Predicting intervention onset in the ICU with switching state space models.

Area of Science:

• Artificial Intelligence in Medicine
• Critical Care Medicine
• Physiological Monitoring

Background:

• The impact of intensive care unit (ICU) interventions is often not fully understood, particularly in diverse patient groups.
• Accurate prediction of ICU treatments remains a challenge due to patient heterogeneity.

Purpose of the Study:

• To develop and evaluate unsupervised machine learning models for capturing dynamic physiological states from vital signs.
• To assess the utility of these learned states in predicting key ICU interventions.

Main Methods:

• Trained unsupervised switching state autoregressive models on vital signs from the MIMIC-III database (36,050 patients).
• Compared the predictive power of learned physiological states against static demographics and raw vital signs.
• Evaluated prediction accuracy for five ICU treatments: ventilation, vasopressor administration, and three types of transfusions.

Main Results:

• Learned physiological states, when combined with demographics and raw vital signs, significantly improved prediction for most ICU interventions.
• Enhanced prediction accuracy was observed up to 4 and 8 hours before the intervention onset.
• The model's performance is competitive with existing methods on a larger and more diverse patient cohort.

Conclusions:

• Unsupervised learning of physiological states from vital signs offers a powerful tool for predicting ICU interventions.
• This approach provides robust patient state representations, enabling more evidence-driven clinical decision-making.
• The learned states have broader applicability across various interventions compared to custom classifiers targeting specific events.