Adaptação de algoritmos híbridos baseados em aprendizagem de máquinas para aplicação em problemas na área de Saúde com bases de dados desbalanceadas

Abstract

Recently, several researchers have used models developed from Machine Learning algorithms as computational tools to help solve problems in medicine and public health. These models can help specialists in decision making regarding events related to healthcare, improving the quality of care and enabling cost reduction. However, these researchers also face some challenges and three of them are addressed in this work: class imbalancing problem; interpretability of models; and classification uncertainties. Thus, this work presents two approachs. The first approach deals with data imbalance and classification uncertainties and is applied in monitoring and preventing falls in the elderly. This approach uses a Swarm Intelligence algorithm with membership windows formed from signals captured by Radio Frequency Identification (RFID) devices to monitor and detect the movement of elderly participants out of bed. The second approach is used to predict Stroke on a highly unbalanced dataset. In the approach, an Artificial Immune System handles data imbalance and a Decision Tree provides an easy-to-understand classification model. Both approaches have better results when compared to state-of-the-art approaches, and this is an important step towards the development of Machine Learning-based technologies applied in healthcare.