Development And Validation Of A Machine Learning Algorithm To Optimise The Dosing Of Unfractionated Heparin.

• By: BARRAS, Michael (University Of Qld, Australia)
• Co-author(s): A/prof Michael Barras (University of Qld, Brisbane, Australia / Princess Alexandra Hospital, Brisbane, Australia)
  Dr Nazanin Falconer (University of Qld, Brisbane, Australia / Princess Alexandra Hospital, Brisbane, Australia)
  Dr Ahmad Abdel-Hafez (Princess Alexandra Hospital, Brisbane, Australia / Queensland University of Technology, Brisbane, Australia)
  Prof Ian Scott (University of Qld, Brisbane, Australia / Princess Alexandra Hospital, Brisbane, Australia)
  Mr Sven Marxen (Logan & Beaudesert Hospitals, Brisbane, Australia)
  Mr Aaron Van Garderen (Princess Alexandra Hospital, Brisbane, Australia)
  Mr Oscar Bonilla (Princess Alexandra Hospital, Brisbane, Australia)
  Mr Stephen Canaris (Princess Alexandra Hospital, Brisbane, Australia)
  
• Abstract:

  Background Information
  
  Unfractionated heparin (UFH) is considered a high-risk medication, with complex dosing. An excessive dose can cause bleeding, while an insufficient dose can lead to a recurrent embolic event. Following initiation of intravenous (IV) UFH therapy, the therapeutic response is monitored using a measure of blood clotting, the activated partial thromboplastin time (aPTT). Clinicians iteratively adjust the dose of UFH to a target aPTT therapeutic range, with the local range between 60 to 100 seconds. Unfortunately, dose estimation for UFH is difficult due to pronounced intra- and inter-patient variability in the pharmacodynamic response. Data across four metropolitan tertiary Australian hospitals showed only 23% of patients were reaching the target aPTT range after the first dose. New dosing methods are required, and advances in the development of machine learning (ML) algorithms offers an exciting opportunity to optimise the dosing of UFH.
  
  Purpose
  
  The aim of this study was to develop and validate a ML algorithm to predict, aPTT within 12 hours after a specified bolus and maintenance dose of UFH.
  
  Method
  
  This was a retrospective cohort study of data obtained over 3-years. The patient population was adult general medicine and surgical patients being administered UFH for an embolic event such as deep vein thrombosis or pulmonary embolism, or for the prevention of embolic events due to atrial fibrillation. Data was collected from electronic health records of five hospitals in Queensland, Australia. Data from four hospitals were used to build and test ensemble models using cross validation, while the data from the fifth hospital was used for external validation. Modelling was performed using H2O Driverless AI® an automated ML tool.
  
  Results
  
  A total of 2691 patients were included from 3019 episodes of care. 17 different experiments were conducted in an iterative process to optimise model accuracy. In predicting aPTT, the best performing experiment produced an ensemble with 4x LightGBM models with a root mean square error (RMSE) of 31.35 (standard deviation (SD)=1.37). This model relied on 93 features. The dataset was re-purposed as a multi-classification task (sub-therapeutic, therapeutic, and supra-therapeutic aPTT result) and achieved a 0.599 (SD=0.029) accuracy and area under the receiver operating characteristic curve (AUC) of 0.735. External validation yielded similar results: RMSE of 30.52 (SD = 1.29) for the prediction model, and accuracy of 0.568 (SD=0.032) and AUC of 0.724 (95% CI 0.714-0.734) for the multi-classification model.
  
  Conclusion
  
  To our knowledge, this is the first ML model applied to IV UFH dosing that has been developed and externally validated in a multisite adult general medical and surgical inpatient setting. The algorithm will be embedded into an application and evaluated in a prospective clinical trial.