Anomaly Detection Semi-Supervised Framework for Sepsis Treatment

Abstract

Sepsis is one of the leading causes of morbidity and mortality in hospitals. Early diagnosis could substantially improve the patient outcomes and reduce the mortality rate. In this paper we propose a machine learning approach for anomaly detection to aid the early detection of sepsis. Using the medical data of over 40,000 patients [1], we use both unsupervised and supervised methods to extract relevant features from the data, and then use standard classification approaches to predict sepsis six hours before clinical diagnosis occurs. To extract features, we used the reconstruction error of an autoencoding neural network trained on control patients free of sepsis, and used random forest classifiers to learn the most important features for the classification of patients. We then combined the features from both of these approaches with a variety of standard classification models. Cross-validation as well as the asymmetric utility function designed for this challenge are used to evaluate the resulting models. We obtained a utility function score for the full unseen dataset of 0.177 (Team Kriss); achieved with a logistic regression classifier. All the implementation is publicly available at https://github.com/ ineskris/SepsisChallenge-Cinc2019.

A Novel Hybrid Methodology for Anomaly Detection in Time Series

“…There is no unique technique that can be sufficient to diagnose sepsis and the combinations of different approaches are needed. A different studies based on patient retrospective data [4] that includes more than 40 different patients features are used for optimal treatments strategy in sepsis disease and sepsis classification [5][6][7]. To the best of our knowledge, no study, based on the waveform vital sign parameters from patient bedside time series data, has been conducted in the context of sepsis disease and anomaly detection.…”