On-device single channel EEG classification on Android smartphones using lightweight machine learning models

Objective. Electroencephalogram (EEG) signals record brain activity through electrical measurements taken from the scalp, providing important insights for both cognitive research and clinical applications. Although deep learning approaches can deliver high classification accuracy for EEG data, they typically depend on large training datasets and significant computational power. To enable mobile deployment with limited data, we present a machine learning pipeline as a first step towards on-device offline EEG classification. The approach efficiently trains lightweight models for EEG signal classification on Android devices. As a case study, the approach was employed to distinguish between eyes open and eyes closed states using EEG signals.Approach.EEG data were collected from ten participants performing eyes-open and eyes-closed tasks using the CameraEEG app. Preprocessing involved the Embedded-Artifact Subspace Reconstruction (E-ASR) algorithm to remove artifacts, followed by power spectral feature extraction. A Support Vector Machine (SVM) classifier was trained using a single-channel occipital electrode and then deployed into the Android app for EEG classification.Main Results.The single-channel model achieved an accuracy of 90%, demonstrating its efficiency and suitability for mobile deployment. Further evaluation using precision, sensitivity, specificity, F1-score, and Matthews Correlation Coefficient (MCC), confirmed the model's robustness. The EEG classifier Android app, tested on devices such as Google Pixel 7 Pro and Samsung S22, successfully performed EEG signal classification.Significance.By enabling EEG classification with limited data on Android devices, this work enhances physiological measurement methodologies in natural environment settings. Beyond eye-state detection, this pipeline can be extended to applications such as cognitive workload monitoring, seizure detection, and mental health assessment. Limitations include dependence on multiple software platforms and non-realtime classification, which we plan to address in future work. This study demonstrates the feasibility of deploying machine learning-based EEG classifiers on smartphones, contributing to scalable and accessible EEG-based monitoring for research and clinical use.