The difficulty of enhancing the detection of schizophrenia from EEG signals using machine learning and sophisticated feature selection is addressed in this work. In order to efficiently reduce the feature set from 1140 to 58 while maintaining important information, the suggested strategy makes use of Monarch Butterfly Optimization (MBO). Preprocessing, class balancing, and noise reduction are applied to the EEG data prior to feature extraction. Lightweight classifier models appropriate for limited contexts are therefore made possible by using MBO to optimize the feature set for computational efficiency. The classifier was assessed using detailed metrics such as ROC curves, precision-recall, and confusion matrices, and it achieved training accuracy of 94.26% and validation accuracy of 89.79%. The system’s ability to detect schizophrenia in real time is demonstrated by its performance, which offers a reliable and scaleless solution for realistic embedded system deployment. Contextualized within the larger field of EEG-based neurological diagnoses, this work tackles issues such as noisy data and limited sample sizes that frequently compromise the dependability of machine learning models in clinical settings.