JMIR Publications

ABSTRACT

Background:

Surgical Site Infection (SSI) is one of the most common healthcare-associated infections, comprising nearly 20% of all Healthcare-Associated Infections (HAI) in hospitalized patients. SSIs contribute to increased hospital length of stay, readmission, and healthcare costs, with an associated mortality rate twice that of non-infected patients.

Objective:

To develop machine learning models to predict potential SSI cases following colon surgeries, thereby improving the efficiency of SSI surveillance and enabling healthcare professionals to better prioritize patient care.

Methods:

Data was extracted from the EPIC Electronic Health Record system at a single academic center focusing on colon surgery patients between 2018 and 2023. The dataset included structured features such as demographics, medications, and vital signs, as well as unstructured clinical notes processed using natural language processing techniques. Logistic Regression, Random Forest, and XGBoost models were developed and trained to predict SSI risk. Cost-sensitive learning and Synthetic Minority Over-sampling Technique (SMOTE) were applied to handle the imbalanced dataset. Model evaluation was conducted using performance metrics including accuracy, precision, recall, and Area Under the Receiver Operating Characteristic curve (AUC-ROC).

Results:

From a cohort of 1508 patients, 66 (4.4%) developed SSIs. The XGBoost model demonstrated the best overall performance with precision of 50%, recall of 38%, and an AUC-ROC score of 0.788. The Random Forest model achieved the highest precision (100%) but with lower recall (23%), while Logistic Regression showed higher recall (46%) but lower precision (10%). Patients who developed SSIs were significantly older (mean age 61.1 vs. 58.5 years), had higher ASA scores (78.8% vs. 45.3% with ASA score 3), more frequently had contaminated wounds (40.9% vs. 27.5%), and more commonly received steroids (83.3% vs. 58.7%).

Conclusions:

Machine learning has shown promising results in enhancing the efficiency of SSI surveillance by enabling healthcare professionals to prioritize high-risk patients. While the models demonstrated strong overall accuracy, challenges related to imbalanced datasets remain. The XGBoost model provided the best balance between precision and recall, making it potentially the most clinically useful approach for SSI surveillance augmentation.