Most of us have experienced the “tip of the tongue” feeling of not quite recalling a name, a face, or items from memory. As part of normal brain and cognitive aging, such “senior moments” increase, especially under stress. The lure of brain training is the possibility to directly enhance neural activities associated with good memory and therefore reduce the moments of memory lapses. A central component of cognitive ability is working memory, i.e., the capacity to hold active information in one’s memory for immediate manipulation. By definition, working memory shares a large amount with other core cognitive functions (e.g., cognitive control and attention), which are strongly associated with performance in intelligence tests (Chen et al., 2019). Measurements of neural activity have become strong predictors of mild cognitive impairments in persons with various kinds of cognitive deficits. Electrophysiological changes during working memory are some of the earliest signals in preclinical risk of mild cognitive impairment (Jiang et al., 2021).

Neurofeedback (NF) is biofeedback that uses real-time sensory reward for brain activity associated with certain performance (e.g., accurate and fast recall). In contrast, traditional cognitive rehabilitation typically applies offline behavioral reward (e.g., money). Growing evidence supports the idea that the ultimate biofeedback is to reward with sensory pleasure (e.g., enhanced visual clarity) in real-time to neural circuits that are associated with a desired performance, such as excellent memory (YuLeung To et al., 2016) and attention (deBettencourt et al., 2015, 2019). Using functional neuroimaging, brain regions that have been consistently implicated in neurofeedback or real-time self-modulation include the anterior insula (cognitive control, self-awareness), and the basal ganglia (sensory motor integration, implicit learning), independent of the targeted region-of-interest (Emmert et al., 2016). There are different networks involved in neurofeedback in cognitive and affective functions (Trambaiolli et al., 2021); this is seen particularly in fMRI-based studies that allow better spatial definition of the structures involved during cognition and affective processes (Paret et al., 2019; Skottnik et al., 2019; Direito et al., 2021).

Since fMRI-based (blood-and-oxygen-level dependent) signals are hundreds of times slower than electrophysiological signals, there is a strong case to be made for the advantages of using electrophysiological signals (Ros et al., 2013, 2014) measured by methods such as electroencephalography (EEG) or magnetoencephalography (MEG). In addition to high temporal resolution, human scalp EEG is non-invasive, affordable, with no movement restriction, and is more suitable to large-scale applications. EEG is a tool that measures the summations of neural postsynaptic potentials at the scalp. EEG-based NF has been successful in enhancing attention in older adults (Jiang et al., 2017). Using a double-blind controlled design in the older brains, seminal work by Angelakis et al. (2007) applied EEG NF in an older population and showed improved processing speed and executive functions. Additional success has been reported using EEG-based NF training in older dementia patients (Surmeli et al., 2016).

Despite exciting progress using NF in brain training to improve memory in young and older adults, most of the NF training studies thus far have been limited to traditional frequencies. NF training using event-related potentials (ERPs), i.e., averaged EEG signals associated with cognitive events, is currently lacking in the literature.

So, once the field gains new knowledge of neuromarkers of specific neurophysiological functions, then neurofeedback training should be better targeted beyond the single frequency approach. The new methods should include frequency interactions and event-related potentials. The challenges include both neurological understanding as well as technical difficulties for real-time closed-loop feedback (Sitaram et al., 2017).

Very few studies have applied memory-related potentials directly in the context of preclinical and clinical Alzheimer’s pathology. ERPs are a well-studied approach for indexing brain responses associated with memory and cognition. These electrophysiological outcome measures can predict individual risk of mild cognitive impairment 5 years before diagnosis (Jiang et al., 2021).

Neurofeedback has been used to improve cognitive and physical performance of humans (Daly and Wolpaw, 2008; Pfurtscheller et al., 2008; Machado et al., 2013; Broccard et al., 2014; Chaudhary et al., 2016). While the efficacy of these methods differs, all have been reported to enhance performance by training attention (another core function of cognition) and working memory tasks. However, evidence for improvement in everyday life utilizing cognition is limited, which provides the impetus for developing better and time-efficient methods to directly train neural processes underlying attention. A closed-loop BCI system can be designed to directly control external devices. Also, it can be utilized as a neurofeedback platform in neurorehabilitation programs to improve and enhance the cognitive abilities (Chaudhary et al., 2016).

