Predicting ICU Admissions for Hospitalized COVID-19 Patients with a Factor Graph-based Model

Yurui Cao, Phuong Cao, Haotian Chen, Karl M. Kochendorfer, Andrew B. Trotter, William L. Galanter, Paul M. Arnold, Ravishankar K Iyer

Research output: Chapter in Book/Report/Conference proceedingChapter

Abstract

This paper presents a factor graph-based model that takes comorbidities and clinical measurements as inputs and predicts intensive care unit (ICU) admissions 3 days and 7 days in advance for hospitalized COVID-19 patients. We applied the proposed model on a COVID-19 cohort from a large medical center in Chicago (with records from March 2020 to August 2021). We used the first occurrence of the Delta variant in the U.S., February 2021, as the threshold to divide the dataset into pre-Delta data (533 patients) and post-Delta data (56 patients). Our model demonstrated 0.82 AUC on the pre-Delta data and 0.87 AUC on the post-Delta data in 7-day predictions. Our contribution is a model that (i) explains relationships between different clinical features and provides interpretations for ICU admissions, (ii) outperforms existing methods for 7-day predictions, and (iii) maintains more robustness than existing models in predictions under the influence of the Delta variant. The proposed model could be used as a predictive tool in clinical practice to help clinicians in decision-making by predicting which patients will need ICU support in the future.

Original languageEnglish (US)
Title of host publicationStudies in Computational Intelligence
PublisherSpringer
Pages245-256
Number of pages12
DOIs
StatePublished - 2023

Publication series

NameStudies in Computational Intelligence
Volume1060
ISSN (Print)1860-949X
ISSN (Electronic)1860-9503

Keywords

  • COVID-19 prognosis
  • Comorbidities
  • Factor graph
  • Predictive biomarkers
  • Probabilistic graphical model

ASJC Scopus subject areas

  • Artificial Intelligence

Fingerprint

Dive into the research topics of 'Predicting ICU Admissions for Hospitalized COVID-19 Patients with a Factor Graph-based Model'. Together they form a unique fingerprint.

Cite this