Analyzing the Performance of Feature Selection on Regression Problems: A Case Study on Older Adults’ Functional Profile

• Autor(a) principal: Rojo, Javier
• Data de Publicação: 2022
• Outros Autores: Pinho, Lara, Fonseca, César, Lopes, Manuel, Helal, Sumi, Hernández, Juan, Garcia-Alonso, Jose, Murillo, Juan Manuel
• Tipo de documento: Artigo
• Idioma: por
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10174/32472; https://doi.org/Rojo, J., Pinho, L.G., Fonseca, C., Lopes, M.J., Helal, A., Hernández, J., Murillo, J., Garcia-Alonso, J. (2022). Analyzing the Performance of Feature Selection on Regression Problems: a Case Study on Older Adults' Functional Profile. IEEE Transactions on Emerging Topics in Computing. https://doi.org/10.1109/TETC.2022.3181679; https://doi.org/10.1109/TETC.2022.3181679
• topic: aging informatics; ehealth; Feature selection; regression; healthcare data analytic; machine learning

Healthcare systems are capable of collecting a significant number of patient health-related parameters. Analyzing them to find the reasons that cause a given disease is challenging. Feature Selection techniques have been used to address this issue—reducing these parameters to a smaller set with the most ”determinant” information. However, existing proposals usually focus on classification problems—aimed to detect whether a person is or is not suffering from an illness or from a finite set of illnesses. However, there are many situations in which health professionals need a numerical assessment to quantify the severity of an illness, thus dealing with a regression problem instead. Proposals using Feature Selection here are very limited. This paper examines several Feature Selection techniques to gauge their applicability to the regression-type problems, comparing these techniques by applying them to a real-life scenario on the functional profiles of older adults. Data from 829 functional profiles assessments in 49 residential homes were used in this study. The number of features was reduced from 31 to 25—with a correlation between inputs and outputs of 0.99 according to the R2 score and a Mean Square Error (MSE) of 0.11—or to 14 features—with a correlation of 0.98 and MSE of 5.73.