| dc.description.abstract | Wearable sensors may obtain reliable physiological signals to diagnose diseases and detect changes. Wearables can measure signs such as electrocardiogram, heart rate, electroencephalogram, electromyogram, or galvanic skin response. All these signals have intrinsic characteristics in a normal state and change if associated with illness. The literature presents the Machine Learning Approaches and Deep Learning Models as alternatives to pattern detection in physiological signals. The state-of-the-art in this area indicates a trend to use wearables for continuous monitoring of patients, whether in a hospital or home environment, as it is a portable and noninvasive option. In addition, many studies point out the low cost of wearable sensors as another advantage compared to traditional hospital medical equipment. Other studies highlight the possibility
of supporting automatic diseases detection, especially chronic diseases, using artificial
intelligence in physiological signals. Based on the review carried out, it is possible to conclude that there are still new development opportunities. The studied papers do not address at the same time aspects of lower cost, greater flexibility, wide use of Machine Learning resources, and communication of results. This work’s main objective is to develop an architecture for multisignal acquisition with wearable sensors for continuous monitoring and stress detection. The architecture comprises wearable sensors and single-board computers—the wearable sensor for data processing and single-board computer to communicate the results to other platforms. The differentials of this architecture consist of the integration of resources for multi-signal acquisition for continuous monitoring of patients with a low implementation cost, flexibility, and ease of use. We developed a prototype in a modular way, and we tested each module of the architecture. These tests aimed to guarantee the independence of the components, carefully evaluating the stability and plausibility of the data. We also carried out two practical stress-inducing experiments.
The first composed a proprietary dataset to generate a Machine Learning model, and the
second allowed full architecture assessment, focusing on real-time detection. The training and classification results of the Machine Learning model showed promising results, with accuracy above 98.72% for binary classification and 92.72% for classification with three classes. When analyzing the real-time classification, we obtained an accuracy of 69.00% for participants in the first round of experiments. The architecture presented excellent communication and operation stability. During the experiments, the architecture performed short and long acquisitions efficiently. The acquired data showed promising results, with plausible and justifiable values within the context of the experiment performed. The classification results obtained when testing the model with participants who were in training, the results were relatively high. | en |
