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

Abstract

Critically ill patients frequently require multiple concurrent interventions with complex interdependencies, yet existing prediction models treat these as independent events. We developed FIRST-ICU (Forecasting Interventions and Risk Stratification in the ICU), a unified deep learning framework integrating a graph neural network encoder, LSTM, and a novel Intervention Interaction Attention Module (IIAM) for joint prediction of seven ICU interventions. FIRST-ICU was developed using MIMIC-IV (n = 23,926) and externally validated on AmsterdamUMCdb (n = 12,603) without retraining. The Temporal Decoder achieved AUC-ROC > 0.98 for all interventions with Brier scores <0.035. The Discrete Decoder outperformed state-of-the-art baselines, achieving the highest Macro Average Precision for six of seven interventions. Gains were largest for vasopressor interventions, with norepinephrine and phenylephrine improving by 28.0% and 32.8%, respectively, over CNN baselines. Ablation analysis showed that IIAM improved AUC-PR, particularly for vasopressor prediction. Cold-start analysis confirmed robust prediction from physiological signals alone (AUC-PR 0.667-0.868). External validation maintained AUC-ROC > 0.90 across intervention categories despite substantial differences in prescribing practices. A Generative Topographic Mapping layer identified six clinically distinct phenotypes, enabling interpretable risk stratification. FIRST-ICU advances multi-intervention prediction through joint modelling of treatment co-occurrence patterns, validated physiological learning, cross-continental generalisability, and interpretable risk stratification. It provides a framework for multi-intervention ICU decision support.