| Summary: | Medication adherence is a real problem among older adults which can lead to serious repercussions on their health and life. Adherence is defined by the World Health Organization as the extent to which the behavior of a person corresponds with recommendations from a health care provider. A low medication adherence to a certain prescription can undermine the treatment benefits in many cases. Moreover, taking wrong medication may lead to unwanted secondary effects, adverse health conditions, and visits to the hospital. This dissertation describes the work focused on the design, development, and research of a solution for monitoring medication boxes using attached sensors. The main contributions of this work include the development of a mobile application, a study on how to classify data from medication box gestures, an implementation of the algorithm that retrieves data from sensor boxes, and an integration of the data classification algorithm into the mobile application. A medication reminder proof-of-concept was developed in the scope of this Master’s project. Sensor data is received by the prototype through a module that integrates the connection and data transference from the sensor boxes via wireless communication. Another module implements metric extraction functions that are applied to the inertial sensor data retrieved from the sensor box. The calculated metrics, herein corresponding to features, are passed to a machine learning algorithm, integrated in the data classification and feature extraction module, for posterior data identification. An in-depth analysis on how to classify inertial data from medication box gestures was conducted during the development of the solution. This in-depth analysis included the creation of two datasets with different characteristics which were preprocessed and fed to several machine learning algorithms. The analysis of the results outputted by the algorithms is included in this document. The dataset collection took place in two different locations, corresponding to a controlled environment and to a non-controlled environment. The obtained results showed that it is possible to identify the gestures in the dataset for the controlled environment, with the best results achieving a true positive rate of 97:9%. The results obtained for the dataset of the non-controlled environment (which was created with target users) showed that there are still many aspects that need to be improved before a final version of the solution is released.
|
|---|