TL;DR
This paper develops and validates machine learning models to predict the risk of surgical site infections (SSI) after minimally invasive transforaminal lumbar interbody fusion (MIS-TLIF). The Naïve Bayes model demonstrated the best performance, offering a promising tool to help clinicians identify high-risk patients and improve decision-making.
What Problem Does This Paper Address?
Surgical site infections (SSI) are a significant complication following spinal surgery, including minimally invasive transforaminal lumbar interbody fusion (MIS-TLIF). While MIS-TLIF offers advantages over traditional open surgery, SSI can lead to prolonged hospital stays, increased costs, and poorer patient outcomes. Current methods for assessing SSI risk are often insufficient, relying on individual risk factors rather than comprehensive predictive models. This study aims to bridge this gap by developing and validating machine learning (ML) algorithms that can predict SSI risk using readily available patient and surgical data, specifically for MIS-TLIF performed via the Quadrant channel.
Methods
The researchers employed a retrospective study design, collecting data from 705 patients who underwent MIS-TLIF between May 2012 and October 2019.
- Feature Selection: Stepwise logistic regression was used to identify potential predictors of SSI from a dataset including clinical characteristics, surgical parameters, and routine laboratory tests.
- ML Model Development: The identified predictors were then fed into six different supervised ML algorithms:
- k-Nearest Neighbor (KNN)
- Decision Tree (DT)
- Support Vector Machine (SVM)
- Random Forest (RF)
- Multi-Layer Perceptron (MLP)
- Naïve Bayes (NB)
- Validation: A 10-fold cross-validation strategy was implemented during the training process to ensure robust model evaluation and prevent overfitting.
- Performance Metrics: The performance of each ML model was assessed using standard metrics: Area Under the Receiver Operating Characteristic curve (AUC), sensitivity, specificity, and accuracy (ACC).
Datasets
The study utilized a single-center dataset comprising 705 patients who underwent MIS-TLIF. Data was extracted from electronic medical records and included:
- Clinical Variables: Age, Body Mass Index (BMI), smoking status, alcohol consumption, diabetes, COPD, stroke, rheumatic disease, CHF, steroid use, diagnosis type, ASA physical status classification.
- Surgical Variables: Operation time, estimated blood loss (EBL), number of fusion segments.
- Laboratory Tests: Pre-operative HbA1c, albumin, globulin, glycated protein, RBC, WBC, thrombocyte count, PT, APTT, fibrinogen, potassium, and calcium.
No new benchmark datasets were introduced; the study focused on internal validation of ML models using existing patient data.
Key Insights and Ideas
- HbA1c as a Dominant Predictor: Pre-operative glycated hemoglobin A1c (HbA1c) emerged as a particularly strong predictor of SSI. This highlights the critical role of long-term glycemic control in surgical outcomes, likely due to its impact on immune function and vascular health.
- Naïve Bayes Shines: Among the tested ML algorithms, the Naïve Bayes (NB) model consistently outperformed others in predicting SSI risk, suggesting its suitability for this specific clinical prediction task.
- Multivariable Approach is Key: While univariate analysis identified many potential risk factors, the study emphasizes the importance of multivariate logistic regression and ML models to capture complex interactions and provide a more accurate prediction of SSI risk.
- Clinical Utility of ML: The research demonstrates that ML can translate complex clinical and laboratory data into actionable risk scores, potentially aiding surgeons in identifying high-risk patients before surgery.
Key Results
- SSI Incidence: The study found an SSI incidence of 4.68% (33 out of 705 patients).
- Significant Predictors: Stepwise logistic regression identified age, BMI, operation time, EBL, pre-operative HbA1c, and pre-operative albumin as significant predictors of SSI.
- ML Model Performance:
- The six ML models achieved an average AUC between 0.60 and 0.80, and an ACC between 0.80 and 0.95.
- The Naïve Bayes (NB) model stood out, achieving an average AUC of 0.78 and an ACC of 0.90.
- Feature Importance (NB Model): The most important predictors identified by the NB model, in descending order, were: pre-operative HbA1c, EBL, operation time, pre-operative albumin, BMI, and age.
Limitations and Future Work
- Retrospective Design: The study's retrospective nature may introduce selection bias.
- Single-Center Study: The findings are based on data from one institution, which might limit generalizability. External validation with larger, multi-center datasets is crucial.
- Missing Variables: Important variables like MRSA carriage were not available and could influence SSI risk.
- Future Directions: The authors suggest enlarging the sample size and conducting external validation to further confirm the model's reliability and clinical utility. They also emphasize the need for future research to investigate other potential risk factors and refine the ML models.
Summary updated