Comparison of Machine Learning Models for Stress Detection from Sensor Data Using Long Short-Term Memory (LSTM) Networks and Convolutional Neural Networks (CNNs)

Stress has become one of the serious concerns in modern society and affects health, both mentally and physically, influencing productivity and quality of life. Chronic stress has grave health consequences, including cardiovascular diseases, anxiety, depression, and suppression of immune responses. Early detection and management are thus of utmost importance for mitigating its adverse effects. The increased popularity of wearable devices and physiological sensors now allows the detection of stressors using real-time data. The presented research is targeted at the study and comparison of the performance of two leading deep learning architectures: LSTM networks and CNN, based on stress detection according to physiological sensor data.

This research is based on a dataset of three important physiological parameters: Heart Rate Variability, Galvanic Skin Response, and Skin Temperature, recorded using wearable devices. All these data are filtered for noise and normalized for quality and consistency. In this paper, LSTM and CNN models have been designed, trained, and tested on the same datasets so that a fair comparison can be performed. Accuracy, precision, recall, F1-score, and computational efficiency were the metrics considered to evaluate the model's effectiveness.

These results emphasize the fact that LSTM networks outperform CNNs in terms of extracting temporal dependencies in time-series data, yielding high accuracy, recall, and F1-score. This underlines the fact that LSTMs are much more appropriate in applications such as detection of stress, which requires the deeper understanding of sequential patterns. On the other hand, CNNs do have more computational efficiency and multi-dimensional feature extraction capability, useful in cases where real-time performance is required using limited resources.

This comparative analysis not only points out the trade-offs of the two models but also serves as a pointer to the potential hybrid approaches which integrate the strengths of LSTM and CNN architectures. The findings of this study give practical insights to the researchers, developers, and practitioners in the domain of health monitoring and stress detection. It therefore enhances the knowledge to date on the ability and limits of the models in the general domain of the application of machine learning in mental health, hence more accurate, scalable, yet efficient automatic stress detection systems are enabl

ed.