Publication: DEVELOPMENT OF A BLEEDING RISK SCORE FOR MALAYSIAN PATIENTS WITH NON-VALVULAR ATRIAL FIBRILLATION REQUIRING ORAL ANTICOAGULANTS
| dc.contributor.author | LEE JENG SHIN | |
| dc.date.accessioned | 2024-10-01T04:39:05Z | |
| dc.date.available | 2024-10-01T04:39:05Z | |
| dc.date.issued | 2023 | |
| dc.description.abstract | Introduction: Bleeding is a significant risk associated with oral anticoagulant (OAC) use, particularly in patients with non-valvular atrial fibrillation (NVAF). Existing bleeding risk scores have shown limited accuracy in predicting clinically relevant bleeding events, necessitating the development of a new predictive model. This study aims to create and validate a novel bleeding risk model that effectively predicts clinically relevant bleeding events. Methods: Using demographic, laboratory, and clinical data collected by Beshir SA et al., five bleeding risk models were developed through machine learning techniques, specifically utilizing recursive feature elimination (RFE) to identify pertinent variables. The dataset was divided into training and internal validation sets at 70:30 ratio. The model performance was assessed with accuracy, precision, sensitivity, specificity, F1 score, and area under the receiver operating characteristic curve (AUROC). Upsampling and downsampling techniques were applied to original dataset with repetition in model building and evaluation. Results: Of 1,017 patients (median age: 67 years, male 52%), 77 patients (7.6%) faced clinically relevant bleeding within first year of observation period. The developed model comprised 10 variables, including renal failure, peripheral vascular disease, history of bleeding, age at diagnosis, hypertension, ischemic heart disease, sex, type II diabetes mellitus, smoking, and race. All five machine learning models had poor predictive performance, with AUROC between 0.49 to 0.50. The SVM-Radial model with upsampling dataset, demonstrated the highest predictive performance, with an AUROC 75.5%, a sensitivity of 86.88%, specificity of 64.18%, accuracy of 75.5.%, precision 70.81%, and F1 score of 78.03%. Conclusions: Feature selection using recursive feature selection – random forest picked 10 clinical predictors for CRB events despite data imbalance in original dataset. Downsampling provided realistic bleeding risk prediction which SVM-Linear performed best among five models. Upsampling improved bleeding risk prediction which SVM-Radial performed best among five models. New techniques to improve bleeding risk prediction are warranted. Keywords: non valvular atrial fibrillation, oral anticoagulants, bleeding, machine learning, bleeding risk score. | |
| dc.identifier.uri | https://hdl.handle.net/20.500.14377/36744 | |
| dc.language.iso | en | |
| dc.publisher | International Medical University | |
| dc.subject | Anticoagulants | |
| dc.subject | Hemorrhage | |
| dc.subject | Machine Learning | |
| dc.subject | Risk Factors | |
| dc.subject | Atrial Fibrillation | |
| dc.title | DEVELOPMENT OF A BLEEDING RISK SCORE FOR MALAYSIAN PATIENTS WITH NON-VALVULAR ATRIAL FIBRILLATION REQUIRING ORAL ANTICOAGULANTS | |
| dc.type | Thesis | |
| dspace.entity.type | Publication | |
| oairecerif.author.affiliation | #PLACEHOLDER_PARENT_METADATA_VALUE# |