SRFP079: Predicting Length of Stay in the Emergency Department with Deep Learning

Ian McGregor, MD

Abstract

Context: Emergency departments in the US are often crowded and have long wait times. A large variety of medical problems are treated in emergency departments from basic primary care to life threatening emergencies. Predicting how long someone may stay in the emergency department is a difficult task with many contributing variables that are difficult to track. Health informatics using machine learning methods, in particular deep learning neural networks, present a possible solution to this prediction. This information could then be used to allocate resources more efficiently and help with discharge/admit planning. Objective: Predict the length of stay for each patient encounter using electronic medical record (EMR) data and machine learning. Study Design: Create a prediction model of a patient’s length of stay in the emergency department using retrospective data from the EMR including event logs, demographic data, emergency department census, and medical data (demographic data, care areas, orders, and symptoms).
Setting: All encounters at a large urban community hospital’s emergency department between April 2018 and March 2019. Population Studied: All 83,483 patient encounters were included; there were no exclusions. Instrument: Python programming language using Keras and Tensorflow was used to develop the prediction model. Main Outcome Measure: Accuracy of the neural network model in predicting the length of stay in the emergency department for each patient encounter. Results: The neural network model predicted the length of stay at an accuracy of 90% for the following time categories of: 0 < 1 hours, 1 < 3 hours, 3 < 6 hours, 6 < 12 hours, and >= 12 hours. Conclusions: Machine learning using deep learning neural networks is a promising method for predicting length of stay in the emergency department and potentially in the hospital as a whole. Increasing the data available to the network and adjusting the algorithm could increase the accuracy of the model.
Leave a Comment
Jack Westfall
jwestfall@aafp.org 11/21/2021

Very interesting research. Great work. Thanks

Jaky Kueper
jkueper@uwo.ca 11/22/2021

Hi! I'm curious how event logs were treated - where and what time interval were these logs taken from and was there any preprocessing done on them?

Diane Harper
harperdi@med.umich.edu 11/22/2021

Present your results in Phoenix next year! Thank you for sharing your work with NAPCRG!

William R. Phillips
wphllps@uw.edu 11/23/2021

Very nice poster demonstrates real world application of AI. I would like to know more about the clinical, cost and patient outcome import of these predictions within these time ranges.

William R. Phillips
wphllps@uw.edu 11/23/2021

#2 - I'd really like to see this method compared to predictions made by warm bodies: ED doctor, ED nurse, front desk clerk. Thanks. - Bill Phillips

Andy Pasternak
avpiv711@sbcglobal.net 11/28/2021

I'll be interested to hear your findings and what the EDs will do with the data

Social Media

Address

NAPCRG
11400 Tomahawk Creek Parkway
Leawood, KS 66211
800.274.7928
Email: napcrgoffice@napcrg.org