The search was performed on Pubmed and bioRxiv with the keywords “music” and “neurofeedback” or “BCI,” finding 103 papers and 112 preprints published from January 2001 up to September 2024. The last search update was made on October 7th, 2024. Additionally, the references of the papers that were selected for full-text eligibility assessment were screened to retrieve further relevant publications.

The title, objective, neuroimaging technique, and eligibility verdict of the identified articles were recorded on a spreadsheet by AS. Published papers in English were initially filtered by reading the abstract to identify which presented original data, excluding reviews, editorial notes, and conference proceedings, and which implemented a neurofeedback experiment (excluding, for instance, references to neurofeedback in future work).

The following steps were performed by two independent reviewers (AS and BD), each considering half of the records. The decision was based on the input of the two reviewers, while any discrepancies in judgments of risk of bias or justifications for judgments were resolved by discussion to reach a consensus between the two reviewers. A full-text analysis excluded papers that measured only behavioral ratings or that did not include music in the neurofeedback loop. Studies that provided auditory feedback of nonmusical sounds (e.g., sea waves, applause) were also excluded. After this, the references of the selected papers were checked, adding three papers to the list.

The final selection includes 15 studies that fulfill all the requirements (Supplementary Table S1). For these, we extracted the following fields—Title, Authors, Year, Music-specific hypothesis/motivation, Paradigm and music characteristics explored, Changes in NF target region/network and neural correlates of music, Main objective, Neuroimaging technique, Methods summary, Results summary, Main conclusion, Total number of subjects, Music used, Music features, Number of NF sessions, Control groups/conditions, Number of EEG channels/ROIs, Location of music in the loop, Control group type, Music-NF impact, Number of sessions with active feedback, Number of participants in the active group.

The main motivation for using music in the context of neurofeedback is explicitly referred to by six of the studies analyzed here: its ability to evoke emotions in the listener can be explored for emotion regulation (Ehrlich et al., 2019; Pino, 2022; Fedotchev, 2018; Ramirez et al., 2015; Pino and La Ragione, 2016; Daly et al., 2016). However, it is unclear if there are specific hypotheses for the mechanism, brain structures, or networks involved that allow music to act on emotion regulation, particularly neurofeedback.

We found clear associations between music and the study objectives. In Takabatake et al. (2021), the objective was to use classical music with different degrees of superimposed white noise to modulate the power of a specific EEG frequency band (in this case, the alpha band as measured in frontal areas) to the point at which some cognitive improvement (namely in a short-term memory test) is achieved. The authors linked the use of music in this NF context with previous evidence from alpha power training methodologies. In a proof-of-concept clinical study (Keller and Garbacenkaite, 2015), the authors mention that for patients with unresponsive awareness syndrome (UWS, a consequence of severe brain injuries), the awareness in the auditory domain was greater than in other domains, inviting the use of auditory stimulation for therapeutic approaches. In this study, three patients listened to favorite songs during a neurofeedback intervention that aimed to determine if they were in some way able to alter their brain activity. In Cordes et al. (2015), the authors recruited patients with schizophrenia and proposed an NF paradigm to enable them to control the activity of their anterior cingulate cortex (ACC), known to be dysfunctional in schizophrenia. Without any previous instruction by the experimenters regarding the mental strategy, the clinical group tended to use the imagery of music as the NF strategy, while the control group tended to use the imagery of sports.

The exploration of music's potential in regulating emotions was also evident in clinical contexts linked to emotion disorders. The pilot study by Ramirez et al. (2015) aimed to allow elderly participants with depression to control two parameters of a set of music performances using their own emotional state, as measured via EEG probes. In Fedotchev (2018), the NF experiment was designed to reduce stress-induced functional disturbances in people who had specialized jobs associated with high workloads.

In our systematic review, we identified two distinct categories of methodological solutions that we report in the following sections. In ten cases, music was utilized as feedback directly, whereas, in the other five, another feedback modality was used and music served as a tool to help participants perform during the NF experiment (Figure 5A). At the end, we also address the matter of control groups in these studies.

The majority of the studies analyzed here report successful results of the intervention or of BCI control achievement. Importantly, we found no reporting of results regarding the use of music as the interface of the neurofeedback—the specificity of the musical interface vs. another type of interface remains untested. We identify success measures in three different domains: modulation of the target brain signal, changes in behavioral/clinical scores, and changes in neural patterns (Figure 6). The number of domains discussed and the level of detail varies among studies. Lastly, we also address the studies that do not associate music feedback with a specific brain signal but rather use music as a complement in the feedback loop.

