Search

Background:Implementation of the infant single ventricle monitoring program (ISVMP) after Stage one palliation, has led to significant decreases in morbidity and mortality at our institution and other US institutions. The aim of this study is to apply machine learning algorithms to predict deterioration in circulatory physiology during home monitoring of SV infants. We hypothesize that na?ve Bayesian (NB) models can predict unplanned hospitalization based on home monitoring data.Methods:We collected data from 312 SV infants enrolled in ISMVP over 8 years from 2011 to 2019 on a weekly basis. Objective data included failure to gain weight of 20 g/day in 3 days or weight loss of >30 g in one day, decreased O2saturations <75% and heart rates. In addition, ?red flag? symptoms of poor oral intake, irritability, vomiting, diarrhea, fever (>100.4?), sleepiness and increased work of breathing were also included. We developed NB models to predict an event (E), defined as unplanned hospitalizations within 2 weeks from the last outpatient follow-up. We used two-week windows before E to run the analysis (Fig. 1). Uneventful (UE) was defined as follow-up without any hospitalization. Predictive performance was evaluated by area under the Receiver Operating Characteristic curve (AUC), specificity, sensitivity and positive predictive value (PPV) .Results:We identified 665 UE windows and 194 E windows. We performed 10-fold cross validation to test NB models. Average AUC was 0.67 (95% confidence interval: 0.53-0.81, specificity 80%, sensitivity 45% and PPV 46%.Conclusions:Na?ve Bayesian models can learn from sparse home surveillance data obtained by ISVMP and can predict critical events. NB models require fewer parameters and can perform well in presence of missing and categorical data. This model is an appealing tool for outpatient ISVMP program, where data collection is performed by the family and can be limited. Further refinements will be added to this pilot study in the future.