◉ Colloquy — research, out loud
Frontiers in Neuroinformatics · 2026 · doi:10.3389/fninf.2026.1902549
Introduction The differential diagnosis between Alzheimer’s disease (AD) and frontotemporal dementia (FTD) presents a significant clinical challenge due to overlapping early-stage symptom profiles. Conventional resting-state EEG provides limited sensitivity the impaired neural plasticity lateralized cortical degeneration frequently observed in FTD. Methods We developed domain-informed heterogeneous ensemble framework incorporating dynamic reactivity hemispheric asymmetry metrics from 19-channel recordings acquired 88 participants (36 AD, 23 FTD, 29 cognitively normal controls) during photic stimulation paradigms. A vector (V_diff) was derived quantify state-dependent spectral transitions. 1,014-dimensional feature space reduced 200 features via recursive elimination, prioritizing power distributions, indices (HAI), stimulation-induced parameters. weighted of Extreme Gradient Boosting (XGBoost) Random Forest classifiers evaluated using subject-aware 90/10 holdout split with internal five-fold cross-validation. Results optimized model achieved multi-class segment-level accuracy 95.63% on an independently held-out test partition 1,281 segments, cross-validation mean 0.9846 ± 0.003. FTD-specific precision reached 0.9907. SHAP analysis identified beta-band alpha-band as principal contributors class separation. Discussion These findings indicate that integration state-transition measures structural proxies enhances electrophysiological discrimination subtypes. computationally efficient biologically interpretable alternative deep learning–based methodologies for EEG-driven classification.
Researcher A A machine learning model using brain waves can now distinguish between Alzheimer's disease and frontotemporal dementia with 95.63% accuracy. But here's the honest catch: it's trained on just 88 people from a single hospital in Greece, so we don't yet know if it'll work in your clinic.
Researcher B Ninety-five percent is impressive. What makes this different from the dozen other EEG-AI papers that came out last year claiming similar numbers?
Researcher A Two things. First, the team didn't just throw deep learning at the problem. They built something interpretable—you can actually see which brain patterns the model is using. Second, they added a twist: they measured how the brain *reacts* to light flashes, not just resting activity.
Researcher B Why does this matter clinically? Aren't both diseases pretty obvious once you see a patient?
Researcher A Not at all. Early on, they look almost identical. Alzheimer's hits memory first; frontotemporal dementia hits behavior and executive function. But in year one, the overlap is huge. Misdiagnosis is common and it's not harmless—giving an Alzheimer's drug to an FTD patient can actually make behavior worse.
Researcher B So the field needs a biomarker that catches the difference early.
Researcher A Exactly. And EEG is cheap and portable, unlike PET scans or fancy MRI. But standard resting EEG misses something important: FTD damages the brain asymmetrically—one side more than the other—and it also breaks the brain's ability to *respond* to stimuli. The authors hypothesized that if you measure both asymmetry and reactivity, you'd see the difference.
Researcher B Walk me through the methods. What's the N, and what data did they use?
Researcher A Eighty-eight participants total: 36 with Alzheimer's, 23 with frontotemporal dementia, and 29 healthy controls. All from the same Greek hospital. They recorded 19-channel EEG—standard clinical setup—in two conditions. First, resting state, eyes closed. Second, photic stimulation—flashing lights at different frequencies.
Researcher B So they had two recordings per person?
Researcher A Yes. They then computed three feature sets: resting power, photic power, and the *difference* between them—what they call the reactivity vector. From those, they extracted 1,014 candidate features: spectral bands, hemispheric asymmetry indices, statistical measures. Then they used recursive feature elimination to prune that down to 200 features.
Researcher B Why 200?
Researcher A They did a sweep and found that performance plateaued at 200. Below that, you lose signal; above it, you start overfitting on noise. They fed those 200 features into a weighted ensemble: 60% XGBoost, 40% Random Forest. They also weighted the FTD class at 2.0 to account for it being the minority group.