While there is fundamental disparity between neurofeedback and BCI (Abiri et al., 2019, 2020), combining both approaches can potentially help to expand achievements of each individual approach.

For instance, the goal in motor-imagery BCIs is to control an external object by inducing and modulating the brainwaves of interest during the training session so that the BCI system can determine the user’s intention in real-time in testing sessions. Users must learn how to regulate their brainwaves which then makes the BCI analogous to neurofeedback. In both, users must learn to modulate their brain activity to acquire a certain feedback, specific output, or greater reward. The similarity may result in triggering similar cognitive and neural processes. Thus, the training process of BCI may expand our knowledge of neurofeedback.

Most of the neurofeedback studies have utilized a few channels of EEG signals or a fixed and narrow range of frequencies (Gruzelier, 2014a,b) while most of the BCI systems, such as motor-imagery BCIs have incorporated multiple EEG channels and individualized and adaptable ranges of frequencies determined by signal processing, machine learning, or deep learning approaches. Embracing those approaches in neurofeedback regiments may help to distinguish and extract individualized features from EEG signals with which we may obtain higher efficacy compared to having a fixed channels and fixed frequency approach (Ang et al., 2012).

There is also research about task-unspecific internal and external confounding variables characterized and studied in the BCI literature. For instance, the attentional state of the user has been quantified with the level of mu and gamma bands in different brain areas. The measure has helped to enumerate the level of attention and motivation in a motor-imagery BCI task (Grosse-Wentrup et al., 2011). Having those measures alongside the neurofeedback therapy can be beneficial to assess how likely a user is motivated to participate in neurofeedback therapy and how one neurofeedback regiment may be adapted and accepted by the users over another one.

The type and form of feedback signal is another instrumental element in neurofeedback. We have five main senses: visual, auditory, somatosensory, olfactory, and gustatory. Although vision is the most prominent sensory feedback in humans compared to other primates, other sensory feedbacks may play a significant role in learning and modulating our brainwaves in response to different phenomena.

There are yet more challenges to applying neurofeedback. One challenge is how to define the reinforcement signals. It might be easier and ethically more appropriate to define the reinforcement signal in non-human experiments. For example, withholding food from a rat and providing it later as a reward when the animal successfully modulates a neurophysiological signal may not be an ethical concern (at least now), but it would be more challenging and clearly unethical to deprive human subjects of food. Another challenge is that there is no guarantee that a human subject would interpret the feedback as a reward. So, the motivational state of the subject results in different interpretation of the feedback signal.

Clinical neurofeedback training has been shown to be an effective treatment for people with a wide range of deficits, including epilepsy (Sterman and Egner, 2006; Strehl et al., 2014), attention deficit hyperactivity disorder (ADHD) (Arns et al., 2009), stroke (Rayegani et al., 2014), autistic spectrum disorder (ASD) (Kouijzer et al., 2009), emotional disorders (Reiter et al., 2016), and tinnitus (Hartmann et al., 2014). Furthermore, neurofeedback has been evaluated as a means to enhance cognitive control in healthy people (Zoefel et al., 2011; Wang and Hsieh, 2013; Gruzelier, 2014a,b). The quality and quantity of changes (effect size) due to the neurofeedback regimen can be enumerated with behavioral responses, measured by cognitive scoring questionnaires, or demonstrated by neural measures of cognition.

The basic EEG components in human adults are the delta, alpha, theta, beta, and gamma frequency bands. Alpha frequency during eyes-closed resting EEG has emerged as an important biomarker in mild cognitive impairment and dementia (Babiloni et al., 2021a; Cecchetti et al., 2021; (Stoiljkovic et al., 2021). Alpha oscillations (8–12 Hz), sourced in frontal sites including the anterior cingulate cortex, are related to working memory and related performance in humans. This EEG wave is the most common wave seen during awake states (Adrian and Matthews, 1934). Global alpha power is more abnormal in younger MCI and AD patients (Babiloni et al., 2021b). Posterior alpha sources are more abnormal in male MCI/AD patients (Babiloni et al., 2021c).

