[Paper Review] DeepAISE -- An End-to-End Development and Deployment of a Recurrent Neural Survival Model for Early Prediction of Sepsis

Supreeth P. Shashikumar, Christopher S. Josef+2 more|arXiv (Cornell University)|Aug 10, 2019

Sepsis Diagnosis and Treatment41 references8 citations

TL;DR

DeepAISE is an end-to-end recurrent neural survival model that predicts sepsis onset in ICU patients using temporal clinical time series, integrating a clinically defined sepsis onset criterion with treatment policies to optimize prediction accuracy. It achieves AUCs of 0.90 (internal) and 0.87 (external) with low false alarm rates (FAR=0.20 and 0.26), while providing interpretable risk representations through joint modeling of survival outcomes and feature interactions.

ABSTRACT

Sepsis, a dysregulated immune system response to infection, is among the leading causes of morbidity, mortality, and cost overruns in the Intensive Care Unit (ICU). Early prediction of sepsis can improve situational awareness amongst clinicians and facilitate timely, protective interventions. While the application of predictive analytics in ICU patients has shown early promising results, much of the work has been encumbered by high false-alarm rates. Efforts to improve specificity have been limited by several factors, most notably the difficulty of labeling sepsis onset time and the low prevalence of septic-events in the ICU. Here, we present DeepAISE (Deep Artificial Intelligence Sepsis Expert), a recurrent neural survival model for the early prediction of sepsis. We show that by coupling a clinical criterion for defining sepsis onset time with a treatment policy (e.g., initiation of antibiotics within one hour of meeting the criterion), one may rank the relative utility of various criteria through offline policy evaluation. Given the optimal criterion, DeepAISE automatically learns predictive features related to higher-order interactions and temporal patterns among clinical risk factors that maximize the data likelihood of observed time to septic events. DeepAISE has been incorporated into a clinical workflow, which provides real-time hourly sepsis risk scores. A comparative study of four baseline models indicates that DeepAISE produces the most accurate predictions (AUC=0.90 and 0.87) and the lowest false alarm rates (FAR=0.20 and 0.26) in two separate cohorts (internal and external, respectively), while simultaneously producing interpretable representations of the clinical time series and risk factors.

Motivation & Objective

• To address the challenge of high false-alarm rates in sepsis prediction by modeling time-to-event outcomes with clinical event definitions.
• To improve prediction specificity by aligning sepsis onset labeling with real-world treatment policies (e.g., antibiotic initiation within one hour).
• To develop an end-to-end system that learns complex temporal and higher-order interactions among clinical risk factors for sepsis.
• To enable real-time deployment of sepsis risk scores in clinical workflows with interpretability.

Proposed method

• The model uses a recurrent neural network (RNN) architecture to process sequential ICU vital signs and lab values over time.
• It applies a survival modeling framework that estimates the time-to-event distribution for sepsis onset, using a likelihood-based objective function.
• A clinical criterion defines sepsis onset as meeting SIRS and organ dysfunction criteria, with treatment initiation within one hour used as a policy anchor.
• Offline policy evaluation is used to rank and select the optimal sepsis onset definition based on predictive utility.
• The model learns predictive representations through joint optimization of survival likelihood and temporal feature interactions.
• Interpretable attention mechanisms are applied to highlight clinically relevant time series patterns and risk factors.

Experimental results

Research questions

• RQ1 Can a recurrent neural survival model improve early sepsis prediction accuracy while reducing false alarm rates compared to baseline models?
• RQ2 How does aligning sepsis onset definition with a real-world treatment policy affect model performance and clinical utility?
• RQ3 To what extent can deep learning models capture higher-order temporal interactions among clinical variables for sepsis prediction?
• RQ4 Can an end-to-end system be successfully deployed in a clinical workflow to deliver real-time risk scores with interpretability?
• RQ5 What is the relative utility of different sepsis onset criteria when evaluated under a consistent treatment policy?

Key findings

• DeepAISE achieved an AUC of 0.90 in the internal validation cohort and 0.87 in the external validation cohort, outperforming four baseline models in discrimination.
• The model recorded false alarm rates of 0.20 (internal) and 0.26 (external), significantly lower than baseline models, indicating improved specificity.
• By using offline policy evaluation, the study identified a clinically grounded sepsis onset criterion that maximized predictive utility.
• The model generated interpretable representations of clinical time series, highlighting key temporal patterns and risk factors contributing to sepsis risk.
• The integration of DeepAISE into a clinical workflow enabled real-time hourly risk scoring, supporting timely clinical decision-making.
• The survival modeling approach effectively captured complex, non-linear temporal dependencies in ICU data, enhancing predictive performance.