PROBLEM STATEMENT: After undergoing a Fontan open-heart surgery to redirect blood flow directly into the lungs, many pediatric patients are readmitted shortly after for the same reasons. The cause for readmissions is unknown, so there is a need to predict the children that will be readmitted to the hospital. If effective, doctors can pursue a different course of action for such patients.

METHODOLOGY: In this study, we incorporated electronic health records (EHRs), such as vital signs, administered medications, and laboratory results, to determine whether a patient will require readmission after undergoing Fontan surgery. The EHR data were extracted for 338 patients admitted to the Children’s Healthcare of Atlanta from May 2009 to August 2020 for Fontan surgery. 38.8% of the patients, or 131 patients, were readmitted after the surgery for the same conditions. The 8 vitals, 104 medications, and 60 laboratory results were taken from the patient’s admission time until the time of discharge. Machine learning models, such as random forest and decision tree, were implemented to make these predictions. A condensed list of medications, split into 14 categories, was used to reduce the number of inputted features. Missing numerical values were imputed with the associated mean value. Binary and categorical features were one-hot-encoded.

RESULTS: 80% of the EHR data went towards training, while the remaining 20% went towards testing. All of the results have been reported for the test data. The random forest classifier had a 92.5% accuracy [85.1%, 98.5%], 85.2% sensitivity [74.1%, 96.3%], 97.5% specificity [90%, 100%], 95.5% PPV [84%, 100%], 90.7% NPV [84.1%, 97.6%], 86.5% AUPRC [75.6%, 96.4%], 91.3% AUROC [83.9%, 98.1%], and 90.2% F1 score [80.8%, 98.1%]. The decision tree classifier had a 91.0% accuracy [83.6%, 97.0%], 88.9% sensitivity [77.8%, 100%], 92.5% specificity [82.5%, 100%], 88.9% PPV [77.8%, 100%], 92.5% NPV [85.4%, 100%], 83.5% AUPRC [71.5%, 94.2%], 90.7% AUROC [83.2%, 96.9%], and 88.9% F1 score [80.0%, 96.3%].

DISCUSSION: Overall, the random forest classifier slightly outperformed the decision tree classifier. Thus, a random forest classifier can be leveraged by pediatricians to suggest an alternative course of action for patients that may not benefit from Fontan surgery. Feature importance, shown in Figure 1, was used to conclude that the 8 vitals (pulse, central venous pressure, blood pressure, oxygen saturation, respiratory rate, arterial blood pressure, mean arterial pressure, and body temperature) were most important in the random forest classifier. The next steps are to include demographic information and try out additional models, such as XG Boost. Feature selection will be used to filter down the number of total features. A binary indicator will be used on the labs to flag the imputed values.

Figure 1: Top 30 Most Important Features