The Signature-Based Model for Early Detection of Sepsis from Electronic Health Records in the Intensive Care Unit

Abstract

Optimal feature selection leads to enhanced efficiency and accuracy when developing both supervised and unsupervised machine-learning models. In this work, a new signature-based regression model is proposed to automatically identify a patient's risk of sepsis based on physiological data streams and to make a positive or negative prediction of sepsis for every time interval since admission to the intensive care unit. The gradient boosting machine algorithm that uses the features at the current time-points and the signature features extracted from the time-series to model the longitudinal effects of sepsis yields the utility function score of 0.360 (officially ranked 1st, team name: 'Can I get your Signature?') on the full test set. The signature method shows a systematic and competitive approach to model sepsis by learning from health data streams.

A Multi-Task Imputation and Classification Neural Architecture for Early Prediction of Sepsis from Multivariate Clinical Time Series

“…The subsequent XGBoost model developed at the follow-up 'hackathon' event, held by the challenge organizers, improved the utility score to 0.342 and achieved an official rank of 2nd place. Table 2 also shows the utility scores achieved by the official challenge winner [8]. 6.…”

“…(7): A complement vector, x c t , is formed by combining both the available feature values with estimated information when the feature value is missing, where represents element-wise multiplication. (8): In addition to predicting missing information the neural architecture also applies a temporal decay, γ t , to the recurrent hidden state, based on the time gap since a feature value was measured, i.e. the δ matrix.…”

Maximizing the use of social and behavioural information from secondary care mental health electronic health records

“…The uniqueness of UK-CRIS is in providing access to the unstructured EHR text to researchers, providing ample opportunity to exploit a wide range of information, not presently captured or availability at the structured level. Novel signature-based machine learning methods are being developed to make use of such rich time-series data that transform the sequential data into useful features which feed into algorithms to identify robust combinations predictive of outcomes (27)(28)(29). The downstream potential for embedded models within EHR systems is to translate unstructured text into a structured format comprising clinically meaningful variables, and important determinants of mental health.…”

An efficient representation of chronological events in medical texts

“…The second order corresponds to the area enclosed by a path and a chord connecting endpoints (Chevyrev and Kormilitzin, 2016). The usefulness of a path signature as a feature map of sequential data was demonstrated theoretically as well as in numerous machine learning applications in healthcare (Morrill et al, 2019;Morrill et al, 2020b;Kormilitzin et al, 2017), finance (Arribas, 2018), computer vision (Yang et al, 2017;Xie et al, 2017), topological data analysis and deep learning (Kidger et al, 2019).…”