QEEG‐based TabNet classifier for dementia pathologies: Alzheiemer’s disease and Lewy body disease

Abstract

AbstractBackgroundEarly screening and prevention of dementia are widely discussed and studied in the current healthcare paradigm. However, about 70 percent of dementia diagnoses are categorized into Alzheimer’s disease (AD), neglecting the presence and influence of Lewy body. The present study aims to distinguish Alzheimer’s disease dementia (pure AD) and Lewy body dementia (pure LBD) through quantitative electroencephalogram (QEEG)‐based classification algorithm. The mixed dementia that exhibits both AD and LBD related propensities were further tested for the verification of the established classification model.Method19 channel EEG data were recorded in accordance with the 10‐20 system in eyes‐closed resting state. The acquired data were bandpass filtered into the range of interest (1‐45Hz) and re‐referenced into common average reference. Artifacts were removed by bad epoch rejection and independent component analysis (ICA) through an automated, cloud‐based QEEG analysis platform iSyncBrain®. Henceforth, sensor‐level features were extracted and computed for the training of TabNet structure. Due to the data imbalance (N = 146, Pure AD = 48, Pure LBD = 102), time window‐based augmentation has been applied to establish the final dataset (N = 7292, Pure AD = 3376, Pure LBD = 3916), which were then split into 8 to 2 ratio for the training and testing of the classification model. In addition, mixed dementia (N = 136) data were used to further validate the model’s performance.ResultThe best classification model yielded test accuracy of 80% with pure AD sensitivity at 80% and pure LBD sensitivity 80%. Moreover, the predicted probability distribution for the mixed dementia data were prominently concentrated around 0.5, which reflect that both pure AD and LBD properties exist.ConclusionAlong with the promising classification performance of the established classification model, the centered probability distribution of the mixed dementia data reflect that the model can well distinguish the characteristics that are specific to pure AD and LBD.