Predictive Models for Pediatric Cardiac Surgery Outcomes

Abstract

This project outlines several iterative attempts to create a predictive model for postoperative outcomes subsequent to Pediatric Cardiac surgery. My analysis is based on a dataset comprised of several lab and observational metrics from the operative hospital stays of 1016 cardiac surgery patients at Boston Children's Hospital. Outlined first is our work to preoperatively predict two intraoperative indicators, the sum of all weight normalized intraoperative transfusions (Category 1), as well as the Thromboelastography reading of Maximum Amplitude at rewarming, which is near the end of surgery. Using a more comprehensive data frame of pre- and intraoperative variables, I then fine tuned models to predict the binary outcome variables Bleeding and Thrombosis. Accurate predictions on the test set were difficult to procure because of the low frequency of these complications within the population. Various regression models, stratification techniques, data augmentation mechanisms, and validation procedures were attempted to bolster the effectiveness of the model. Many of these techniques proved useful in helping the model identify the most likely indicators in the dataset for these complications. Finally, we discussed a set of clinically relevant potential continuous surrogates for these outcomes based on their proven relation to post operative morbidity and complications and their relevance over numerous readings within time stratified blocks. Initial predictions on these features within the first twenty-four hours were made, with an acknowledgement of their potential power within a pediatric cardiac surgery predictor. Overall, this work accurately identifies key features that can indicate postoperative complications, and works to apply these findings towards strong predictions. The methodology described, especially if bolstered by additional data in the future, outlines a hopeful process for contributing information towards physicians' decision making processes.