Machine Learning Prediction of Prolonged Mechanical Ventilation

Monday, March 23, 2026

10:15 AM - 11:15 AM Central Time

Location: Connections Central - RST 03

Kaitlin Blotske, PharmD (she/her/hers)

University of Colorado Health, Colorado

Disclosure(s): No relevant financial relationship(s) to disclose.
Brian P. Murray, PharmD, BCCCP

University of Colorado Skaggs School of Pharmacy, Colorado

Disclosure information not submitted.
BZ

Bokai Zhao, N/A

Student
University of Georgia

Disclosure information not submitted.
ZC

Zhetao Chen, n/a

Graduate Student
UGA

Disclosure information not submitted.
Susan E. Smith, PharmD, BCCCP, FCCM

University of Georgia College of Pharmacy, Georgia

Disclosure information not submitted.
YK

Yanlei Kong, MS

PhD Candidate
Renmin University of China

Disclosure information not submitted.
YS

Ye Shen, PhD

Professor
University of Georgia

Disclosure information not submitted.
XC

Xianyan Chen, PhD

Associate Professor
UGA

Disclosure information not submitted.
Andrea Sikora, Select Member HP

University of Colorado School of Medicine, Colorado

Disclosure information not submitted.

Introduction: Medications are associated with prolonging the duration of mechanical ventilation; however, current models used to predict prolonged mechanical ventilation (PMV) have little to no incorporation of those causal medications. This study developed and validated prediction models for PMV using both traditional and machine learning techniques and specifically incorporated medication data. We hypothesized that machine learning models, using medication data, would have higher predictive performance.

Methods: This retrospective, multicenter cohort study was conducted from October 2015 to October 2020 for model development and January 2020 to June 2023 for model validation. Patients were included if they were 18 years or older and admitted to the ICU on mechanical ventilation for 24 hours or greater. Two separate data cohorts for validation of prediction models were collected from the University of North Carolina (UNC) and Oregon Health Sciences University (OHSU) via trained data analysts. Prediction models for prolonged mechanical ventilation (> 5 days) were developed using regression and supervised machine learning (XGBoost, Random Forest, and Support Vector Machine (SVM)) models. The primary outcome was to assess the performance of models using the area under the receiver operating characteristic (AUROC).

Results: A total of 318 randomized patients were included from UNC data pools in the development of the PMV prediction model and validated against two separate cohorts from UNC and OHSU. Among prediction models, the base model’s AUROC scores improved from 0.67 to 0.75 when adding medication regimen complexity and severity of illness scores. The strongest performing model was the Random Forest with an AUROC of 0.78, with similar performance achieve in the validation cohort.

Conclusions: Prolonged mechanical ventilation prediction model performance was improved when incorporating medication-related clinical factors in a validated cohort.