Can Personality Predict Neurofeedback Success in ADHD?

This post explores a new emerging research with novel insights: an explainable machine learning (XAI) framework developed by Hoseini and Shalbaf (2025) to predict who will respond to neurofeedback treatment for attention-deficit/hyperactivity disorder (ADHD) using a subsample from the TDBRAIN database. Rather than focusing on brainwaves directly, the authors use demographic, behavioral and personality questionnaire data to build a transparent prediction model for neurofeedback outcomes.

ADHD is a heterogeneous neurodevelopmental condition with symptoms of inattention, impulsivity, hyperactivity and emotional dysregulation that often persist into adulthood. Traditional treatments are still dominated by medication, yet neuromodulation approaches like neurofeedback are increasingly used as non-pharmacological options. Neurofeedback, in general terms, is a form of EEG biofeedback where individuals learn to self-regulate specific aspects of their brain activity through real‑time feedback and reinforcement. More broadly, biofeedback uses physiological signals such as heart rate, breathing, skin conductance or muscle tension to teach self‑regulation of the body’s responses. Both approaches sit at the intersection of learning theory and neuroplasticity: the brain and body gradually change how they operate as people repeatedly practice regulation with feedback.

Despite encouraging evidence, neurofeedback does not work equally well for everyone. ADHD itself is neurobiologically diverse, and even "standard" protocols vary (for example, theta/beta training, sensorimotor rhythm [SMR] training, or slow cortical potential [SCP] neurofeedback). This raises a critical clinical question: can we predict in advance who is likely to benefit, using information that is easy to collect in routine practice? The study reviewed here tackles that question using an explainable machine learning approach, aiming not only for accurate prediction but also for clarity about why the model makes its decisions.

Methods

The authors drew on the TDBRAIN database, a large clinical archive including EEG and clinical information from 1274 psychiatric patients collected over two decades. For this specific study, they focused on 72 individuals with ADHD (aged 6 to 68 years, average around 24) who had received qEEG‑informed neurofeedback. Neurofeedback protocols were assigned based on each person’s quantitative EEG assessment and clinical profile, following a personalized medicine approach.

Three main EEG‑neurofeedback protocol families were used:

• Theta/beta ratio (TBR) training – typically aiming to reduce excessive slow theta activity and normalize beta activity over fronto‑central regions, to improve attention and reduce hyperactivity/impulsivity.

• Sensorimotor rhythm (SMR) training – training 12–15 Hz activity over the sensorimotor strip, often at C3, C4 or Cz, to support motor inhibition, sleep quality and stable attention.

• Slow cortical potential (SCP) training – teaching regulation of very slow shifts in cortical excitability, often at midline sites (e.g., Cz or FCz), to modulate activation thresholds and improve self‑control.

Sessions lasted about 20–30 minutes and were delivered two to three times per week. Alongside neurofeedback, participants also received sleep hygiene management and coaching, reflecting a multimodal, behaviorally informed treatment environment. Treatment outcome was measured using the ADHD Rating Scale (ADHD‑RS) before and after the course of neurofeedback. Responders were defined as those showing at least a 50% reduction in ADHD‑RS total score, yielding 47 responders and 25 non‑responders in the analyzed sample.

Interestingly, although resting‑state EEG and event‑related potential data were available in TDBRAIN, this particular analysis deliberately did not include EEG features in the machine learning models. Instead, the authors focused on variables that are very easy to gather in most clinical settings:

• Demographics: age, sex, years of education.

• Clinical/behavioral data: hours of sleep last night, number of neurofeedback sessions, time since last meal, coffee, alcohol or drugs, self‑rated “feeling today,” baseline ADHD‑RS scores and task performance measures.

• Personality: responses to all 60 items of the NEO Five‑Factor Inventory (NEO‑FFI), covering neuroticism, extraversion, openness, agreeableness and conscientiousness.

