Project Summary:

With the emergence of COVID-19 there has been increasing interest in identifying risk factors for mortality. While several studies have evaluated scoring systems for all hospitalized or all patients diagnosed with COVID-19 these models have not been validated in ICU patients. Furthermore commonly used scoring systems used in the ICU have not been validated in COVID-19.

We identified a cohort of ICU patients in a large health-system in the Mid-West and validated the performance and the SOFA score and 4C in predicting mortality when applied to clinical data from the first 24 hours of the ICU stay.

We also developed a machine learning model to evaluate which features could be added or modified in the current models to improve performance.

Relevant Skills:

• Clinical Medicine
• EHR Data Validation and Cleaning in Python (using PANDAS)
• Machine Learning (scikit-learn, XGBoost)
• Data visualizations (Matplotlib)
• Feature Importance and Explainability (SHAP)
• Statistical Methods: ROC, Sensitivity, Specificity, PPV, Bootstrapping, Permutation Testing

Validation:

4C Score was found to outperform the SOFA score in predicting mortality based on information from the first 24 hours of ICU admission highlighting the importance of disease specific models. Examination of the feature importance of the machine learning model suggests that the 4C score may be able to be modified to improve performance in a critically ill population.