AN AI ENABLED DIAGNOSTIC SYSTEMS FOR CARDIOVASCULAR DISEASE AND DIABETES

Abstract

Noncommunicable diseases (NCD), or chronic diseases which include diabetes, cancer, respiratory disorders, cardiovascular diseases (CVD), cancer etc., can be difficult to fully recover from and they are responsible for 71% of worldwide deaths. The NCDs may stem from a combination of lifestyle, genetics, and environmental factors. In this dissertation, we focused on developing AI-enabled methods using multi-modal dataset to diagnose two most prevalent NCDs in Qatar: CVD and diabetes. For this purpose, we actively collaborate with Qatar Biobank (QBB) and Hamad Medical Corporation (HMC) at Qatar. In the first study, we investigated lifestyle related risk factors for CVD patients in Qatar by analyzing their food, drink, smoking, and physical activity habits. Statistical analysis revealed that the CVD group consumed less fast food, soft drinks, snacks, and meat than the control group. However, smoking rates were higher among the CVD group, and their consumption of healthy breakfast items was lower. Our findings suggest that the Qatari CVD cohort mostly follows standard guidelines but should reduce smoking and adjust moderate physical activity based on physician recommendations. Second, we developed an AI-enabled method that can detect CVD onset with 93% accuracy by integrating a multimodal clinical dataset from QBB. This study included the largest collection of biomedical measurements representing anthropometric measurements, clinical biomarkers, bioimpedance, spirometry, VICORDER readings, and behavioral factors of the CVD group from QBB. Our analysis revealed that physio-clinical and bioimpedance measurements were the most effective in distinguishing these two groups. Moreover, we confirmed known CVD risk factors and proposed potential novel risk factors linked to CVD-related comorbidities such as renal disorder, atherosclerosis, hypercoagulable state, and liver function. Proposed method outperformed the clinically approved scales such as the Framingham scale and ASCVD scale that used in clinical settings for CVD diagnosis plans. Third, we developed a novel, fast, and non-invasive method for CVD diagnosis based on retinal images and Dual-energy X-ray Absorptiometry (DXA) scans. Retinal images provide quick diagnosis of CVD by measuring the diameter and tortuosity of retinal blood vessels. DXA, recently FDA-approved for CVD diagnosis, can measures bone health of patients. Our AI-enabled system fuses information from both sources to diagnose CVD, achieving 78.3% accuracy. Using gradient class activation map, we showed that the model focuses on hemorrhages, a sign of CVD, in the center of retinal images. Moreover, DXA data showed better bone health condition in the CVD group. Finally, We proposed DiaNet, a deep learning model to predict diabetes from retinal images with over 90% accuracy based on the retinal images collected from QBB and HMC. The model identifies relevant regions on the retina images and highlights the distinguishing capability of retinal images for diabetes patients in multiple cohort. Overall, we concluded that, compared to traditionally used clinical data, retinal images contain sufficient information to distinguish diabetes cohort from the control group with high accuracy suggesting their inclusion in clinical diagnosis setup is warranted in the future. Currently we are working on the implementation of the proposed DiaNet system at HMC.

Date of Award	2023
Original language	American English
Awarding Institution	HBKU College of Science and Engineering