A Machine Learning Framework for Automatic and Continuous MMN Detection with Preliminary Results for Coma Outcome Prediction

Mismatch Negativity (MMN) is an important brain response measured using EEG and is widely used to predict recovery in patients who are in a coma.Traditional methods for identifying MMN require long EEG recordings and careful analysis by specialists, which makes the process slow and less effective i1n1real clinical situations. This study proposes an automated machine learning approach that can detect MMN continuously using short EEG segments of only two minutes. The system analyzes auditory oddball EEG signals by extracting statistical, frequency-based, and wavelet features. A Localized Feature Selection (LFS) method is then used to build reference models based solely on EEG data from healthy individuals. EEG data from coma patients is compared with these reference patterns using a similarity-based scoring approach. Experimental results using leave-one-subject-out validation showed an accuracy of 92.7% on healthy subjects. When tested on two coma patients, the model successfully detected MMN-positive time periods and accurately predicted their recovery, even when conventional averaging techniques failed. The results indicate that the proposed method offers a reliable, fully automated, and clinically useful solution for MMN detection and coma outcome prediction.