Advancing Education, Research, and Quality of Care for the Head and Neck oncology patient.
Introduction: Recurrent laryngeal nerve (RLN) injury is a significant complication of thyroidectomy that can result in vocal fold paralysis, dysphonia, shortness of breath, and impaired quality of life. Accurate prediction of RLN injury risk can guide preoperative planning and surgical decision-making, ultimately improving patient safety and outcomes. This study evaluates the performance of machine learning models—specifically RandomForest and Logistic Regression with SMOTE—to predict RLN injury.
Methods: Data from 46,371 thyroidectomy patients from the 2016-2022 NSQIP database were collected and analyzed. Predictive models were developed using RandomForest and Logistic Regression with Synthetic Minority Over-sampling Technique (SMOTE) to handle class imbalance. Sixteen clinical variables were utilized, including but not limited to primary indication for surgery, prior neck surgery, presence of neoplasm, tumor characteristics, and surgical approach. Data was split into training and testing sets (80% training, 20% testing), and models were evaluated on accuracy, sensitivity, specificity, NPV, PPV, F1 score, and area under the receiver operating characteristic curve (ROC AUC). Hyperparameter tuning was performed for the RandomForest model to achieve optimal performance.
Results: The RandomForest model outperformed Logistic Regression with SMOTE across all metrics. The RandomForest model achieved an accuracy of 83.1%, a PPV of 80.6%, a sensitivity of 87.2%, a specificity of 78.9%, a negative predictive value (NPV) of 86.1%, an F1 score of 83.7%, and an ROC AUC of 92.2%. In comparison, the Logistic Regression model demonstrated an accuracy of 68.6%, a PPV of 68.9%, a sensitivity of 67.7%, a specificity of 69.4%, an NPV of 68.3%, an F1 score of 68.3%, and an ROC AUC of 76.8%.
Discussion: The improvement of all performance metrics with the RandomForest model highlights its potential clinical utility, and its ability to handle data with complex non-linear relationships better than standard methodologies such as a logistic regression. With an ROC AUC of 92.2%, the model demonstrates excellent discrimination between patients who are at risk of RLN injury and those who are not. Additionally, the model's specificity of 78.9% and NPV of 86.1% indicate that it performs well in ruling out low-risk patients, reducing unnecessary interventions.The RandomForest model's superior sensitivity further emphasizes its capability to reduce the likelihood of missing high-risk patients, thereby decreasing postoperative complications and improving patient outcomes.
This is particularly important in a surgical setting, where identification of high-risk patients can lead to targeted preoperative planning, such as adjusting the surgical technique to minimize nerve damage and prioritizing resource allocation before surgery. Additionally, this tool could be helpful in calculating then stratifying the risk profiles individually for each patient, and in turn, aid in pre-operative counseling and guidance.
Conclusion: The RandomForest model presents a robust tool for predicting RLN injury in thyroidectomy patients, significantly outperforming traditional methods like Logistic Regression with SMOTE. Incorporating predictive analytics in thyroid surgery can lead to individualized metrics based on patient characteristics that can be used to decrease surgical morbidity and improve quality of life.