The EEG theta rhythm (4–8 Hz) is also related to memory performance. Theta rhythms have been shown to play a role in encoding episodic memories and are also correlated with behavioral performance (Hasselmo and Stern, 2014). EEG delta rhythms (0.5–3.5 Hz) are somewhat different than the other three as an increase in delta band activity leads to inhibition (Harmony, 2013). This inhibition is useful however, as it blocks interference from other centers in the brain, allowing one to concentrate on the mental task at hand. Delta waves are prominent in deep stages of sleep. Finally, beta rhythms (13–35 Hz) are related to brain consciousness and motor functions. Both alpha and beta waves are common during awake states, but beta waves are also present in states of drowsiness (Nayak and Anilkumar, 2020). Finally, gamma rhythms (30–80 Hz) are higher frequency EEG rhythms that have demonstrated a role in visual and sensory processing. Due to the low amplitude of the gamma rhythms, these EEG frequencies have been historically difficult to measure but have recently been linked to other rhythms and functions. However, rapidly growing new literature on gamma rhythms and its underlying biological basis is fascinating (e.g., Milekovic et al., 2019). NF using high frequency components are a new frontier needing research.

Increasing evidence suggests the coupling of two different rhythms is also used to perform specific functions. Coupling of the brain’s theta-gamma rhythms (θ–γ) has demonstrated a link between the cortical regions (prefrontal areas and the cingulate cortex) which plays a major role in working memory (Schack and Klimesch, 2002). Delta-theta coupling has been shown to be important in decision making and cognitive processing (Adams et al., 2019). Interestingly, changes in theta-gamma coupling have been shown to be earlier neuromarkers than traditional AD biomarkers. AD and MCI patients demonstrate the lowest level of θ–γ coupling in a verbal working memory task in comparison with healthy participants (Goodman et al., 2018). An impairment of θ–γ coupling that increases in parallel to the progression of the MCI has been recently reported in human patients (Musaeus et al., 2020). These new findings suggest that coupling of brain oscillations is critical for proper cognitive functioning and will likely be another neuromarker for applying NF treatment.

To determine the most successful way of performing a NF study, samples of previous studies of successful and unsuccessful studies were analyzed. Twelve studies involving older adults (> 60 years old) and eight involving younger adults (< 35 years old) were selected for using several different parameters with varied success reported in each study. Investigating the differences seen between different age groups, however, provided important insight into the quality of studies performed in the past. As much of human cognitive development occurs before the age of twelve, there is a rational fear that once a body reaches a certain stage, mental growth and training is impossible. Although there is increasing evidence that the brain remains plastic past its “peak years,” the physiological degeneration due to brain aging is an obstacle to training older adults with NF (Mattson and Arumugam, 2018).

After a review of successful past studies, it was demonstrated that the aging brain can in fact be trained and improved, despite the anatomic and pathological limitations faced. There were seven studies involving younger individuals investigated, five of which were successful in using NF to train the brain. Of the eleven studies involving patients over sixty, nine were successful in training the brain with NF, two of which involved patients with MCI. This demonstrates that not only can the healthy aging brain be trained successfully with NF, but a degenerating aging brain can be trained as well. Although there are many different factors that potentially change the results in these studies, it is promising to see evidence that the aging brain can still be changed and improved upon, as this provides an avenue into treating more severe diseases such as Alzheimer’s disease. The factors that potentially confound these studies are compared below.

Each study used a slightly different method in controlling for their BCI group, so it was important to investigate the different controls used (Nicolas-Alonso and Gomez-Gil, 2012; Sorger et al., 2019). Between the studies that used older and younger populations, there were three common types of control groups used: a “waitlist” control, a sham NF control, and a group that received cognitive training in place of the NF (some used a combination of these or no control at all, but these were the most common occurring groups). The waitlist control groups received no training between the first and last testing sessions. The sham NF groups received the NF training sessions but did not receive the NF protocol to change any specific EEG band. The cognitive training groups used other cognitive training methods such as brain teasers and memory games in place of the NF. Control groups will be discussed in much more detail later in this review.

