Predicting Stress Using Time-Domain Feature Engineering via Smartphone Soft-Typing Data

People with significant intellectual disabilities frequently struggle to articulate their stress, which makes timely carer reactions difficult. Stress, a frequent problem in modern life, is often overlooked, emphasising the importance of early detection. Smartphones, which are continually used, generate a variety of behavioural data, making them perfect for stress detection. This paper offers an autonomous real-time stress detection system that use wearable sensors to capture physiological signals and machine learning to detect stress-related changes. Our study uses data from smartphone sensors, including linear acceleration, gravity, gyroscope, and touchscreen sensors, to detect stress through soft keyboard typing behaviours. We capture nuances in phone interactions under stress by extracting statistical time domain features from raw sensor data, such as mean, variance, standard deviation, and energy. Using feature selection techniques, we reduced the dataset to the most significant features, minimizing model complexity and improving accuracy. We applied advanced machine learning methods like bagging, boosting, and stacking classifiers. The AdaBoost classifier achieved a 94.21% accuracy, a 5.58% improvement using the selected feature subset compared to the full feature set. The optimized model also significantly reduced the model-building time to 1.73 seconds, demonstrating the efficiency and practicality of our approach for real-time stress detection.