Deep learning and feature based medication classifications from EEG in a large clinical data set

Area of Science:

• Neuroscience
• Computational Biology
• Medical Informatics

Background:

• Increasing availability of human phenotypic data, including electroencephalographical (EEG) signals.
• Limited understanding of inferential capabilities from phenotypic data using advanced statistical methods.
• Emerging applications of machine learning in neurological signal analysis for diagnostics.

Purpose of the Study:

• To assess the predictability of patient medication status from EEG data.
• To evaluate the performance of deep learning and feature-based methods in classifying medication use.
• To investigate correlations between EEG signals and physician-reported medication data.

Main Methods:

• Utilized the Temple University EEG corpus, linking EEG records with physician reports.
• Applied and compared deep learning and feature-based machine learning approaches.
• Trained algorithms to distinguish between patients taking Dilantin (phenytoin), Keppra (levetiracetam), or no medications.

Main Results:

• Machine learning models successfully differentiated patients on anticonvulsants from those not taking medication.
• Algorithms could also distinguish between patients taking Dilantin (phenytoin) and Keppra (levetiracetam).
• Different analytical approaches showed varying effectiveness for distinct classification tasks.

Conclusions:

• EEG data holds significant potential for predicting medication status using machine learning.
• The study validates the use of EEG analysis in identifying patient treatment regimens.
• Findings suggest tailored machine learning strategies may optimize diagnostic accuracy in neurological applications.