Comprehending the impact of deep learning algorithms on optimizing for recurring impediments associated with stress prediction using ECG data through statistical analysis

Despite the myriad of stress related research studies, there has been very few studies which focused on the

complexity of the ECG signal/data, prior to predicting stress. In order to counter the problem of “data complexity

and overfitting”, we innovated ML (machine learning) approaches using transfer learning and autoencoder

techniques, in order to predict stress/not stress (2 classes) with high precision from WESAD dataset. We then

assessed the bias and variance associated with our algorithms through various statistical tests, in order to un-

derstand their ability to generalize well on newer data. Our proposed algorithms were able to achieve 98.99%

(CNN) and 98.92% (VGG16) accuracy through 10-fold cross validation, while maintaining a very low bias,

variance, akaike information criterion (AIC) of and Bayesian information criterion (BIC) scores, substantiating

their ability to predict stress with very high accuracy without overfitting. Results illustrates their ability to

generalize well on any stress related data, irrespective of their complexity. Our algorithms performed better than

every other related studies. Although we were able to reduce the time and space complexity of the algorithms

through transfer learning and autoencoder techniques, the algorithms still require more time and computational

power than simpler algorithms, something which will require more attention for our future work.