Answers are grounded in this paper and cite their sources.
Many existing ICU risk prediction approaches do not explicitly represent physiological relationships and offer limited insight into how predictions are formed. In this study, patient trajectories are represented using a temporal graph structure that combines physiological measurements with information on ongoing therapeutic interventions, allowing patient stability to be examined in relation to clinical support. Ablation experiments indicate that removing therapeutic context substantially reduces predictive performance, suggesting that treatment information plays an important role in risk estimation. In the absence of this context, distinguishing between apparent physiological stability and actual clinical improvement becomes more challenging. Compared with deep learning baselines, the graph-based formulation exhibits more consistent calibration behavior and lower sensitivity to structural perturbations. Analysis of longitudinal attention patterns further shows forms of instability that are not captured by models excluding therapeutic context. Benchmarking against baseline models, including tabular, tree-based, sequence, and graph-based approaches, shows clear differences in model behavior. While tree-based and sequence models achieve strong discrimination performance, sequence models exhibit substantial miscalibration, with high ECE values across prediction horizons. In contrast, the proposed graph-based model achieves lower raw discrimination but produces probability estimates that can be effectively improved through post-hoc calibration, resulting in highly reliable predictions. These results highlight an important trade-off between predictive accuracy and reliability. Models that achieve high discrimination can still produce overconfident or unstable probability estimates. As a result, the proposed approach yields more consistent predictions by capturing relationships between clinical variables, making it more suitable for settings where reliable risk estimation is required.
No takes yet. Share an insight, caveat, or question.
Joseph et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: