Fairness-Aware Graph Neural Networks for ICU Length of Stay Prediction in IoT-Enabled Environments

This study introduces a novel end-to-end framework designed to optimize patient outcomes and operational efficiency in intensive care units (ICUs) through fairness-aware and accurate length of stay (LOS) predictions within an IoT-enabled environment. Motivated by the concept of “fairness...

Full description

Saved in:
Bibliographic Details
Main Authors: Angelos Christos Maroudis, Konstantina Karathanasopoulou, Charithea C. Stylianides, George Dimitrakopoulos, Andreas S. Panayides
Format: Article
Language:English
Published: IEEE 2025-01-01
Series:IEEE Access
Subjects:
Online Access:https://ieeexplore.ieee.org/document/10963712/
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1850179684772347904
author Angelos Christos Maroudis
Konstantina Karathanasopoulou
Charithea C. Stylianides
George Dimitrakopoulos
Andreas S. Panayides
author_facet Angelos Christos Maroudis
Konstantina Karathanasopoulou
Charithea C. Stylianides
George Dimitrakopoulos
Andreas S. Panayides
author_sort Angelos Christos Maroudis
collection DOAJ
description This study introduces a novel end-to-end framework designed to optimize patient outcomes and operational efficiency in intensive care units (ICUs) through fairness-aware and accurate length of stay (LOS) predictions within an IoT-enabled environment. Motivated by the concept of &#x201C;fairness through unawareness,&#x201D; our proposed framework employs a demographic feature exclusion strategy, preventing access to potentially discriminatory information, and thus enforcing an inductive bias that redirects learning toward non-discriminatory pattern dependencies. To address the loss of static information, we introduce a custom graph neural network that dynamically reconstructs patient relationships over time, adapting from static demographics to evolving inter-patient correlations via multi-modal embeddings (e.g., medications, procedures, vitals, conditions) and learned feature-driven edge formation. We evaluate our approach on MIMIC-IV across diverse time-series formats and demonstrate its superiority over top performing LOS prediction methods, showcasing notable performance improvements while maintaining comparable computational complexity. Additionally, we present a performance analysis highlighting our approach&#x2019;s computational and predictive scalability across different graph sizes at inference, while underscoring its intrinsic ability to improve performance when deprived of demographic features by dynamically readjusting its expressive power, unlike all evaluated methodologies in this study. Finally, we present a comparative study against the literature&#x2019;s top-performing LOS prediction approaches that highlights our method&#x2019;s predictive superiority, demonstrating significant performance gains in predicting LOS exceeding 3 and 7 days, with improvements of up to 13.1% AUC and 10.3% ARP, and 14.1% AUC and 42.0% ARP, respectively. The code is available at <uri>https://github.com/icsa-hua/mimic_fairness_aware_gnn</uri>
format Article
id doaj-art-069b5a8bdcdb4ffcaf363044132e82f7
institution OA Journals
issn 2169-3536
language English
publishDate 2025-01-01
publisher IEEE
record_format Article
series IEEE Access
spelling doaj-art-069b5a8bdcdb4ffcaf363044132e82f72025-08-20T02:18:27ZengIEEEIEEE Access2169-35362025-01-0113645166453310.1109/ACCESS.2025.356018010963712Fairness-Aware Graph Neural Networks for ICU Length of Stay Prediction in IoT-Enabled EnvironmentsAngelos Christos Maroudis0https://orcid.org/0000-0002-9447-5837Konstantina Karathanasopoulou1Charithea C. Stylianides2https://orcid.org/0009-0002-3568-3449George Dimitrakopoulos3https://orcid.org/0000-0002-7424-8557Andreas S. Panayides4https://orcid.org/0000-0001-9829-7946Department of Informatics and Telematics, Harokopio University of Athens, Athens, GreeceDepartment of Informatics and Telematics, Harokopio University of Athens, Athens, GreeceCYENS Centre of Excellence, Nicosia, CyprusDepartment