Toward Real-Time Cognitive State Sensing in Everyday Work: A Lightweight CNN-Transformer for 4-Channel Wearable EEG

Knowledge work involves frequent transitions between cognitive modes such as focused effort, engagement, relaxation, and sustained immersion. Detecting these shifts in real time could support reflective productivity tools, adaptive work interfaces, and early intervention for cognitive overload. However, most EEG-based cognitive-state classifiers rely on research-grade hardware, handcrafted features, or narrowly defined binary states, limiting their suitability for everyday interactive systems. We present a real-time cognitive-state sensing pipeline for consumer-grade 4-channel wearable EEG that combines Short-Time Fourier Transform (STFT) spectrograms with a lightweight CNN-Transformer architecture. Using four channels corresponding to a 10-20 international system (T3, T4, O1, O2), our approach classifies four target states: Focus, Relaxation, Engagement, and Flow, without manual feature engineering. Using a subjectstratified dataset of 6,220 four-second segments, the model achieved 94.69% test accuracy, while operating with an initial latency of approximately 4.2 seconds and 0.5-second update intervals. Beyond model performance, we contribute a design-oriented analysis of what minimal-channel EEG can and cannot support in interactive workplace settings. Our results suggest that end-to-end spectrogram learning can approach the performance of handcrafted-feature pipelines while reducing engineering overhead and enabling faster adaptation to new devices or cognitive-state definitions. At the same time, we show that high performance under controlled task-derived labels should not be conflated with validated inference of lived psychological states, especially for Flow. We discuss implications for human-centred cognitive sensing, ecological validity, calibration, interpretability, and future deployment in adaptive and reflective work systems.