Alzheimer’s disease and current state-of-the-art

By 2050, it is anticipated that more than 132 million individuals will be affected by dementia globally [1]. The associated annual costs are projected to reach 2.8 trillion dollars in 2050 compared to 1.3 trillion in 2019 [2]. While several disorders can lead to dementia, the most common cause of dementia is Alzheimer’s disease (AD) [3]. Alzheimer’s dementia manifests with cognitive decline which progresses over time at the detriment of patients and relatives.

Recently, there has been increasing success in the development of disease-modifying pharmacological therapies with monoclonal antibodies that target the removal of Amyloid-beta (A). However, the effects observed are limited to about 30% reduction in disease progression, and the treatment is associated with some inconvenience and potential risk of brain oedema and brain haemorrhage from the A-clearance process [4].

Concurrently, there is growing interest in the interpretation of AD as a dysfunction of the brain’s neural networks, where the electrophysiological abnormalities are investigated as pathological markers and prospective treatment targets (this topic is vast and has been covered extensively by others, including but not limited to [5–8]). As a results of this, investigations have emerged into the potential for using 40 Hz brain stimulation based on visual flicker for the purpose of AD therapy [9–19]. However, little is understood about the mechanism of action (MoA) for how 40 Hz stimulation could slow progression of AD, and part of this understanding begs the question: Why exactly 40 Hz? In some cases, 40 Hz stimulation is compared to (relatively) low- and high-frequency stimulation [9] as a kind of placebo control. However, such comparisons to frequencies outside the mechanistically relevant vicinity of 40 Hz would not reveal nuances about the specificity of 40 Hz.

This study investigates the ability of a principally novel light technology, invisible spectral flicker [20] (ISF), to evoke a cortical response at other frequencies in the vicinity of 40 Hz. Invisible spectral flicker is distinct from luminance flicker (LF) used in [9–13, 15, 16], in which the brightness is modulated in time, and from heterochromatic flicker (HF), in which the light alternates between two colours. By definition, ISF generalises to alternating between two different but arbitrarily complex spectra of light. If the two spectra are composed to be perceived as sufficiently similar, it may reduce the sensation of flicker or render it entirely invisible under some circumstances (see e.g. [21] for a review of flicker fusion). We have previously demonstrated that the implementation of ISF investigated in [17, 18] and in the ongoing clinical trial ClinicalTrials.gov ID NCT05260177 is associated with significantly lower perception of flicker and self-reported discomfort compared to LF and red/green HF [22]. Selecting the appropriate spectra for stimulation is not trivial, but we have previously investigated which parts of the visual spectrum are most important to represent [23]. From that study, comparing the evoked cortical responses to all unique HF combinations of six different colours, we concluded that it is more important to include either extreme of the visual spectrum (i.e. blue or red) in at least one of the flicker phases, than it is for the two phases to be very spectrally distinct.

While flickering visual stimuli at frequencies in the 40 Hz range have have been demonstrated previously [24, 25], we aim to widen the basic understanding of cortical responses to ISF at various frequencies. As detailed below, two observations may favour modification of the stimulation frequency: (1) hypotheses for the mechanism of action (MoA) are not tied to exactly 40 Hz, and (2) the perception of flicker is frequency dependent.

Motivation for changing stimulus frequency

There are multiple hypotheses for the MoA of 40 Hz stimulation AD therapy. The work conducted by Cardin et al. [26] shows a key role of fast-spiking (FS) parvalbumin-positive (PV+) inhibitory interneurons in generating gamma oscillations in the cortical network with notable sensitivity to 35-55 Hz modulated optogenetic activation. This contrasts with the peak sensitivity of pyramidal neurons at approximately 8 Hz. Additionally, targeting PV+ and somatostatin-positive (SST+) interneurons responsible for gamma oscillations seems to be the right approach, as there is a cell-specific dependency on the propagation of visual flicker [27]. This is aligned with [28], highlighting a hippocampal circuit resonating at 39 3 Hz (mean sd) when activated with a choline agonist. Furthermore, PV+ and SST+ interneurons have been linked to neurological disorders like AD and could be a causal step in the disease pathogenesis [29]. However, none of these systems expose a necessity for stimulation to be at exactly 40 Hz. On the contrary, they may provide evidence of a wider dynamic range of applicable frequencies.

