Predicting sepsis with a recurrent neural network using the MIMIC III database.

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Critical Care Medicine

Background:

• Sepsis prediction is crucial for timely intervention and improved patient outcomes in intensive care units.
• Existing algorithms like InSight have shown promise but may have limitations in capturing complex temporal patterns.
• Recurrent neural networks (RNNs) offer potential for analyzing sequential data, making them suitable for time-series prediction tasks like sepsis onset.

Purpose of the Study:

• To develop and evaluate a recurrent neural network (RNN) for predicting sepsis onset.
• To compare the performance of the RNN against the previously proposed InSight algorithm.
• To investigate the impact of sequence length and definition of sepsis on prediction accuracy.

Main Methods:

• Retrospective analysis of adult ICU patients from the MIMIC III database.
• Utilized Area Under the Receiver Operating Characteristic Curve (AUROC) to measure prediction performance.
• Evaluated RNN performance with varying look-back periods (3, 6, and 12 hours) and compared it to InSight.

Main Results:

• The RNN achieved a higher AUROC of 0.81 (95% CI: 0.78-0.84) at 3 hours prior to sepsis onset, compared to InSight's 0.72 (95% CI: 0.69-0.75).
• At 90% sensitivity, the RNN achieved 47.0% specificity, significantly outperforming InSight's 31.1% specificity.
• Prediction performance was significantly influenced by the look-back window length.

Conclusions:

• Recurrent neural networks demonstrate superior predictive performance for sepsis onset compared to the InSight algorithm.
• The ability of RNNs to capture temporal dependencies is likely responsible for the improved prediction accuracy.
• Further research is needed to refine sepsis onset detection methods for retrospective analysis.