The machine learning pipeline followed a hierarchical feature selection strategy. First, each of the 78 candidate features was evaluated with receiver operating characteristic (ROC) analysis and an exact binomial test. Only features with area under the ROC curve (AUC) greater than 0.60 were kept, yielding 20 preliminarily informative predictors (including several NEO items, age, education and sleep).

From that point, four more advanced feature selection methods were applied: mutual information, ReliefF, minimal‑redundancy–maximal‑relevance (mRMR), and sequential forward selection (SFS). In parallel, five classifier types were trained and tuned using grid search and stratified 5‑fold cross‑validation: logistic regression, support vector machine (SVM), random forest (RF), artificial neural network (ANN) and adaptive boosting (AdaBoost). Class imbalance between responders and non‑responders was handled using SMOTE (synthetic minority oversampling technique) within the training folds.

Finally, the best‑performing model was made explainable using SHAP (Shapley Additive exPlanations) via TreeExplainer. SHAP provides global feature importance (which predictors matter most overall) and local explanations (why the model made a specific prediction for a given individual), allowing the authors to move beyond a black‑box classifier.

Results

Using the initial set of 20 statistically pre‑selected features, all classifiers performed reasonably well, but the random forest model stood out. With these 20 features, RF reached an accuracy of about 85%, with an AUC around 0.92, and balanced sensitivity and specificity. Support vector machines and neural networks also showed good performance, but RF provided the best overall trade‑off between accuracy, robustness and interpretability.

Feature selection further sharpened the model. Sequential forward selection (SFS) identified a compact set of just seven optimal features that maximized predictive performance when used with the random forest classifier. These seven features were:

• NEO‑FFI item 6 – feeling inferior to others (linked to neuroticism).

• NEO‑FFI item 8 – sticking to a chosen way of doing things (conscientiousness).

• NEO‑FFI item 11 – feeling like falling apart under great stress (neuroticism).

• NEO‑FFI item 15 – describing oneself as not very methodical (reverse‑scored conscientiousness).

• NEO‑FFI item 28 – often trying new and foreign foods (openness to experience).

• NEO‑FFI item 44 – being hard‑hearted and tough‑minded (low agreeableness).

• Education – years of formal education.

With just these seven variables, the RF model achieved an average accuracy of 88.3% (±6.8), sensitivity of 87.6% and specificity of 89.6%. In other words, nearly 9 out of 10 cases were correctly classified as likely responders or non‑responders based solely on personality and education data.

SHAP analysis confirmed and refined these findings. Globally, the most influential predictors for distinguishing responders from non‑responders included NEO items 11, 6, 60, 15, 26, 55, education, 8, 25 and 43. Many of these overlap with the SFS‑selected features, supporting the internal consistency of the model. SHAP summary plots showed that higher values on several items (for example, endorsing feeling inferior or falling apart under stress, or having higher education) tended to push the prediction toward responder status. In contrast, higher endorsement of striving for excellence in everything (NEO‑FFI item 60) and some goal‑oriented items were associated with non‑response.

At the individual level, SHAP force plots illustrated why a given patient was predicted to respond or not. For a typical responder, high scores on items reflecting emotional vulnerability (e.g., feeling inferior, feeling like falling apart under stress) and certain openness‑conscientiousness combinations contributed strongly to a positive prediction. For a typical non‑responder, low scores on these vulnerability items – combined with certain other personality patterns – shifted the model toward predicting no meaningful improvement.

The authors also compared their model’s performance to previous machine learning work predicting medication response in ADHD (methylphenidate or atomoxetine). Their neurofeedback‑focused model showed similar or better accuracy and AUC than many medication‑based prediction studies, despite the modest sample size and the absence of EEG features.

Important limitations were acknowledged

The sample was relatively small (72 participants), and SMOTE was used to synthetically balance responders and non‑responders, which may inflate performance estimates. Neurofeedback protocols were heterogeneous (SMR, TBR, SCP) and tailored per individual, so the model predicts overall response to semi-personalized neurofeedback rather than to any specific protocol. Finally, EEG biomarkers and early treatment response – two powerful predictors in other work – were not included, and the model will need to be validated in larger, independent cohorts.

