An Empirical Study on KDIGO-Defined Acute Kidney Injury Prediction in the Intensive Care Unit DOI Creative Commons
Xinrui Lyu, Bowen Fan, Matthias Hüser

et al.

medRxiv (Cold Spring Harbor Laboratory), Journal Year: 2024, Volume and Issue: unknown

Published: Feb. 3, 2024

Motivation Acute kidney injury (AKI) is a syndrome that affects large fraction of all critically ill patients, and early diagnosis to receive adequate treatment as imperative it challenging make early. Consequently, machine learning approaches have been developed predict AKI ahead time. However, the prevalence often underestimated in state-of-the-art approaches, they rely on an event annotation solely based creatinine, ignoring urine output. Methods We construct evaluate warning systems for multi-disciplinary ICU setting, using complete KDIGO definition AKI. propose several variants gradient-boosted decision trees (GBDT)-based models, including novel time-stacking approach. A LSTM-based model previously proposed prediction used comparison, which was not specifically evaluated settings yet. Results find optimal performance achieved by GBDT with time-based stacking technique (AUPRC=65.7%, compared model’s AUPRC=62.6%), motivated high relevance time since admission this task. Both models show mildly reduced limited training data perform fairly across different subco-horts, exhibit no issues gender transfer. Conclusion Following official substantially increases number annotated events. In our study GBDTs outperform LSTM prediction. Generally, we both types are robust variety arising data.

Language: Английский

FAMEWS: a Fairness Auditing tool for Medical Early-Warning Systems DOI Creative Commons

Marine Hoche,

Olga Mineeva,

M. Burger

et al.

medRxiv (Cold Spring Harbor Laboratory), Journal Year: 2024, Volume and Issue: unknown

Published: Feb. 8, 2024

Abstract Machine learning applications hold promise to aid clinicians in a wide range of clinical tasks, from diagnosis prognosis, treatment, and patient monitoring. These potential are accompanied by surge ethical concerns surrounding the use Learning (ML) models healthcare, especially regarding fairness non-discrimination. While there is an increasing number regulatory policies ensure safe integration such systems, translation practices remains open challenge. Algorithmic frameworks, aiming bridge this gap, should be tailored application enable fundamental human-right principles into accurate statistical analysis, capturing inherent complexity risks associated with system. In work, we propose set impartial checks adapted ML early-warning systems medical context, comprising on top standard metrics, analysis outcomes, screening sources bias pipeline. Our further fortified inclusion event-based prevalence-corrected as well tests measure biases. Additionally, emphasize importance considering subgroups beyond conventional demographic attributes. Finally, facilitate operationalization, present open-source tool FAMEWS generate comprehensive reports. reports address diverse needs interests stakeholders involved integrating practice. The has reveal critical insights that might otherwise remain obscured. This can lead improved model design, which turn may translate enhanced health outcomes.

Language: Английский

Citations

2
An Empirical Study on KDIGO-Defined Acute Kidney Injury Prediction in the Intensive Care Unit DOI Creative Commons
Xinrui Lyu, Bowen Fan, Matthias Hüser

et al.

medRxiv (Cold Spring Harbor Laboratory), Journal Year: 2024, Volume and Issue: unknown

Published: Feb. 3, 2024

Motivation Acute kidney injury (AKI) is a syndrome that affects large fraction of all critically ill patients, and early diagnosis to receive adequate treatment as imperative it challenging make early. Consequently, machine learning approaches have been developed predict AKI ahead time. However, the prevalence often underestimated in state-of-the-art approaches, they rely on an event annotation solely based creatinine, ignoring urine output. Methods We construct evaluate warning systems for multi-disciplinary ICU setting, using complete KDIGO definition AKI. propose several variants gradient-boosted decision trees (GBDT)-based models, including novel time-stacking approach. A LSTM-based model previously proposed prediction used comparison, which was not specifically evaluated settings yet. Results find optimal performance achieved by GBDT with time-based stacking technique (AUPRC=65.7%, compared model’s AUPRC=62.6%), motivated high relevance time since admission this task. Both models show mildly reduced limited training data perform fairly across different subco-horts, exhibit no issues gender transfer. Conclusion Following official substantially increases number annotated events. In our study GBDTs outperform LSTM prediction. Generally, we both types are robust variety arising data.

Language: Английский

Citations

0