SMITH - Smart MonITor Health system

Abstract

Diabetes is a huge health problem that is a ecting more and more people over the time. When it comes to diagnosing such disease in people, doctors make their diagnoses based on some proper tests and may not take into consideration other factors that are related to the disease. The creation of tools that can analyse information about the current health status of patients can support doctors by providing more information for the diagnosis. Since there is still no cure for this disease, a person who has been diagnosed with diabetes has to control his blood sugar level between some thresholds. This is extremely important as non-controlled level of glucose can lead the subject to severe health complications and compromise his lifestyle. Tools that forecast the subject's glucose level within a prediction horizon may let the individual take preemptive actions to avoid crossing the normal thresholds. This thesis aims to investigate machine learning methods for such problems. These types of methods are already being used in the medicine and will allow us to come up with computational models that o er more relevant data to support the medical team when it comes to diagnosing diabetes in people and avoid thresholds overpassing when patients are controlling their glucose level. Both these problems represent highly challenging tasks. For the diabetes diagnosis problem, we built several models, tuned and tested them using the PIMA dataset. A key contribution of this work is the diverse methods introduced, analysed and tested to handle missing values present in the dataset. The method of substituting the missing values by the mean of the features considering the class they belong to along with Random Forest yielded the best results with an accuracy of 87.66%. Regarding the glucose level ii predictions we also created various models based on real patient datasets provided by Associa c~ao Protectora dos Diab eticos de Portugal(APDP), which are an important asset. Besides, two prediction methods namely direct and iterative prediction methods were investigated and tested. From the computational experiments, the Linear Regression with direct prediction method is the most advantageous combination resulting on an RMSE average of 14.25 mg/dL and 23.46 mg/dL for 30 and 60 minutes ahead prediction. Both these domains represent highly challenging tasks and our methods demonstrate that we can attain excellent performance on these tasks. From an application standpoint, there remains many challenging problems in both Diabetes Diagnosis and Prediction of Glucose. Indeed, in the advent of the Internet of Things by combining many sensors available with our methods, we will come to reach a Smart Monitor Health System able to prevent, diagnosis, treatment and after care for the society in general.