Two primary modalities used in the NF studies were investigated: visual and auditory. A majority of the studies utilized a visual NF protocol. In the projects that used visual feedback (Krause et al., 2017), a visual reward such as a motivating progress bar or pleasing image was given upon activating the specific EEG band. The auditory feedback was similar except it used a pleasing tone for achieving the goal range and an unpleasant tone for missing the EEG power range. The combination study was unsuccessful at achieving its goal. The fifteen visual modality studies were successful where one out of the two auditory modality studies were successful. More research needs to be performed to determine whether auditory only NF protocols are useful in training the brain, as two studies are not a large enough sample size to draw a conclusion.

Follow up time was the major limitation found throughout each study included in this review. Only one study reported longitudinal findings (Marlats et al., 2020). This group had participants return and re-test after 1 month. They demonstrated that scores for several different memory tasks remained elevated from preliminary testing after receiving NF training. Follow up is vital to the integrity of any study like this as it demonstrates the clinical power EEG NF training can have on those with declining cognitive function. Having at least a follow up after 1 month should be an integral portion of any NF study moving forward; continuing follow ups through the year will increase the power of the studies by showing prolonged cognitive improvements.

Another component to the follow up that requires consideration is whether there is a need for re-training. Spaced repetition is a useful technique for memorizing large amounts of information and could also be a useful tool when considering the long-lasting effects of EEG based NF. Doing a certain number of sessions throughout a year could prove to have even more benefits than performing 10–20 sessions within the span of 1 month. This, as well as the vitality of the training during follow up sessions, will need to be investigated in the future.

Neurofeedback did not work in some studies listed in Tables 1, 2. We investigated individual differences in brain signals for answers. Strong individual differences, including but not limited to genetic and environmental differences, have been reported in working memory and performance (Parasuraman and Jiang, 2012), and how cognitive training effects are preserved and learning is transferred (Greenwood and Parasuraman, 2015). Learning transfer means in cognitive training in one domain (working memory) can transfer in similar tasks (short-transfer) and untrained tasks (long-transfer). However, there are strong individual differences in applying learning transfer to important daily tasks (Sinotte and Coelho, 2007; Westerberg et al., 2007; Cicerone et al., 2011; Kuo et al., 2014). Attention, mood, and motivation are important factors (Kadosh and Staunton, 2019).

Cognitive aging is caused by synaptic, metabolic, and structural changes during brain aging which slowly lead to loss of full cognitive function. EEG measures synchronized synaptic functions and network at the scalp. Recent work has identified neurosynaptic changes as one of the earliest biomarkers of preclinical AD, appearing before onset of tau-mediated neuronal injury or brain structure changes (Jack et al., 2011; Sperling et al., 2011). Among the most common early symptoms of dementia are deficits in working memory. The exact neural mechanisms subserving working memory are under debate (Mi et al., 2017). There are three types of hypotheses: First, the Frequency model suggested that WM involves periodic reactivation of memory representations at each gamma cycle within gamma-theta nested oscillations in the hippocampus, mediated by slow after depolarizations with a time constant that should be matched to the theta period. Assuming each memory is activated exactly once during a theta burst cycle, the same memories are repeatedly reactivated over subsequent theta cycles (Lisman and Idiart, 1995; Siegel et al., 2009). Deficits in higher frequency such as gamma rhythms are associated with cognitive deficits. Therefore, brain stimulation studies (Chan et al., 2021; Mimenza-Alvarado et al., 2021), applying gamma frequency treatment in mild cognitive impairment and mild Alzheimer’s disease, set foundations for future neurofeedback applying gamma rhythm.

The question remains why manipulating lower frequencies such as delta and alpha improves working memory performance. New evidence suggests that glial astrocyte reactivity and Ca + channel related slow waves altered in the brains of persons with dementia might be the cause of the low frequency alterations (Sardinha et al., 2017). Astrocytic signaling supports hippocampal-prefrontal theta synchronization and cognitive function (Lenk et al., 2020).