According to the IEEE Recommended Practices for Modulating Current in High-Brightness LEDs for Mitigating Health Risks to Viewers, the risk of biological (side) effects such as seizures, headaches, or malaise is highest from exposure to visible flicker [30]. Gamma stimulation applications that make use of the sensory pathway(s) may be especially affected by the stimulation frequency. This includes 40 Hz visual flicker as used in [13–19], in which the amount of perceived flicker is affected by the flicker frequency [21]. As such, modifying the stimulation frequency could enhance the utility and treatment adherence of such therapies, though the clinical implications are speculative.

Under the speculative assumption that small deviations from 40 Hz can be tolerated without loss of potential clinical benefit, it is worth reconsidering the frequency. If the sensitivity distribution for input frequencies to the FS-PV+ interneurons presented in [26] were to guide the choice of stimulation frequency, it would suggest increasing the stimulation frequency. This is important as the perception of flicker is reduced at higher frequencies which can be exploited to improve the user experience. Contrarily, the resonance frequency distribution of hippocampal circuits presented in [28] would point to a slight reduction in the preferred frequency. This would have the opposite implications for the amount of perceived flicker.

Previous studies have shown the ability to visually evoke steady-state potentials (SSVEPs) with invisibly flickering stimuli [24, 31–33]. The point at which flicker becomes invisible—the critical flicker fusion frequency (CFF)—has been studied extensively for decades. It is known to depend on many parameters including but not limited to the perceiver, ambient lighting, and light modulation depth, duty cycle, luminance, and varies across a wide frequency band depending on circumstances[30, 31, 34, 35]. Reconsidering stimulation frequency may be especially valuable for invisible spectral flicker (ISF) [20] (as used in [17, 18]) which aims to reduce the amount of perceivable flicker to the point of being “merely perceptible” by alternating between two complex spectra of light, not limited to monophasic spectra of single-colours (see Fig. 2). The results of [22] would indicate that 40 Hz is close to the CFF of ISF for some observers (participants) under the reported circumstances (though this was not explicitly tested) and that increasing the frequency could make the flicker even less perceptible or imperceptible. Thus, it is relevant to verify that ISF can evoke SSVEPs above 40 Hz and evaluate whether any frequency preference is evident from the evoked power.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Overview of experimental design:

Nine ISF stimulation conditions with frequencies varying from 36–44 Hz in 1 Hz increments were repeated five times in a randomised complete block design (RCBD). Within each block, the nine frequency conditions were shown once. The stimulation trials had a duration of 30 seconds and were separated by a three second inter-trial interval. Before and after the RCBD, a 1-minute resting-state baseline was recorded—30 seconds with eyes closed and 30 seconds with eyes open.

https://doi.org/10.1371/journal.pone.0321633.g001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Invisible spectral flicker light specification:

Invisible spectral flicker alternates at 40 Hz between two phases of spectrally different white lights which ease the fusion and reduce the perception of flicker. A: Spectral profiles of individual phases and the composite which defines ISF. One phase is a colder white, while the other is a warmer white, but both span broad bands of the visible spectrum. The combined spectrum is the average of the two phases achieved by 50% duty cycle between them. B: Schematic of 40 Hz temporal colour mixing with square waveforms and 100% modulation depth to achieve ISF.

https://doi.org/10.1371/journal.pone.0321633.g002

In this study, we investigate the ability of ISF to evoke cortical responses at several stimulation frequencies in the range of 36–44 Hz. This is proximal to and symmetrically distributed around the current choice of 40 Hz used in [17, 18] (and the ongoing study NCT05260177), overlaps with the interneuron sensitivity- and hippocampal resonance frequency distributions from [26, 28], and is identical to the clinically relevant band used by [36–38] for differentiating AD patients from geriatric controls. The study design and objectives are similar to those of [25], but the visual stimulation paradigms are qualitatively different. They investigate the [32–50] Hz band (in two-hertz increments) using visual stimulation with single colour luminance flicker of white, red, green, or blue, thus making it an excellent study for comparison.

Experimental design

Twenty healthy volunteers (10 males, 10 females) with an age between 21 to 64 years (an average of 30 years) were included in an electroencephalography (EEG) experiment (approved by the institutional review board at DTU Compute under the protocol number: COMP-IRB-2021-01) between May 28th and July 24th of 2021. Participants were recruited conveniently by word-of-mouth and were not reimbursed for participating. The data was collected in accordance with the Declaration of Helsinki, and all participants gave written informed consent prior to data collection. All subjects had either normal or corrected to normal vision, and the exclusion criteria included history of migraine, photosensitive epilepsy, or other known neurological disorders. Subjects were situated in front of a visual stimulation device (Light Therapy System (LTS), OptoCeutics ApS, Denmark) at a distance of 60 cm and were requested to sit calmly, minimise unnecessary movement and to look at the device while the stimulation light was on. The EEG recording was conducted with a 19 channel dry electrode system (Zeto Inc., CA, USA) with a sampling rate of 500 Hz and mounted according to the 10–20 system, with linked mastoid reference. Recording was done using cloud-based software (the Zeto Cloud Platform), a 24-bit analogue-to-digital converter (ADC), and no additional live filtering after the ADC.

