A Comparison of Machine Learning Gesture Recognition Techniques for Medication Adherence

Abstract

Every year, many poor health outcomes are the result of patients missing their medication, as prescribed by their healthcare providers. Guidance and reminders to these patients would result in better health outcomes and significant financial savings to the economy. This thesis utilizes accelerometers and gyroscopes, which are widely available inside devices (e.g., smart phones and watches) to actively monitor patient activities, including those related to adherence to medication regimens. Different machine learning techniques are compared for recognizing when a pill bottle has been opened. Such actions could remind the patient to take their medication if an opening were not detected. An artificial neural network (ANN) model will be compared with a support vector machine (SVM) and a K-nearest neighbor (KNN) classifier. The models are trained on data collected by former University of Oklahoma students. Raw (normalized) sensor data is used, without extensive data processing or feature extraction. A neural network proves the most promising with an accuracy of 98.12%, as well as the greatest flexibility in data pre-processing requirements. KNN achieved high accuracy, although results were likely due to overfitting limited data with the simple model. SVM did not perform as well as the others, however; it did achieve similar results to previous research utilizing the approach (e.g., ~95% accuracy). Data collected from a greater number of gestures and additional test subjects is needed to verify generalization. A medication adherence system utilizing the developed model would be an acceptable approach.