Loading...
Publication
Análise e avaliação de algoritmos para reconhecimento de posturas com base em matrizes de pressão
| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 2.8 MB | Adobe PDF |
Authors
Abstract(s)
O processo de controlar e monitorizar o estado do paciente durante as suas horas de sono tem vindo a ajudar no estudo de diminuição de riscos de saúde, nomeadamente em situações nas quais os maus hábitos noturnos ou mau posicionamento do corpo causam diferentes tipos de problemas no seu respetivo bem-estar. Vários estudos têm sido feitos neste âmbito, nomeadamente para a classificação de posturas das pessoas que se encontram acamadas, cada um deles seguindo diferentes abordagens. Por exemplo, técnicas que utilizam matrizes de pressão obtidas de sensores posicionados sobre o colchão, juntamente com algoritmos de machine learning (ML), têm alcançado precisões de até 99% (e.g., [1]) na classificação das posturas principais (virado para cima, virado para baixo, lateral esquerdo e lateral direito).
Este trabalho centra-se na utilização de algoritmos de ML aplicados a dados obtidos por uma matriz de sensores de pressão posicionada por baixo do colchão, de forma a minimizar o desconforto para o paciente monitorado. Usando este conjunto de dados, são aplicados e avaliados algoritmos para a classificação das posturas principais e as suas variações, com o objetivo de desenvolver uma solução eficaz e menos invasiva para monitorização da postura.
Os experimentos incluíram 10 algoritmos de ML, que alcançaram uma accuracy média entre 79,14% e 98,93% com validação Group K-Fold e entre 80,03% e 97,14% com a técnica Leave-One-Group-Out (LOGO) para classificar as quatro posturas principais. O estudo também expandiu a classificação para 28 variações de postura (7 variações para cada postura principal), com o algoritmo SVM alcançando uma accuracy de 65,18% na validação Group K-Fold. Isso representa uma melhoria significativa em relação a estudos anteriores, especialmente ao considerar o maior número de posturas. Comparações com outras abordagens que utilizam sensores de pressão mostram que esta solução, com sensores posicionados por baixo do colchão, oferece uma accuracy comparável e, em alguns casos, superior. Esses resultados sugerem que o uso de sensores por baixo do colchão é uma alternativa eficaz e menos intrusiva para a classificação de posturas.
Abstract: The process of controlling and monitoring a patient's condition during their sleeping hours has been helping to reduce health risks, particularly in situations where poor nighttime habits or poor body positioning cause different types of problems to their well-being. Several studies have been carried out in this area, particularly for the classification of postures of people who are bedridden, each of them following different approaches. For example, techniques that use pressure matrices obtained from sensors positioned on the mattress, together with machine learning (ML) algorithms, have achieved accuracies of up to 99% (e.g., [1]) in the classification of the main postures (face up, face down, left side and right side). This work focuses on the use of ML algorithms applied to data obtained by a matrix of pressure sensors positioned under the mattress, to minimize discomfort for the monitored patient. Using this dataset, algorithms for classifying the main postures and their variations are applied and evaluated, with the aim of developing an effective and less invasive solution for posture monitoring. The experiments included 10 ML algorithms, which achieved an average accuracy between 79.14% and 98.93% with Group K-Fold validation and between 80.03% and 97.14% with the Leave-One-Group-Out (LOGO) technique to classify the four main postures. The study also expanded the classification to 28 posture variations (7 variations for each main posture), with the SVM algorithm achieving an accuracy of 65.18% in Group K-Fold validation. This represents a significant improvement over previous studies, especially when considering the larger number of postures. Comparisons with other approaches using pressure sensors show that this solution, with sensors positioned under the mattress, offers comparable and, in some cases, superior accuracy. These results suggest that the use of sensors under the mattress is an effective and less intrusive alternative for posture classification.
Abstract: The process of controlling and monitoring a patient's condition during their sleeping hours has been helping to reduce health risks, particularly in situations where poor nighttime habits or poor body positioning cause different types of problems to their well-being. Several studies have been carried out in this area, particularly for the classification of postures of people who are bedridden, each of them following different approaches. For example, techniques that use pressure matrices obtained from sensors positioned on the mattress, together with machine learning (ML) algorithms, have achieved accuracies of up to 99% (e.g., [1]) in the classification of the main postures (face up, face down, left side and right side). This work focuses on the use of ML algorithms applied to data obtained by a matrix of pressure sensors positioned under the mattress, to minimize discomfort for the monitored patient. Using this dataset, algorithms for classifying the main postures and their variations are applied and evaluated, with the aim of developing an effective and less invasive solution for posture monitoring. The experiments included 10 ML algorithms, which achieved an average accuracy between 79.14% and 98.93% with Group K-Fold validation and between 80.03% and 97.14% with the Leave-One-Group-Out (LOGO) technique to classify the four main postures. The study also expanded the classification to 28 posture variations (7 variations for each main posture), with the SVM algorithm achieving an accuracy of 65.18% in Group K-Fold validation. This represents a significant improvement over previous studies, especially when considering the larger number of postures. Comparisons with other approaches using pressure sensors show that this solution, with sensors positioned under the mattress, offers comparable and, in some cases, superior accuracy. These results suggest that the use of sensors under the mattress is an effective and less intrusive alternative for posture classification.
Description
Keywords
Matrizes de pressão Algoritmos de reconhecimento Mantas de sensores Posturas na cama Machine learning Pressure matrices Recognition algorithms Sensor blankets Bed postures