CoChat

AI-Driven Predictions of Readmission and Mortality for Improved Discharge Decisions in Critical Care: A Retrospective Study

Open access

The DOI matches its Crossref record; 26 of the 26 references it cites were found in a registry or an index; 7 of the 12 sources its claims cite were checked for support.

Abstract

Background/Objectives: The transition from the intensive care unit (ICU) to the hospital ward is a critical high-risk period for patients. Early ICU discharge reduces costs and frees up ICU resources but can lead to readmission or unexpected death if patients are discharged prematurely. Despite the availability of risk stratification tools such as the Stability and Workload Index for Transfer (SWIFT) score, predicting ICU readmission remains challenging and inconsistent. However, artificial intelligence (AI) and machine learning (ML) techniques have recently shown promise in improving clinical decision support systems, particularly in the ICU. This study aimed to identify the risk factors and assess the performance of AI models in predicting readmission or death within seven days of ICU discharge using the MIMIC-IV (between 2008 and 2019) and Kangwon National University Hospital (KNUH, between 1 January 2016 and 28 February 2023) databases. Methods: This retrospective cohort study utilized the MIMIC-IV database for model training and internal validation and the KNUH database for external validation. Various machine learning and deep learning models have been developed to predict ICU readmission or death within seven days of discharge. The performance of the primary model, GRU-D++, was compared to the SWIFT score. Statistical analysis focused on the area under the receiver operating characteristic curve (AUROC) data to evaluate model accuracy. Results: The GRU-D++ model outperformed the SWIFT score, achieving AUROC of 0.802 and 0.756 for internal and external validations, respectively. Both datasets demonstrated that the GRU-D++ model provided better predictive performance for ICU readmission or death within seven days than the traditional SWIFT score. Conclusions: Our findings suggest that the GRU-D++ deep learning model is a valuable tool for the early detection of patient deterioration after ICU discharge, potentially aiding the prevention of ICU readmission. This study highlights the potential of AI to improve clinical decision-making in intensive care settings.

Overview

CoChat Summary

## TL;DR

This retrospective study trained and externally validated a GRU-D++ deep learning model to predict ICU readmission or death within 7 days of discharge, using MIMIC-IV for training/internal validation and KNUH for external validation. GRU-D++ outperformed the SWIFT score and achieved AUROC 0.802 internally and 0.756 externally.

## Key findings

  • The study included 50,261 ICU stays from MIMIC-IV and 4,852 ICU stays from KNUH; the outcome was readmission to the ICU or death within 7 days of ICU discharge.
  • GRU-D++ achieved the best predictive performance among compared models, with AUROC 0.802 in internal validation and AUROC 0.756 in external validation.
  • Compared with SWIFT, GRU-D++ performed better in both datasets: 0.82 vs 0.69 AUROC in MIMIC and 0.76 vs 0.68 AUROC in KNUH, with p < 0.001 by DeLong test.
  • SHAP analysis identified BUN, GCS, and WBC as the most influential features for predicting post-discharge readmission/mortality.

## Methods

  • Retrospective cohort study using MIMIC-IV v2.2 for model development/internal validation and KNUH data for external validation; included adult ICU stays with ICU length of stay > 24 h and excluded in-ICU deaths and transfers to other specialized units/OR.
  • Started with 48 clinical variables, averaged daily, then selected 26 variables with <50% missingness; numerical variables were normalized to zero mean and unit variance.
  • Compared GRU-D++ against nine other ML/DL models and the SWIFT score; internal validation used 10-fold cross-validation, with performance assessed mainly by AUROC (and also AUPR).

## Limitations

  • The authors state that MIMIC-IV is disproportionately White (81.8%) and note that applying MIMIC-only findings to Asian populations may be limited; they addressed this with KNUH fine-tuning and external validation.
  • They state that important system-level factors (e.g., ICU bed occupancy, nighttime/weekend discharge, step-down unit availability, rapid response teams) were not available in the retrospective datasets, so predictions may reflect institutional practice as well as physiology.
  • They state that the composite endpoint combines ICU readmission and post-ICU death, which may represent different clinical trajectories; the text also makes evident a limitation of class imbalance (about 7% failure rate) and retrospective design with need for prospective multi-center validation.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.3390/diagnostics16060874

