EURO-Online login
- New to EURO? Create an account
- I forgot my username and/or my password.
- Help with cookies
(important for IE8 users)
2841. GIMI: Generative iterative multiple imputation for improving health trajectory estimation in the ICU
Invited abstract in session TB-31: Analytics and the link with stochastic dynamics II, stream Analytics.
Tuesday, 10:30-12:00Room: 046 (building: 208)
Authors (first author is the speaker)
| 1. | Anahita Khojandi
|
| University of Tennessee-Knoxville | |
| 2. | Matt Baucum
|
| Florida State University | |
| 3. | Senne Van Steelandt
|
| Thalamus | |
| 4. | Rama Vasudevan
|
| Oak Ridge National Laboratory | |
| 5. | Robert Davis
|
| University of Tennessee Health Science Center | |
| 6. | Nadeem Shafi
|
| University of Tennessee Health Science Center |
Abstract
As the field of precision medicine grows, accurately forecasting patients' health trajectories is essential for developing data-driven treatment models. Yet researchers must still contend with high levels of missingness due to issues such as missed bedside readings, data entry error, or in the case of outpatient monitoring, inconsistent patient follow-ups. When forecasting patient health trajectories, even small imputation errors can compound over the given time horizon, complicating the task of missing data imputation and jeopardizing model accuracy. To address this, we develop a novel on-training imputation algorithm - Generative Iterative Multiple Imputation (GIMI) - which estimates missing health data by directly maximizing the accuracy of patients' predicted health trajectories. Rather than separating the imputation and forecasting steps, GIMI leverages state-of-the-art generative neural networks to simultaneously forecast patients' missing and observed health trajectories, and is trained in an end-to-end fashion to maximize forecast accuracy. Using a large clinical dataset of intensive care unit (ICU) patients, we show that GIMI outperforms commonly used benchmarks in forecasting patient health trajectories in the presence of missing data. Notably, we show that the imputation errors of several existing imputation techniques compound over the health trajectory time horizon, whereas GIMI's prediction accuracy remains stable.
Keywords
- Analytics and Data Science
- Health Care
- Machine Learning
Status: accepted
Back to the list of papers