Improving Early Prediction of Abnormal Recovery after Appendectomy in Children using Real-world Data from Wearables

Postoperative complications are primary concerns during early recovery from pediatric appendectomy. Identifying complications or symptoms of abnormal recovery typically relies on intermittent and subjective assessments from children and their caregivers, which may result in delayed diagnosis. Wearable devices can capture continuous and objective health measurements, which can be mined for early biomarkers of complications or symptoms using machine learning models. However, real-world datasets from wearables often have missing and imbalanced data, which can affect model performance and utility. We have recorded real-world Fitbit data from 93 children during the first 21 days following appendectomy for complicated appendicitis. This dataset included missing data (37.0% across all participants) and imbalanced data (2.7% of total days recorded from children exhibiting abnormal recovery). Aiming to improve early prediction of abnormal recovery, we extracted 143 daily features from the data, including Fitbit metrics of activity, heart rate, and sleep, as well as metrics derived from clinical knowledge. We trained a Balanced Random Forest classifier and tested different early prediction strategies for identifying abnormal recovery (complications or abnormal symptoms) 1-3 days before they were clinically diagnosed. The best-performing model predicted abnormal recovery three days before diagnosis at an accuracy of 87.5%, two days before at 76.4%, one day before at 85.7%, and on the day of diagnosis at 78.8%. The overall prediction accuracy was improved 10.1% compared to a previous study. With further development, this approach could be used to generate near real-time alerts of abnormal postoperative recovery to enhance pediatric care and clinical decision making.