Glymphatic flow resistance in humans measured by multifrequency impedance dispersion

System Architecture

The Applied Cognition device is a wearable multi-sensor acquisition system that uses an STM32WB5MMG microcontroller (ST Microelectronics, NV). It consists of a primary module housing the main electronics board and a 430 mAh Li-ion battery, two earpieces each housing one electrode, one accelerometer and one photoplethysmogram sensor, and two separate mastoid electrodes. A device schematic and visualization of signal outputs is provided in Figure 2a-f. Sensor data is stored in on-board FLASH memory and is downloaded via a USB port for offline analysis. The relevant sub-systems are further described below.

Electrical Impedance Spectroscopy

Electrical impedance spectroscopy (EIS) was used to measure electrical impedance using the AFE4500 analog front-end (AFE) from Texas Instruments. Four electrodes – two in-ear and two mastoid – were used to deliver the excitation current across the head and sense the resulting voltage. The two mastoid current injection sources were initially co-located inside the ear next to the Ag/AgCl sensors but because of salt-bridging were externalized to the mastoids. In a departure from traditional four-electrode measurements, where excitation and sense electrodes are fixed, the AFE4500 was programmed to implement a technique developed by Texas Instruments where multiple excitation and sense configurations are used to determine the contact impedance and body impedance accurately.

Measurements were successively performed at eighteen frequencies ranging from 1 kHz to 256 kHz following a logarithmic series, all with excitation current below 50 μA root-mean-square (rms). Frequencies were in alternating order from high-low to minimize any impact on impedance drift during each scan. Sensed voltages were quadrature-demodulated, filtered and digitized internally by the AFE4500 to provide in-phase and out-of-phase components. Measurement time for each frequency was adjusted to give consistent signal-to-noise ratio across the frequency range, resulting in a scan time of 106 seconds.

Electroencephalogram

Electroencephalogram (EEG) was measured with an ADS1299-4 AFE from Texas Instruments, between the two in-ear electrodes (differential measurement). A third electrode (left mastoid) was used to drive the common mode. While these electrodes were shared with the EIS, the two measurements were time-multiplexed and a pair of analog multiplexers (TMUX1136, Texas Instruments, Inc.) effectively decoupled EEG and EIS sub-circuits and respective measurements. For EEG, both in-ear electrodes were also buffered (OPA376, Texas Instruments, Inc.) in the earpiece. These buffers were bypassed (analog multiplexer TMUX1136, Texas Instruments, Inc.) for EIS or electrode impedance measurement.

The ADS1299-4 was programmed to have a sampling rate of 250 samples-per-second (24-bit samples) and an input range of ±180 mV. The common mode drive was actively derived from the in-ear electrodes. Both mastoid electrodes were also connected to an input channel so that the low-frequency (31.25 Hz) impedance of each electrode could be measured using the internal current source of the ADS1299-4. Such impedance measurements were performed before starting a recording session, and then automatically between EEG periods and EIS. EEG measurement periods were 170 seconds long, and the sequence EEG-EIS-Electrode impedance repeated until the end of the recording session. A sample EEG spectrogram and hypnogram, derived from the device EEG is shown in Figure 2b.

Impedance Plethysmogram

In addition to impedance spectroscopy, real-time bioimpedance was acquired simultaneously with the EEG, using the MAX30001 AFE (Analog Devices, Inc.). The four-electrode configuration shared electrodes with the EEG, with the two mastoid electrodes used for current injection and the two in-ear electrodes for voltage sensing. A 96 μA current at 82 kHz was used for excitation. Such high frequency excitation is effectively filtered by the ADS1299-4 and does not corrupt the EEG signal. The MAX30001 built-in synchronous demodulator filtered and ADC recovered the in-phase component of the sensed voltage and digitized it at 64 samples-per-second (20-bit resolution). The resulting impedance measurements had a noise floor of 3 mν rms (100 ν load), enabling the detection of blood pulsations. A sample trace from a study participant, derived from the device IPG is shown in Figure 2d. Note that the impedance shifts measured in either the respiratory or cardiac frequency bands range between 50 to 150 mν. In contrast, impedance shifts detected by EIS were at least an order of magnitude greater.