Five studies report changes in both behavioral/clinical and imaging outcomes. At the end of the intervention of Ramirez et al. (2015), the behavioral depression score applied showed an overall improvement, and the EEG data analysis revealed a decrease in relative alpha activity on the left frontal lobe. While frontal alpha asymmetry is often suggested as a depression biomarker, its clinical diagnostic value is limited, with cross-sectional analyses revealing significant heterogeneity (Van Der Vinne et al., 2017) and very low reproducibility (Kolodziej et al., 2021). Most of these studies were not preregistered, leaving the robustness of this finding yet to be confirmed, particularly considering the analytical flexibility and heterogeneity involved in EEG data analysis. In Fedotchev (2018), the occipital alpha rhythm power increased after the NF intervention session, but significantly so only when the feedback loop was turned on. Both treatment alternatives (with and without feedback) resulted in positive changes in psychological tests. Moreover, post-treatment reports demonstrate acceptance and value in the proposed musical interface, but no additional neural correlates were analyzed. The results presented by Keller and Garbacenkaite (2015) indicate that two out of the three UWS patients were able to drop the theta/beta ratio, measured at the Cz scalp location, during the intervention period, which was mainly a consequence of a decreased theta amplitude (the authors suggest that these results reveal a shift of the dominant rhythm into the alpha band and reflect some brain function recovery). The weekly assessment of the JFK Coma Recovery Scale-Revised (CRSR) increased specifically in the auditory function, motor function, and arousal subscales but, since the sample size was three, the authors only provide a descriptive analysis of the time courses. Nevertheless, this study provides the first evidence that NF can be used in patients with this specific syndrome, but also the potential of music in such an extreme case. In Takabatake et al. (2021), the data analysis from two groups of healthy participants (five subjects per group, receiving real or random feedback) revealed a significant difference in the alpha power, measured at the forehead, achieved during the 4 weeks of intervention when comparing real with random feedback. Cognitive functions were also assessed before and after the intervention (digit span test, standard verbal paired-associate learning test, simple calculation task, N-back test, and Test of Variables of Attention). The results were assessed based on a binary classification of responders and non-responders (based on the regression analysis of a success measure corresponding to the modulation of the target brain signal). The authors report an improvement in short-term memory in responders compared to non-responders. The crossover design and control conditions limit conclusions regarding the causality of the music interface. The training program described in Van Boxtel et al. (2024) was successful in increasing the alpha activity of the participants, as measured by EEG, but also in improving both their performance on task switching and mental rotation tasks and sleep duration, as the results of a self-reported questionnaire revealed.

The authors also report that the participants in the experimental group showed a significant improvement in their performance in a cognitive task compared to the control group. The results suggest that the training program was effective in increasing the alpha power in the occipital lobe of the participants, as measured by EEG. The authors also report that the participants in the experimental group showed a significant improvement in their performance in a cognitive task compared to the control group. The results suggest that the training program was effective in increasing the alpha power in the occipital lobe of the participants, as measured by EEG. The authors also report that the participants in the experimental group showed a significant improvement in their performance in a cognitive task compared to the control group.

In the distinct case of Dekker et al. (2014), the results do not show changes in imaging markers based on the EEG data: a positive change in the occipital alpha power was detected, but it was not significantly different between groups. However, behavioral questionnaires indicated improvements in the level of mental balance, focus ability, and thought control. Although objectively measuring changes in athletes' performance was unfeasible, this study was the first attempt to target the mental capacities of athletes, such as mental shape and focus, with a simple music-based neurofeedback training protocol.

In Ehrlich et al. (2019), changes were found only in the imaging outcomes: the efficacy of the music algorithm was tested on 11 participants in the first study, and the algorithm was embedded in a real-time BCI architecture to investigate affective closed-loop interactions in five participants in a second pilot study. The results suggest that participants were able to intentionally modulate the musical feedback by self-inducing emotions, indicating that the system can capture the listener's current affective state in real-time and potentially provide a tool for listeners to mediate their own emotions by interacting with music. The proposed concept offers a tool to study emotions in the loop, potentially shedding light on emotion-related brain research and clarifying the interactive, spatio-temporal dynamics underlying affective processing in the brain.

