Towards Portable Mental Stress Detection Systems: A TinyML Approach for EEG Analysis

Electroencephalogram (EEG) signals are a reliable way to understand brain activity, especially to detect mental stress, a condition that can cause serious health problems if left unchecked. Although deep learning models show promise in identifying stress from EEG data, running these models on portable devices is difficult due to their computational demands. To remedy this problem, this paper proposes a novel QAT_FFNN model, a lightweight Feed-Forward Neural Network optimised using tiny machine learning techniques such as Quantisation Aware Training and Pruning. The proposed QAT_FFNN model achieves state-of-the-art 99.74% accuracy on the EEGMAT dataset, with a 91.3% model size reduction and 84.77 ms inference latency, enabling real-time feasibility on smartphones and smartwatches. By balancing high performance with efficiency, QAT_FFNN bridges the gap between laboratory-grade accuracy and edge-device practicality, offering a foundation for accessible, continuous stress monitoring to support proactive mental healthcare interventions.