(1359) Bayesian Estimates of Risk Mortality on ICU Admission: Including Uncertainties in Predictions

Introduction: Predictive scoring systems have been created to predict the risk of in-hospital death for adults admitted to the intensive care unit. The three most common systems in use are the Acute Physiologic and Chronic Health Evaluation (APACHE version IVa), Simplified Acute Physiology Score (SAPS version 3), and the Sequential Organ Failure Assessment (SOFA). There are substantial limitations in existing models including: Temporal changes in the significance of chronic health conditions, variable calibration by race and clinical conditions, and lack of communication about prediction uncertainty.

Methods: We used the EICU-CRD database to assemble a cohort of 166,355 first ICU admissions and extracted age, admission details (unit type, admission source), laboratory values (sodium, potassium, albumin, anion gap, chloride, bicarbonate, bilirubin, glucose, creatinine, blood urea nitrogen, hemoglobin, platelet count, white blood cell count, hemoglobin, prothrombin time, and internationalized normalized ratio), and therapy data (ventilator, oxygen, dialysis, and vasopressor use). All data came from the first 24 hours of hospitalization, and we used the worst extreme laboratory values from the same period. We used variational inference to estimate a Bayesian hierarchical logistic regression model with clustering by hospital and unit. Age and laboratory values were included as spline terms. We used 80% of the data for training and 20% for model evaluation. Our analysis was conducted in Python using PyMC.

Results: The model had an evaluation set area under the receiver operating curve of 0.85 and an average precision of 0.42. The Brier skill score (the percentage improvement in brier score of the model compared to a ‘no-skill’ baseline model) was 20.2%. These results were similar across race and gender subgroups. The median width of the 94% highest density posterior interval for the predicted outcomes on the logit scale was 0.12 (IQR (interquartile range) 0.04 – 0.32).

Conclusions: A Bayesian model for predicting the risk of death is feasible and accurate on internal validation. The resulting predictions include a measure of uncertainty that may be helpful when scores are used to compare patients. Further exploration and validation of this technique is warranted.