Second, the persistent firing model hypothesizes that working memory is supported by persistent activity of certain neurons active in the prefrontal cortex, hippocampus, parietal, and other cortical and subcortical networks, in the absence of direct perceptual stimulations. For instance, enhanced EEG activity during memory retrieval is observed in frontal sites, which are different from visually evoked potentials during perception. The attractiveness of this model is that it explains the memory network supporting working memory functions via synchronized oscillations at the same frequency in different brain regions. Recent evidence in animal models and with neuroimaging also points to brain connectivity networks as novel neuro-markers for indexing early deficits in AD risk. Aß peptides disrupt neural activity at the synaptic level and induce aberrant activity patterns in neural network circuits within and between brain regions in animal models (Palop and Mucke, 2010). Patterns of functional brain connectivity in humans are highly predictive of cognitive performance (Hachinski et al., 2006; Finn et al., 2015). The blood-oxygen-dependent level brain signals measured by functional MRI show that brain connectivity (particularly bilateral parietal and frontal-temporal) correlates with CSF AD biomarkers (Aß and tau), particularly during working memory tasks (Jiang et al., 2016). Overall synaptic synchronization during an intrinsic spontaneous resting state measured by EEG output has also been linked to neuroinflammation (astrocytes reactivity and calcium signaling; Wang et al., 2021), which provides a potential treatment direction.

Third, Synaptic theory of working memory is a synaptic-based theory for short-term information storage in neural circuits (Mongillo et al., 2008; Lundqvist et al., 2011). In this model, memory is retained by an item-specific pattern of synaptic facilitation. This mechanism does not require neurons to fire with elevated rates for the whole duration of the memory task, resulting in a robust and metabolically more efficient scheme (Mi et al., 2017). Resting EEG and network oscillations are correlated to working memory performance (Borhani et al., 2019) and cognitive impairment (McBride et al., 2013, 2014, 2015).

Finally, the Neurofeedback to frontal and pre-frontal cortices enhance goal-directed behaviors and executive functions. The goal-directed behaviors are strongly associated with the brain activity in the prefrontal cortex (Miller and Cohen, 2001). Executive functions encompass the mental processes that allow individuals to take control over automatic responses of the brain to produce goal-oriented behaviors (Garon et al., 2008). The prefrontal cortex functions comprise planning, goal setting, decision making, voluntary attention, task switching, set shifting, behavioral and perceptual inhibitions, voluntary regulation, and error correction. In therapeutical neurofeedback, especially the ones that target cognitive control, some of the functions seem to be fundamental to set up intrinsic reward, to integrate feedback information, and to self-regulate behavior. Most of those functions interact with attention, a broad concept that can be defined as the set of processes dealing with the allocation of working memory to the different neural representations available in the brain (Knudsen, 2007). Attention plays a critical role in allocating the brain resources to working memory while high-level cognition relies on working memory to learn tasks (Cowan et al., 2005). There are studies indicating a shared neural mechanism that supports both attention and working memory (Ikkai and Curtis, 2011; Gazzaley and Nobre, 2012). The shared neural mechanisms support the theory that there is also a mutual connection between the two brain functions. Thus, event-related (memory and/or attention) brainwave patterns should be targets of NF training.

Electroencephalography recordings directly measure post-synaptic potentials. Recording the averaged EEG signals, i.e., event-related potentials (ERP), during cognitive events known as cognitive ERP is a promising but less often investigated approach for indexing brain mechanisms underlying cognition and memory (Olichney et al., 2008, 2011, 2013; Li et al., 2017; Jiang et al., 2021). Memory-related neuromarkers are sensitive to general cognitive decline before conventional biomarkers of AD can be detected by CSF/PET methods and behavioral performance changes.

Using longitudinal follow-up of healthy older adults over 10 years, Jiang et al. (2021) revealed that the different brainwaves between working memory responses to Targets and Non-targets in three left frontal electrodes (F7, F5, or F3) or averaged from three left frontal sites are the best predictor for diagnosis of cognitive impairment 5 years before diagnosis. Intriguingly, memory non-targets or distractors did not predict well.

