Interpretable AI for neural signal decoding in dementia: an EEG ensemble approach to differential diagnosis

Artificial intelligence has long been accused of being a black box, especially in medicine, where a wrong answer can change a life. Now a research team spanning Pakistan, Saudi Arabia and the United Kingdom has built an AI system that not only distinguishes Alzheimer’s disease from frontotemporal dementia with remarkable accuracy using a cheap, non-invasive brain recording, but can also explain exactly which electrical signatures of the brain drove every decision. The study, published in Frontiers in Neuroinformatics, reports a multi-class classification accuracy of 95.63 percent on independently held-out EEG data, with FTD-specific precision reaching an extraordinary 0.9907, meaning the model almost never falsely labels someone as having frontotemporal dementia.

The clinical problem the team attacked is one of the most stubborn in neurology. Alzheimer’s disease and frontotemporal dementia are two of the most common neurodegenerative conditions worldwide, and while they have distinct pathological fingerprints, Alzheimer’s typically involves degeneration of the medial temporal lobes and progressive episodic memory loss, whereas frontotemporal dementia strikes the frontal and anterior temporal regions, producing executive dysfunction and behavioral disturbances, their early symptoms overlap so heavily that diagnostic ambiguity and misclassification are frequent. With global dementia prevalence projected to triple by 2050, improving subtype differentiation has become a critical clinical priority. Getting it wrong is not merely an academic error: misdiagnosis can lead to contraindicated treatments, such as acetylcholinesterase inhibitors given to FTD patients, which may actually worsen behavioral symptoms, along with significant psychological trauma and pharmacological mismanagement.

Electroencephalography, the humble 19-channel scalp recording that has existed for nearly a century, is an economically accessible window into functional brain dynamics. But conventional resting-state EEG analyses rely primarily on static spectral power measures, which the authors argue inadequately capture two things that separate these diseases: impaired neural plasticity and thalamocortical responsiveness, often seen in FTD, and lateralized cortical degeneration, which is not always fully reflected in global spectral metrics. Their solution was to stop looking at the brain only at rest and instead measure how it reacts when stressed, and to quantify whether the two hemispheres are electrically balanced or drifting apart.

The framework was built on publicly available data from 88 participants, 36 with Alzheimer’s disease, 23 with frontotemporal dementia and 29 cognitively normal controls, recorded at AHEPA University Hospital of Thessaloniki and hosted on OpenNeuro. Each participant underwent two paradigms: eyes-closed resting-state EEG and eyes-open photic stimulation, in which stroboscopic light flickers at multiple frequencies to probe how the visual cortex and thalamocortical circuits respond. From these paired recordings the team derived a neural reactivity vector, called V_diff, defined as the spectral power differential between the photic stimulation and resting conditions. The underlying logic is that neurodegenerative impairment is often more visible during state transitions than in isolation; a slowing of the neural response when shifting from baseline to photic stimulation serves as a sensitive indicator of impaired thalamocortical integrity and reduced neural plasticity.

In parallel, the team introduced a Hemispheric Asymmetry Index, computed as power differentials across seven homologous electrode pairs, including Fp1-Fp2, F7-F8, F3-F4, T3-T4, C3-C4, P3-P4 and O1-O2, spanning frontal, temporal, central, parietal and occipital regions. This measure operationalizes the lateralized degeneration that structural MRI studies have quantified in FTD through a Cortical Asymmetry Index, which showed that FTD manifests distinctly more asymmetric atrophy than Alzheimer’s disease. The electrophysiological proxy is elegant in its simplicity: Alzheimer’s subjects maintain near-zero variance in temporal asymmetry, whereas FTD subjects exhibit significantly elevated variance, establishing asymmetry as a functional proxy of structural degeneration.

Feature extraction was exhaustive. Power spectral density was computed using Welch’s method for five standard clinical bands, delta, theta, alpha, beta and gamma, plus six fine-grained sub-bands such as low and high theta, low and high alpha, and low and high beta, all chosen to capture FTD-specific signatures. Statistical descriptors, global relative power, alpha/theta ratios, a Frontal Slowing Index and the asymmetry indices were added, yielding 338 distinct features per condition. Applied across three conditions, REST, PHOTO and the reactivity differential DIFF, this produced an initial feature vector of 1,014 dimensions per segment. That high dimensionality, however, is a trap. With only 88 subjects, the model was susceptible to the Hughes phenomenon, a manifestation of the curse of dimensionality whereby classifier performance improves with increasing features only up to a point and then deteriorates as data become sparse and overfitting takes hold.

