Leveraging Neural Networks for Real-Time Blood Analysis in Critical Care Units

Abstract

We apply deep neural networks to analyze blood samples in real-time, targeting clinical care practices in the evolving field of critical care medicine. These practices currently rely on flow cytometry, the throughput and skill requirement of which often lead to delays, causing essential physicians of care to use less accurate methods and rely on various heuristics for decision-making. Our neural networks, designed to specifications, can help alleviate these real-time needs by being served as a simple service on portable devices with real-time capability.

While previous use of deep learning for CBC estimation from small amounts of data or shallow learning has shown only marginal improvement over or equivalence to the usage of off-the-shelf equipment, we expand the applicability to a diverse array of costs and complexities, estimating all five types of white blood cells at once with 96.8% correlation and 94.8% accuracy over a wide range while also achieving robust judgment for labeling, which is delicate when performed manually, thereby avoiding the current time consumption and the measuring problems given by flow cytometry.