Photoplethysmogram and Acceleration

In-ear reflective photoplethysmogram (PPG) was measured with two miniature, fully-integrated optical sensors MAXM86161 (Analog Devices, Inc.) located in each earpiece, facing the anterior wall of the ear canal. One side was programmed to use green light (530 nm), while the other used alternatively red (660 nm) and infrared (880 nm) lights. To achieve a tight synchronization with EEG samples, the sensors ran at 1024 Hz for each color, and the most recent samples available at the time of the EEG read were averaged, resulting in a 250 Hz output rate and a phase variation <1 ms. A sample trace of the left (red) and right PPG (green) from a study participant is provided in Figure 2f.

In-ear acceleration (X,Y,Z) was measured using two miniature MEMS Inertial Measurement Unit (IMU) LSM6DSOX (ST Microelectronics, NV) located in each earpiece. Similar to the PPG, the sensors were run at 833 Hz (each axis) and averaged to ensure tight time synchronization between sensors. A noise floor of 1 mg rms enabled the resolution of cardiac/blood pulsations, in addition to posture and activity. From these accelerometer data, a ballistocardiogram is derived, as shown for a study participant in Figure 2e.

Electrodes

In-ear electrodes were laser-cut from silver sheets 100 microns thick and 99.9% purity. The electrodes were 4.5 mm diameter plates with a 1 mm-wide stem. The electrodes were abraded and sonicated in distilled water, dip-coated with Ag/AgCl ink provided by Creative Materials Inc (SKU EXP 2653-138-1) and cured at 100°C for 60 min. Following curing, they were sintered at 427°C for 60 minutes under low-flow argon gas.

Once fabricated, all electrodes were characterized according to the ANSI/AAMI EC12 standard for AC impedance⁵⁷, DC offset voltage, combined offset instability and internal noise, and bias current tolerance (Supplemental Table 7). Pairs of electrodes were also connected to the earpiece circuits with clip-on wires and tested on an agar phantom as described in the Device Testing section.

Electrical Impedance Spectroscopy Validation

Measurement accuracy was first evaluated using a test fixture simulating the load and the four electrodes (Supplemental Figure 2). The load was either purely resistive (R, 20-100 ν), or slightly reactive (2R1C, 20-100 ν in parallel with 470 ν and 22 nF in series) to mimic the brain parenchyma. The electrodes were implemented as either (i) low impedance (270 ν resistor in series with a 1 kν resistor and 180 nF capacitor in parallel), or (ii) high impedance (620 ν resistor in series with a 3.3 kν resistor and 82 nF capacitor in parallel), in order to simulate the frequency-dependence of typical Ag/AgCl electrodes, but also to ensure a minimal impedance at high frequencies. Two 10 ν resistors were added between the electrodes connected to the same side of the load to simulate the tissue impedance between these electrodes as well. Supplemental Table 8 describes the benchtop tests performed and Supplemental Table 9 and Supplemental Table 10 contain the results.

• Download figure
• Open in new tab

Supplemental Figure 2.

a, Circuit model of transcranial measurements using four-point electrical impedance spectroscopy and models for each of the four electrode contact impedances. b, Electrode impedance models comprised a low impedance and a high impedance model, representing the range of in-human impedances. c, Calibration loads varied the resistance R_s of the free fluid or interstitial fluid compartment from 20 to 100 Ω in increments of 20 Ω, with and additional R_pC_p in parallel to simulate complex loads. The realistic load parameters were chosen to create a frequency dispersion that approximated the in-human measured dispersions. d, Electrode impedance dispersion over the measured frequencies for the low and high impedance model. e, Load impedance dispersion over the measured frequencies using the Real Load model.

Device Testing

In addition to the EIS test fixture, stability and reproducibility of EIS measurements over twelve hours were evaluated with an agar phantom. A slab (5-10 mm thick) of conductive agar was first cast using a solution of 2.5% weight agar (Living Jin USA, Inc.), 0.64% weight NaCl (RND Center, Inc.) and distilled water. Electrodes (mastoids and earpieces for the final testing) were then coated with conductive gel (Electro-Gel, Electro-Cap International) to reduce the contact resistance, and placed on the agar slab in a pattern geometrically similar to the electrode location on the head. A 12-hour recording was performed, resulting in a total of 450 minutes of EEG recording, 159 EIS cycles and 160 electrode impedance checks. The results of a 12 hr agar phantom EEG test is shown in Supplemental Figure 3.

