Predicting Clinical Deterioration in Mechanically Ventilated Children using High-Frequency Physiologic Data - PROJECT SUMMARY
Significance: Clinical deterioration in the PICU, as defined by new or progressive organ dysfunction, occurs in
up to 5% of critically ill children during their first week of admission. Importantly, patients with progressive
organ dysfunction have a mortality rate as high as 50%, with over 20% of survivors experiencing moderate to
severe disability. Earlier recognition of patients at high risk of clinical deterioration is critically needed to inform
treatment decisions and personalize health care delivery, which would lead to improved patient outcomes.
However, timely prediction is challenging due to limitations in existing models. The few models that try to
predict clinical deterioration do not leverage metrics from high-frequency physiological monitor data, such as
heart rate variability (HRV), despite evidence that these data are highly predictive of deterioration. For the
proposed K23 award, I will examine clinical deterioration in the context of new or worsening cardiovascular
and/or respiratory (cardiorespiratory) dysfunction in mechanically ventilated children. My central hypothesis is
that a longitudinal, multivariable prediction model of clinical deterioration that incorporates high-frequency
physiological monitor data and clinical variables will accurately identify children at risk of new or worsening
cardiorespiratory dysfunction and that this model will be useful and acceptable to clinicians.
Aim 1. Derive, validate, and compare longitudinal prediction models of new or worsening
cardiorespiratory dysfunction in critically ill children requiring mechanical ventilation.
Aim 2. Design for the implementation of a clinical deterioration model to improve use in practice.
Innovation: Our innovative approach includes (1) high-frequency physiological data, (2) advanced machine
learning applications, and (3) a user-centered design approach to optimize the implementation plan.
Approach: Aim 1: I will derive and compare machine learning-based prediction models using physiological
and EHR-based clinical data in a single center. I will then perform internal and external validation of the best
performing model. Aim 2: Using user-centered design principles with critical care clinicians, I will elicit the ideal
packaging and implementation plan for incorporating a clinical deterioration model into practice. Then, I will
lead iterative usability testing of the design and implementation plan to assess acceptability, usefulness, and
desirability. This approach will identify the ideal packaging and implementation plan for the clinicians who will
ultimately use the model.
Anticipated Outcomes and Public Heath Relevance: This study will provide the foundation to: (1) develop
accurate prediction models of deterioration that leverage continuous physiological measures and clinical
variables; (2) implement useful prediction models at the bedside to support clinical decision making; and (3)
design intervention trials to test the impact on patient outcomes.
