Dynamic Risk Profiling in Patients with Cardiac Inflammatory Syndrome

Team: Lilac

• Program: Biomedical Engineering
• Course: Precision Care Medicine

Project Poster

Open full size poster in new tab (PDF)

Project Post Summary:

Cardiac inflammatory syndrome (CIS) is a life-threatening condition that is difficult to diagnose and can cause long-term consequences, particularly in children. Fortunately, early ICU admission has shown to significantly increase the survival rates for the patients. Using clinical data collected by the International Kawasaki Disease Registry Consortium (IKDR), we developed machine learning models that can predict ICU admission early to improve clinical outcomes. Previously-developed models typically follow a “snapshot” paradigm, using simple time-point data to predict ICU admission. We developed both Snapshot models and Window models, utilizing engineered features from time-series data, to predict ICU admissions within 48
hours of evaluation. Several classification algorithms, including Random Forest and XGBoost, were evaluated. The resulting models accurately predicted ICU admission within 48 hours for at-risk pediatric patients with CISs, with Window models incorporating time-series engineered features yielding better performance. Evaluation of feature importance revealed clinical symptoms that were highly predictive of deteriorating outcomes in patients with cardiac inflammatory syndromes.