Classify contaminated blood tests: a neural approach

Automated laboratories analyze thousands of blood tests every day. Some samples can become contaminated by improper collection, producing incorrect results. Identification of such samples is critical and requires a labor intensive expert review. In our previous work, we demonstrated a useful machine learning approach to automated detection, but a severe class imbalance in the dataset led to unstable results. Using a modified data augmentation strategy and an attentional neural architecture, we show significant improvement over previously published methods in all binary classification metrics, with an absolute F1-score improvement of 13%. This work represents the state-of-the-art in automated contamination detection and demonstrates that a machine learning pipeline based on our neural architecture can perform on par with human experts.