A0662
Title: Supervised and ANN classification to physical activities performed by pregnant women recorded via accelerometer data
Authors: Rafael Biagioni Fazio - FMRP-USP (Brazil)
Christoph Michael Mitschka - ICMC-USP (Brazil)
Gleici Perdona - USP (Brazil) [presenting]
Abstract: Physical activity, even if just of light intensity, is helpful to mother and fetal health during pregnancy and after delivery. It is critical to analyze the amount of physical activity among pregnant women from various socioeconomic backgrounds and lifestyles, to better understand the elements that influence their physical activity habits. This research aims to define and classify physical activities undertaken by 150 pregnant women in the Ribeiro Preto city in Brazil at the Unified Health System (SUS), based on data generated using accelerometers and an application to record the performed physical activities. To assess the need for physical activity, a virtual assistant is being developed in the project https://eva.fmrp.usp.br/ - EVA, which aims to follow pregnant women during this period and analyze their need for physical activity and based on the results, make recommendations. Consequently, caring for the health of the pregnant woman. For this purpose, data cleaning and processing techniques were applied and then, for supervised classification, were considered LightGBM (tree-based gradient boosting) and artificial neural networks of the type of Long short-term memory (LSTM). Among the conclusions, training using a 30-second period is pointed out as the approach with the best accuracy metrics. Some weaknesses of this approach were also identified, and possible improvements were derived.
