Construction and validation of a machine learning prediction model for facial nerve functional outcomes following resection of Koos grade 3～4 vestibular schwannomas
KUAI Guohu
LI Yandong
ZHANG Gaocai
AIHEMAITI Aierken
ZHU Guohua
MAIMAITILI Mijiti
GENG Dangmurenjiafu
Abstract：Objective To construct a prediction model for facial nerve function outcomes after resection of Koos grade 3～4 vestibular schwannomas using machine learning,and to evaluate its efficacy.Methods The clinical data from 55 patients with Koos grade 3～4 vestibular schwannomas who underwent surgical treatment in the Department of Neurosurgery at our institution between August 2020 and May 2022 were retrospectively analyzed.Facial nerve function was assessed 6 months postoperatively using the House-Brackmann(H-B)grading system,with H-B gradesⅠ～Ⅱ defined as favorable prognosis and grades Ⅲ～Ⅵ as unfavorable prognosis.Feature variables were initially screened using the Random Forest(RF)algorithm.Combined with results from multivariate logistic regression analysis,the final variables were incorporated into the prediction model.The subjects were randomly divided into a training set and a testing set in a 7:3 ratio.Five machine learning methods were employed for modeling:Support Vector Machine(SVM),Naive Bayes(NB),XGBoost,Gradient Boosting Decision Tree(GBDT),and LightGBM.Results Among the 55 patients,38 had a favorable facial nerve functional prognosis and 17 had an unfavorable prognosis.The prognostic factors screened by the RF algorithm,ranked by feature importance score,were:degree of tumor adhesion to the facial nerve(32.3%),tumor consistency(26.1%),tumor diameter(22.6%),age(11.4%),and preopera-tive facial nerve function grade(7.7%).Multivariate logistic regression analysis identified cystic tumor consistency,larger tumor diameter,and severe tumor adhesion to the facial nerve as independent risk factors for unfavorable facial nerve functional out-comes in Koos grade 3～4 vestibular schwannomas(P<0.05).Integrating results from the RF algorithm and multivariate logistic regression,the top four prognostic factors(degree of tumor adhesion to the facial nerve,tumor diameter,tumor consistency,and age)were selected for machine learning model development.In the training set,the XGBoost and GBDT models performed best,both with an area under the ROC curve(AUC)of 1.000.However,in the testing set,their AUCs were 0.764 and 0.786,respec-tively,indicating suboptimal performance and a tendency towards overfitting.In the testing set,the SVM model demonstrated the best predictive performance.To ensure model stability and generalizability,the SVM model was ultimately selected as the final prediction model,as it showed good performance in both the training set(AUC=0.989)and the testing set(AUC=0.971).Conclusion This study demonstrates the value of machine learning in developing predictive models for facial nerve functional outcomes following resection of Koos grade 3～4 vestibular schwannomas,with the SVM model showing the best predictive effi-cacy in this cohort.
Keywords：vestibular schwannomafacial nerve functionmachine learningprediction Model
Publication Date：2025-11-25
Online Publishing Date：2026-01-12(First online date of this platform, not the publication date of the document)
Pages：5( 667-671 )