FIRST-ICU: forecasting interventions and risk stratification in the ICU using graph neural network autoencoders

Area of Science:

• Artificial Intelligence
• Critical Care Medicine
• Machine Learning

Background:

• Critically ill patients often require multiple, interdependent interventions.
• Existing prediction models typically fail to account for these complex interdependencies.
• Accurate prediction of multiple interventions is crucial for effective ICU decision support.

Purpose of the Study:

• To develop a unified deep learning framework, FIRST-ICU, for the joint prediction of seven intensive care unit (ICU) interventions.
• To integrate graph neural networks, LSTMs, and a novel Intervention Interaction Attention Module (IIAM) for enhanced prediction.
• To enable interpretable risk stratification through phenotype identification.

Main Methods:

• Developed FIRST-ICU using MIMIC-IV dataset (n=23,926) and externally validated on AmsterdamUMCdb (n=12,603).
• Integrated a graph neural network encoder, LSTM, and the IIAM for joint intervention prediction.
• Employed Temporal and Discrete Decoders, alongside Generative Topographic Mapping for phenotype discovery.

Main Results:

• Achieved AUC-ROC > 0.98 for all interventions with Brier scores < 0.035 using the Temporal Decoder.
• The Discrete Decoder outperformed baselines, achieving the highest Macro Average Precision for six of seven interventions.
• Significant improvements observed for vasopressor interventions (norepinephrine +28.0%, phenylephrine +32.8%); IIAM enhanced AUC-PR, particularly for vasopressors.

Conclusions:

• FIRST-ICU advances multi-intervention prediction by jointly modeling treatment co-occurrence and physiological signals.
• The model demonstrates robust generalizability across different healthcare settings and prescribing practices.
• FIRST-ICU offers a framework for interpretable risk stratification and improved multi-intervention decision support in the ICU.