Pearson correlation was then Fisher Z-transformed and assigned to the corresponding voxel location for the subject (Fig. 4b). The process was then repeated for each voxel in the FOV (Fig. 4c), and for all participants (stroke patients and older controls). To establish an average older control similarity map (upper map of Fig. 5a), the mean Fisher-Z value for a given voxel of each older control subject similarity map (n = 11) was calculated and then assigned to the respective voxel. This was performed for each seed location in the field of view. The same process was used to produce the mean stroke patient similarity map (n = 13) (Fig. 5).
To compare the mean similarity between stroke patients and older healthy controls, the mean similarity for each subject was calculated by averaging all z values in each subject’s similarity map to produce a single similarity summary metric per subject. These values were then compared using an independent samples t-test. We tested for differences in variance using Levene’s test, which was evaluated as non-significant (p = 0.363), and so equal variances were assumed. To evaluate the effect size of the group comparison of fc-map similarity, Cohen’s-d was calculated. The distribution of similarity values for a given stroke patient was calculated using the skewness of similarity values (Fig. 5). The relationship between the skewness of the similarity with the NIHSS of the stroke patients was calculated using Pearson correlation. For the nine patients with identifiable infarct on CT scan, a Pearson correlation was performed between the infarct volume and the NIHSS.