Abstract Number: PB0855
Meeting: ISTH 2021 Congress
Background: Diagnosing heparin-induced thrombocytopenia (HIT) at the bedside is challenging, and current diagnostic algorithms expose patients to a considerable risk of overtreatment and delayed diagnosis.
Aims: We conducted a prospective multicenter study detailedly acquiring clinical and laboratory variables to assess the diagnostic performance of these variables and to develop an easy-to-apply clinical prediction model.
Methods: Consecutive patients with suspected HIT were included in 11 study centers and detailed clinical data were collected. Heparin-induced platelet activation assay (HIPA; reference standard) and various immunoassays were conducted at the central laboratory. Variables with a p-value < 0.05 for each level in a multivariable logistic regression were selected for the final model. Using 75% of the patients, logistic regression, penalized logistic regression, two random forest, and gradient boosting machine models were trained. The models were evaluated on the remaining 25% (validation set). The performance of the model with the best c-statistic was then compared to the current clinical practice.
Results: To date, we enrolled 1’182 patients with suspected HIT; the prevalence of HIT was 9.3%. Variables selected for the final model were: platelet nadir, use of unfractionated heparin, timing of thrombocytopenia, presence of other causes of thrombocytopenia, and immunoassay test result. Applied to the validation set and using an IgG-specific ELISA, the c-statistic of the random forest model was 98.8% (95% confidence interval [CI]: 97.7, 99.9), the sensitivity was 96.0% (95% CI: 79.6, 99.8) and the specificity 97.3% (95% CI: 93.0, 98.1). In contrast, the sensitivity of the currently recommended diagnostic algorithm was 80.0% (95% CI: 59.3, 93.2), and the specificity 89.1% (95% CI: 84.6, 92.6).
Conclusions: Using detailed clinical and laboratory data and machine-learning algorithms, we developed and validated an accurate clinical prediction model for the diagnosis of HIT. This model has the potential to relevantly reduce overtreatment and delayed diagnosis in clinical practice.
To cite this abstract in AMA style:Nilius H, Studt J-, Tsakiris DA, Greinacher A, Mendez A, Schmidt A, Wuillemin WA, Gerber B, Vishnu P, Graf L, Bakchoul T, Nagler M. Diagnosing Heparin-induced Thrombocytopenia Using Machine Learning Algorithms: First Data of the TORADI-HIT Study [abstract]. Res Pract Thromb Haemost. 2021; 5 (Suppl 1). https://abstracts.isth.org/abstract/diagnosing-heparin-induced-thrombocytopenia-using-machine-learning-algorithms-first-data-of-the-toradi-hit-study/. Accessed September 24, 2021.
« Back to ISTH 2021 Congress
ISTH Congress Abstracts - https://abstracts.isth.org/abstract/diagnosing-heparin-induced-thrombocytopenia-using-machine-learning-algorithms-first-data-of-the-toradi-hit-study/