Layer-Wise Relevance Propagation in Multi-label Neural Networks to Identify COVID-19 Associated Coinfections

Abstract

COVID-19 has been affected worldwide since the end of 2019. Clinical studies have shown that a factor that increases its lethality is the existence of secondary infections. Coinfections associated with the infection SARS-CoV-2 are classified into bacterial infections and fungal infections. A patient may develop one, both, or neither. From a machine learning point of view, this is considered a multi-label classification problem. In this work, we propose a multi-label neural network able to detect such infections in a patient with SARS-CoV-2 and thus provide the medical community with a diagnosis to guide therapy in these patients. However, neural networks are often considered a "black box" model, as their strength in modeling complex interactions, also make their operation almost impossible to explain. Therefore, we propose three adaptations of the Layer-wise Relevance Propagation algorithm to explain multi-label neural networks. The inclusion of this post-hoc interpretability stage made it possible to identify significant input variables in a classifier output.