A Novel Alternating Joint Longitudinal Model for Post-ICU Hemoglobin Prediction

TL;DR

This paper introduces a Bayesian joint longitudinal model that predicts post-ICU anemia and readmission by accounting for hospital events and patient history, validated on a large retrospective cohort.

Contribution

The novel model explicitly incorporates hospital admissions and discharges with patient history to improve anemia prediction accuracy.

Findings

Accurately estimates hemoglobin levels over time.

Predicts anemia status with AUC of 0.82.

Predicts 30-day readmission with AUC of 0.72.

Abstract

Anemia is common in patients post-ICU discharge. However, which patients will develop or recover from anemia remains unclear. Prediction of anemia in this population is complicated by hospital readmissions, which can have substantial impacts on hemoglobin levels due to surgery, blood transfusions, or being a proxy for severe illness. We therefore introduce a novel Bayesian joint longitudinal model for hemoglobin over time, which includes specific parametric effects for hospital admission and discharge. These effects themselves depend on a patient's hemoglobin at time of hospitalization; therefore hemoglobin at a given time is a function of that patient's complete history of admissions and discharges up until that time. However, because the effects of an admission or discharge do not depend on themselves, the model remains well defined. We validate our model on a retrospective cohort of 6,876 patients from the Rochester Epidemiology Project using cross-validation, and find it accurately estimates hemoglobin and predicts anemic status and hospital readmission in the 30 days post-discharge with AUCs of .82 and .72, respectively.