Resumen de Development of machine learning techniques for diabetic retinopathy risk estimation
Emran Saleh Ali Ali
Diabetic retinopathy (DR) is a chronic illness. It is one of the main complications of diabetes, and an essential cause of vision loss among people suffering from diabetes. Diabetic patients must be periodically screened in order to detect signs of diabetic retinopathy development in an early stage. Early and frequent screening decreases the risk of vision loss and minimizes the load on the health care centres. The number of the diabetic patients is huge and rapidly increasing so that makes it hard and resource-consuming to perform a yearly screening to all of them.
The main goal of this Ph.D. thesis is to build a clinical decision support system (CDSS) based on electronic health record (EHR) data. This CDSS will be utilised to estimate the risk of developing RD.
In this Ph.D. thesis, I focus on developing novel interpretable machine learning systems. Fuzzy based systems with linguistic terms are going to be proposed. The output of such systems makes the physician know what combinations of the features that can cause the risk of developing DR.
In this work, I propose a method to reduce the uncertainty in classifying diabetic patients using fuzzy decision trees. A Fuzzy Ransom forest (FRF) approach is proposed as well to estimate the risk for developing DR.
Several policies are going to be proposed to merge the classification results achieved by different Fuzzy Decision Trees (FDT) models and obtain the final decision. These policies will be graded from a conservative policy to permissive policy.
To improve the final decision of our models, I propose three fuzzy measures that are used with Choquet and Sugeno integrals to aggregate the output of the FDTs rules. The definition of these fuzzy measures is based on the confidence values of the rules. In particular, one of them is a decomposable fuzzy measure in which the hierarchical structure of the FDT is exploited to find the values of the fuzzy measure. Out of this Ph.D. work, we have built a CDSS software (desktop application with GUI and a web service API) that may be installed in the health care centres.
