An AI model with less than one megabyte of memory can detect Alzheimer's disease in EEG recordings with 97 percent accuracy. A study published in Scientific Reports in 2026 documents the performance, and points to a new path for early dementia detection that is affordable, non-invasive and portable.
The Problem: Alzheimer's Is Recognized Too Late
More than 55 million people worldwide live with Alzheimer's disease or another form of dementia. No cure exists. What does exist are ways to slow the course of the disease in its early stages. For this, the disease must be recognized early, ideally before irreversible damage to the brain occurs.
The established diagnostic tools are expensive and invasive. A PET scan to detect amyloid plaques in the brain costs several thousand dollars and is not widely available. A lumbar puncture, which extracts cerebrospinal fluid from the spinal cord, is burdensome for many patients. Both procedures are rarely used in routine practice before unambiguous symptoms appear.
The Model: 1,609 Parameters, 97 Percent Accuracy
The research team behind the Scientific Reports study combined a Temporal Convolutional Network with a Long Short-Term Memory network. The result: a model with only 1,609 parameters and less than one megabyte of memory footprint, able to analyze EEG recordings in real time and detect Alzheimer's with 97.1 percent accuracy. Previous EEG-based approaches reached 70 to 80 percent.
Particularly relevant is its ability to discriminate. The model distinguishes not only between healthy participants and patients, but also separates Alzheimer's from frontotemporal dementia. That is clinically significant, because the two conditions have different courses and require different therapies.
A second variant of the model uses federated learning: several clinics train the system jointly without sharing patient data. Global model performance reaches 96.9 percent, compliant with privacy regulations and practical in everyday use. The concept allows smaller clinics to contribute to training without centralizing medical data.
Prevention Rather Than Medication
AI early detection takes a different approach from the new Alzheimer's drugs that have recently drawn criticism. A review analysis covering 20,000 patients showed that lecanemab and donanemab slow clinical decline by around 30 percent, but in absolute terms make a difference of only a few months.
AI early detection starts earlier. Detecting the disease before serious damage occurs opens the door to lifestyle changes, prevention trials and earlier use of medication that becomes ineffective in advanced stages. For the millions of people diagnosed with Alzheimer's every year, earlier diagnosis could have concrete consequences.
On the European level, the Predictom project is running in parallel. It combines MRI scans, EEG patterns and digital speech analysis into a multimodal risk profile. The idea: not a single signal, but the combination of several biomarkers should raise predictive accuracy further.
What Is Still Missing
The path from the lab to the general practitioner's office is long. The study in Scientific Reports was validated on controlled datasets, not in clinical routine. Before an EEG-based AI screening becomes a recognized guideline, further clinical studies are needed.
The long-term vision is clearer: portable EEG headsets, combined with privacy-preserving AI models, could make dementia screening as accessible as at-home blood pressure measurement.