07 May 2024
10:00 Master's Defense IC3 Auditorium
Theme
Sedentary lifestyle detection via biosignals, machine learning and visual analysis
Student
Alexis Aldo Mendoza Villarroel
Advisor / Teacher
Anderson de Rezende Rocha
Brief summary
A sedentary lifestyle is linked to several health problems, including type 2 diabetes, metabolic syndrome and cardiovascular disease. Sedentary behavior involves minimal energy expenditure during activities such as sitting or reclining. Machine learning offers promise for assessing physical activity intensity, but challenges remain, such as a lack of standardized datasets and model interpretability. We propose a novel self-attention-based neural network composed with 1D convolutional layers. It uses an attention mechanism that leverages sensor modalities and domain embeddings to create a feature representation for evaluating the intensity of physical activity. This allows us to observe the importance of different modalities for model predictions. Furthermore, we developed an interactive framework that provides an overview of datasets and visualizes attention weights. This framework makes it easier for healthcare experts to interpret model decisions, allowing them to better understand how sensor data contributes to predictions of activity intensity. We perform experiments on publicly available datasets and explore the ability of smartwatch sensor data (IMU and physiological signals) to detect sedentary behavior. Our model achieved superior performance compared to existing methods for physical activity intensity estimation, especially when evaluating across datasets. This work demonstrates the effectiveness of our approach in capturing the importance of sensor modalities and positioning, while providing valuable insights for healthcare experts monitoring patient behavior for personalized interventions.
Examination Board
Headlines:
| Anderson de Rezende Rocha | IC / UNICAMP |
| João Paulo Papa | FC / UNESP |
| Jacques Wainer | IC / UNICAMP |
Substitutes:
| Levy Boccato | FEEC / UNICAMP |
| Celmar Guimaraes da Silva | FT / UNICAMP |