Text

theses

Methicillin-resistant Staphylococcus aureus (MRSA) is the leading cause of antibiotic resistance related mortality in surgical patients. Effective prediction of MRSA and MRSA-related SSI would facilitate the prophylactic use of appropriate antibiotics or application of other prevention techniques, which have been shown to improve clinical outcomes. While there is a range of factors that have been shown to increase the risk of MRSA-related infections, research is less clear on the best approaches to developing predictive models for incorporation into a clinical decision support system. This study compared two common modeling approaches — logistic regression (LR) and artificial neural networks (ANN) — for the prediction of MRSA infection and MRSA-related SSI in patients undergoing major surgical procedures (MSPs) in the United States.

The data source for analysis is the National Inpatient Sample, which contains approximately 7 million discharges each year. A descriptive analysis was performed to identify potential predictors for each of three research hypotheses and ANN and LR models were developed and evaluated for the prediction of: (1) MRSA infection in patients undergoing MSPs; (2) MRSA-related SSI in patients undergoing MSPs; and (3) MRSA-related SSI in patients with S. aureus infection.

The ANN model performed best for Hypothesis 1, with an AUC of 0.87, sensitivity of 0.86 and specificity of 0.74; the LR model achieved an AUC of 0.85, sensitivity of 0.79 and specificity of 0.75. For Hypothesis 2, the ANN model achieved an AUC of 0.86, sensitivity of 0.73 and specificity of 0.87; the LR model achieved an AUC of 0.85, sensitivity of 0.77 and specificity of 0.76. For Hypothesis 3, the ANN model achieved an AUC of 0.67, sensitivity of 0.57 and specificity of 0.67; the LR model achieved an AUC of 0.68, sensitivity of 0.61 and specificity of 0.64.

This study assessed the feasibility of LR and ANN for the prediction of MRSA-related infections in surgical patients using a range of demographic, clinical, procedural, and hospital-related factors. The results showed that both algorithms are effective modeling approaches with reasonable sensitivity and specificity and suggest that a clinical decision support tool based on either model could be informative in clinical practice. ?

ETD_10081

Ph.D.

Includes bibliographical references

doi:10.7282/t3-gmn8-t474

ETD doctoral

The author owns the copyright to this work.