• Download figure
• Open in new tab

Supplemental Figure 3.

a, Periodogram of a device over a 12 hour test using four Ag/AgCl sintered disk electrodes on an Agar phantom with EEG gel. For reference, human EEG beta power during NREM sleep is above −10dB on the device. b, EEG recording over an epoch shows 0.5 µV peak-to-peak noise. c, Power spectral density shows a noise floor at −150dB/Hz with a 1/f increase to −140db/Hz below 10Hz.

Signal Processing – Parenchymal Resistance

A python neuroprocessing pipeline developed by Applied Cognition performed the data analysis of the signals captured by the investigational device. The neuroprocessing pipeline runs in the Amazon Web Services cloud on a distributed architecture allowing for fast parallel execution of participant device readings.

An EIS scan by the investigational device measured impedances Z(ω) at 18 frequencies f (ω = 2πf) ranging from 1,000 Hz to 256,000 Hz and took 106 seconds. The frequencies were measured in the following permuted order for each scan (all in Hz): 2276, 102400, 1600, 64000, 4551, 51200, 12800, 85333, 3200, 25600, 1000, 32000, 6400, 8533, 128000, 204800, 256000. During the EIS scan, all other device sensors were powered off except the left-in-ear IMU that sampled at 1Hz to detect head motion and position. The EIS scans were duty-cycled throughout the recording period with 165 second neurophysiology scans that captured data from the other sensors: EEG, PPGs and IMUs. Four electrode impedance checks at 30Hz and lasting 7 seconds were performed during the neurophysiology scan and immediately preceding the EIS scan.

The Cole-Cole model^58–60 is commonly used to analyze EIS data. The analysis is based on the four parameters contained in the Cole equation R₀, R_∞, α and τ. As the excitation frequency ω (ω = 2πf) increases to large values, the capacitive cell membranes are invisible to the excitation frequency and the impedance Z(ω) approaches R_∞. The four excitation and response electrodes are positioned trans-cranially and at large ω measure total intracranial brain volume. As the excitation frequency ω decreases approaching DC values, the capacitive cell membranes prevent any transmembrane conduction and the impedance Z(ω) approaches R₀ the measure of electrical resistance through extracellular fluid.

The value of τ in the Cole-Cole model is the inverse of tissue characteristic frequency 2πf_c. The Cole-Cole α describes the divergence of a measured dielectric dispersion from the ideal dispersion exhibited by a Debye type of dielectric relaxation, and is widely assumed to be related to a distribution of the relaxation times in the system involved. The value of α ranges from 0.5 to 1 with a value of 1 reducing the Cole-Cole model to the Debye model.

For each EIS scan, the four parameters were estimated using a non-linear least-squares fit (Python SciPy’s least_squares) with the Trust Region Reflective algorithm that is a robust method suitable for large sparse problems with bounds. The α parameter was found to be close to one in the estimations and thereafter set to one, reducing the Cole-Cole model to the Debye model. The remaining three parameters were estimated with bounds of 0 to 90Ω for R₀, 0 to 70Ω for R and bounds for determined by requiring the characteristic frequency f to lie between 16kHz and 160kHz. The parameter bounds where confirmed to encompass neurophysiological values in all participants by plotting all EIS scan resistance (real values of impedances Z(ω)) and reactance (imaginary values of impedance Z(ω)). Furthermore, no parameter solutions lay on a boundary for each scan across the range of dynamic impedance spectroscopy measurements for all participants.

Prior to fitting the Debye model to the EIS scan data, to compensate for stray capacitance, time delay effects from signal transmission and electrical components, the impedances Z from each scan were rotated by multiplying by the factor e^jωT with the value T chosen so that the reactance at the highest frequency of the scan was zero⁵⁹. This rotation was independently validated using EIS scans recorded on the benchtop tests using known impedance loads (Supplemental Table 10). The root-mean-squared-error (RMSE) of the Debye model fits to the benchtop tests (Supplemental Table 8) are reported in Supplemental Table 11.

Each EIS scan measures R_P at a point in time and to evaluate glymphatic function the Debye model fits need to detect small changes in R_P. With the device time-multiplexing settings, the device completes approximately 84 scans over a 7-hour night of sleep. The statistical models require significance testing of the mean change in R_P between sleep and wake, and testing significance of predictors of R_P in linear mixed models. These tests depend on the SD of the Debye model dispersion and the number of EIS scans (Supplemental Table 12). Details of these analyses are found in Supplemental 4.

