July 29, 2023Open Access
Adazd-Net: Automated adaptive and explainable Alzheimer’s disease detection system using EEG signals
Key Result
The Adazd-Net model achieved an accuracy of 99.85%, sensitivity of 99.75%, and specificity of 100% in detecting Alzheimer's disease from EEG signals using ten-fold cross-validation.
Structured PICO
Does the Adazd-Net model improve the accuracy and explainability of Alzheimer's disease detection using EEG signals compared to traditional machine learning models?
P
Population
23 older adults (mean age 72.8 years, 48% female), comprising 12 with Alzheimer's disease and 11 with normal cognition, whose EEG signals were retrospectively analyzed to develop a diagnostic model.
I
Intervention
Adazd-Net (adaptive flexible analytic wavelet transform combined with an explainable boosting machine) for automated EEG signal analysis.
C
Comparator
Traditional machine learning models (SVM, KNN, decision tree, ANN, random forest) and traditional flexible analytic wavelet transform (FAWT).
O
Outcome
Accuracy of Alzheimer's disease detection using EEG signals.
The Adazd-Net model provides a highly accurate (99.85%) and explainable method for detecting Alzheimer's disease using EEG signals, potentially aiding clinicians in early diagnosis.
Limitations
- The study used only a single dataset comprising 663 epochs belonging to a small sample size of 23 subjects.
Abstract
Alzheimer’s disease (AZD) is a degenerative neurological condition that causes dementia and leads the brain to atrophy. Although AZD cannot be cured, early detection and prompt treatment can slow down its progression. AZD can be effectively identified via electroencephalogram (EEG) signals. But, it is challenging to analyze the EEG signals since they change quickly and spontaneously. Additionally, clinicians offer very little trust to the existing models due to lack of explainability in the predictions of machine learning or deep learning models. The paper a novel Adazd-Net which is an adaptive and explanatory framework for automated AZD identification using EEG signals. We propose the adaptive flexible analytic wavelet transform, which automatically adjusts to changes in EEGs. The optimum number of features needed for effective system performance is also explored in this work, along with the discovery of the most discriminant channel. The paper also presents the technique that can be used to explain both the individual and overall predictions provided by the classifier model. We have obtained an accuracy of 99.85% in detecting AZD EEG signals with ten-fold cross-validation strategy. We have suggested a precise and explainable AZD detection technique. Researchers and clinicians can investigate hidden information concerning changes in the brain during AZD using our proposed model. Our developed Adazd-Net model can be employed in hospital scenario to detect AZD, as it is accurate and robust.