Development and validation of a machine learning-based interpretable model for predicting sepsis by complete blood...

Area of Science:

• Medical Informatics
• Clinical Pathology
• Machine Learning in Healthcare

Background:

• Sepsis is a life-threatening infectious disease with high mortality rates.
• Early detection and prompt intervention are critical for improving patient outcomes.
• Complete blood cell (CBC) parameters offer a readily available data source for sepsis assessment.

Purpose of the Study:

• To develop and validate a clinical prediction model for early sepsis detection.
• To leverage machine learning algorithms for analyzing CBC parameters.
• To identify key CBC parameters that predict sepsis onset.

Main Methods:

• Utilized data from 572 patients (215 sepsis, 357 local infection) at West China Hospital.
• Analyzed 57 CBC parameters using LASSO, RF, SVM, and XGBoost for feature selection.
• Developed a Logistic Regression model and evaluated performance using AUC, calibration, and decision curves on discovery and validation cohorts.
• Employed SHAP and BD profiles for model interpretability.

Main Results:

• The LASSO-based model achieved the highest diagnostic performance.
• Achieved an AUC of 0.9446 in the discovery cohort and 0.9001 in the validation cohort (P < 0.001).
• LY-Z, MO-Z, and PLT-I were identified as the most impactful CBC parameters for sepsis prediction.

Conclusions:

• A predictive model utilizing CBC parameters is effective for early sepsis detection.
• Machine learning enhances the utility of routine blood tests for critical care diagnostics.
• This approach offers a promising tool for timely sepsis identification and management.