During EIS, tissue behaves as a linear, time-invariant and causal system allowing the use of Kramers-Kronig relations to identify electrical impedance measures that are corrupted by artifacts from motion, electrode impedances and other sources^31,32. The Kramers-Kronig transform allows the real part of the electrical impedance measure to be derived from the imaginary part and vice versa³¹. These relations are used as validity tests for measured spectrums. When the real and imaginary impedance measures do not satisfy the Kramers-Kronig relation the measure is discarded. The Python PyEIS1.0.10 repository was used for Kramers-Kronig impedance validation. The validation uses the method of Schönleber³² to avoid ambiguities in the linear Kramers-Kronig validity tests due to under- and over-fitting.

In addition to the Kramers-Kronig validity tests, EIS measures were discarded if for any of the four electrode impedances, the difference across successive time measures exceeded 1kΩ or if the EIS scan was collected during a wake interval on a participant randomized to sleep. The first condition ensured that the electrode impedances were maintained within a narrow range during the EIS scans and the second condition ensured that the EIS scans were representative of sleep physiology for participants randomized to sleep.

Signal Processing – Electroencephalography

The raw EEG tracings from the device and the commercial PSG (Philips Respironics Alice 6 LDx Diagnostic Sleep System) were time-synchronized by time-aligning a 100-ms square pulse generated by the device every 10 seconds and stored both in the device flash memory and in the PSG recording through as an unused auxiliary analog input channel. Each raw tracing was notch filtered at 60 Hz using a second-order infinite impulse response (IIR) notch digital filter (Python SciPy iirnotch). The digital filter was applied to the signal forward and backward with a combined filter phase of zero (Python SciPy filtfilt). The signal was then bandpass filtered between 0.3 Hz and 50 Hz using a finite impulse response (FIR) filter with a Hann window (Python SciPy firwin). The length of the low-pass filter was 501 and the length of the high-pass filter was 2,401.

During the observation periods, participants were not allowed to touch devices that were plugged to power outlets. Other non-physiologic sources included excessive head motion that led to signal artifacts from the electrode-skin interface. Physiologic sources of signal artifact included electrooculogram, electrocardiogram and electromyogram. The signals where partitioned into time-aligned 30-second epochs. Epochs containing one or more peak-to-peak signal amplitude exceeding 350 µV or maximum power in the Welch power spectrum exceeding 1,000 µV²/Hz were removed. In addition, the device left in-ear and right-in ear IMU signals were used to filter excessive motion by filtering epochs with peak-to-peak values on an x-, y-, or z-axis on either ear that exceeded 100 milli-g during sleep and 200 milli-g during wake.

The power spectral density of each 30-second epoch was computed using Welch’s method⁶¹ (Python Scipy welch) using 10-second segments and 50% overlap across successive segments. Relative power bands were computed for delta power (1-4 Hz), theta power (4-8 Hz), alpha power (8-12 Hz), sigma power (12-15 Hz), beta power (15-30 Hz) and low gamma power (30-50 Hz), each normalized to the total power in the power spectral density. Simpson integration of the power spectral density was used for computing the total power and respective band powers.

Hypnogram staging of each 30-second epoch used automated scoring (Python yasa 0.6.3 SleepStaging) trained and validated on 3,000 nights of data from the National Sleep Research Resource⁶². The automated scoring uses a single EEG derivative and a single EOG channel. For device hypnogram staging the single in-ear transcranial derivative was used and for the commercial PSG, hypnogram staging using C3-A2 and C4-A1 derivatives were recorded separately. The PSG left EOG channel was used for commercial PSG hypnogram staging. An additional T3-T4 transcranial PSG derivative was used for hypnogram staging as another comparison to the device’s transcranial in-ear derivative. The internal agreement of commercial PSG and agreement between investigational device-based EEG and PSG are presented in Supplemental Table 6.

Signal Processing – Heart Rate

The photoplethysmography (PPG) left and right raw signals were bandpass filtered between 0.5 Hz and 10 Hz using a finite impulse response (FIR) filter with a Hann window (Python SciPy firwin). Empirical mode decomposition was used to identify peaks in the raw signals (Python emd 0.6.2) using a box-car rolling window of 100 observations⁶³. Heart rate variability metrics were computed after removing outliers, ectopic beats and interpolating missing values (Python Aura-healthcare hrvanalysis). Both time-domain features (standard deviation (SD) of the inter-beat-interval of normal sinus beats, sdnn) and frequency-domain features using the Lomb-Scargle periodogram (low-frequency signal power lf between 0.04-0.15 Hz considered a mixture of sympathetic and parasympathetic activity, high-frequency signal power hf between 0.15-0.4 Hz representing beat-to-beat changes from parasympathetic vagal activity, and the low-frequency to high-frequency ratio (lf-hf-ratio) were computed⁶⁴.

