Ranking the Importance of Spatiotemporal Windows of EEG Signals Results in a Better Alzheimer's Disease Prediction

The integration of Electroencephalogram (EEG) measurements with machine learning holds the promise of enhancing diagnostic accuracy and providing personalized insights into the progression of neurodegenerative diseases (NDs) and Alzheimer's disease (AD) in particular. The complex nature of EEG signals, influenced by individual variability and noise, poses difficulties in interpreting the rich and dynamic embedded information, thus requiring algorithms capable of discerning meaningful patterns. In this work, we develop a novel approach for ranking the importance of spatiotemporal EEG information based on the Smart Aggregation Framework (SAF) framework in which each spatiotemporal window is weighted non-linearly using the Boltzmann distribution with a hyperparameter, analogous to temperature. We validate our model on a dataset that includes EEG recordings of 65 healthy and AD subjects. We rank the significant spatiotemporal windows for each subject and show that the features of the top-ranked windows provide significant separability between the AD and healthy subjects. We determine the most significant electrode and show that taking only the top two electrodes provides a better classification of the AD patients compared with taking all the electrodes or a random pair. Besides providing cutting-edge accuracy in classifying AD, our work provides an interpretability framework for ranking spatiotemporal information in EEG signals that can be harnessed to enhance the diagnostics of other neurodegenerative conditions.