A Data-Driven Approach to Predicting Septic Shock in the Intensive Care Unit.

Area of Science:

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

Background:

• Early sepsis and septic shock diagnosis significantly reduces mortality and improves patient outcomes.
• A critical need exists for automated clinical tools for large-scale screening of high-risk patients.
• Current diagnostic methods for sepsis and septic shock have limitations in early detection.

Purpose of the Study:

• To develop a novel algorithm for identifying patients at high risk of septic shock 24 hours prior to clinical diagnosis.
• To evaluate the performance characteristics of this predictive algorithm.
• To assess if the novel algorithm can improve current sepsis evaluation metrics.

Main Methods:

• Utilized publicly available intensive care unit data to create septic shock and control patient cohorts.
• Employed Bayesian networks to infer causal relationships between clinical variables (diagnoses, procedures, labs, demographics).
• Developed a predictive model for septic shock based on inferred causal networks and augmented existing sepsis risk scores.

Main Results:

• A novel predictive model achieved an area under the curve of 0.81, a negative predictive value of 0.87, and a positive predictive value of 0.65 for predicting septic shock 24 hours in advance.
• The model improved the specificity of the Quick Sequential Organ Failure Assessment (qSOFA), Systemic Inflammatory Response Syndrome (SIRS), and Modified Early Warning Score (MEWS).
• No improvement in Sequential Organ Failure Assessment (SOFA) score performance was observed when augmented with the novel model.

Conclusions:

• A data-driven, expert-agnostic algorithm was developed for early septic shock detection with strong predictive performance.
• The novel model enhances the specificity of common clinical sepsis screening tools.
• This work provides a foundation for real-time patient screening using electronic medical record data.