Identifying surgical site infections in electronic health data using predictive models

Area of Science:

• Medical Informatics
• Surgical Infections
• Predictive Analytics

Background:

• Surgical site infections (SSI) are a significant concern following ambulatory surgery.
• Effective surveillance methods are needed to identify patients requiring further investigation.
• Electronic Health Record (EHR) data offers a potential resource for SSI detection.

Purpose of the Study:

• To prospectively derive and validate a prediction rule for detecting surgical site infections (SSI) after ambulatory surgery.
• To assess the utility of EHR data, including free text, in developing an accurate SSI prediction model.
• To achieve a sensitivity exceeding 80% and a positive predictive value (PPV) above 10%.

Main Methods:

• Analysis of EHR data from 7910 ambulatory surgeries across 4 facilities.
• Development of prediction rules using regularized logistic regression and random forests on a derivation dataset (30 months).
• Validation of the prediction rule on a subsequent dataset (10 months), evaluating models with and without free text data.

Main Results:

• An optimal random forest prediction rule incorporating free text data demonstrated a sensitivity of 0.9 and a PPV of 0.28.
• Antibiotic prescription within 60 days showed improved sensitivity (0.84) and PPV (0.28) in the validation set.
• 234 SSIs were identified among 7910 surgeries analyzed.

Conclusions:

• EHR data, particularly when including free text, can effectively support SSI surveillance.
• The derived prediction rule shows promise for identifying children at risk of SSI after ambulatory surgery.
• This approach can enhance early detection and management of SSIs, improving patient outcomes.