Evaluation of postoperative bleeding risk after dental extractions in patients on antithrombotic medication: A comparison of machine learning and clinical experience

Abstract

Objectives: The aim of this study was to identify high-risk dental extractions in patients taking antiplatelet (AP) medication or anticoagulants (ACs) and to compare an experienced surgeon's decisions with machine learning (ML) algorithms.

Materials and methods: Our study included 2000 procedures, of which 1788 were conducted in patients under monotherapy with AP medication, vitamin K antagonists (VKAs), heparin, or direct oral anticoagulants (DOACs), 426 were performed under dual therapy, and 27 under triple therapy. Four algorithms, logistic regression (LR), eXtreme gradient boost (XGB), random forest (RF), and K-nearest neighbors (KNN), were trained with 80% (1600 procedures) of the derived data. Afterwards, an experienced oral surgeon and the algorithms were tested on the remaining 20% (400 procedures) of the data to evaluate the predictive power with respect to bleeding incidents.

Results: The incidence of hemorrhagic events was low (4.35%). Dual anticoagulation significantly affected the risk of bleeding. Evaluating the results of the predictions, all four algorithms outperformed the surgeon in terms of balanced accuracy (LR: 58%; RF: 59%; XGB: 61%; KNN: 62%; surgeon: 53%).

Conclusions: Decision-making based on various parameters influencing bleeding risk is complex, and surgeons tend to overestimate this risk. Both the algorithms and the surgeon had a share of false positive predictions; however, in a medical context, preventive overcaution does less damage than underestimation.

Clinical relevance: Algorithms can provide an objective assessment of bleeding risk and help determine risk profiles, uncover variables with the highest predictive power, and serve as guidance on postoperative observation periods.

Trial registration: This study was approved by the Ethics Committee of the Medical Faculty of RWTH Aachen (Decision Number 24-353). This was a retrospective clinical study designed to analyze postoperative bleeding after dental extractions in patients under antithrombotic medication and to evaluate the prediction of bleeding events by different algorithms and human experience.

Keywords: Antiplatelet therapy; Dental extraction; Machine learning algorithm; Oral anticoagulation; Postoperative bleeding.

Fig. 1

Machine learning workflow and evaluation process

Fig. 2

Accuracy of the different algorithms (a–d) and the senior surgeon (e) Classification outcomes for each diagnostic method. (a) Logistic Regression (LR): 4 true positives, 354 true negatives, 29 false positives, and 13 false negatives. (b) Random Forest (RF): 6 true positives, 318 true negatives, 65 false positives, and 11 false negatives. (c) eXtreme Gradient Boosting (XGB): 6 true positives, 331 true negatives, 52 false positives, and 11 false negatives. (d) K-Nearest Neighbors (KNN): 9 true positives, 269 true negatives, 114 false positives, and 8 false negatives. (e) Senior surgeon: 5 true positives, 291 true negatives, 92 false positives, and 12 false negatives