Julia Clemente

Stress is one of the primary triggers of serious pathologies (e.g., depression, obesity, heart attack). Prolonged exposure to it can lead to addictive substance consumption and even suicide, without ignoring other adverse side effects in the economic, work and family spheres. Early detection of stress would relax the pressure of medical practice exercised by the population affected and result in a healthier society with a more satisfying quality of life. In this work, a convolutional-neural-network (CNN) model is proposed to detect an individual’s stress state by analyzing the diffusive dynamics of the photoplethysmographic (PPG) signal. The characteristic $(p,q)$-planes of the 0–1 test serve as a framework to preprocess the PPG signals and feed the CNN with the dynamic information they supply to typify an individual’s stress level. The methodology follows CRISP-DM (Cross Industry Standard Process for Data Mining), which provides the typical steps in developing data-mining models. An adaptation of CRIPS-DM is applied, adding specific transitions between the usual stages of deep-learning models. The result is a CNN model whose performance amounts to 97% accuracy in diagnosing the stress level; it compares with other published results.