Development and Application of Machine Learning Classifiers and Explainable Artificial Intelligence for Predicting Long-Term Facial Nerve Function after Vestibular Schwannoma Surgery

Introduction: Vestibular schwannomas (VSs) represent the most prevalent tumors in the cerebellopontine angle. While microsurgery remains the sole curative approach, it carries inherent risks, including neurological deficits, particularly facial nerve paresis.

In this study, we harnessed the power of machine learning, specifically the Extreme Gradient Boosting (XGBoost) classifier, to perform binary classification of long-term facial nerve outcomes after VS surgery, distinguishing between satisfactory and bad outcomes (House-Brackmann grades 1–3 and 4–6, respectively). Additionally, we employed two methods for model interpretability, namely, Local Interpretable Model-Agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP).

Materials and Methods: We conducted a retrospective analysis of 345 patients who underwent sporadic VS surgery between 1990 and 2018. This group consisted of 133 men (38.5%) and 212 women (61.5%), with a mean age of 47 (range: 18–76). The average tumor size was 29.3 mm (range: 8–72), and the mean tumor volume was 13.6 cm³ (range: 0.2–113). The retrosigmoid approach was predominantly employed (339 cases, 98.3%).

To optimize the dataset, we removed highly correlated and redundant variables, resulting in a dataset comprising a binary target variable (satisfactory vs. bad outcomes, defined as House-Brackmann grades 1–3 and 4–6, respectively); 37 various pre-, intra-, and postoperative predictors (features); and 297 patient records.

We divided the data into training (80%) and testing (20%) sets and employed the XGBoost package in Python for the classification task. Model performance metrics included the area under the curve (AUC), receiver operating characteristic (ROC), classification accuracy (CA), F1 score, precision, recall, and the Matthews correlation coefficient (MCC). Model explanations were provided using SHAP and LIME.

Results: The majority of patients achieved satisfactory outcomes (n = 230, 77%), leading to an imbalanced target distribution. Notably postoperative eye closure (tau = −0.48) exhibited correlation with long-term facial nerve function. The XGBoost model demonstrated an accuracy of 0.83, a ROC AUC score of 0.72, and an MCC score of 0.60.

SHAP values facilitated both global and local explanations. Complete eye closure was strongly associated with positive outcomes, while high tumor volume and increasing patient age during surgery were linked to adverse outcomes.

Local interpretations using LIME reinforced the importance of postoperative eyelid closure. Additionally, LIME highlighted tumor size, with values less than 17.53 mm³ being indicative of good outcomes.

Conclusions: Our study introduces an effective machine learning model for classifying long-term facial nerve outcomes following vestibular schwannoma surgery. This model holds promise for clinical applications, enabling the evaluation of patients and providing treatment recommendations. Postoperative eye closure is a key consistent risk factor for predicting long-term facial nerve function.

Publikationsverlauf

Artikel online veröffentlicht:
05. Februar 2024