The $50 Brain Scan That Rivals PET Imaging: How AI Is Turning EEG Into a Dementia Detective

Differentiating between Alzheimer's disease and frontotemporal dementia has long required expensive PET scans, invasive spinal taps, or specialized plasma biomarker tests—technologies often unavailable in resource-limited settings. Now, researchers at Aydın Adnan Menderes University in Turkey have demonstrated that electroencephalography (EEG), already present in most neurology departments worldwide, can achieve near-90% accuracy in distinguishing these two conditions when paired with explainable artificial intelligence.

The breakthrough, published in Neural Computing and Applications, combines brain wave recordings with clinical data to create a diagnostic tool that approaches the performance of far more expensive methods while remaining accessible, painless, and interpretable to clinicians.

Multimodal Machine Learning Achieves Clinical-Grade Performance

Researchers Ahmet Çağdaş Seçkin and Fatih Soygazi developed an XGBoost-based machine learning system that achieved 89.8% classification accuracy with an area under the curve (AUC) of 0.971 using leave-one-out cross-validation. The model analyzes publicly available EEG recordings originally curated by Miltiadous et al., drawing from patients with Alzheimer's disease, frontotemporal dementia, and healthy controls.

The system employs multimodal feature engineering, extracting both time-domain and frequency-domain characteristics from EEG signals and combining them with demographic and cognitive data. To prevent overfitting and reduce dimensionality, the researchers applied Fast Correlation-Based Filter feature selection using the Yu and Liu method. The team tested multiple competing algorithms including Random Forest, AdaBoost, artificial neural networks, and Gradient Boosting, with XGBoost emerging as the top performer.

Explainability Framework Addresses Clinical Adoption Barriers

What distinguishes this approach from typical "black box" AI diagnostics is its integration of SHAP (SHapley Additive exPlanations), an explainability framework that shows clinicians which neurophysiological and clinical features drive each individual diagnosis. This transparency directly addresses regulatory and clinical adoption barriers that have historically limited AI deployment in medical settings, where physicians need to understand and trust algorithmic recommendations.

The clinical distinction between these dementias is crucial: Alzheimer's disease begins in the medial temporal lobes and hippocampus, primarily affecting memory, while frontotemporal dementia attacks the frontal and anterior temporal lobes, disrupting personality, behavior, and language. Current gold standards for differential diagnosis—PET imaging, cerebrospinal fluid analysis, and plasma biomarkers—are costly, have limited availability, or require invasive procedures.

Democratizing Early Dementia Diagnosis

EEG offers compelling advantages: it's inexpensive, painless, widely available globally, and provides millisecond temporal resolution. By repurposing routine EEG equipment already deployed in most neurology departments, this research potentially democratizes access to early differential diagnosis in regions lacking advanced neuroimaging infrastructure.

For healthcare providers, the system offers a pathway to earlier, more confident differential diagnosis that could improve clinical trial enrollment, treatment planning, and family counseling without adding expensive imaging modalities. For health AI companies developing dementia diagnostics, the study validates the commercial viability of multimodal, feature-engineered approaches over pure deep learning for small-sample clinical problems.

The research demonstrates that sophisticated AI need not require exotic data sources or prohibitive costs to deliver clinical value—sometimes the most transformative innovations come from reimagining the tools already at hand.

Sources

1. Brain Waves and Machine Learning Join Forces to Tell Alzheimer's and FTD Apart — Scienmag.com