The device inertial-measurement units (IMU) were located inside the left and right ear canal and provided left and right cartesian coordinates of acceleration measured in milli-g where 1,000 milli-g is the acceleration caused by earth’s gravitational force. Each in-ear component from the device is uniquely and reproducibly oriented on account of the ellipsoidal shape of the human ear-canal and the matching shape of the in-ear component. From this orientation, head position and motion (acceleration) were resolved from each set of IMU samples that included gross positions such as supine, reclined, upright, left and right.

Participant Recruitment

Following distribution of study flyers in public-facing locations in senior recreation and/or local medical centers, the Benchmarking Study enrolled 34 healthy participants 56-66 years of age and the Replication Study enrolled 14 healthy participants 49-63 years of age. Participants were cognitively intact and had no history of clinical depression, confirmed during screening using the Montreal Cognitive Assessment (MOCA) and Geriatric Depression Scale (GDS). Other exclusion criteria included a history of diabetes, hypertension, coronary artery disease, pulmonary disease, neurological disease, or anxiety. Participants planning travel to alternate time zones within two weeks of study participation were also excluded. Consolidated Standards of Reporting Trials (CONSORT) diagram for the Benchmarking Study and Replication Study is provided in Figure 3 (bottom), and participant exclusions are detailed in the Results section.

Demographic data including age, sex, APOE4 allele status, MOCA and GDS for the Benchmarking Study, Replication Study, and Combined Studies are provided in Supplemental Table 1.

Study Protocol

As shown in the study schematic (Figure 3, top), this is a cross-over study in which participants took part in two overnight study visits, one undergoing normal sleep and the other undergoing overnight sleep deprivation, at a 2-4 week interval. Each overnight study visit consisted of three study periods beginning at 1900 hrs, 0700 hrs, 1330 hrs separated by two intervals – an overnight interval (2300-0630 hrs), and a morning interval (0800-1130 hrs).

Magnetic Resonance Imaging Approach

MRI scanning at The Villages site was conducted on a 3T Siemens Vida system at Lake Medical Imaging, while scanning at the University of Washington site was conducted on a 3T Philips Ingenia Elition X at the Diagnostic Imaging Sciences Center. In the present report, to define the relationship between brain R_P and glymphatic function, we utilize CE-MRI from Benchmarking Study participants following the overnight sleep condition. We have included a description of the other MRI-based approaches within the overall protocol for completeness. However, the relationship between these different MRI measures of glymphatic function, Investigational Device outputs, sleep parameters, cognitive tests, and fluid biomarker levels will be reported elsewhere.

Statistical Approach

The Benchmarking study was designed to have 80% power at a significance level of 5% to detect a low-to-moderate correlation r (r=0.5) between the device measurements and CE-MRI measurement of glymphatic function. The Replication study was designed to have 80% power at a significance level of 20% to detect a low-to-moderate correlation r (r=0.5) between the device measurements and non-contrast MRI measurement of glymphatic function and to detect overnight sleep/wake differences in brain parenchymal resistance.

The parenchymal resistances R_P computed from each EIS scan from the device were normalized to the sleep/wake onset values to allow comparison across visits and participants. To reduce measurement variability of the sleep/wake onset value, the value was inferred from a linear regression to the initial R_P values. The overnight change in R_P per participant visit was taken as the average of the normalized values for that visit. For example, a value of 1 indicated that there was no overnight change from the onset value and a value of 0.8 indicated a 20% overnight reduction from onset. A Student’s T-test group comparison between sleep/wake visits, group mean and SD were calculated.