To escape that trap, the team applied Random Forest-based Recursive Feature Elimination, ranking features by their Gini impurity reduction and pruning the space to the 200 most discriminative dimensions. Critically, to prevent data leakage, the exact composition of the 200-feature subset was recalculated independently within the training split of each cross-validation fold, so the reported cross-validation accuracy reflects performance on entirely unseen test subjects. The classification core was a heterogeneous ensemble: Extreme Gradient Boosting, configured with 1,500 estimators, a maximum depth of 12 and a conservative learning rate of 0.02, fused with a Random Forest of 1,500 estimators and depth 30 via weighted soft-voting, assigning a 0.6 coefficient to XGBoost and 0.4 to Random Forest. The rationale, grounded in stacked generalization theory, is that combining learners with different inductive biases, the sequential error-correction of boosting and the parallel feature-subsampling of bagging, jointly mitigates both underfitting and overfitting. A custom cost-sensitive loss assigned a penalty weight of 2.0 to the minority FTD class, encouraging the model to prioritize learning minority-class spectral patterns.

The results are striking. On an independently held-out test partition of 1,281 segments, grouped by participant so that no subject contributed data to both training and test sets, the optimized model achieved segment-level accuracy of 95.63 percent, with internal five-fold cross-validation yielding a mean of 0.9846 plus or minus 0.003. Alzheimer’s sensitivity reached 97.0 percent, FTD recall 95.1 percent with precision of 0.9907, and cognitively normal classification 94.0 percent accuracy. Cohen’s Kappa of 0.9145 confirmed that agreement between predicted and true labels substantially exceeded chance. Area under the ROC curve exceeded 0.99 for all three classes, 0.998 for FTD, 0.995 for cognitively normal and 0.993 for Alzheimer’s. The model correctly identified 428 of 450 FTD segments, misclassifying only 4 as cognitively normal, with residual errors primarily involving subtle overlaps between early-stage FTD and Alzheimer’s spectral signatures, mirroring the known clinical difficulty of differentiating these pathologies by electrophysiology alone.

What elevates the work beyond raw numbers is its transparency. SHAP analysis, a game-theoretic feature attribution method, identified resting-state beta-band hemispheric asymmetry as the highest-gain discriminative feature, consistent with prior reports of asymmetric cortical disconnection in frontotemporal dementia, while alpha-band reactivity to photic stimulation emerged as a principal contributor to class separation, aligning with established findings of attenuated photic-driving responses and reduced alpha reactivity in Alzheimer’s disease, reflecting impaired thalamocortical and visual pathway integrity. These core physiological markers demonstrated extremely high selection stability, appearing in 100 percent of the fold-wise feature subsets. The authors are candid that the SHAP attributions were derived from a model refit on the full 88-subject cohort and therefore constitute exploratory biomarker signals rather than held-out validated findings, and that the single-site, modestly sized cohort constrains statistical power and generalizability, warranting replication on larger, multi-site and demographically diverse populations before clinical translation.

Benchmarked against recent deep learning approaches on comparable EEG datasets, the ensemble holds its own: a shallow CNN reported 94.54 percent, an SVM with functional connectivity 93.50 percent, multilayer network methods 81.10 percent for AD versus FTD, and a hybrid TCN-LSTM 80.34 percent, while the present framework reached 95.63 percent with far fewer computational parameters. In practical terms, the XGBoost and Random Forest ensemble requires no GPU infrastructure for training or inference, and even with a GPU, full model training completes in under seven minutes, making it potentially suitable for resource-constrained clinical environments where neuroimaging is inaccessible. Future priorities include fusing EEG-derived reactivity and asymmetry markers with structural MRI indices, external validation across sites with varying acquisition protocols, longitudinal prognostic modeling of neural stiffness and lateralized atrophy, and cohort expansion. The authors suggest that the beta-band hemispheric asymmetry index could serve as a computationally lightweight screening flag within a clinical decision-support system, alerting clinicians to atypical lateralization patterns, while attenuated alpha reactivity may function as a quantifiable indicator of thalamocortical integrity. The message is clear: carefully engineered, biologically grounded, interpretability-first machine learning may close the specificity gap in dementia subtyping without the opacity or computational appetite of deep learning, offering a scalable, non-invasive path toward earlier and safer diagnosis.

Subject of Research: Interpretable AI for neural signal decoding in dementia: an EEG ensemble approach to differential diagnosis

Article Title: Interpretable AI for neural signal decoding in dementia: an EEG ensemble approach to differential diagnosis

Article References: Muhammad, F., Usmani, I. A., Aamir, M., Alduailij, M., Marzougui, M., & Nawaz, R. (2026). Interpretable AI for neural signal decoding in dementia: an EEG ensemble approach to differential diagnosis. Frontiers in Neuroinformatics, 20, Article 1902549. https://doi.org/10.3389/fninf.2026.1902549

Image Credits: AI Generated

DOI: 10.3389/fninf.2026.1902549

Keywords: Interpretable, neural, signal, decoding, dementia, ensemble, approach, differential, diagnosis, scientific research