The EEG consisted of an initial baseline recording of one minute, of which the first 30 seconds was with eyes closed (EC) and the latter 30 seconds with eyes open (EO). Following this, a stimulation paradigm ensued, during which subjects were exposed to ISF at nine integer frequencies from 36–44 Hz with no concomitant task. The experiment followed a randomised complete block design as sketched in Fig. 1, where the order of stimulus frequencies was randomised within each of the five blocks, resulting in a total of 45 stimulation trials per subject. Trials were separated by a three second inter-trial-interval without stimulation (light off) to avoid leakage between them. The duration between blocks was identical to the inter-trial-interval, tying the five blocks into one continuous run. When including the baseline EO and EC trials, the total number of trials per subject was 49, resulting in a grand total of 980 trials. The total duration of the experiment was about 29 minutes.

With exception of the varying temporal modulation frequency, the light specifications of the ISF stimulation was similar to the parameters presented in [20] and used in [17, 18] with a luminance of . The waveform was controlled with pulse-width modulation (PWM) operating at and a modulation depth of 100%. The PWM dimming values were defined individually for each of the coloured LEDs in two different sets to achieve the two colour profiles needed to fuse the alternating phases of ISF light. The individual and combined spectra of the the two phases of ISF and the 40 Hz temporal mixing are shown in Fig.2. While the term invisible spectral flicker was coined in [20] to reflect a subtype of chromatic flicker with reduced flicker fusion frequency, no tests were conducted in to estimate the participants’ perceptibility of the flicker as a function of frequency.

EEG processing

Processing of the EEG was done using custom scripts written in Python (version 3.8) using the open-source MEG+EEG analysis package MNE-Python (version 1.3.0). All recordings were manually inspected and marked for periods with high-amplitude noise. All trials containing excessively noisy segments were removed, reducing the total number of trials from 980 to 597. This high number of rejections resulted in slight imbalance of the factors and non-completion among the subject-stimulus combinations. As the noise typically was persistent over a couple of consecutive trials, sub-segmenting and rejection of shorter durations of bad data did not improve the fraction of rejected data.

Each recording was filtered with a finite impulse response (FIR) Hamming notch filter at 50 Hz with 0.5 Hz band width and 0.5 Hz transition bandwidth. Subsequently, they were highpass filtered with an FIR Hamming filter with lower cutoff of 1 Hz and transition bandwidth of 1 Hz. No lowpass filtering took place. Finally, the channels were re-referenced to a common average.

From the pre-processed EEG data, the power spectral density (PSD) of each channel m in trial n was estimated by a 30 s periodogram given by (Eq 1):

(1)

in which L is the number of samples in the trial, j is the unit imaginary number, f_s = 500 Hz is the sampling frequency, and Hz is the Nyquist frequency. The PSDs of the occipital channels (O1, O2) were averaged according to (Eq 2) to obtain the mean occipital PSD for trial n:

(2)

The signal-to-noise ratio (SNR) was calculated as the power at the stimulus frequency, f_n, for trial n normalised to the average power in the neighbouring bands (frequency bins within Hz, but excluding the nearest Hz) according to (Eq 3) and sketched in Fig. 3:

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Estimating signal-to-noise ratio.

The SNR of a given trial, n, was estimated as the occipital power at the stimulation frequency, f_n, normalised to the noise defined as the average power in the neighbouring bands (see Eqs 1 to (5).

https://doi.org/10.1371/journal.pone.0321633.g003

(3)

where the denominator noise term given by (Eq 4) estimates the mean power on either side of f_n within the trial:

(4)

where is the Hadamard product, and k(f,f_n) is the kernel that selects the noise neighbourhood according to (Eq 5):

(5)

For the baseline condition, in which no stimulation took place, the SNR was calculated at 40 Hz under the assumption that it according to Eq 3 would be 0 (or at least equivalent) for any frequency relative to its neighbours. This assumption is fair as the PSD in the baseline conditions at rest is primarily aperiodic and thus follows approximately a 1/f^a shape.

