AI-Based Early Warning System for Mental Fatigue

Abstract

Detecting mental fatigue before its effects become operationally dangerous is a critical challenge in modern knowledgeintensive workplaces. This paper presents an AI-Based Early Warning System that infers cognitive fatigue from passively collected keyboard and mouse interaction signals, requiring no wearable sensors or physiological instrumentation. A hybrid Convolutional Neural Network–Long Short-Term Memory (CNN-LSTM) architecture is trained on nine behavioral features extracted from oneminute sliding windows applied to the SWELL-KW dataset. Class imbalance is addressed via weighted binary cross-entropy loss. An optimised classification threshold of 0.56 yields a test accuracy of 73.14% with a ROC-AUC of 0.7616. A Streamlit desktop application delivers real-time three-tier risk alerts and batch prediction on any standard personal computer.