Publication: Developing an artificial intelligence (AI) model in predicting the occurrence of abdominal aortic aneurysm (AAA)
Loading...
Date
2022
Authors
Khee, Saw Shier
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Background: AAA is usually asymptomatic until complications occurred such as aorta dissection and ruptured AAA. Early detection and proper management of AAA reduces the overall death rate from rupture. Therefore, an AI model is proposed to tackle this problem: to predict the occurrence of AAA, also to assist physicians for early detection which leads to appropriate management, thereby reducing complication and mortality. Objective: To develop an AI model to predict occurrence of AAA and to determine the accuracy of AI model in predicting the occurrence of AAA in screening population Methods: There are two phases in this study: (i) Phase 1 - retrospective data consisting of AAA’’s risk factors were extracted from two institutions and was used to train the model; (ii) Phase 2 – prospective data were collected and was used to validate the trained model. During the training process, the AI model was trained with different combinations of the features. Two SVM models were trained: (i) normal vs diseased (small and big AA) and (ii) small vs large AA, which is defined by AA size. Accuracy, sensitivity and specificity were reported. Results: For normal vs diseased AA, AI model 1 achieved an accuracy of 86.1% and 79.4% in Phase 1 and 2, respectively. We found that important parameters to differentiate between normal and diseased AA were age, gender, hypertension, systolic and diastolic blood pressure, dialysis status and years of hypertension. For small vs large AA, AI model 2 achieved an accuracy of 86.7% and 60.8% in Phase 1 and 2, respectively. Additional parameters to distinguish between small and large AA include dyslipidemia, smoking status, lung disease and kidney diseases. Conclusions: We showed that the AI model has the potential to predict AA size with satisfactory accuracy using only the patient’s demographics and medical history. The models developed could be useful for physicians who do not have the accessibility of imaging facilities in the district area to predict AA size without radiography images.
Description
Keywords
-