469 Use of data-driven technology to optimise DOACs in morbidly obese patients

Abstract

Abstract Introduction Despite the advantages of Direct Oral Anticoagulants (DOACs) over older classes of anticoagulants, clinical experience is limited in special populations; data from landmark trials on safety and efficacy are relatively scarce (compared to warfarin). This makes it challenging for clinicians to prescribe the right DOAC at the right dose for such patients (e.g., morbidly obese patients) (1). Insights derived from analysing real-world data have proven to be a vital source of clinical evidence backing the recommendation of medications (2). Therefore, data-driven technologies like machine learning can harness big data in electronic health records (EHRs) to optimise DOAC therapy and improve clinical outcomes. Aim The study aims to accurately predict clinical outcomes in morbidly obese patients, and identify the key variables in the model for optimising the safety and efficacy of Direct Oral Anticoagulant (DOAC) doses. Methods An observational, retrospective cohort study was carried out in partnership with an NHS Trust. Based on eligibility criteria, the dataset of morbidly obese patients on DOACs was extracted from EHRs, pre-processed and analysed considering the access granted. After partitioning the entire dataset into a 70:30 split, the training dataset (70%) was run through selected machine learning (ML) classifiers (Random Forest, decision trees, K-nearest neighbours (KNN), bootstrap aggregation algorithm, gradient boosting classifier, support vector machines, and logistic regression) to rank variables, and derive predictions which were evaluated against the test dataset (30%). A multivariate regression model was used to adjust for confounders and to explore the relationships between DOAC regimens and clinical outcomes. Results We identified 4,275 morbidly obese patients out of n=97,413 records overall. The bootstrap aggregation, decision trees, and random forest classifiers (from the ML algorithms tested) achieved superior prediction accuracies (98.6%, 97.9%, and 98.3%, respectively) for the individual DOAC doses, with excellent values for precision, recall, and F1 scores (performance metrics). The most important characteristics in the model for predicting mortality and stroke were age, treatment days, and length of stay. Among DOACs, apixaban (84%) was the most frequently prescribed DOAC followed by rivaroxaban (15%). Apixaban 2.5 mg (twice daily) received the highest ranking for relevance to mortality, while it raised the mortality risk (OR 1.430, 95% CI: 1.181, 1.1.732, p=0.001). There were mixed results for apixaban 5mg (twice daily), the most widely prescribed dose of apixaban (54%), with significantly reduced risk of mortality (OR 0.751, 95% CI: 0.632, 0.905, p=0.003), but significantly increased risk of stroke events (OR 32.457, 95% CI: 17.083-61.664, p=0.001). Conclusion Given the large sample size—a strength of our study, data-driven technologies were successfully employed in predicting the safety and efficacy of DOACs in morbidly obese patients using the real-world dataset; the key variables in the model for optimising clinical outcomes were identified. However, the limitations in our study, such as reporting errors, selection bias, and confounding bias, were not ruled out. Therefore, confirmatory studies (e.g., external validation with prospective data) are needed to confirm findings and provide a sound basis for universal deployment in clinical settings.