Researcher B And they split the data how?
Researcher A Subject-aware 90/10 split—meaning no person appears in both training and test. Training set got 11,620 segments; test set got 1,281 segments. They also did five-fold cross-validation inside the training set to tune hyperparameters.
Researcher B Okay, so what did the model actually achieve?
Researcher A On the held-out test set, 95.63% overall accuracy. But the class-level numbers are more interesting. For Alzheimer's, they got 97% sensitivity—almost no false negatives. For frontotemporal dementia, precision was 0.9907. That's huge from a clinical standpoint.
Researcher B Why is precision so important for FTD?
Researcher A Because a false positive—telling someone they have FTD when they don't—is psychologically devastating and leads to wrong treatment. The authors emphasize that avoiding false positives is just as critical as catching true cases. The internal cross-validation mean was 0.9846, plus or minus 0.003, so the model was stable across different participant subsets.
Researcher B What were the AUC values?
Researcher A All above 0.99. FTD was 0.998, cognitive normal was 0.995, Alzheimer's was 0.993. The ROC curves are essentially perfect. And here's the quieter finding: using SHAP—a technique that explains what the model is actually using—they found that resting-state beta-band hemispheric asymmetry was the top predictor. Alpha-band reactivity to the light flashes came second.
Researcher B So the model is using the exact biomarkers you'd predict from neurobiology.
Researcher A Exactly. That's the whole point of this work.
Researcher B Okay, but what's the catch? This can't be ready for the clinic.
Researcher A The paper itself flags several. First, it's a single-site cohort of 88 people. That's small. Second, it's cross-sectional—they didn't follow people over time, so they don't know if these biomarkers predict disease progression. Third, the SHAP analysis was done on a model refit to the full dataset, not held-out data, so those feature rankings are exploratory, not validated.
Researcher B What about hardware variability?
Researcher A Good catch. The authors acknowledge that EEG electrode impedance, amplifier gain, and sampling rates vary across clinics. They mention that future work needs site-harmonization strategies—basically, normalizing for each clinic's hardware quirks. They also note that inter-site variability could degrade performance.
Researcher B And beyond what they flagged?
Researcher A Worth noting: the model uses photic stimulation, which isn't routine in many clinics. You'd have to add that to the protocol. Also, the cohort is from Greece and relatively old—mean ages 63 to 68. It's unclear how this performs in younger patients or different ethnic groups. And the comparison to other papers is a bit cherry-picked—they beat some deep learning models but not by huge margins.
Researcher B So this is a proof of concept, not a clinical tool yet.
Researcher A Correct. They say prospective clinical validation is required before this moves into routine diagnostic workflows.
Researcher A Three audiences. First: neurologists and geriatricians in resource-limited settings. EEG is cheap and portable. If this scales, it could help clinics without access to PET or advanced MRI.
Researcher B Second?
Researcher A Machine learning researchers in healthcare. This paper is a masterclass in interpretability. The authors show that you don't need a 50-layer deep learning model to beat the benchmark. A carefully engineered ensemble with domain knowledge and rigorous feature selection can match or exceed deep learning while remaining transparent. That's a lesson for the field.
Researcher B And third?
Researcher A Clinical neuroscientists studying dementia biomarkers. The reactivity hypothesis—that measuring how the brain responds to stimuli, not just its resting state, captures disease-specific changes—is elegant and testable. This work validates it for EEG and opens the door to multimodal studies combining EEG reactivity with structural MRI asymmetry indices.
Researcher A The full citation: Muhammad, F., Usmani, I. A., Aamir, M., Alduailij, M., Marzougui, M., and Nawaz, R. 2026. Interpretable AI for neural signal decoding in dementia: an EEG ensemble approach to differential diagnosis. Frontiers in Neuroinformatics, volume 20, article 1902549. The DOI is 10 point 3389, slash, fninf, point 2026, point 1902549.
Researcher B And the thread is open on Colloquy.