Predicting Patient ICU Readmission Using Recurrent Neural Networks With Long Short-Term Memory
<p dir="ltr">Intensive care is quite expensive, so patients’ transfer to lower-level wards must be carefully considered. A patient who is discharged from the ICU too soon carries the danger of receiving insufficient care, which results in readmission. Readmission prediction tasks are...
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | , , |
| Published: |
2025
|
| Subjects: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | <p dir="ltr">Intensive care is quite expensive, so patients’ transfer to lower-level wards must be carefully considered. A patient who is discharged from the ICU too soon carries the danger of receiving insufficient care, which results in readmission. Readmission prediction tasks are less extensively researched, and the techniques employed have not produced satisfactory results. Conventional machine learning methods with non-sequential data have been employed in the majority of the studies. The goal of this research is to propose a solution that can accurately predict the likelihood of an unplanned ICU readmission within 30 days. The proposed method addresses the issues of missing values and data imbalance, in particular. Our method combines low-dimensional embeddings of medical concepts, such as diseases labeled using the ICD-9 code, with temporal features extracted from chart events data. Three alternative models are trained using Long-Short-Term Memory (LSTM), which employ ICU stay data from the last 24 hours, 48 hours, and 72 hours. The Medical Information Mart for Intensive Care (MIMIC-III) dataset is used to train and validate models. We tested our proposed methods on the unseen data of the MIMIC-III dataset to assess their efficacy. The model that was trained using the last 48-hour ICU data performed better than other models, reaching an 84.2% area under the receiver operating characteristic curve (AUC-ROC). The ROC scores of the other models are 82.6%, 78.9%, 77.4%, 75.6%, and 71.4% for Gated Recurrent Unit (GRU), Support Vector Machine (SVM), Logistic Regression (LR), Naïve Baise (NB), and K-Nearest Neighbor (KNN), respectively. The findings indicate that the suggested approach can outperform the state-of-the-art methods.</p><h2 dir="ltr">Other Information</h2><p dir="ltr">Published in: IEEE Access<br>License: <a href="https://creativecommons.org/licenses/by/4.0/deed.en" target="_blank">https://creativecommons.org/licenses/by/4.0/</a><br>See article on publisher's website: <a href="https://dx.doi.org/10.1109/access.2025.3601117" target="_blank">https://dx.doi.org/10.1109/access.2025.3601117</a></p> |
|---|