of Informatics and Telematics, Harokopio University of Athens, Athens, GreeceCYENS Centre of Excellence, Nicosia, CyprusThis study introduces a novel end-to-end framework designed to optimize patient outcomes and operational efficiency in intensive care units (ICUs) through fairness-aware and accurate length of stay (LOS) predictions within an IoT-enabled environment. Motivated by the concept of &#x201C;fairness through unawareness,&#x201D; our proposed framework employs a demographic feature exclusion strategy, preventing access to potentially discriminatory information, and thus enforcing an inductive bias that redirects learning toward non-discriminatory pattern dependencies. To address the loss of static information, we introduce a custom graph neural network that dynamically reconstructs patient relationships over time, adapting from static demographics to evolving inter-patient correlations via multi-modal embeddings (e.g., medications, procedures, vitals, conditions) and learned feature-driven edge formation. We evaluate our approach on MIMIC-IV across diverse time-series formats and demonstrate its superiority over top performing LOS prediction methods, showcasing notable performance improvements while maintaining comparable computational complexity. Additionally, we present a performance analysis highlighting our approach&#x2019;s computational and predictive scalability across different graph sizes at inference, while underscoring its intrinsic ability to improve performance when deprived of demographic features by dynamically readjusting its expressive power, unlike all evaluated methodologies in this study. Finally, we present a comparative study against the literature&#x2019;s top-performing LOS prediction approaches that highlights our method&#x2019;s predictive superiority, demonstrating significant performance gains in predicting LOS exceeding 3 and 7 days, with improvements of up to 13.1% AUC and 10.3% ARP, and 14.1% AUC and 42.0% ARP, respectively. The code is available at <uri>https://github.com/icsa-hua/mimic_fairness_aware_gnn</uri>https://ieeexplore.ieee.org/document/10963712/Time series analysisgraph neural networkslength of ICU stayMIMIC datasetInternet of Thingspatient outcome prediction
spellingShingle Angelos Christos Maroudis
Konstantina Karathanasopoulou
Charithea C. Stylianides
George Dimitrakopoulos
Andreas S. Panayides
Fairness-Aware Graph Neural Networks for ICU Length of Stay Prediction in IoT-Enabled Environments
IEEE Access
Time series analysis
graph neural networks
length of ICU stay
MIMIC dataset
Internet of Things
patient outcome prediction
title Fairness-Aware Graph Neural Networks for ICU Length of Stay Prediction in IoT-Enabled Environments
title_full Fairness-Aware Graph Neural Networks for ICU Length of Stay Prediction in IoT-Enabled Environments
title_fullStr Fairness-Aware Graph Neural Networks for ICU Length of Stay Prediction in IoT-Enabled Environments
title_full_unstemmed Fairness-Aware Graph Neural Networks for ICU Length of Stay Prediction in IoT-Enabled Environments
title_short Fairness-Aware Graph Neural Networks for ICU Length of Stay Prediction in IoT-Enabled Environments
title_sort fairness aware graph neural networks for icu length of stay prediction in iot enabled environments
topic Time series analysis
graph neural networks
length of ICU stay
MIMIC dataset
Internet of Things
patient outcome prediction
url https://ieeexplore.ieee.org/document/10963712/
work_keys_str_mv AT angeloschristosmaroudis fairnessawaregraphneuralnetworksforiculengthofstaypredictioniniotenabledenvironments
AT konstantinakarathanasopoulou fairnessawaregraphneuralnetworksforiculengthofstaypredictioniniotenabledenvironments
AT charitheacstylianides fairnessawaregraphneuralnetworksforiculengthofstaypredictioniniotenabledenvironments
AT georgedimitrakopoulos fairnessawaregraphneuralnetworksforiculengthofstaypredictioniniotenabledenvironments
AT andreasspanayides fairnessawaregraphneuralnetworksforiculengthofstaypredictioniniotenabledenvironments