Cites

26 of 26 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1 Flow diagram of MIMIC-IV and KNUH datasets (N = MIMIC-IV/KNUH). Abbreviations: ICU, Intensive care unit; MIMIC, Medical Information Mart for Intensive Care; KNUH, Kangwon National University Hospital.
Figure 1 Flow diagram of MIMIC-IV and KNUH datasets (N = MIMIC-IV/KNUH). Abbreviations: ICU, Intensive care unit; MIMIC, Medical Information Mart for Intensive Care; KNUH, Kangwon National University Hospital.CC BY · PubMed Central
Figure 2 Overview of the study design and predictive model. Abbreviations: ICU, Intensive care unit; GRU, Gated Recurrent Unit.
Figure 2 Overview of the study design and predictive model. Abbreviations: ICU, Intensive care unit; GRU, Gated Recurrent Unit.CC BY · PubMed Central
Figure 3 Comparison of predictive performance using GRU-D++ and SWIFT. (A) MIMIC-IV and (B) KNUH. Abbreviation: AUROC, area under the receiver operating characteristic; SWIFT, stability and Workload Index for Transfer.
Figure 3 Comparison of predictive performance using GRU-D++ and SWIFT. (A) MIMIC-IV and (B) KNUH. Abbreviation: AUROC, area under the receiver operating characteristic; SWIFT, stability and Workload Index for Transfer.CC BY · PubMed Central
Figure 4 Feature importance and impact on post-discharge deterioration (readmission or death) prediction using Shapley values in ICU patients. The color represents the variable value: red indicates a high value of a specific clinical variable, while blue indicates a low value. Abbreviations: HFNC, high-flow nasal cannula; MBP, mean blood pressure; HR, heart rate; RR, respiratory rate; BT, body temperature; SpO2, pulse oxygen saturation; WBC, white blood cell count; BUN, blood urea nitrogen; GCS, Glasgow Coma Scale; SAS, Sedation-agitation scale; LOS, length of stay.
Figure 4 Feature importance and impact on post-discharge deterioration (readmission or death) prediction using Shapley values in ICU patients. The color represents the variable value: red indicates a high value of a specific clinical variable, while blue indicates a low value. Abbreviations: HFNC, high-flow nasal cannula; MBP, mean blood pressure; HR, heart rate; RR, respiratory rate; BT, body temperature; SpO2, pulse oxygen saturation; WBC, white blood cell count; BUN, blood urea nitrogen; GCS, Glasgow Coma Scale; SAS, Sedation-agitation scale; LOS, length of stay.CC BY · PubMed Central
Figure 5 Average prediction scores, where a higher score indicates a higher risk of ICU discharge. The blue line represents the average prediction scores for the discharge success group, and the orange line represents the discharge failure group, both with 95% confidence intervals. The dashed gray line indicates the optimal threshold.
Figure 5 Average prediction scores, where a higher score indicates a higher risk of ICU discharge. The blue line represents the average prediction scores for the discharge success group, and the orange line represents the discharge failure group, both with 95% confidence intervals. The dashed gray line indicates the optimal threshold.CC BY · PubMed Central
Figure 6 Trends in patients’ discharge prediction scores based on the developed model. (A) Random sampling of patients from the MIMIC-IV data. Patient A1–A5: Discharge success group; Patient A6–A10: Discharge failure group. (B) Random sampling of patients from the KNUH data. Patient B1–B5: Discharge success group; Patient B6–B10: Discharge failure group.
Figure 6 Trends in patients’ discharge prediction scores based on the developed model. (A) Random sampling of patients from the MIMIC-IV data. Patient A1–A5: Discharge success group; Patient A6–A10: Discharge failure group. (B) Random sampling of patients from the KNUH data. Patient B1–B5: Discharge success group; Patient B6–B10: Discharge failure group.CC BY · PubMed Central

Verification

Claims the paper makes

5 Valid · 4 Not valid · 3 Not checked · 71 Own results and interpretations

40 more claims

Create a free account to see all 83 claims and to verify the 3 references still open.

See all claims — free account
  • It may state two separate claims as one. 25
  • Probably not a substantial claim. 19

Create a free account to see these claims.

Citations

Cites

Research works printed but not linked: 3

  1. Shi, 2022, Predicting Unplanned 7-Day Intensive Care Unit Readmissions with Machine Learning Models for Improved Discharge Risk Assessment, AMIA Summits Transl. Sci. Proc., 2022, 446
  2. Royalty, J.P. (2021). Machine Learning Time-to-Event Mortality Prediction in MIMIC-IV Critical Care Database. [Doctoral Dissertation, Texas A&M University].
  3. Lundberg, S.M., and Lee, S.-I. (2017). A Unified Approach to Interpreting Model Predictions. Advances in Neural Information Processing Systems, Curran Associates, Inc.

Metadata

Authors
Heo, Y., Kim, M., Han, S.-S., Kim, T.-H., Heo, J., Kim, D., Kim, W. J., Lee, S.-J., Kwon, O. B., Kim, Y., Choi, H.-S., & Moon, D. H.
Affiliations
  • Kangwon National University (KR)
  • Department of Research & Development, Ziovision Co., Ltd., Chuncheon 24341, Republic of Korea
  • Institute of Medical Science, School of Medicine, Kangwon National University, Chuncheon 24341, Republic of Korea
  • Department of Pulmonology, Chinjujeil Hospital, Jinju 52709, Republic of Korea
  • Seoul National University of Science and Technology (KR)
Corresponding author
  • Choi, Hyun-Soo
  • Moon, Da Hye
Venue
Diagnostics
Publisher
MDPI AG
Published
2026-03-16
Volume
16
Issue
6
Pages
874
ISSN
2075-4418 (ISSN-L)
OpenAlex
W7137124956
Cited by
0
References
23
FWCI
0.00
Type
journal article
Language
en
Open access
Locations
  • Diagnostics (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
  • Korea Health Technology R&D Project through the Korea Health Industry Development Institute
  • Ministry of Health & Welfare, Republic of Korea (RS-2021-KH114109)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Machine Learning in Healthcare
  • Health Sciences › Health Professions › Health Information Management › Artificial Intelligence in Healthcare
  • Health Sciences › Medicine › Critical Care and Intensive Care Medicine › Sepsis Diagnosis and Treatment
Keywords
ICU readmission prediction, post-ICU mortality, deep learning, clinical decision support, risk stratification, MIMIC-IV

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.