Machine Learning Model for Sepsis Prediction in Prolonged and Chronic Critical Illness: Development and Validation...

Area of Science:

• Critical Care Medicine
• Machine Learning in Healthcare
• Sepsis Prediction

Background:

• Existing machine learning (ML) models for sepsis prediction are not specifically designed for patients with prolonged or chronic critical illness (PCI/CCI).
• This gap highlights the need for specialized ML tools to accurately predict sepsis in this vulnerable patient group.

Purpose of the Study:

• To develop and validate a machine learning-based sepsis prediction model tailored for patients with PCI/CCI.
• To evaluate the generalizability of a PCI/CCI-focused model versus a universal model trained on mixed ICU populations.

Main Methods:

• Analysis of ICU admissions from the Russian Intensive Care Dataset (RICD) for PCI/CCI patients and public PhysioNet datasets for acute critical illness.
• Development of ML models using tree-based algorithms (XGBoost, LightGBM, Random Forest, AdaBoost) within a right-aligned prediction framework with a 6-hour window.
• Internal and external validation, including subgroup analyses for sepsis phenotypes and comparison of PCI/CCI-focused and universal models.

Main Results:

• The PCI/CCI-focused XGBoost model achieved an AUROC of 0.82 internally but failed external validation (AUROC 0.47).
• A universal model showed reduced discrimination in PCI/CCI patients (AUROC mean difference 0.02, p = 0.0012).
• Key predictors included respiratory rate, heart rate, body temperature, and age; performance was better for hypoinflammatory sepsis (AUROC 0.84).

Conclusions:

• A right-aligned ML model for PCI/CCI demonstrates robust internal sepsis exclusion capabilities but limited cross-population generalizability.
• The findings emphasize the necessity for population-specific prediction models.
• Prospective validation is crucial before implementing these ML models in clinical practice for PCI/CCI patients.