Improving Surgical Site Infection Prediction Using Machine Learning: Addressing Challenges of Highly Imbalanced Data
Salha Al-Ahmari1,2, Farrukh Nadeem1
1Department of Information Systems, Faculty of Computing and Information Technology, King Abdulaziz University, Jeddah 21589, Saudi Arabia.
Diagnostics (Basel, Switzerland)
|February 26, 2025
Summary
Machine learning models effectively predict surgical site infections (SSIs). Random Forest with SMOTE resampling achieved the highest accuracy, offering a promising tool for clinical risk assessment and improved patient outcomes.
Area of Science:
- Healthcare Informatics
- Medical Machine Learning
- Clinical Data Science
Background:
- Surgical site infections (SSIs) are a major cause of hospital readmissions and increased healthcare costs globally.
- Machine learning (ML) shows promise for predicting SSIs, but class imbalance remains a significant challenge.
- Accurate prediction of SSIs is crucial for mitigating patient harm and optimizing resource allocation.
Purpose of the Study:
- To evaluate and enhance ML model predictive capabilities for SSIs.
- To assess the impact of feature selection, resampling techniques, and hyperparameter optimization on SSI prediction accuracy.
- To identify optimal ML strategies for handling imbalanced datasets in SSI surveillance.
Main Methods:
- Utilized a dataset of 64,793 surgical patients from Saudi Arabian hospitals, with 1632 developing SSIs.
- Tested seven ML algorithms: Decision Tree, Gaussian Naive Bayes, Support Vector Machine, Logistic Regression, Random Forest, Stochastic Gradient Boosting, and K-Nearest Neighbors.
- Employed feature selection, resampling techniques (including SMOTE and IHT), and grid search cross-validation for hyperparameter optimization.
Main Results:
- Random Forest (RF) demonstrated the highest performance with a Matthews Correlation Coefficient (MCC) of 0.72.
- Synthetic Minority Oversampling Technique (SMOTE) generally improved model performance, except for Logistic Regression and Gaussian Naive Bayes.
- Instance Hardness Threshold (IHT) provided a computationally efficient undersampling alternative, albeit with potential performance trade-offs.
Conclusions:
- ML models are effective tools for assessing SSI risk, meriting further clinical investigation.
- Advanced ML techniques and robust validation, including MCC, provide reliable SSI prediction even with imbalanced data.
- Optimized ML approaches can significantly enhance the accuracy and reliability of SSI risk prediction, leading to better patient outcomes.
