DeepSOFA: A Continuous Acuity Score for Critically Ill Patients using Clinically Interpretable Deep Learning.

Area of Science:

• Critical Care Medicine
• Artificial Intelligence in Healthcare
• Biomedical Informatics

Background:

• Traditional illness severity scores in ICUs are time-consuming and error-prone.
• Existing methods fail to utilize real-time electronic health record data and dynamic patient patterns.

Purpose of the Study:

• To introduce DeepSOFA, a novel acuity score framework utilizing temporal data and interpretable deep learning.
• To assess illness severity dynamically throughout an ICU stay.

Main Methods:

• Developed DeepSOFA, a deep learning model, to analyze temporal physiological measurements.
• Compared DeepSOFA's predictive accuracy against Sequential Organ Failure Assessment (SOFA) baseline models using identical input data.

Main Results:

• DeepSOFA significantly outperformed SOFA in predicting in-hospital mortality across the entire ICU stay.
• A DeepSOFA model achieved a mean AUC of 0.90, compared to SOFA models with AUCs of 0.79 and 0.85.
• The model demonstrated strong performance in both a public database and an institutional cohort.

Conclusions:

• DeepSOFA offers a more accurate and timely assessment of illness severity in critically ill patients.
• Deep learning models can identify patients at risk for adverse events, enabling proactive interventions.
• DeepSOFA can enhance shared decision-making regarding patient care and resource allocation in the ICU.