Automated Detection Of Surgical Adverse Events From Retrospective Clinical Data

Abstract

The Detection of surgical adverse events has become increasingly important with the growing demand for quality improvement and public health surveillance with surgery. Event reporting is one of the key steps in determining the impact of postoperative complications from a variety of perspectives and is an integral component of improving transparency around surgical care and ultimately around addressing complications. Manual chart review is the most commonly used method in identification of adverse events. Though the manual chart review is the most commonly used method that is considered the “gold-standard” for detecting adverse events for many patient safety studies (research setting), it could be very labor-intensive and time-consuming and thus many hospitals have found it too expensive to routinely use. In this dissertation, aiming to accelerate the process of extracting postoperative outcomes from medical charts, an automated postoperative adverse events detection application has been developed by using structured electronic health record (EHR) data and unstructured clinical notes. First, pilot studies are conducted to test the feasibility by using only completed EHR data and focusing on three types of surgical site infection (SSI). The built models have high specificity as well as very high negative predictive values, reliably eliminating the vast majority of patients without SSI, thereby significantly reducing the chart reviewers’ burden. Practical missing data treatments have also been explored and compared. To address modeling challenges, such as high-dimensional dataset, and imbalanced distribution, several machine learning methods haven been applied. Particularly, one single-task and five multi-task learning methods are developed and compared for their detection performance. The models demonstrated high detection performance, which ensures the feasibility of accelerating the manual process of extracting postoperative outcomes from medical chart. Finally, the use of structured EHR data, clinical notes and the combination of these data types have been separately investigated. Models using different types of data were compared on their detection performance. Models developed with very high AUC score have demonstrated that supervised machine learning methods can be effective for automated detection of surgical adverse events.