Predicting Adverse Events After Surgery

Predicting risk of adverse events (AEs) following surgical procedure is of significant interest, as that may guide in better resource utilization and an improved quality of care. Currently available comorbidity indices are largely inaccurate to predict adverse events other than death, as well as off-the-shelf machine learning models do not typically account for the temporal sequence of events to enable predictive analytics. We propose a study to improve the current techniques for assessing and predicting the risk of adverse events (AEs) associated with multiple chronic conditions by designing machine learning models that account for and incorporate the temporal sequence and timing of conditions. We formalize the task as a binary classification problem. Our technical contributions include devising novel sequence based feature discovery techniques to augment existing supervised classification algorithms, as well as formalizing the classification task as a Markov Chain Model (MCM) that captures the temporal sequence of prior chronic conditions/events. Finally, we design a hybrid or multi-classifier that combines prediction from the aforementioned classification models to finally predict AE. Our experimental results, conducted using the Truven Health MarketScan Research Databases with more than 27 million of claim records on two different surgery types, discover interesting insights that can guide patient-centered decision-making and can direct healthcare teams to adjust techniques and interventions. We also extensively compare the performance of our solutions to appropriate baselines.