Four other studies report the development of a new implementation for music-based NF. The device of Deuel et al. (2017) has been tested on 15 novice users, and the results show that most were able to hit target notes with a level of accuracy significantly higher than random. The Encephalophone is a novel instrument that uses EEG control to create music, and the study suggests that with continued training, users could significantly improve their accuracy and ability to use the instrument. The article discusses the potential benefits of the Encephalophone for both music therapy and neurological rehabilitation. The instrument may be particularly useful for patients who have lost motor function due to conditions such as stroke or traumatic brain injury, as it allows them to generate music using different regions of the brain that are still functioning normally. Overall, the study suggests that the Encephalophone is a promising new technology that could serve as a movement-free musical instrument and as a therapeutic biofeedback device for patients with motor deficits.

Machine learning algorithms were used in two studies to assess imaging data in real time. In Daly et al. (2016), the authors found that the affective music-based BCI can change the majority of participants' affective states to make them happy, calm, or de-stressed. The selected features used by the affective state detection method contain both neuronal and physiological features, indicating the importance of physiological features in identifying affective states. The feasibility of the experiment of Lorenzetti et al. (2018) was proven, with the results showing distinct and relevant brain network fMRI activation for each of the two targeted emotions/trials (tenderness and anguish). As a limitation of the study, the authors present the fact that the same music excerpts were used for all participants. This decision did not account for inter-individual differences in music taste, an important factor that could have affected the intensity of the emotions experienced by each participant. While the behavioral ratings showed that the excerpts helped induce the proposed emotions, the use of personalized audio tracks could have been more effective, as the authors also recognize.

Lastly, Trost et al. (2024) found evidence that live music performances can be used in a neurofeedback system, as the results suggest that the participants' emotional responses to the live music performances were significantly influenced by the performers' ability to adjust the pleasantness of the music in real-time. While the implementation of novel experimental paradigms with live music performances is challenging, the study suggests, based on state-of-the-art connectivity analyzes, that the left amygdala is a key region (a hub) of a dense functional network that is especially sensitive to the emotional content of live music, a feature that could be further explored in future music-based NF interfaces.

Another three studies are based on multimodal approaches and do not link music feedback to brain signals. Pino (2022) combines a repetitive visual and auditory stimulation feedbacking individual's EEG signals as flicker light so that a continuous closed loop can be obtained. The authors showed differences in some sub-scales of the neuropsychological assessment [reduction in depressive symptoms (HAM-D), increase in cognitive function (IQ)] between the two groups, namely when comparing before and after the intervention. Cordes et al. (2015) found that different cognitive strategies were used during neurofeedback targeting the ACC by the two participant groups—patients with schizophrenia reported using the imagery of music. Specifically, imagery of music was used by eight patients and only four controls, while sports was applied by only three patients, but seven controls. However, evidence suggests that the different strategies did not contribute to the differences found in neural activation. Lastly, the authors of Leite et al. (2018) found that the performance of individuals playing their game was not significantly different between the version with background music and the other versions without it. Volunteers reported that background music was almost irrelevant but did not disturb them, suggesting that background sounds do not significantly impact the performance of individuals in this BCI scenario.

NF interfaces based on music share the conceptual understanding that NF training can support the regulation of functional and structural patterns, promoting neuroplasticity by operant conditioning (Chen et al., 2022; Lubianiker et al., 2019).

The neural correlates of NF training are often divided into two dimensions of NF paradigms, Strategy Execution and Feedback Processing. The Anterior Cingulate Cortex (ACC), the anterior Insula (aI), and the Basal Ganglia (BG) are brain areas associated with the neural mechanisms of NF-assisted self-regulation (Dewiputri et al., 2021). The ACC is functionally connected to the insula and is part of the salience network—a network associated with task-transitioning and monitoring linking cognition and emotion or interoception. The involvement of the BG in NF learning is a well-established concept, and the mechanism involves the dopaminergic pathway. The BG is a target of midbrain dopaminergic neurons, that convey reward prediction error—a neurophysiological signal that relates to how unexpected or surprising an outcome is. This signal is then relayed to cortical structures that evaluate and reinforce or punish behaviors (Skottnik et al., 2019; Paret et al., 2018; Emmert et al., 2016).

The specific mechanisms behind music-based NF interfaces may be related to the emotional impact of music and its capacity to evoke intense pleasure responses. Music-induced pleasantness is directly linked to musical surprises, as it explores reward-related predictive processes, via recruitment of the mesolimbic system (feedback monitoring component) and its connections with the auditory cortex (sensory component) (Shany et al., 2019; Salimpoor et al., 2013). The patterns and pattern variations (surprise) associated with music stimuli elicit prediction errors and reward prediction errors, which trigger pleasure-related neural networks and the release of dopamine in the reward centers of the brain (Koelsch, 2020; Singer et al., 2016; Ferreri et al., 2019; Salimpoor et al., 2011). In Singer et al. (2016), a link was found between the activation of limbic regions, such as the amygdala and the hippocampus, and the affect and variations in temporal information in music. In Trost et al. (2024), the authors found that the amygdala acts as a central node in a dense functional network that is triggered while listening to live music.