Discussion

This study offers a fascinating window into how relatively simple, low‑cost baseline data can meaningfully predict neurofeedback outcomes for ADHD. Instead of complex neuroimaging features, the model leans heavily on personality traits, basic demographics and a few behavioral variables. The core message is that who the person is – how they tend to feel, cope and organize their life – may be just as important as which protocol we choose when it comes to neurofeedback response.

From a clinical perspective, this is appealing. In most practices, administering a NEO‑FFI or a similar personality inventory, asking about sleep and education, and reviewing baseline functioning are far easier than acquiring large EEG datasets for machine learning. If validated, a tool like this could act as a triage assistant: flagging clients who are particularly likely to benefit from qEEG‑informed neurofeedback, and gently warning when expectations should be more cautious or when additional support (for example, psychotherapy or coaching) might be necessary alongside training.

The pattern of predictors is also clinically meaningful. Many of the NEO‑FFI items that predict response describe emotional sensitivity, self‑doubt or feeling overwhelmed under stress. This might sound counterintuitive at first – why would more fragile individuals respond better? One possibility is that these clients have a larger “window of opportunity” for change: physiological self‑regulation may directly address their heightened emotional reactivity, and they may be more motivated to engage deeply with treatment because their distress is close to the surface. Similarly, education level may reflect not just cognitive resources but also familiarity with learning environments, persistence and the ability to stick with repetitive training tasks over weeks or months.

For people considering neurofeedback, one practical implication is that baseline personality and life context matter. The findings suggest that clients who are emotionally reactive, self‑critical or prone to feeling overwhelmed might actually do very well if they are supported, encouraged and given clear feedback during training. Conversely, individuals who are highly perfectionistic and relentlessly driven to “excel at everything” might need extra framing: helping them see neurofeedback not as a performance test to ace, but as a gradual skill‑building process where experimentation and patience are more important than immediate success.

For professionals who refer to neurofeedback – such as psychologists, psychiatrists, pediatricians or educators – the study underscores the value of integrating psychometric and lifestyle data into referral decisions. Rather than treating neurofeedback as a generic adjunct to medication or therapy, this approach points toward more nuanced questions: Does this client have the attentional bandwidth, emotional readiness and environmental support (sleep, routines, family involvement) to benefit from a learning‑based intervention? Could neurofeedback be prioritized for those who either cannot tolerate medication or who show certain personality profiles linked to better response?

For neurofeedback practitioners, the implications are perhaps the richest. The study reminds us that protocol choice (SMR vs TBR vs SCP) is only one part of personalization. Equally important is how the person learns: their tolerance for frustration, their reward sensitivity, their relationship to structure and novelty. Clients who “fall apart under stress” may benefit from protocols that stabilize arousal (for example, SMR training at C3/C4 or Cz to promote sensorimotor stability and improve sleep) combined with gentle shaping of strategies and frequent reinforcement for process rather than outcome. Those who are less emotionally reactive but highly perfectionistic may need explicit coaching around experimentation, accepting micro‑failures and letting go of rigid control during sessions.

A broader theme emerging from this work is the convergence between psychological models of learning and modern explainable AI. The SHAP analyses effectively give us a kind of “mechanistic hypothesis” about how personality traits interact with an intervention like neurofeedback. Instead of only knowing that a model is 88% accurate, we see that, for example, higher emotional vulnerability, certain patterns of conscientiousness and higher education collectively push the probability of benefit upward. That picture fits well with reinforcement‑learning views of neurofeedback, where motivation, expectation and reward history shape how quickly and robustly people acquire self‑regulation skills.