A participant wearing a wireless EEG headset performs a visual memory task (Figure 1). Recorded EEG signals are analyzed in real-time, e.g., EEG wavelets analysis (frequency, power, and time) or ERP memory-related potentials. Neurofeedback is given by rewarding brain patterns that are associated with accurate and fast memory retrieval.

Existing studies on ERP modulation using NF have revealed insights into designing reward functions. These studies mainly investigate the decoding method of single-trial responses (Brandmeyer et al., 2013) as well as the neural mechanism underling reward stimuli (Moser et al., 2014; Zioga et al., 2019; Wang et al., 2021). Research in the literature demonstrates broad applications of ERP-based NF training in cognitive deficits, including Attention-deficit/hyperactivity disorder (ADHD), Post-traumatic stress disorder (PTSD), and Subjective cognitive decline (SCD) (Askovic et al., 2020; Pei et al., 2020; Deiber et al., 2021). The main challenge of developing NF training using ERP is designing reward functions associated with components of ERP as the adjustment target in single-trial responses. In a recent study, Wang et al. (2021) compared the change of event-related desynchronization (ERD) power attenuation using ERP-based NF vs. motor imagery (MI)-based NF and showed that only ERP based study presents significant differences in the frequency bands of EEG signals between the neurofeedback and non-feedback groups.

Important for clinical applications of neurofeedback, we reviewed clinical trials related to mild cognitive impairment and dementia using NF training around the world. Table 5 includes studies using only neurofeedback to improve cognitive ability in older adults, in Alzheimer’s disease, and related dementia (AD/ADRD). Importantly, there are different EEG and ERP variation in MCI induced by subtypes of ADRD (Güntekin et al., 2021). Note three times more neurofeedback trials are being conducted to improve mood, depression, and attention at the clinicaltrials.gov site that are not included in the current review.

Most current Neurofeedback training improves memory and cognition to a certain extent. Large scale and optimal clinical applications of neurofeedback are limited by several challenges. Debate remains about the frequency and length of optimal NF effects, outcome measures, and long-term effects.

A well-known issue in such training is that some people simply do not respond to neurofeedback. Thus, we also reviewed the literature of individual differences in placebo effects and non-responses. Future work needs to focus on an individual based approach.

Sleep EEG has also emerged as a very important biomarker (Whitehurst et al., 2020), especially for outcome measures in neurofeedback studies in the future (Alfini et al., 2020). D’Atri et al. (2021) reported the importance of time of day showing differences in Delta and alpha signals during PM and AM wake studies (lack of difference in patients with cognitive impairment and AD). Rapid Eye Movement (REM) EEG slowing during sleep showed the strongest correlation with cognitive decline than other wake EEG in their study. Sleep EEG also has stable signals. If daytime nap EEG can be managed (Alger et al., 2012), it should be an ideal outcome measure for effectiveness in NF training.

The range of EEG features for NF training are as follows:

Our review brings overall good news of the potential effectiveness of neurofeedback brain training. Most training studies typically take multiple weeks (e.g., 5–20 weeks) with 2–3 sessions per week. We review various neurofeedback reward strategies (visual and auditory methods) and outcome measures. Our review recognizes that individuals’ neural responses are on a continuous spectrum. We recommend that “neural modulation sensitivity” instead of “BCI illiteracy” is to be considered as the preferred terminology in brain training. Future directions include much needed research in mild cognitive impairment, in non-Alzheimer’s dementia populations, and neurofeedback using EEG features during resting and sleep for memory enhancement and as sensitive outcome measures.

YJ wrote the first draft and oversaw all aspects of the revisions. WJ wrote the reviews summarized in Tables 1, 2. SH wrote the reviews summarized in Tables 3, 4. SB wrote the closed-loop real-time NF section. ZL wrote the related review summarized in Table 5. XZ conceptualized and wrote the BCI portion. LP and WH contributed to conceptualization and interpretation of EEG-based NF training. JS and SC-S contributed to conceptualization of clinical implications, critical revision, overall edits, and to Figure 1.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.