Advanced Risk Models to Predict Surgical Site Infections Using Machine Learning Approaches and Risk-Adjusted Control Charts

Date of Award

12-2021

Degree Name

Doctor of Philosophy

Department

Industrial and Entrepreneurial Engineering and Engineering Management

First Advisor

Dr. Lee Wells

Second Advisor

Dr. Steven Butt

Third Advisor

Dr. Tycho Fredericks

Fourth Advisor

Dr. Ikhlas Abdel-Qader

Keywords

Surgical site infection, quality control, risk-adjusted control charts, machine learning, deep neural network

Abstract

Even though most infections are treatable with antibiotics, Surgical Site Infections (SSIs) remain a significant healthcare problem as they increase hospital stays, require additional treatments, reduce a patient’s quality of life, and increase morbidity and mortality. To reduce the risks and to encourage the highest quality healthcare, it is important to measure, predict, and monitor the risk of SSIs. Preoperative prediction of SSIs is essential to facilitate the implementation of preventative strategies for high-risk patients. The overarching goal of this dissertation is to investigate new opportunities of using machine-learning (ML) to improve the accuracy and use of pre-operative prediction models for SSIs. This dissertation consists of three research efforts aimed at accomplishing this goal. Using the American College of Surgeons National Surgical Quality Improvement Program (ACS NSQIP) database, the first research effort focused on identifying the best prediction models for SSIs, considering multiple state-of-the-art machine learning (ML)approaches. The results indicate that the deep neural network (DNN) approach performed the best, compared to the other machine learning approaches considered. The second research effort expanded upon the first effort and focused on identifying the best prediction models for specific surgery types (e.g., general, and vascular). The accuracy of these models was also compared to the model developed in the first research effort, which was developed considering any surgery type. The results indicated that the DNN model that considers any surgery type outperformed the models that were developed for specific surgery types. The third research effort investigated the performance gains of the superior prediction model developed in the first effort when implemented in a risk-adjusted control chart. The results of this effort show that the DNN prediction model allows for the quicker detection of shifts in SSI rates, as compared to the more traditional logistic regression-based risk-adjusted control charts.

Comments

Fifth advisor: Dr. Robert Sawyer

Access Setting

Dissertation-Open Access

This document is currently not available here.

Share

COinS