Applying Machine Learning to Blood Count Data Predicts Sepsis with ICU Admission.

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Clinical Diagnostics

Background:

• Timely sepsis diagnosis is critical for effective treatment.
• Existing machine learning (ML) models for sepsis prediction are often complex and limited in application.
• This study focused on developing an ML model utilizing only complete blood count (CBC) diagnostics.

Purpose of the Study:

• To develop and validate a machine learning model for early sepsis prediction using readily available complete blood count (CBC) parameters.
• To assess the model's performance in non-intensive care unit (non-ICU) settings and its robustness through external validation.
• To compare the efficacy of a CBC-based model with a model that includes procalcitonin (PCT) for sepsis prediction.

Main Methods:

• A boosted random forest model was trained using patient age, sex, and CBC parameters (hemoglobin, platelets, MCV, WBC, RBC).
• Data were collected from a German tertiary care center (2014-2021), with external validation using data from another center and the MIMIC-IV database.
• An additional model incorporating procalcitonin (PCT) was trained for comparative analysis.

Main Results:

• The CBC-only model achieved an AUROC of 0.872, with external validation AUROCs of 0.805 and 0.845.
• The model incorporating PCT showed a higher AUROC (0.857) compared to PCT alone (0.790).
• Over 1.3 million laboratory requests were analyzed, identifying 2016 sepsis cases.

Conclusions:

• Routine CBC results, when analyzed with ML, can significantly enhance sepsis diagnosis in non-ICU patients.
• The developed CBC model demonstrates high robustness across external validations, facilitating early sepsis prediction.
• Implementation in clinical decision support systems can provide a critical time advantage, improving patient safety.