Continuous Prediction of Emergency Department Disposition with Multimodal Deep Learning

Poster Number: P79
Presentation Time: 05:00 PM - 06:30 PM

Abstract Keywords: Deep Learning, Clinical Decision Support, Information Extraction, Machine Learning, Precision Medicine
Primary Track: Applications
Programmatic Theme: Clinical Informatics

Background: Emergency Department (ED) physicians face challenges in rapidly determining patient disposition, leading to prolonged stays and overcrowding. This study aims to improve disposition accuracy and reduce decision-making time using static and continuous patient information.

Methods: We developed a transformer-based model using a multimodal dataset from 169,244 ED visits (2020-2023), incorporating triage data, orders, continuous vitals, lab results, and radiology results to predict discharge, hospital admission, or ICU admission throughout a patient's ED stay.

Results: Our model achieved an Area Under the Receiver Operating Characteristic (AUROC) curve of 0.80 at triage time and 0.91 at final disposition time.

Conclusion: A multimodal machine learning model for continuous prediction of patient disposition in the ED has the potential to assist providers in real-time management of patient flow, improving efficiency and patient outcomes.

Speaker(s):
Kevin Tang, BS
Albert Einstein College of Medicine

Author(s):
Kevin Tang, BS - Albert Einstein College of Medicine; Julia Reisler, BS - Stanford University; Tom Jin, MS - Stanford University; David Kim, MD PhD - Stanford University;