File Download

There are no files associated with this item.

  Links for fulltext
     (May Require Subscription)
Supplementary

Conference Paper: USE OF ADVANCED MACHINE LEARNING MODEL TO PREDICT POST-COLONOSCOPY THROMBOEMBOLIC EVENTS IN PATIENTS ON ANTITHROMBOTIC AGENTS

TitleUSE OF ADVANCED MACHINE LEARNING MODEL TO PREDICT POST-COLONOSCOPY THROMBOEMBOLIC EVENTS IN PATIENTS ON ANTITHROMBOTIC AGENTS
Authors
Issue Date1-May-2023
PublisherElsevier
Abstract

Introduction: Patients on anti-thrombotic agents are at risk of both bleeding and thromboembolic (TE) events during colonoscopy. There is a need to identify high-risk patients, who may need close monitoring during peri-colonoscopy period. This study evaluates the role of advanced machine learning (ML) models in predicting risk of post-colonoscopy TE events.
Methods: We included a cohort (n=1505) of patients who were on various anti-thrombotic agents and underwent colonoscopy between Jan 2016 and Aug 2020 in our hospital to train nine available ML models including Logistic Regression (LR), Linear Discriminant Analysis (LDA), Gradient Boosting Classifier (GBC), LightGBM (LightGBM), Naive Bayes (NB), Quadratic Discriminant Analysis (QDA), CatBoost Classifier (Catboost), Extreme Gradient Boosting (XGboost) and K Neighbors Classifier (KNN). A total of 25 baseline clinical parameters were used to build the ML model. All models were then validated in a temporal validation colonoscopy cohort from our hospital (between Sep 2020 and May 2021 n=237) as well as an external validation colonoscopy cohort performed in 42 other local public hospitals in 2021 (n=10,430). All patients in the two validation cohorts were on antithrombotic agents. The primary outcome was post-colonoscopy TE event, which were defined as any hospitalization for TE and ischaemic events within 30 days of the index colonoscopy including acute myocardial infarction, angina pectoris, deep vein thrombosis, pulmonary embolism and infarction, ischaemic stroke and transient ischaemic attack. Accuracy was determined by the area under the receiver operating characteristic curve (AUROC), with conventional CHA2DS2-VASc score as the reference.
Results: The incidence of TE events was 0.7%, 1.7% and 1.1% in the training, temporal and external validation cohort, respectively. Table 1 showed the performance of all ML models and CHA2DS2-VASc score. The Catboost model had the highest AUC and was significantly higher than the CHA2DS2-VASc score in both temporal (0.89 vs 0.77, p <0.01) and external validation cohort (0.85 vs 0.72, p<0.01; ). Both Catboost and CHA2DS2-VASc had very high negative predictive value (0.998 and 0.997, respectively), but the Catboost model had a significantly lower false negative rate (11.4% vs 15.7%, p <0.001).
Conclusion: Advanced ML model could help to predict the risk of post-colonoscopy TE events in patients who are on anti-thrombotic.


Persistent Identifierhttp://hdl.handle.net/10722/340492
ISSN
2023 Impact Factor: 25.7
2023 SCImago Journal Rankings: 7.362

 

DC FieldValueLanguage
dc.contributor.authorLui, Thomas Ka-Luen-
dc.contributor.authorChen, Justin C-
dc.contributor.authorLi, Yan Kiu-
dc.contributor.authorGuo, Chuan-Guo-
dc.contributor.authorLiu, Sze Hang Kevin-
dc.contributor.authorCheung, Ka Shing-
dc.contributor.authorLeung, Wai Keung-
dc.date.accessioned2024-03-11T10:45:02Z-
dc.date.available2024-03-11T10:45:02Z-
dc.date.issued2023-05-01-
dc.identifier.issn0016-5085-
dc.identifier.urihttp://hdl.handle.net/10722/340492-
dc.description.abstract<p><strong>Introduction:</strong> Patients on anti-thrombotic agents are at risk of both bleeding and thromboembolic (TE) events during colonoscopy. There is a need to identify high-risk patients, who may need close monitoring during peri-colonoscopy period. This study evaluates the role of advanced machine learning (ML) models in predicting risk of post-colonoscopy TE events.<br><strong>Methods: </strong>We included a cohort (n=1505) of patients who were on various anti-thrombotic agents and underwent colonoscopy between Jan 2016 and Aug 2020 in our hospital to train nine available ML models including Logistic Regression (LR), Linear Discriminant Analysis (LDA), Gradient Boosting Classifier (GBC), LightGBM (LightGBM), Naive Bayes (NB), Quadratic Discriminant Analysis (QDA), CatBoost Classifier (Catboost), Extreme Gradient Boosting (XGboost) and K Neighbors Classifier (KNN). A total of 25 baseline clinical parameters were used to build the ML model. All models were then validated in a temporal validation colonoscopy cohort from our hospital (between Sep 2020 and May 2021 n=237) as well as an external validation colonoscopy cohort performed in 42 other local public hospitals in 2021 (n=10,430). All patients in the two validation cohorts were on antithrombotic agents. The primary outcome was post-colonoscopy TE event, which were defined as any hospitalization for TE and ischaemic events within 30 days of the index colonoscopy including acute myocardial infarction, angina pectoris, deep vein thrombosis, pulmonary embolism and infarction, ischaemic stroke and transient ischaemic attack. Accuracy was determined by the area under the receiver operating characteristic curve (AUROC), with conventional CHA<sub>2</sub>DS<sub>2</sub>-VASc score as the reference.<br><strong>Results: </strong>The incidence of TE events was 0.7%, 1.7% and 1.1% in the training, temporal and external validation cohort, respectively. Table 1 showed the performance of all ML models and CHA<sub>2</sub>DS<sub>2</sub>-VASc score. The Catboost model had the highest AUC and was significantly higher than the CHA<sub>2</sub>DS<sub>2</sub>-VASc score in both temporal (0.89 vs 0.77, p <0.01) and external validation cohort (0.85 vs 0.72, p<0.01; ). Both Catboost and CHA<sub>2</sub>DS<sub>2</sub>-VASc had very high negative predictive value (0.998 and 0.997, respectively), but the Catboost model had a significantly lower false negative rate (11.4% vs 15.7%, p <0.001).<br><strong>Conclusion:</strong> Advanced ML model could help to predict the risk of post-colonoscopy TE events in patients who are on anti-thrombotic.</p>-
dc.languageeng-
dc.publisherElsevier-
dc.relation.ispartofGastroenterology-
dc.titleUSE OF ADVANCED MACHINE LEARNING MODEL TO PREDICT POST-COLONOSCOPY THROMBOEMBOLIC EVENTS IN PATIENTS ON ANTITHROMBOTIC AGENTS-
dc.typeConference_Paper-
dc.identifier.doi10.1016/S0016-5085(23)03699-5-
dc.identifier.volume164-
dc.identifier.issue6-
dc.identifier.eissn1528-0012-
dc.identifier.issnl0016-5085-

Export via OAI-PMH Interface in XML Formats


OR


Export to Other Non-XML Formats