Machine Learning Algorithmic and System Level Considerations for Early Prediction of Sepsis

• Main Authors: Narayanaswamy, Lakshman, Garg, Devendra, Narra, Bhargavi, Narayanswamy, Ramkumar
• Format: Conference Proceeding
• Language: English
• Subjects: Data models; Hospitals; Machine learning; Mathematical model; Predictive models; Recurrent neural networks; Training
• ISSN: 2325-887X
• DOI: 10.22489/CinC.2019.161

This study presents a machine learning (ML) model that predicts onset of sepsis earlier in time than what is possible using common severity scoring systems. Our study's focus is on building solutions that maximizes sepsis prediction, is real-world implementable and usable by care providers particularly in developing countries like India. We have selected features based on the observation that patient vitals are available on an hourly basis, whereas lab results if available are less frequent. To capture the time series nature of the data, we trained the model using long short term memory (LSTM), a version of recurrent neural network (RNN) architecture. To capture locale specific pathology baseline, we have engineered features using two methods. We define a minimum & maximum value for vitals and lab tests and normalize the incoming data against this min-max value. Secondly, to leverage sparsely available lab data that signal increased sepsis risk, we define a synthetic "risk" feature. This risk feature is assigned a higher score when certain lab values are available and exceed a threshold. Our solution achieved an official utility score of 0.179 on the full test under the team name LDBR. Finally, we present practical considerations we discovered from our interactions with local hospitals and health-care providers.