Contrast-enhanced MRI of sleep-related glymphatic function was analyzed using a random intercept linear mixed effect model (R, lme in the lmne package). Glymphatic function was defined by measuring brain parenchymal contrast enhancement, the % change in T1-weighted signal intensity between 10-240 min post-GBCA injection (100% * (T₂₄₀-T₁₀)/T₀) at each of eight regions of interest (ROIs): frontal cortical gray and white matter, parietal cortical gray and white matter, temporal cortical gray and white matter, and occipital gray and white matter. Only sleep visits were analyzed leading to a one level of grouping, the eight ROIs, for the participant sleep visit. All models included regressors in the mixed model for the T₁₀-T₂₄₀ change in blood signal within the internal carotid artery, and the T₁₀-T₂₄₀ change in CSF signal within the cerebral lateral ventricles to compensate for the influence of vascular and CSF contrast on parenchymal enhancement. Potentially confounding biological variables age, gender, and APOE ε4 status were also included in all models.

To evaluate different sleep-related contributions to glymphatic function, overnight changes from sleep onset for each of the four recorded EEG relative band powers (delta, theta, alpha, beta) and for HR were computed. The recorded values were normalized to their respective value at sleep onset and a single statistic, the average of the overnight normalized values, computed per participant visit. The four hypnogram sleep stages (N1, N2, N3, REM) and WASO were each summed into single total sleep stage durations and total WASO duration per participant visit. The device R_P normalization and averaging has been previously described. Models were run separately for each of the four EEG relative band powers (delta, theta, alpha, beta), for each of the four hypnogram sleep stages (N1, N2, N3, REM), for WASO, for HR and for device R_P. Two full models were then fit to the data, with the EEG band power and the hypnogram sleep stages separated into distinct models because of collinearity across those predictors. The variance inflation factor (R ‘vif’ function from the ‘car’ package) was used to test for predictor collinearity using a threshold of 5. The EEG powerbands included were limited to delta, theta and beta because of collinearity among the four relative band powers. HR and the device R_P were included in both full models. Each full model was reduced to a selected model using backward elimination of predictors with a threshold Wald’s p-value of 0.05.

Model selection stability, or robustness of the selected model to perturbations in the dataset, was tested using resampling-based multi-model inference²⁹. The dataset was bootstrapped resampled with replacement to generate 500 bootstrapped datasets. For each of these datasets the two full models were estimated and reduced to a selected model using backward elimination. The bootstrap inclusion frequency (the percent of bootstrap datasets that included the predictor), the RMSD ratio (the RMSD of the predictor bootstrap estimate divided by the standard error of that coefficient in the full model and expressing the variance inflation or deflation caused by variable selection), the relative conditional bias (quantifying the variable-selection-induced bias to expect if a predictor is selected) and the bootstrap median of each predictor coefficient were reported.

The analyses of device R_P values with EEG band powers, heart rate and hypnogram staging were performed on the overnight sleep recordings. Analyses were grouped into REM sleep and NREM sleep (N1, N2 and N3). The device resistances were computed repeatedly throughout the night from each EIS scan. During sleep recordings, these data were non-stationary. After normalization to sleep onset as previously described, first-order differences ΔR_P and ΔX for X one of the EEG band powers or HR were incorporated into individual random intercept linear mixed models to evaluate their relationship. Threshold regression models were used in the linear mixed model. A custom estimation procedure described in Supplemental Information was used to jointly estimate the threshold regression for the full positive and negative range of the predictors. The hinge threshold model was analytically reduced to estimation by linear regression which allowed it to be used in a random intercepts linear mixed effects regression model. The autocorrelation function and augmented Dickey Fuller tests were used to confirm stationarity of the differenced variables ΔR_P and ΔX and also of the fitted model residuals (R, acf and adf functions).

In addition to the analyses of ΔR_P and ΔX for X one of the EEG powerbands or HR during REM sleep and NREM sleep, these were analyzed at hypnogram transitions NREM to REM, REM to NREM, sleep to wake and wake to sleep. The analysis was similar, using threshold regression models in a random intercept mixed linear model.

Univariate outliers were trimmed at the 0.5 and 99.5 percentile. For multivariate analysis, the Mahalonobis distance was computed and outliers trimmed at the 0.5 and 99.5 percentile of the Mahalonobis distance. All regressions were adjusted for potential confounding effects of age, sex and APOE genotype. All tests were considered significant at a 5% level if the 95% confidence interval (CI) did not contain the null hypothesis. All statistical analyses were performed in R version 4.2.1 (2022-06-23) or Graphpad Prism 8.0.0.

We reported the pooled findings of the Benchmarking and Replication study for both the investigational device transcranial resistance measurements and EEG parameters.