Data-driven diabetic retinopathy (DR) prediction model with the assistance of a score-based diffusion model

Diabetic Retinopathy (DR) is a common microvascular diabetes complication in clinics, which may result in vision loss due to damage of blood vessels in the back of the eye. Currently, it is also the leading cause of blindness among all American adults. Thus, with the available electronic health records (EHR) data, establishing an effective classification model to accurately predict the risk of DR can significantly benefit the diabetic patients who have not been affected by DR yet. However, one of the biggest challenges is that the collected EHR data are usually imbalanced, as in general the number of patients who are diagnosed positive for DR is much less than the negative. For instance, the imbalanced ratio of positive to negative patients may be less than 1:10, and then it will result in significant model training bias and reduce the prediction accuracy of the classifiers. Therefore, to handle the data imbalanced issue, this thesis aims to develop a data augmentation approach with the integration of machine learning models based on the EHR data collected from diabetic patients’ primary care visits. This research leverages a score-based diffusion generative model to generate high-quality EHR samples for DR patients and improve classification performance. Diffusion models are an emerging generative machine learning model that are used for generating more accurate images compared to other data augmentation approaches like Generative Adversarial Networks (GAN). This thesis aims to test its effectiveness and potential feasibility on non-image data, particularly for EHR data, since diffusion models have been proven that they can generate image data effectively. The classification results show an approximate 7% improvement in terms of the F1 score, which demonstrates that the proposed diffusion model-based EHR data augmentation approach is promising for augmenting non-image data for better predictive modeling like EHR-based disease risk prediction.