Machine Learning to Predict Brain Amyloid Pathology in Pre-dementia Alzheimer’s Disease Using QEEG Features and Genetic Algorithm Heuristic
The use of positron emission tomography (PET) as the initial or sole biomarker of β-amyloid (Aβ) brain pathology may inhibit Alzheimer’s disease (AD) drug development and clinical use due to cost, access, and tolerability. We developed a qEEG-ML algorithm to predict Aβ pathology among subjective cog...
Guardado en:
| Autores principales: | , , , |
|---|---|
| Formato: | article |
| Lenguaje: | EN |
| Publicado: |
Frontiers Media S.A.
2021
|
| Materias: | |
| Acceso en línea: | https://doaj.org/article/434d140441cf4c15a15ea6ef0ba9e065 |
| Etiquetas: |
Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
|
| Sumario: | The use of positron emission tomography (PET) as the initial or sole biomarker of β-amyloid (Aβ) brain pathology may inhibit Alzheimer’s disease (AD) drug development and clinical use due to cost, access, and tolerability. We developed a qEEG-ML algorithm to predict Aβ pathology among subjective cognitive decline (SCD) and mild cognitive impairment (MCI) patients, and validated it using Aβ PET. We compared QEEG data between patients with MCI and those with SCD with and without PET-confirmed beta-amyloid plaque. We compared resting-state eyes-closed electroencephalograms (EEG) patterns between the amyloid positive and negative groups using relative power measures from 19 channels (Fp1, Fp2, F7, F3, Fz, F4, F8, T3, C3, Cz, C4, T4, T5, P3, Pz, P4, T6, O1, O2), divided into eight frequency bands, delta (1–4 Hz), theta (4–8 Hz), alpha 1 (8–10 Hz), alpha 2 (10–12 Hz), beta 1 (12–15 Hz), beta 2 (15–20 Hz), beta 3 (20–30 Hz), and gamma (30–45 Hz) calculated by FFT and denoised by iSyncBrain®. The resulting 152 features were analyzed using a genetic algorithm strategy to identify optimal feature combinations and maximize classification accuracy. Guided by gene modeling methods, we treated each channel and frequency band of EEG power as a gene and modeled it with every possible combination within a given dimension. We then collected the models that showed the best performance and identified the genes that appeared most frequently in the superior models. By repeating this process, we converged on a model that approximates the optimum. We found that the average performance increased as this iterative development of the genetic algorithm progressed. We ultimately achieved 85.7% sensitivity, 89.3% specificity, and 88.6% accuracy in SCD amyloid positive/negative classification, and 83.3% sensitivity, 85.7% specificity, and 84.6% accuracy in MCI amyloid positive/negative classification. |
|---|