This paper presents TypeState, a privacy-preserving framework for detecting cognitive load from keystroke micro-rhythms using machine learning. The system operates on timing-derived features (flight time and rolling variance) without storing or analyzing text content, enabling privacy-preserving inference. A bidirectional LSTM model is trained on keystroke sequences and compared with conventional baselines such as Random Forest and SVM. Experiments are conducted on a pilot dataset collected from student typing sessions under relaxed and time-pressured conditions. Results indicate that sequential modeling improves classification performance over static approaches, while preliminary analysis suggests that stress may correspond to reduced variance in typing rhythms. This is a preliminary pilot study with a limited dataset and simplified experimental design. Results require validation through larger, controlled studies with subject-wise evaluation and validated stress measures.
Atharv Khare (Thu,) studied this question.