The authors also highlight important caveats that clinicians should keep in mind before rushing to use any such model in practice. The modest sample size, reliance on SMOTE, and heterogeneity of neurofeedback protocols all mean that the reported accuracy is likely an upper bound. Real‑world performance will probably be lower until the model is tested in larger, more diverse samples and ideally extended to include EEG biomarkers (such as individual alpha peak frequency) and markers of early treatment response. For now, it is safer to treat these findings as hypothesis‑generating rather than as prescriptive rules.

Still, even at this stage, the study supports a more holistic, multidisciplinary approach to ADHD treatment. Neurofeedback should not be seen as either a magical cure or an all‑or‑nothing add‑on, but as one component within a broader ecosystem that includes education about sleep, coaching around habits, psychotherapeutic work on emotional regulation and, when appropriate, medication. In that ecosystem, explainable models that integrate personality, behavior and physiology could help us better match the right person to the right mixture of tools.

Brendan’s perspective

When you first read a paper that predicts “responders” and “non‑responders,” it’s easy to slip into a slightly fatalistic mood: does this mean neurofeedback just doesn’t work for some people? Maybe. But I don’t actually think that’s the most useful interpretation.

What this study really shows, in my view, is that a standardised, protocol‑driven approach to ADHD neurofeedback will fit some people quite well and others quite poorly. The model isn’t testing whether a human nervous system can learn from feedback; it’s testing who is likely to benefit from a particular way we currently deliver qEEG‑informed training in a particular context. That’s a very different question.

Seen through that lens, the so‑called “non‑responders” might not be people for whom neurofeedback is doomed to fail. They might be people for whom the usual recipe – a certain number of sessions, fairly typical SMR/TBR/SCP protocols, standard levels of coaching and sleep advice – simply isn’t sufficiently personalised. The model is waving a flag and saying: if you work with clients like this, you probably can’t just lean on the protocol manual and hope for the best.

This is exactly where my own practice tends to live: at the intersection of three streams of information that have to be harmonised if we want to move beyond one‑size‑fits‑all work:

• objective biomarkers (qEEG patterns, spectral features, network organisation, sleep metrics),

• structured psychometrics (personality traits, symptom scales, executive function profiles), and

• rich, ongoing subjective experience and observation (how the client describes their inner world, what families notice, what we see in the room week by week).

If you take any one of these in isolation, you’re flying partially blind. A beautiful qEEG map without personality or context can seduce us into thinking the right electrode and frequency band will magically solve everything. Symptom scales without physiology can lead to endless cognitive work for a nervous system that is simply too unstable or over‑aroused to implement it. Subjective narratives without any anchors can float off into endless story‑making without changing underlying self‑regulation.

Bringing them together is where we get leverage.

In practical terms, here’s how a study like this nudges me to refine ADHD neurofeedback in the clinic.

First, I treat the qEEG as a map of constraints and opportunities. Elevated frontal theta/beta ratio? Great, that points me toward TBR training at Fz or Cz to support sustained attention. Low SMR with fragmented sleep? That suggests rewarding 12–15 Hz at C3/C4 while inhibiting excess theta and high beta, especially in the evening. These choices matter. But before I lock them in, I run them through what the psychometrics and the intake conversations are telling me.

If personality data suggest high emotional vulnerability – endorsing items like feeling inferior or falling apart under stress – I assume this nervous system will need a gentler learning curve. I might prioritise SMR at C3/C4 or Cz for the first block of sessions, with relatively short training epochs (10-12 x 2 minutes) and very forgiving thresholds, combined with explicit work on sleep routines. Only when we see some stabilisation – fewer meltdowns, less bedtime chaos, a bit more morning bandwidth – do I lean harder on frontal TBR for impulse control.

On the other hand, if the psychometrics and interview tell me we’re dealing with a highly conscientious, perfectionistic, “I must excel at everything” profile, I’m on alert for a different kind of risk: over‑control and frustration. Here, the map of biomarkers might still push me toward excessive frontal high beta or low posterior alpha, but the way we inhabit those protocols changes:

• I frame entrainement en neurofeedback explicitly as a messy, iterative learning process – more like learning a musical instrument than taking a test.

