Improved Interpretability Without Performance Reduction in a Sepsis Prediction Risk Score.

Area of Science:

• Clinical Medicine
• Biostatistics
• Health Informatics

Background:

• Sepsis is a critical condition causing significant hospital mortality.
• Machine learning (ML) models offer superior sepsis prediction compared to traditional risk scores.
• Limited clinical adoption of ML models stems from their lack of interpretability.

Purpose of the Study:

• To enhance the interpretability of ML-based sepsis prediction models.
• To maintain predictive performance in non-ICU settings.
• To develop an interpretable sepsis prediction tool.

Main Methods:

• A logistic regression model was developed for sepsis onset prediction.
• The model was converted into an integer point system (STEWS) using regression coefficients.
• STEWS performance was compared against the logistic regression model at 90% sensitivity using Positive Predictive Value (PPV).

Main Results:

• STEWS demonstrated statistically equivalent performance to the logistic regression model (0.051 vs. 0.051; p = 0.004).
• The conversion to an integer score did not compromise predictive accuracy.
• The study confirmed the robustness of the STEWS system.

Conclusions:

• The STEWS system achieves comparable performance to logistic regression for non-ICU sepsis prediction.
• Converting ML models into interpretable formats like STEWS is feasible without performance loss.
• STEWS offers a promising, interpretable alternative for clinical sepsis prediction.