Sepsis Detection Using Missingness Information

Abstract

Sepsis is one of the leading causes of death in hospital. An early detection is crucial to reduce its consequences and mortality. The challenge of Computing in Cardiology 2019 is addressing this issue by providing about 40,000 records from intensive care unit patients. As clinical measurements are collected at irregular frequencies, this dataset is missing many observations. Simply discarding missing values is counterproductive. Indeed, it has been observed that missing data patterns hold relevant information regarding the patient health state. To take advantage of this information, we propose a sepsis detection model incorporating representations of missingness information. This model is a recurrent neural network network composed of two gated recurrent unit (GRU) layers to capture long-term dependencies and a sigmoid layer to output a probability of sepsis. First, the model is trained by simply imputing missing values in the dataset. Then, the dataset is extended with the pattern of missing values. Finally, a GRU cell modi?ed to that take into account missing data is evaluated. Our best model achieves an utility of 0.00 on the grading dataset.