Electroencephalographic Biomarker-Guided Early Detection of Alzheimer’s Disease via Cortically Subdivided Neurodynamic PINN
Figure 1: Detailed network structure of our proposed CodimNet.
Figure 2: EEG cortical power spectral density (PSD) topographical distributions across distinct cognitive states, delineating oscillatory frequency-dependent neurophysiological divergences among HC, MCI, and AD.
Figure 3: Visualize the multi-branch structure.
Figure 4: Visualize the ND-PINN structure.
The proposed CodimNet is an advanced model for early diagnosis of Alzheimer's disease (AD) based on EEG. This network combines multi-branch feature extraction, designed according to the anatomical segmentation of the international 10-20 system, with the Neurodynamic Physics-Informed Neural Network (ND-PINN) to better capture the characteristics of EEG signals in AD patients. It outperforms nine other methods on the CAUEEG dataset, achieving state-of-the-art accuracy in AD identification.
We run CodimNet and previous methods on a system running Ubuntu 22.04, with Python 3.8, PyTorch 2.1.0, and CUDA 12.1.
Figure 5: Comparison of CodimNet and other methods for AD diagnosis on the CAUEEG dataset using a fixed hold-out validation approach with EEG data.
CodimNet attaining peak performance across key classification metrics—Acc (76.27%), AUC (85.27%), and Sp (88.28%)—outperforming its closest competitor, EEGNet, with a Se enhancement of 0.29 percentage points, an F1-score improvement of 2.07 percentage points, and a G-Mean augmentation of 1.89 percentage points, indicative of its enhanced robustness in mitigating class imbalance effects intrinsic to medical datasets.
Figure 6: Comparison of CodimNet and other methods for AD diagnosis on the non-overlapping version of the CAUEEG dataset using a four-fold cross-validation approach with EEG data.
Figure 7: Quantitative assessment of regional contributions to AD classification efficacy through cortical branch-specific ablation analysis.
Figure 8: Quantitative evaluation of model classification efficacy across AD EEG tri-classification under varying instantiations of ND-PINN.
Figure 9: Quantitative evaluation of ND-PINN instantiation across distinct intermediate-layer constraints.
Figure 10: Quantitative assessment of the impact of ND-PINN removal across distinct cortical branches on classification performance.
Figure 11: Quantitative classification performance across distinct recurrent neural network architectures instantiated within a multibranch paradigm.
Figure 12: Performance metrics of ND-PINN under differential brain rhythm constraints, evaluating the impact of rhythm-specific spectral ablation on AD EEG classification.
The effectiveness of the proposed module was systematically evaluated through the four-fold cross-validation ablation experiment, and its applicability and advantages in related tasks were verified.
In order to verify the effectiveness of CodimNet model in extracting EEG features from patients with Alzheimer's disease, a comparative analysis diagram and heat map including time-frequency graph, original PSD graph and ND-PINN treated PSD graph were constructed based on the hook function's forward propagation layer feature visualization technology.
Figure 13: Comparison of time-frequency spectral features of middle-layer embeddings between CodimNet and RNN.
Figure 14: Comparison of PSD plots with and without ND-PINN.
Figure 15: The heatmap generated by CodimNet shows the neuronal activation intensity of AD, MCI, and HC.
if you have any questions, please contact 'zhengliang.zhang@hdu.edu.cn'