Enhancing Alzheimer's Detection through Late Fusion of Multi-Modal EEG Features

Alzheimer s disease (AD) is a progressive neurodegenerative disorder characterized by cognitive decline, where early detection is essential for timely intervention and improved patient outcomes. Traditional diagnostic methods are time-consuming and require expert interpretation, thus, automated approaches are highly desirable. This study presents a novel deep learning framework for AD diagnosis using Electroencephalograph (EEG) signals, integrating multiple feature extraction techniques including alpha-wave analysis, Discrete Wavelet Transform (DWT), and Markov Transition Fields (MTF). A late-fusion strategy is employed to combine predictions from separate neural networks trained on these diverse representations, capturing both temporal and frequency-domain patterns in the EEG data. The proposed model attains a classification accuracy of 87.23%, with a precision of 87.95%, a recall of 86.91%, and an F1 score of 87.42% when evaluated on a publicly available dataset, demonstrating its potential for reliable, scalable, and early AD screening. Rigorous preprocessing and targeted frequency band selection, particularly in the alpha range due to its cognitive relevance, further enhance performance. This work highlights the promise of deep learning in supporting physicians with efficient and accessible tools for early AD diagnosis.

💡 Research Summary

The paper presents a novel deep‑learning framework for early detection of Alzheimer’s disease (AD) using electroencephalography (EEG) signals. Recognizing that AD progression is associated with characteristic slowing of the alpha rhythm—often observable before structural changes appear on MRI or CT—the authors aim to exploit EEG’s high temporal resolution for non‑invasive screening. Existing EEG‑based AD studies typically rely on a single feature set (e.g., connectivity metrics, power spectral density) or early‑fusion strategies, limiting their ability to capture the full richness of the signal. To address this, the authors propose a multi‑modal approach that extracts three complementary representations from the same EEG recordings: (1) raw alpha‑band time‑series, (2) discrete wavelet transform (DWT) coefficients, and (3) Markov Transition Field (MTF) images.

Data and preprocessing: The study uses a publicly available dataset comprising 88 subjects (36 AD, 23 frontotemporal dementia (FTD), 29 cognitively normal controls). Recordings were obtained with a 19‑channel 10‑20 system at 500 Hz during eyes‑closed resting state. The preprocessing pipeline includes a 0.5–45 Hz Butterworth band‑pass filter, Artifact Subspace Reconstruction (ASR) with a conservative threshold (SD = 17), ICA (RunICA) followed by ICLabel‑based removal of eye and jaw components, trimming of the first and last 500 samples, segmentation into 6‑second windows, and Z‑score normalization.

Feature extraction:

1. Alpha‑band time‑domain – A third‑order Butterworth filter (8–12.5 Hz) isolates the alpha rhythm, which is then down‑sampled to 25 Hz. These sequences are fed into a convolutional neural network (CNN) that learns temporal dynamics directly.
2. Discrete Wavelet Transform – After down‑sampling the alpha band to 50 Hz, a single‑level DWT using a 16‑tap Daubechies wavelet is applied. Only the approximation coefficients (low‑frequency component) are retained, producing a compact time‑frequency representation that is processed by a second CNN (ResNet‑style).
3. Markov Transition Field – The down‑sampled alpha series is quantized into Q bins; a first‑order Markov transition matrix is computed for each channel, and the resulting matrix is rendered as an image (MTF). A third CNN learns spatial patterns reflecting state‑transition probabilities.

Model architecture and late fusion: Each modality is trained independently with its own neural network, allowing each to specialize in capturing distinct signal characteristics (temporal, spectral, and state‑transition). At inference, the softmax probabilities from the three networks are combined via weighted averaging (late‑fusion), effectively performing decision‑level ensemble. This strategy preserves modularity, reduces overfitting, and leverages complementary information without requiring early concatenation of heterogeneous features.

Results: Using a subject‑wise validation scheme, the fused model achieves an overall classification accuracy of 87.23 %, precision of 87.95 %, recall of 86.91 %, and F1‑score of 87.42 % for distinguishing AD patients from non‑AD subjects (FTD + controls). These metrics compare favorably with prior works that report accuracies ranging from ~80 % (traditional ML on connectivity or PSD) to ~97 % (single‑modal deep models evaluated with cross‑validation). The authors attribute the performance gain to (i) focused analysis of the cognitively relevant alpha band, (ii) the inclusion of both time‑frequency (DWT) and state‑transition (MTF) representations, and (iii) the robustness conferred by late‑fusion.

Discussion and limitations: The study demonstrates that multi‑modal EEG representations can be effectively integrated via late‑fusion to improve AD detection. However, the dataset is relatively small (88 subjects), limiting the assessment of generalizability. The classification task is binary (AD vs. non‑AD), whereas clinical practice often requires differentiation among multiple dementias (e.g., AD, FTD, MCI). Hyper‑parameter selection and the specific weighting scheme for fusion are not exhaustively explored, which may affect reproducibility. Moreover, the computational cost of training three separate CNNs could be a barrier for real‑time clinical deployment.

Future directions: The authors suggest expanding the approach to larger, multi‑center cohorts, incorporating additional modalities (e.g., MRI, CSF biomarkers) for true multimodal fusion, exploring self‑supervised pre‑training to mitigate data scarcity, and developing lightweight architectures for on‑device inference.

Conclusion: By simultaneously harnessing temporal dynamics, spectral content, and Markovian state transitions of the EEG alpha rhythm, the proposed late‑fusion deep‑learning framework offers a promising, non‑invasive tool for early Alzheimer’s screening, potentially augmenting clinicians’ diagnostic workflows and enabling scalable population‑level monitoring.