Using the Electronic Health Record To Identify Children Likely To Suffer Last-Minute Surgery Cancellation

Surgical procedures can provide dramatic health benefits or important diagnostic information. While late surgery cancellation is infrequent, these cancelled cases are a leading source of perioperative wastage. They are a significant source of unutilized healthcare resources, including high-cost operative staff and facilities. Surgical cancellation can also negatively impact family members, causing distress, frustration, and an overall negative experience.

To address this, the research team developed a predictive model of last-minute surgery cancellations using machine learning technologies. Machine learning uses computerized algorithms to identify patterns among a dataset that can be used to predict an outcome. The model implemented patient-specific and contextual data from two pediatric surgical sites at Cincinnati Children’s Hospital Medical Center (CCHMC) to observe key predictors affecting last-minute surgery cancellations among young patients.

The specific aims of this research were as follows:

• Develop computerized models for predicting surgery cancellation. 
• Identify key predictors of last-minute cancellation of surgery from the electronic health record (EHR) and online data resources. 
• Establish a scalable, last-minute surgery cancellation prediction system. 

A 5-year dataset from 2012-2017 was extracted from the CCHMC EHR to observe the surgical activity. The data used included patient demographics, recent healthcare activity, insurance information, cancellation patterns, infection risk based on environmental factors, and weather conditions. Machine learning classifiers were formed to predict patient-related cancellations. The most frequent cancellation causes were identified as patient illness, “no show,” nothing-by-mouth (NPO) violations (i.e., no eating or drinking), and surgical refusal by patient or family. The model performance was measured based on accuracy, precision, recall, specificity, negative predictive value, and measure of balance between recall and specificity.

The findings found that machine learning to predict surgery cancellation is feasible, particularly with prediction of “no show” and NPO violations. The model uncovered useful insight into root causes of surgery cancellation, as indicated by the machine learning classifiers. Using the findings, the research team integrated the predictive models into the CCHMC EHR system as a quality-improvement measure to identify patients of highest risk of cancellation. Early identification of last-minute surgery cancellations, with timely intervention, can provide new opportunities for reducing healthcare costs and improving medical efficiency.