EARLY IDENTIFICATION OF ADVERSE EVENTS DURING THE PERIOPERATIVE PERIOD IN PEDIATRIC POPULATION

Abstract

Paper 1: ABSTRACT BACKGROUND: Early identification of patients at high risk for severe acute pain could enable specific pain management strategies designed to alter pain trajectories after surgery. With this study we applied machine learning techniques to electronic health records (EHR) and a patient questionnaire with the goal of creating time-series algorithms to predict severe acute pain after spinal fusion surgery in patients with adolescent idiopathic scoliosis (AIS). METHODS: Perioperative data were obtained from two databases at Boston Children’s Hospital (BCH) between 2011 and 2019. The primary outcome of interest was severe acute pain in the first 3 days after spinal fusion surgery. K-means clustering was used to create patient cohort cluster assignment. We identified patients with severe acute postoperative pain through the combination of high opioid consumption and high pain scores. To predict patients who would experience severe pain, machine learning algorithms and logistic regression models were developed using preoperative predictors only (Phase 1), preoperative and intraoperative predictors (Phase 2), and preoperative, intraoperative and postoperative variables (Phase 3). Model performance was evaluated using area under the receiver operating characteristic curve (AUROC) and accuracy. RESULTS: Among 643 cases analyzed, 79 (12.3%) patients were identified as experiencing severe acute pain after spinal fusion surgery. Gradient boosting consistently outperformed all other models across all 3 phases. The AUROC for gradient boosting were 0.71 (95% confidence interval (CI), 0.59 – 0.82) in Phase 1, 0.74 (95% CI, 0.61 – 0.84) in Phase 2, and 0.76 (95% CI, 0.64 – 0.87) in Phase 3. The corresponding accuracy were 71.88% (95% CI, 64.95% – 78.11%), 69.79% (95% CI, 62.77% – 76.19%), and 76.56% (95% CI, 69.92% – 82.36%), respectively. CONCLUSIONS: Gradient boosting model had better discrimination and calibration ability to predict severe acute pain after spinal fusion surgery in patients with AIS when analyzing high dimensional perioperative EHR data and patient questionnaire compared with logistic regression model. Future studies are warranted to externally validate the algorithms and apply them to real-world clinical work as a potential aid in identifying the high-risk patients and customizing their care. Paper 2: ABSTRACT BACKGROUND: Delivery of pediatric sedation outside the operating room (OR) is a potentially high-risk practice. Given the high volume and demand over the past two decades, sedation is delivered by a wide range of providers. Despite the use of advanced monitoring techniques and the adherence to the practice guidelines for pediatric procedural sedation, the incidence of severe adverse events (SAE) during pediatric sedation outside the OR has slightly increased in recent years. There is a great need to develop a triage system with the purpose of helping sedation services with facilitate risk stratification and allocate healthcare resources. METHODS: Data were obtained from the PSRC database from November 10, 2011 to December 31, 2019. The primary outcome of interest was airway obstruction (complete and partial) during the sedation outside the OR. Four modeling approaches, including logistic regression model, Least Absolute Shrinkage and Selection Operator (Lasso), elastic net, and gradient boosting, were developed to predict airway obstruction using 195 variables that were routinely collected. Model performance was evaluated using area under the receiver operating characteristic curve (AUROC), accuracy and decision curve analysis on the test set and external validation dataset. RESULTS: Of 288,510 eligible sedation cases for model development, 10994 (3.8%) had recorded airway obstruction as an outcome. Four algorithms were developed on the training set. Gradient boosting had the best discrimination and calibration ability on the test set and the temporal validation dataset with AUROC 0.796 (95% CI, 0.789 – 0.803), PPV 0.095 (95% CI, 0.089 – 0.098), accuracy 72.94% (95% CI, 72.64% – 73.23%), and with AUROC 0.765 (95% CI, 0.755 – 0.776), PPV 0.097 (95% CI, 0.088 – 0.109), accuracy 71.14% (95% CI, 70.72% – 71.55%), respectively. Decision curve analysis showed gradient boosting had the greatest net benefit over the range of threshold probabilities. On the geographic validation dataset, gradient boosting had the lowest net benefit. CONCLUSIONS: Gradient boosting model of Machine Learning achieved better predictive performance on the test set and temporal validation dataset, but not on the geographic validation dataset, indicating that broader generalizability is needed before we apply the forecasting algorithm to the triage system.