ESC Journals
Novel oral anticoagulants (NOAC) is important in preventing thromboembolism in atrial fibrillation (AF) patients. Bleeding risk was evaluated by HAS-BLED score traditionally. Data mining is a relatively new discipline that has sprung up at the confluence of several other disciplines, driven primarily by the growth of large databases.
This study aimed to find a useful predictive model by data mining to assess the risk of rivaroxaban, an antithrombotic drug that causes bleeding in AF patients. The seven parameters of the HAS-BLED score were used to predict the effect of rivaroxaban on bleeding tendency in AF patients and may provide clinicians with appropriate treatments to avoid complications from bleeding events and reduce the incidence of health damage.
Through conducting a multicenter retrospective study, we identified patients with AF who were treated with rivaroxaban for more than 1 month between December 1, 2011 and November 30, 2016. After preprocessing, the established data were used for training and testing of data mining models. This study evaluated four models, including association rules, neural networks, Bayesian classification, and decision trees.
Of the 872 enrolled cases, 432 were in any of the bleeding groups and 432 were in the non-bleeding randomized control group. After comparing the overall classification accuracy, omission error and over-prediction error, the decision tree proved to be the most accurate model for bleeding prediction. The overall classification accuracy is 77%, the omission error is 15%, the over-prediction error is 21.9%, and the AUC score is 0.84. The results show that the model has good discriminative ability and visibility of decision rules.
Among several data mining models, decision tree proved to be the most accurate model for bleeding prediction. The conclusion of this study can be used as a reference for supporting decision making before anticoagulation treatment and suggest future research to compare efficacy of bleeding prediction between HAS-BLED score and decision tree.
Data mining comparison
Model | Omission error | Commission error | Overall accuracy | AUC score | Ranking |
---|---|---|---|---|---|
Decision tree | 15.0% | 21.90% | 77.00% | 0.84 | 1 |
Association rules | 16.8% | 27.20% | 76.50% | 0.81 | 2 |
Neural networks | 12.0% | 26.40% | 78.20% | 0.83 | 3 |
Bayesian classification | 16.1% | 27.50% | 76.50% | 0.83 | 4 |