[Paper Review] An Improved Multi-Output Gaussian Process RNN with Real-Time Validation for Early Sepsis Detection

Joseph Futoma, Sanjay Hariharan+6 more|arXiv (Cornell University)|Aug 19, 2017

Sepsis Diagnosis and Treatment25 references75 citations

TL;DR

This work develops an improved MGP-RNN model that uses streaming lab results, vitals, and medications to predict sepsis before onset, with a new real-time validation scheme. The approach outperforms clinical baselines and earlier models, and is designed for deployment in a real-time dashboard.

ABSTRACT

Sepsis is a poorly understood and potentially life-threatening complication that can occur as a result of infection. Early detection and treatment improves patient outcomes, and as such it poses an important challenge in medicine. In this work, we develop a flexible classifier that leverages streaming lab results, vitals, and medications to predict sepsis before it occurs. We model patient clinical time series with multi-output Gaussian processes, maintaining uncertainty about the physiological state of a patient while also imputing missing values. The mean function takes into account the effects of medications administered on the trajectories of the physiological variables. Latent function values from the Gaussian process are then fed into a deep recurrent neural network to classify patient encounters as septic or not, and the overall model is trained end-to-end using back-propagation. We train and validate our model on a large dataset of 18 months of heterogeneous inpatient stays from the Duke University Health System, and develop a new "real-time" validation scheme for simulating the performance of our model as it will actually be used. Our proposed method substantially outperforms clinical baselines, and improves on a previous related model for detecting sepsis. Our model's predictions will be displayed in a real-time analytics dashboard to be used by a sepsis rapid response team to help detect and improve treatment of sepsis.

Motivation & Objective

• Motivate and address the challenge of early sepsis detection using rich, irregular clinical time series from EHRs.
• Develop a flexible predictive model that can handle irregular sampling and missing data while incorporating medication effects.
• Impute and denoise multivariate physiological time series with a multi-output Gaussian process and feed latent states into a deep RNN classifier.
• Enhance learning with realistic validation schemes, including real-time validation and case-control matching, to evaluate performance.
• Deploy a real-time analytics dashboard to assist sepsis rapid response teams with interpretable risk scores.

Proposed method

• Model patient clinical time series with a multi-output Gaussian process (MGP) to de-noise and impute observations onto a shared grid while maintaining uncertainty.
• Allow the MGP mean to depend on past medication administration to capture treatment effects on physiological trajectories.
• Relax the separable kernel assumption by using a sum of separable kernels (Linear Model of Coregionalization) to model variable temporal correlations across outputs.
• Feed latent MGP function values into a deep recurrent neural network (LSTM) classifier to predict sepsis, using backpropagation to train end-to-end.
• Incorporate target replication by supervising the RNN at multiple time points around sepsis events to mitigate label noise.
• Append missingness indicators to RNN inputs to exploit informative missingness patterns in labs/vitals.

Experimental results

Research questions

• RQ1 Can a flexible MGP-RNN framework improve early sepsis detection compared to clinical baselines and prior models?
• RQ2 Does incorporating medication effects and a more expressive kernel improve multivariate time series modeling for sepsis prediction?
• RQ3 How does real-time validation compare to traditional retrospective validation in assessing early warning performance?
• RQ4 What is the impact of targeting replication and missingness indicators on predictive performance?

Key findings

• The proposed MGP-RNN and its extensions substantially outperform clinical baselines and prior models for sepsis detection.
• Extensions including medication effects, non-separable kernels, target replication, and missingness indicators consistently improve performance.
• Real-time validation shows the model reduces false alarms compared with NEWS across a range of sensitivities.
• At practical early prediction horizons, the approach achieves favorable precision-recall and ROC characteristics relative to baselines.
• The approach is demonstrated on a large 18-month Duke Health System dataset with extensive irregular time series and missing data.