Over the years, several mechanisms have been proposed to mediate the link between music and emotion (e.g., Juslin and Västfjäll, 2008), and various musical features and attributes have been highlighted in determining the affective responses to music, such as tempo and loudness' role in arousal, and the musical mode and dissonance level in the level of pleasantness. Our recent study (Sayal et al., 2024) found evidence for a correlation between distinct musical features (notably expressive features such as vibrato and tonal and spectral dissonance) and the brain networks of valence, arousal, and reward. Intriguingly, few of the music-NF studies utilized these known relationships to manipulate the affective tone of the musical pieces introduced in the loop.

Attempts to decode the brain patterns that are revealed when interpreting valence and arousal (Sayal et al., 2024; Daly, 2023; Putkinen et al., 2021) or complex emotions (Koelsch et al., 2021) in music may also provide important insights on the neural mechanisms that could support using music as feedback, guiding design decision in future music-based NF experiments.

Nevertheless, the majority of the studies included in this analysis do not focus on the nature of the feedback and do not evaluate the specificity of the feedback modality. Most studies recognize music's ability to evoke and modulate emotions in the listener, but its specificity and mechanism of action have not been controlled for. The absence of reports on information outside the brain signal of interest limits our ability to evaluate other potentially important patterns such as music-related ones. Measuring whole-brain activation patterns (as Trost et al., 2024 did with a dynamic effective connectivity metric based on fMRI data) will help further reveal the underlying success mechanisms behind music-based NF and control for its specificity by comparing it with other feedback modalities.

We identified four main dimensions in the implementation of music as the interface in neurofeedback paradigms: (i) the exploration of music attributes, (ii) how they link to the feedback loop, (iii) the imaging modality, and iv. the experimental design and control conditions.

The assessment of success in neurofeedback studies usually includes three analysis stages: (i) assess the ability to modulate the target brain signal, (ii) verify the clinical/behavioral effects of the NF training, i.e. the definition and evaluation of outcome measures, and iii. detailed reporting of experimental design variables. The characterization of success is critical to NF literature as many studies have also reported a considerable number of non-responders among participants. A review of 11 studies reports a percentage of non-responders that ranges from 16 to 57%, with a mean percentage of 38% (Alkoby et al., 2018). Considering the sample in our study, we found no consensus regarding the success measure, either in imaging and/or behavioral outcomes (Figure 6). It is important to note that the studies here considered are in different phases of validation, from more exploratory/proof-of-concept (Dekker et al., 2014; Takabatake et al., 2021; Fedotchev, 2018) to more mature studies focusing on clinical validation of a hypothesis (Pino, 2022; Cordes et al., 2015).

First, determining the relationship between the ability to modulate the target brain region and the mechanisms involved in such success is key to a complete comprehension of the framework (Bassett and Khambhati, 2017). For example, network neuroscience provides a flexible and generalizable approach to describe the neurofeedback framework and an explanation of the heterogeneous interaction patterns between its elements. In this sense, a complete characterization of the music stimuli and their neurobehavioral effects is critical to better understand the effect associated with the music-based feedback interface. Study design should consider the neural correlates of music stimuli and their interaction with the target brain signal, neurofeedback, and reward mechanisms to optimize the feedback loop (Singer et al., 2023). Ultimately, this comprehension may lead to the optimization of the framework or contribute to the translation to more naturalistic/clinical setups (e.g. translation from fMRI to EEG using the electrical fingerprint).

