Early Prediction of Adverse Clinical Events and Optimal Intervention in Intensive Care Units

Adverse clinical events like cardiopulmonary arrest and multiple organ dysfunction are life-threatening events with potentially debilitating health outcomes and are a leading cause for mortality in intensive care unit (ICU) patients. Traditionally, clinical decision making is based on periodic review of physiologic monitoring data by attending clinical staff. Modern hospitals collect an enormous amount of information every minute, but making inferences consistently on a continuous-time basis is impractical for clinicians. Instead, data-driven prediction models can be employed to extract value from this continuous stream of information and provide early warning of adverse events to facilitate timely therapeutic intervention and allow efficient patient monitoring. For patients already in critical condition, patient-specific dynamical system models with feedback control can suggest optimal intervention strategies. This dissertation has four contributions. The first three are the development and validation of prediction models for early recognition of increased risk of (a) cardiac arrest in infants with congenital heart disease, (b) pediatric multiple organ failure, and (c) need for invasive mechanical ventilation due to respiratory failure in pediatric ICU patients using machine learning methods. In all three, we observed that the time-varying risk in individual patients increased several hours before an observed event with a median early warning time of 17, 23, 10 hours respectively. We then investigated the presence of groupings within the positive predictions and found 2-3 distinct clusters where the high-risk group had >90% precision. Additionally, we also demonstrated a novel approach for combining medication history-based features to improve prediction performance. The fourth contribution was the application of system identification to learn the dynamics of oxygen saturation in COVID-19 patients with acute respiratory distress syndrome (ARDS) and implement optimal control strategies for maintaining desired levels of oxygen saturation while minimizing prolonged use of mechanical ventilators to avoid lung injury. In this study, we explored two linear time-invariant and one linear parameter-varying model. We finally showed simulated examples of model predictive control for maintaining 95% oxygen saturation in hypoxemic patients