FCA-Based Online Classifier for Medical Data Streams Under the MOA Framework

Continuous monitoring of physiological signals requires models that learn online, adapt to concept drift, and remain interpretable for clinical decision support [1], [4], [5]. Electroencephalography (EEG) is widely used in healthcare for applications such as seizure detection and fatigue monitoring, where real-time analysis enables timely intervention. Ensemble methods achieve strong accuracy on data streams [9]–[11], whereas Formal Concept Analysis (FCA) provides transparent concept rules and has demonstrated competitive performance in batch settings [2], [14]. We introduce On lin eFCA, a weighted nominal-concept classifier that extends FCA to streaming conditions through a two-phase process: a grace period for initialization followed by incremental updates with weighted sampling and rejection handling. The method is implemented in the Massive Online Analysis (MOA) framework [3], ensuring reproducibility and direct comparison with existing stream learners. Experiments on the EEG Eye State dataset demonstrate that OnlineFCA outperforms baselines such as Naive Bayes, kNN, and Hoeffding Tree [8], while remaining below optimized ensembles such as Adaptive Random Forest [9]. Crucially, OnlineFCA produces explicit and human-readable rules, providing interpretability that is essential for real-time clinical monitoring. These findings highlight the promise of FCA-based classifiers for medical data streams, where transparent decision support is as critical as predictive accuracy [14].