The detection and classification of stress is essential for advancing mental health diagnosis and improving human well-being. Accurate stress assessment can lead to early intervention, personalized treatment plans, and improved quality of life. By identifying levels and types of stress, healthcare professionals can better understand and treat the underlying causes, promoting mental resilience and overall health. This study presents a novel method for stress detection and classification using electroencephalogram (EEG) data combined with neural network modeling. We propose a multi-layer neural network architecture optimized to analyze EEG frequency bands and extract stress-related biomarkers with high precision. The methodology includes preprocessing EEG signals to reduce noise using blind source separation (BSS) techniques, followed by feature extraction focusing on frequency bands associated with stress responses. The neural network is trained on labeled EEG datasets to classify stress levels, demonstrating significant accuracy and outperforming conventional classifiers. This method is implemented on an embedded Raspberry Pi system to capture and analyze data in real-time. The results, stemming from the integration of BSS, neural networks, and the embedded system, indicate that this approach offers a reliable and efficient means for stress detection, with potential applications in mental health monitoring and adaptive biofeedback systems. This work contributes to the field by introducing an innovative, data-driven model that enhances the precision and scalability of EEG-based stress classification.