Fuzzy-based machine learning methods for continuous diagnosis and prognosis of Diabetic Retinopathy

Identificador:  TDX:4385

Handle: https://hdl.handle.net/20.500.11797/TDX4385

Autores:  Pascual Fontanilles, Jordi

Resumen:
Disease diagnosis may be supported by Clinical Decision Support Systems (CDSS) that take advantage of existing medical knowledge and patients' information. Such systems are built using diverse AI and machine learning techniques and can be effective in reducing manual time-consuming tasks, analysing patients health records, or supporting non-expert clinicians in a field. This work focuses on Diabetic Retinopathy, a severe complication of Diabetes Mellitus, a chronic, widespread disease. As a consequence of diabetes, a patient might suffer vision loss and even blindness if not detected and treated at an early stage. The current screening procedure is based on images of the eye-fundus, which is time-consuming and costly. On the contrary, Retiprogram is a CDSS based on Fuzzy Random Forests (FRF) to help in the early diagnosis of DR using patients' clinical data. In this PhD thesis, we aim to study how a FRF classification model can take advantage of data in conditions of dynamic changes. The first contribution is to improve the current results of a binary FRF classifier by taking advantage of the data of the new patients that are treated at the hospital. The second contribution adapts the binary FRF classification procedure to the case of ordinal multiclass. This is particularly helpful to detect the severity of the disease, such as in DR, where ophthalmologists differentiate between different severity degrees of retinopathy. The third contribution is focused on the detection of DR in long-term diabetic patients. Due to continuous controls and medications, long-term diabetic people improve some clinical factors, which makes it much harder to predict the appearance and progression of DR. We propose a method to exploit the electronic health record history data to construct a temporal dataset to improve the grading of DR in long-term patients.