• In initial sessions, I pick feedback that is continuous and process‑oriented, not win/lose. Think smooth animations or soundscapes that respond subtly to brain shifts, instead of highly gamified displays that invite score‑chasing.

• I deliberately build in blocks where the “task” is to stop trying to control the screen and simply notice internal sensations while the brain does its thing. For some perfectionistic clients, this is the hardest but most important skill. Sometimes, playing the "opposite game" of inhibiting a reward or rewarding an inhibit can be useful and appropriate.

The third stream – subjective experience and observation over time – is what keeps everything honest. I want to know: when we train SMR at C4 with theta and high‑beta inhibition, does the client actually sleep better, or do they just get a nicer graph in the software? When we shift TBR training from Cz to Fz, does that translate into fewer impulsive comments in school, or does the teacher see no difference? If the data on paper say “responder” but the family experience says “nothing meaningful has changed,” the model has to lose, not the family.

From that perspective, a model that labels some people as probable non‑responders isn’t a verdict – it’s a warning that the default recipe is unlikely to be enough on its own. For those clients, we may need:

• more intensive work on sleep and daily structure alongside SMR,

• tighter integration with psychotherapy to process emotional themes that are activated as regulation improves,

• different feedback modalities (for example, movement‑based or auditory‑only feedback for clients who find screens dysregulating),

• or even a different sequencing of protocols – perhaps starting with posterior alpha or occipital low‑beta to calm visual processing before asking frontal networks to work harder.

Where this study really aligns with what I teach is in its implicit message: prediction is not the enemy of personalisation; it’s a tool for deeper personalisation. A good predictive model doesn’t tell us “don’t bother” with certain clients. It tells us “don’t assume the usual way of doing things will work – bring more of your creativity, nuance and integration to this case.”

So rather than reading the 88% accuracy as: “neurofeedback works for these people and not for those,” I read it as: “when we deliver qEEG‑informed neurofeedback in a fairly standardised way, certain personality‑context combinations are a better fit than others.” The next step – and this is where practice can leap ahead of research – is to ask: what happens to the so‑called non‑responders when we deliberately tune protocols, feedback, pacing and adjunctive supports based on their psychometrics and lived experience, not just their EEG?

My guess, based on day‑to‑day work, is that many “non‑responders” become partial responders – and many partial responders become strong responders – when we fully harmonise biomarkers, psychometrics and subjective feedback over time. That’s the version of personalised neurofeedback I’m most interested in: not just choosing the right frequency at the right site, but creating the right learning relationship for this particular brain in this particular life.

Conclusion

Hoseini and Shalbaf’s explainable machine learning study adds an important piece to the ADHD neurofeedback puzzle. By showing that personality traits, sleep and education can predict neurofeedback response with relatively high accuracy, it supports a shift toward genuinely personalized, psychologically informed treatment planning. The use of XAI tools like SHAP moves us beyond opaque “black box” models and toward transparent, clinically interpretable decision support.

For clients and families, the take‑home message is hopeful but realistic: neurofeedback is not a magic bullet, but its success is not random either. Who you are – how you cope with stress, how you learn, how your life is structured – can meaningfully shape how much benefit you get from training. For clinicians, the study is a call to integrate psychometrics, lifestyle factors and EEG findings into a single, coherent formulation, rather than treating each in isolation.

As the field evolves, combining qEEG‑guided protocols with explainable predictive models and rich clinical judgment may help us offer the right form of neurofeedback to the right person at the right time. That is a future worth training for.

References

Hoseini, R., & Shalbaf, A. (2025). An explainable machine learning‑based approach to predicting treatment response for neurofeedback in ADHD. Scientific Reports, 15, 43162. https://doi.org/10.1038/s41598-025-27246-9