Maintenance Notice

Due to necessary scheduled maintenance, the JMIR Publications website will be unavailable from Wednesday, July 01, 2020 at 8:00 PM to 10:00 PM EST. We apologize in advance for any inconvenience this may cause you.

Who will be affected?

Accepted for/Published in: JMIR Biomedical Engineering

Date Submitted: Apr 26, 2023
Date Accepted: Dec 29, 2023

The final, peer-reviewed published version of this preprint can be found here:

A Deep Learning Framework for Predicting Patient Decannulation on Extracorporeal Membrane Oxygenation Devices: Development and Model Analysis Study

Fuller J, Abramov A, Mullin D, Beck J, Lemaitre P, Azizi E

A Deep Learning Framework for Predicting Patient Decannulation on Extracorporeal Membrane Oxygenation Devices: Development and Model Analysis Study

JMIR Biomed Eng 2024;9:e48497

DOI: 10.2196/48497

PMID: 38875691

PMCID: 11041448

A Deep Learning Framework for Predicting Patient Decannulation on Extracorporeal Membrane Oxygenation Devices: Development and Model Analysis Study

  • Joshua Fuller; 
  • Alexey Abramov; 
  • Dana Mullin; 
  • James Beck; 
  • Philippe Lemaitre; 
  • Elham Azizi

ABSTRACT

Background:

Venovenous extracorporeal membrane oxygenation (VV ECMO) is the ultimate therapy for patients with refractory respiratory failure. The decision to remove someone from the device, or decannulate, is chosen through a mixture of weaning trials and clinical intuition. To date, there are no prognostication scores to help clinicians determine who will be successfully weaned and decannulated.

Objective:

To remedy this, we present the Continuous Evaluation of VV ECMO Outcomes (CEVO), a deep learning-based model for predicting decannulation results on VV-ECMO. The running metric can be applied daily to determine high-risk and low-risk patients for decannulation. Physicians may then consider patients for a weaning trial using a combination of biological factors and the model's result.

Methods:

Using an LSTM-based network, CEVO is the first ECMO model capable of integrating discrete and categorical clinical information, such as age, sex, and BMI, with continuous data collected from the ECMO machine. Model results are calibrated and turned into risk groups, ranging from zero (high risk) to three (low risk). Two synthetic data sets were then created through Gaussian process regression in order to investigate the model's predictive ability.

Results:

In data collected from 118 VV ECMO supported patients at Columbia University Irving Medical Center, CEVO demonstrates consistently superior classification performance compared to contemporary models. That being said, the area under the receiver operator characteristic (AUROC) demonstrates that the model's patient-by-patient predictive power may be too low to be integrated into a clinical setting (95% confidence interval for AUROC was [0.6822, 0.7055]). However, the patient risk classification system, ranging from high to low risk, shows more promise. When measured at seventy-two-hours, the high-risk group would go on to have a successful decannulation rate of 58% (7/12), while the low-risk group would go on to have a successful decannulation rate of 92% (11/12) (P=.039). When measured at ninety-six hours, the groups had an eventual successful decannulation rate of 54% (6/11) and 100% (9/9) respectively (P=.01). To evaluate CEVO's performance edge, two synthetic datasets were created using a Gaussian process regression model. We hypothesized that CEVO's improved performance was due to its ability to efficiently capture short-term and transient temporal patterns. Indeed, CEVO exhibited improved performance on synthetic data with inherent temporal dependencies (P<.001) compared to logistic regression and a Dense Network.

Conclusions:

For prognostication on VV ECMO machines, the ability to interpret and integrate real-time information is paramount for creating accurate models predictive of patient outcomes. Integrating information as it comes in allows CEVO to retain usefulness throughout the run. Our framework can guide future incorporation of CEVO into more comprehensive intensive care systems.


 Citation

Please cite as:

Fuller J, Abramov A, Mullin D, Beck J, Lemaitre P, Azizi E

A Deep Learning Framework for Predicting Patient Decannulation on Extracorporeal Membrane Oxygenation Devices: Development and Model Analysis Study

JMIR Biomed Eng 2024;9:e48497

DOI: 10.2196/48497

PMID: 38875691

PMCID: 11041448

Download PDF


Request queued. Please wait while the file is being generated. It may take some time.

© The authors. All rights reserved. This is a privileged document currently under peer-review/community review (or an accepted/rejected manuscript). Authors have provided JMIR Publications with an exclusive license to publish this preprint on it's website for review and ahead-of-print citation purposes only. While the final peer-reviewed paper may be licensed under a cc-by license on publication, at this stage authors and publisher expressively prohibit redistribution of this draft paper other than for review purposes.