Statistial analysis

The relationship between ISF stimulation frequency (stimulus) and occipital response SNR was investigated using a linear mixed-effects model. The initial model used SNR as the dependent variable and three main factors as independent variables: stimulus, subject, and block, as well as all their two-way interactions. The stimulus frequency was considered a fixed factor, while the subjects and blocks were modelled as random effects with variable intercepts.

The initial linear mixed effects model of the experiment was described by (Eq 6), in which all terms were present prior to model reduction. Here, refers to the SNR response to the trial with stimulus i, subject j, and block k:

(6)

where

and , a, and b are the stimulus, subject, and block main effects, respectively, while c, d, and f are the subject-block, stimulus-block, and stimulus-subject interaction effects, respectively, and denotes a Gaussian distribution with mean and variance . Note that the 0th index for stimulus and block correspond to the baseline condition.

We implemented the initial model defined by (Eq 6) in the statistical computing software R (version 4.3.0) using the ‘lme4‘ package (version 1.1-33) [39] and fit the model to the data using a restricted maximum likelihood. The initial model fit was checked for variance homogeneity, normality of residuals and random effects, independence of residuals, and uncaptured residual structure. Finally, the influence of single observations was checked by means of the Cook’s distance. One observation was found to be particularly influential (). However, through re-analysis removing this observation from the data, it was confirmed that the significant factors were unchanged (though the estimated mean SNR for the baseline condition was closer to the expected value of 0 in the re-analysis without this observation). Ultimately, the observation was included in the final analysis, as no reason was found to exclude it. One subject had SSVEP responses consistently an order of magnitude higher than the rest, so the influence of this subject on the model was tested. We fitted a corresponding model to the data without that subject to check its influence and verify that the significant terms did not change. As exclusion of the subject did not alter the significant terms, the subject was ultimately included in the analysis.

The random effects structure of our model was simplified by way of likelihood ratio tests which test the null hypothesis against the alternative , where was a given random effect variance. We adjusted the p-values by a factor 1/2 to account for testing on the boundary of the positive variance constraint. Non-significant random effects were eliminated in an iterative manner, where the effect with the highest p-value above was removed, and the reduced model was re-fitted. With this approach, the block main factor was found to be non-significant (, see Table 1), but it was kept in the model, as it was both an integral part of the experimental design and involved in significant interactions. The stimulus-block interaction was not significant and was removed from the model.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Initial model factor reduction.

https://doi.org/10.1371/journal.pone.0321633.t001

Factors in the initial models are tested for significance to reduce model complexity using likelihood ratio tests (LRT) for random factors and maximum likelihood estimates for the fixed effect. Degress of freedom for a given factor is denoted DF. The Block factor was not significant, but as it was an integral part of the design and is implicit in significant interactions, it was decided to keep it in the model. The Block-Stimulus interaction was subsequently removed. As the LRT is a test on the boundary on the positive variance constraint, the p-values for random factors were corrected by a factor 1/2.

We proceeded to investigate the fixed effects structure by maximum likelihood estimation and found that the stimulus fixed effect was significant, thus deciding on Eq 7 as our final model to summarise the results:

(7)

where

is the fixed stimulus effect, a and b are the random subject and block main effects, respectively, while c and f are the random subject-block and stimulus-subject interaction effects, respectively. The model assumptions were re-assessed for the final model by the same criteria as the initial model to verify that model reduction had not led to violation thereof.

In post-hoc analysis, the stimuli were compared in a pair-wise manner to each other and the baseline SNR estimates using paired t-tests with the Kenward-Roger degrees of freedom (DF) approximation. The p-values were adjusted by the Tukey method for a family of 45 hypotheses when comparing all stimuli, while they were corrected using Dunnett’s test when comparing them to the baseline.

Limitations of the study

Importantly, this study investigated the acutely evoked power during stimulation, and not the clinical implications for people with AD. It is currently unknown to which degree (if any) the evoked power is associated with clinical efficacy. This uncertainty has been addressed previously [45] and will require retrospective comparisons of clinical trial outcomes.

This study was conducted in a healthy and conveniently sampled cohort from the Copenhagen area of Denmark. Thus, interpretation of the results is limited by the homogeneity of the sample. This is particularly important in the context of interventions for AD, as our sample is demographically dissimilar to that population.

Our results clearly indicate an ability to evoke potentials at all stimulation frequencies, but the study was not powered to discern differences in evoked power between frequencies.