Clinical Decision Support for Early Detection of Deterioration in Hospitalized Children

PROJECT SUMMARY Hospitalized children who experience cardiopulmonary deterioration are at increased risk for mortality and long- term morbidity. Because timely intervention increases survival in children, it is critically important to identify cardiopulmonary deterioration events as early as possible. However, the current paradigm for detecting these events in advance has several gaps. First, existing risk prediction methods can lead to fragmented care, as each unit employs different tools for predicting specific outcomes. For example, risk prediction within the emergency department (ED) is targeted toward triage; ward-based tools predict the risk of being transferred to the intensive care unit (ICU), while the ICU focuses on determining the likelihood of death or cardiac arrests. Transitioning from multiple, siloed risk assessment tools to a single, hospital-wide cardiopulmonary deterioration prediction model could significantly improve outcomes for children. A second critical gap is that current prediction model outputs are not accompanied by helpful explanations, a need unmet by standard machine learning (ML) explainers due to inherent limitations. Developing new algorithms that provide real-time interpretations of model outputs may increase situational awareness, decrease diagnostic delay, and enable better treatment selection. Third, to ensure high usage and effective decision-making, any new model should be accompanied by a user interface explicitly designed using human factors engineering principles. The long-term goal is to improve outcomes among children experiencing cardiopulmonary deterioration by enabling better quality of care. The overall objective of this project is to develop a new clinical decision support (CDS) tool that is accurate, interpretable, and actionable for early detection of cardiopulmonary deterioration events in children. In Aim 1, we will use electronic health record (EHR) data from pediatric admissions to four academic hospitals and ML to derive and externally validate a new hospital-wide cardiopulmonary deterioration prediction model and compare performances to our preliminary model. In Aim 2, we will develop novel algorithms that provide physiological explanations and clinical context for model predictions for a given patient. Finally, in Aim 3, we will create a new CDS tool that embeds the best-performing prediction tool and explainer algorithm outputs within a graphical user interface purposefully designed to facilitate increased user interaction. The proposed research is innovative because it incorporates deep learning-based pediatric risk prediction, real-time explainable algorithms with highlighted clinical context, and human factors engineering for developing the CDS tool. In addition, the proposed work is significant because it will result in a new, accurate, interpretable, efficient, and user-friendly CDS tool for risk assessment throughout a pediatric hospital. Ultimately, this powerful tool will enable early recognition of pediatric cardiopulmonary deterioration events, facilitating timely diagnosis and intervention to improve outcomes among hospitalized children.