Identification of Risk Factors and Prediction of Sepsis in Pregnancy Using Machine Learning Methods
Georgy Kopanitsa, Oleg Metsker, David Dokkaevich Paskoshev, Sofia Greschischeva
Abstract
Maternal sepsis is a cause of 11% of maternal lethal cases worldwide. It is the third most common reason of maternal death. In recent years, healthcare observes the emergence of machine learning as a new tool to analyze large amounts of data to reveal hidden correlations and regularities. Considering the prospective of machine learning in predicting and risk management of sepsis, we are developing real world evidence data driven models that will be able to identify a risk of sepsis in pregnancy on the early stages of its development. We have retrospectively analyzed electronic health records from the perinatal Center of the Almazov perinatal medical center in Saint-Petersburg, Russia. In the first phase of the work, data of 74,000 episodes was extracted from the medical information system with a total of 15,000 birth cases. The resulting performance of the model on the test data set was 95% AUC. In comparison to the previous study on sepsis, we have revealed new risk factors and dependencies between them. We also managed to build a pregnancy specific model with a similar performance as general sepsis models. The benefit of our approach is that it is based rather on vital signs and anamneses of a pregnant woman, whereas most of the models described in the literature require many laboratory test results, which may not be ordered or available in time if the situation when sepsis is no suspected by clinicians. In our study we managed to preserve a prediction accuracy of the model while decreasing the number of predictors. This can simplify the clinical application of the model.