FIRST-ICU: a deep learning framework for multi-intervention prediction and risk stratification in critical care with cross-continental external validation
N Aikodon, I Olier, B W Johnston, I D Welters, G Y H Lip, S Ortega-MartorellAbstract
Introduction
Critically ill patients frequently require multiple concurrent interventions, including mechanical ventilation and vasoactive medications, with complex interdependencies between treatments. Current prediction models address interventions independently, failing to capture clinically important co-administration patterns such as vasopressor escalation in refractory shock.
Purpose
To develop and externally validate a unified deep learning framework (FIRST-ICU) that jointly predicts multiple ICU interventions while explicitly modelling treatment co-dependencies, and to identify clinically actionable risk phenotypes.
Methods
FIRST-ICU integrates a graph neural network encoder with an Intervention-Aware Attention Module (IAAM) that enables interventions to attend to each other. The framework predicts seven interventions: invasive and non-invasive ventilation, and five vasoactive agents. Development used MIMIC-IV (n=23,926, USA; mean age 64.8, SD 16.12; 41.77% female); external validation used AmsterdamUMCdb (n=12,603, Netherlands; mean age 62.14, SD 15.9; 37.06% female) without retraining. Cold-start analysis evaluated prediction in patients with no prior intervention exposure. Generative Topographic Mapping provided probabilistic risk stratification.
Results
External validation maintained AUC-ROC >0.90 for all intervention categories despite substantial prescribing differences between continents (Figure 1). Cold-start analysis confirmed robust prediction from physiological signals alone (AUC-PR 0.67-0.87), validating genuine early warning capability. FIRST-ICU outperformed baselines, with the largest improvements for vasopressors: norepinephrine +28% and phenylephrine +33%. Six clinically distinct phenotypes were identified with markedly different mortality, ranging from 12.5% (stable) to 50.6% (multi-organ failure with refractory shock) (Figure 2).
Conclusion
FIRST-ICU demonstrates that joint modelling of intervention co-dependencies improves prediction of haemodynamic and respiratory support needs. Cross-continental generalisability, validated physiological learning, and interpretable risk phenotypes address key barriers to translating AI into critical care decision support.Figure 1Figure 2