Second, verifying the clinical and behavioral effects of NF training requires determining whether these effects are specifically due to NF, i.e., establishing causality between NF learning and changes in clinical or behavioral outcomes. The NF literature has proposed several measures to evaluate the success of NF (Thibault et al., 2018), including assessing the modulation of the target brain signal relative to baseline, the first feedback trial, controls, or as a longitudinal trend in multi-session designs. These imaging-based success measures should, in turn, be associated with improvements in behavioral or clinical outcomes. To strengthen such evaluations, a comprehensive framework should incorporate cognitive, behavioral, and neurophysiological measures. Self-report scales, such as the Profile of Mood States (POMS) and the State-Trait Anxiety Inventory (STAI), could be used to assess changes in emotional and affective states, while cognitive performance might be evaluated using tasks like the n-back task or the Stroop test, which measure working memory and attentional control. Behavioral outcomes, such as stress resilience and relaxation responses, could be captured through validated instruments like the Perceived Stress Scale (PSS) or sleep quality indices. Research must establish whether changes in target brain activity (and its associated networks) causally drive behavioral changes. Appropriate control conditions are critical for determining this relationship. Notably, among the studies reviewed, no statistical evidence has established such causality. This is particularly relevant when considering the role of the music interface, as no control condition has been proposed to isolate its effects. Also, the characterization of the sample regarding music training, for instance, may serve as an important confound for these interventions' success (Zentner and Strauss, 2017). Expanding the range of evaluation indicators would enable a more comprehensive assessment of music neurofeedback's effects and provide a richer reference for future research.

Lastly, to improve the reporting and experimental design standards in the field, the NF research community has proposed the standardization of reporting: the tool “consensus on the reporting and experimental design of clinical and cognitive-behavioral neurofeedback studies” (CRED-NF checklist) (Ros et al., 2020). One item of this checklist addresses preregistration and registered reports, which have previously shown to increase the chance of publishing non-significant findings, effectively decreasing publication bias (Scheel et al., 2021; Allen and Mehler, 2019). Templates and guides for these preregistrations have been proposed for functional MRI (Beyer et al., 2021), EEG (Govaart et al., 2022), and fNIRS (Schroeder et al., 2023) studies.

Most NF paradigms analyzed in this study use EEG frequency band power characteristics, particularly alpha, beta, and mu rhythms, and reported good results regarding modulation ability with music-based interfaces. fMRI studies are still scarce, but they allow for a more spatially detailed analysis of the brain regions involved in the NF loop. This could lead to searching network-level metrics that are associated with the success of NF training. In fact, one of the studies reports on the network characteristics associated with the closed-feedback loop (Trost et al., 2024), by analyzing dynamic functional connectivity patterns and concluding that the amygdala may have a hub/central role in the music-evoked emotional experience and feedback. This is key to optimizing music-based NF designs: functional connectivity and brain network definition have developed rapidly in NF research across neuroimaging techniques, with some studies finding evidence for connectivity measures that are related to networks of success monitoring in NF paradigms (Pereira et al., 2023; Trambaiolli et al., 2022).

Neurofeedback training based on EEG is particularly sensitive to transients and noise. Moreover, EEG signals measured by channels are the combination of different sources and the exact neural correlate to the task is often not clear. The solution to determine the signal sources, also known as the inverse problem, requires computationally intensive methods that were not addressed in these studies. Alternatives such as the electrical fingerprint have been proposed in the literature, favoring network-based approaches for NF and highlighting the consideration of individual differences in brain function (Singer et al., 2023; Meir-Hasson et al., 2014; Gurevitch et al., 2024). To improve the signal-to-noise ratio, multi-modal signal acquisition might also be used. Synchronous acquisition of different neuroimaging techniques, such as EEG-fNIRS or EEG-fMRI, allows the combination of the advantages of both techniques— increased time and spatial resolution. Additionally, physiological data can also be combined (e.g. galvanic skin response, electromyography, electrocardiography) not only to regress noise confounds but also as a complementary biofeedback signal [e.g. galvanic skin response is known to be associated to autonomic responses signaling emotional states (Ribeiro et al., 2019; Markiewicz et al., 2022)].

An alternative to volitional neurofeedback, as proposed in the included studies, is the use of non-volitional neurofeedback, also known as covert or implicit neurofeedback. The participant is simply given positive or negative feedback whenever a specific target brain pattern occurs. The idea is to directly reinforce spontaneously emerging brain states of interest (Ramot and Martin, 2022). On the one hand, this approach limits the ability of participants to learn mental strategies to control the activation of specific brain areas. On the other hand, through implicit feedback, we can manipulate spontaneous activity at the network level. The dynamic nature of music feedback, as a naturalistic, highly individual, but also customizable stimuli, could make music an excellent candidate for implicit feedback interfaces.

We found no neurofeedback study exploring the link between music and pain. The effects of music listening on pain perception and reduction are well documented in the literature (Arnold et al., 2024), besides the fact that its clinical relevance is still underexplored (Werner et al., 2023; Sihvonen et al., 2022). The use of music as a neurofeedback interface in future clinical trials targeting pain could be a promising approach to explore the mechanisms of music-induced analgesia and its neural correlates.