Clinical Translation of a Heparin-Induced Thrombocytopenia predictor for Intensive Care Patients

The emergence of novel digital data capture and analysis tools has vastly expanded the methods available to diagnose diseases at earlier stages or even predict their future appearance. Heparin-Induced Thrombocytopenia (HIT) is a as an adverse drug reaction to heparin, a frequently used anticoagulant. Accurate and prompt diagnosis of HIT is important to guide appropriate treatment strategies to avoid thromboembolic complications. This paper presents a machine learning-based HIT classifier, which can be used to predict HIT in intensive care patients. We used MIMIC-IV to develop a Light Gradient Boosting Machine learning model. The model yielded a Mathews Correlation Coefficient of 0.38 and 0.33 for the blind test and 10-fold Cross validation test respectively. The positive and negative likelihood ratios were 13.19 and 0.74 respectively, establishing that a positive test will signal an increased risk of HIT, whereas a negative test will not significantly modify the post-test probability, meaning that patient can follow usual care.