Development and characterization of neural network-based multianalyte delta checks

Background: In laboratory medicine, mislabeled specimens (MLS) are pre-analytical errors where a tube of blood is given the wrong patient ID. When pre-analytical methods fail to catch these errors, post-analytical methods are employed to identify MLS. Inspired by delta checks, we employed neural networks to create a variety of novel algorithms aimed to detect MLS. In this approach, the neural networks evaluate the analytical results from two different time points to determine if they belong to the same patient. Methods: Raw analytical data was collected from clinical practice. Analytes were grouped by patient and collection time, and groups of specimens were linked to create a dataset of properly labeled specimen pairs. The data was processed to simulate MLS by reassigning the second set of analytes to a different patient. In one approach, neural networks were created where both groups of analytes were set panels [e.g., basic metabolic panels (BMP) or comprehensive metabolic panels (CMP)]. In a second approach, neural networks were created where the groups of analytes could consist of any combination of analytes. For both approaches, different neural networks were created for different ranges of time-deltas, and these were evaluated using receiver operator characteristics curve analysis. The performance of each neural network was compared to the current delta check strategy used at our institution. Results: The best performing neural network was the one that compared analytical results from CMPs for time-deltas <1.5 days. This achieved an area under the curve (AUC) of 0.994 and significantly outperformed the classic delta check methods. The best performing open neural network achieved an AUC of 0.964 when 13 analytes were used time deltas <1.5 days. Conclusions: The implementation of machine learning-based protocols to detect MLS with fewer false-positive errors may have a dramatic impact on patient care and health care costs.