Background: Diagnostic workup of Heparin-Induced Thrombocytopenia (HIT) involves functional testing of platelet activation induced by pathogenic anti-heparin-PF4 antibodies. This usually relies on bulk assays providing average values of markers of platelet activation –  aggregation, serotonin release or agglutination – over a sample. Instead, flow cytometry provides multiparametric information-rich content made of tens of thousands of single platelet measurements of 4 to 20 variables, in seconds.

Aims: Classification is at the core of both medical diagnostic and machine learning (ML). The study assessed the performance of ML classifiers – algorithms generating discriminatory (predictive) models – to automatically classify (label) a flow cytometry result as positive or negative.

Methods: We used data from a former assessment study of a functional HIT flow cytometry assay (FCA) and  five types of classifiers: logistic regressions, decision trees, support vector machines, neural networks, and nearest shrunken centroids. The base models generated by these classifiers were tuned and compared over 75% of the data set that was repeatedly split into training and validation subsets. Predictions from the base models were subsequently combined to form an ensemble model providing an enhanced classifier, whose accuracy was estimated by cross validation over the entire dataset – including the remaining 25% of the data set that was kept hidden from the classifiers – and compared with FCA’s previously reported accuracy.

Results: Best performances resulted from a gradient boosting ensemble model whose automated data-driven performance was slightly better than operator-driven FCA’s, with accuracy, sensitivity, and specificity of 96%, 96% and 96%, versus 94%, 90% and 94%, respectively. The model was shown to be robust using additional data acquired by two other cytometers, suggesting potential method generalization.

Conclusions: Artificial intelligence may “augment” reliability and accuracy of flow cytometry-based HIT testing by automatically extracting the relevant information from the whole set of captured data.

« Back to ISTH 2021 Congress

ISTH Congress Abstracts - https://abstracts.isth.org/abstract/breakthrough-in-laboratory-functional-testing-of-hit-using-flow-cytometry-and-artificial-intelligence-towards-augmented-diagnostics/