A PRE-PROCESSING-FREE MENTAL STATE DETECTION MODEL USING NOISY ECG PLOTS AND DEEP TRANSFER LEARNING

In recent years, the prevalence of mental disorders, such as depression and stress, has been on the rise, yet a large number of individuals do not receive timely treatment. Addressing mental health concerns involves the evaluation of an individual’s mental state, which can be influenced by a variety of factors. Technological advancements have introduced smart wearable devices that enable real-time monitoring of vital signs, offering potential applications for self-care in mental health. However, the current methodology utilized by most of these devices relies on hand-crafted features and demands time-consuming pre-processing. To address this limitation, our research aims to develop a pre-processing-free model for real-world application, focusing on noisy electrocardiograph (ECG) signals for four-class mental state detection. For this purpose, we used an available wearable stress and affect detection dataset. We took raw ECG signals and transformed them into two-second plots, which were then fed into seven pre-trained convolutional neural networks (AlexNet, GoogLeNet, EfficientNetB0, VGG16, VGG19, XceptionNet, and InceptionV3). Through our experimentation, the fine-tuned VGG16 model emerged as the most effective, outperforming other techniques in accurately detecting baseline, stress, amusement, and meditation states, achieving an impressive accuracy of 99.35%. This achievement stands significantly higher than existing literature, making our model a suitable option for classifying mental states even in noisy raw ECG signals. Furthermore, it exhibits reduced computational complexity when compared to other state-of-the-art studies.