Three bottlenecks that small EEG cohorts create
Most EEG-based Alzheimer's studies rest on cohorts of roughly 20 to 100 subjects, because collecting EEG from patients with AD is costly and slow. The paper frames three consequences: too little data to learn robust representations, weak generalization to unseen subjects, and models that break when recording length, electrode layout, or sampling rate changes between datasets. The second problem is the clinically decisive one. Models can look strong when training and testing share subjects, then drop substantially under subject-independent evaluation, where inter-subject variability and subject-specific shortcuts can be confused with the diagnostic label.
LEAD's response is to build scale first and adapt second. The authors curate an EEG-AD corpus of 2,238 subjects, which they describe as the largest to date for this purpose, and use it to pretrain a single backbone that is then fine-tuned and tested on five separate AD datasets. The claim to novelty is therefore twofold: the corpus, and the first foundation model built for end-to-end subject-level EEG-based AD detection.
Gated attention across time and electrodes
Rather than resampling every recording to one fixed rate, LEAD downsamples signals above 200 Hz to 200, 100, and 50 Hz, then cuts them into fixed-length samples with 50% overlap. Each sample carries a sampling-rate label, and the model adds a sampling-rate embedding alongside a temporal positional embedding and a 3D channel embedding built from electrode coordinates. This is how one model can ingest recordings with different lengths, channel configurations, and rates without dataset-specific preprocessing.
Inside the gated temporal-spatial Transformer, self-attention runs in parallel along the temporal dimension and along the channel dimension. A learnable gate then fuses the two branches, so the model can weight temporal context against spatial context per input rather than committing to one fixed mixture. The final token of each channel becomes the representation passed downstream. An ablation supports the spatial design: replacing the 3D channel embedding with fixed or learnable embeddings lowered subject-level F1 on ADFTD from 93.95 to 84.64 and 88.01 respectively, though fixed embeddings were slightly better on a few other metrics, so the benefit is dataset-dependent.
Optimizing for subjects, not individual samples
Most prior EEG work classifies individual samples, but a clinical decision is made per person. LEAD therefore aggregates sample-level predictions by majority voting to produce one label per subject, and trains with cross-entropy at both the sample and subject levels. Two engineering choices make that subject-level objective workable. Multi-scale segmentation multiplies the number of samples per subject, which feeds both training and the later vote. Index group shuffling regroups sample indices by subject each epoch so that a mini-batch is likely to contain several samples from the same person; without it, subject-level contrastive learning degenerates toward sample-level learning.
Pretraining uses medical contrastive learning on 13 datasets: 4 AD datasets and 9 non-AD neurological disorder datasets. Samples from the same subject form positive pairs and samples from different subjects form negative pairs, which the authors argue suits subject-level representation learning better than masked autoencoding. The pretraining corpus deliberately includes non-AD disorders, so the encoder is not tuned only to Alzheimer's contrasts.
Best average ranking across 20 evaluations
The comparison spans 5 downstream datasets and 16 baselines, scored at both sample and subject level with F1 and AUROC, giving 20 evaluations. LEAD has the best average ranking, and the authors report it far surpasses the second-best model, LaBraM, while using substantially fewer pretraining resources. The margins are largest where the task is harder. On ADFTD, a three-way healthy/AD/frontotemporal dementia task with 88 subjects and 167,083 samples, LEAD reached subject-level F1 93.95 and AUROC 99.17, against 91.81 and 99.44 for LaBraM and 88.10 and 97.10 for BIOT. On ADSZ, subject-level F1 was 100.00 versus 82.21 for CBraMod.
Several datasets are effectively saturated at subject level: ADFSU, ADSZ, and APAVA reach 100.00 F1 for LEAD, and many baselines also hit 100.00 on ADFSU. The authors themselves caution that these near-perfect numbers reflect controlled in-lab conditions, with expert-selected recordings and severe artifacts manually rejected, and with relatively balanced sample counts per subject that stabilize majority voting. Those conditions are less likely to hold in routine clinical recording, so the saturated cells should not be read as evidence of universal clinical accuracy.
Fine-tuning wins, but pretraining matters most on small cohorts
Comparing training from scratch, linear probing with a frozen encoder, and full fine-tuning isolates what pretraining contributes. Fine-tuning is best in most cases. The more informative pattern is where linear probing beats fully supervised training: on APAVA, which has only 23 subjects, linear probing reached subject-level F1 73.33 and AUROC 95.00 against 36.67 and 75.00 for training from scratch. On the larger ADFTD and CNBPM datasets, linear probing did not consistently beat full supervision and sometimes did worse. The practical reading is that a pretrained encoder is most valuable when downstream AD data are limited, which the authors note is the common situation in this field.
Resting-state and photic-stimulation recordings are complementary
ADFTD subjects were recorded under two paradigms, resting-state eyes-closed and photic-stimulation eyes-open, with different recording lengths. Evaluated separately, neither paradigm supports strong three-way classification: sample-level F1 was 47.54 for resting-state over 19.40 hours and 55.60 for photic stimulation over 7.25 hours. Combined over 26.64 hours, sample-level F1 rose to 81.01 and AUROC to 94.03, with subject-level F1 at 93.95. The gain is far larger than the extra recording time alone would suggest, which makes the choice of paradigm, and whether both are available, a first-order factor in how well this model performs rather than a minor acquisition detail.
MCI remains the unresolved boundary
The clearest limit is disease stage classification. On CNBPM, separating healthy controls, MCI, and AD, LEAD's subject-level F1 was 65.67 with AUROC 80.37, and the confusion matrix shows MCI as the weak class: subject-level recall for MCI was 54.29, with 40.00% of MCI subjects predicted as AD and 5.71% as healthy. The authors explain this as label ambiguity rather than only model weakness. Some MCI patients already carry AD-related features that are invisible without longitudinal follow-up, while others remain stable or revert, making them hard to place near either AD or healthy controls. Their stated priority for future work is longitudinal EEG that tracks progression.
A second scope condition sits in the problem formulation: each subject is assumed to have a single label, with no diagnostic transitions during data collection and no comorbid conditions such as AD together with vascular dementia. Real patients can violate both assumptions, and the paper simplifies them away rather than testing them.
Match the task and the data before the model
The paper's own results sort its value by scenario rather than by a single accuracy figure. LEAD is strongest where the question is AD versus healthy controls or AD versus frontotemporal dementia, where subject-independent evaluation is used, and where both resting-state and photic-stimulation recordings are available; the jump from roughly 47.54 to 81.01 sample-level F1 when paradigms are combined is the sharpest illustration that acquisition design can matter more than architecture choice. It is weakest where the question requires separating MCI from AD and from healthy controls, at about 66% subject-level F1, and where downstream cohorts are so small that even linear probing on pretrained features outperforms training from scratch. The most decision-relevant comparison is therefore not LEAD versus the next baseline on saturated datasets, but pretrained versus from-scratch training on small AD cohorts, and two-paradigm versus single-paradigm recording on the same subjects.
Sources used in this answer
LEAD: An EEG Foundation Model for Alzheimer's Disease Detection
LEAD, pretrained on 13 EEG datasets with medical contrastive learning, achieves the best average ranking across all 20 evaluations on 5 downstream AD datasets, outperforming 16 baselines including state-of-the-art EEG foundation models, while three-way healthy/MCI/AD staging remains limited at about 66% subject-level F1.