Abstract
Motivation: In recent times, mental stress detection using physiological signals have received widespread attention from the technology research community. Although many motivating research works have already been reported in this area, the evidence of hardware implementation is occasional. The main challenge in stress detection research is using optimum number of physiological signals, and real-time detection with low complexity algorithm. Objective: In this work, a real-time stress detection technique is presented which utilises only photoplethysmogram (PPG) signal to achieve improved accuracy over multi-signal-based mental stress detection techniques. Methodology: A short segment of 5s PPG signal was used for feature extraction using an autoencoder (AE), and features were minimized using recursive feature elimination (RFE) integrated with a multi-class support vector machine (SVM) classifier. Results: The proposed AE-RFE-SVM based mental stress detection technique was tested with WeSAD dataset to detect four-levels of mental state, viz., baseline, amusement, meditation and stress and to achieve an overall accuracy, F1 score and sensitivity of 99%, 0.99 and 98% respectively for 5s PPG data. The technique provided improved performance over discrete wavelet transformation (DWT) based feature extraction followed by classification with either of the five types of classifiers, viz., SVM, random forest (RF), k-nearest neighbour (k-NN), linear regression (LR) and decision tree (DT). The technique was translated into a quad-core-based standalone hardware (1.2 GHz, and 1 GB RAM). The resultant hardware prototype achieves a low latency (~0.4 s) and low memory requirement (~1.7 MB). Conclusion: The present technique can be extended to develop remote healthcare system using wearable sensors.
Unsupervised Learning for Mental Stress Detection