E-ISSN: 1019-5157
ISSN: 2651-5024
Research
Prognostic Factors in Sphenoid Wing Meningiomas: Retrospective Analysis With Machine Learning and Comprehensive Literature Review in a Single Center Series.
Amanda Lima Leite✉ ,
Ricardo Santos de Oliveira ,
Rodrigo Inácio Pongeluppi ,
Leandro Lima Leite ,
Júlia de Paula Gonçalves ,
Matheus Ballestero ,
Benedicto Oscar Colli
Division of Neurosurgery, Department of Surgery and Anatomy, Universidade de São Paulo
DOI: 10.5137/1019-5149.JTN.49833-26.3
Article in Press
Corresponding Author:
Amanda Lima Leite (amandalimaleite@usp.br)
Abstract
Aim
The surgical complexity of sphenoid wing meningiomas (SWMs) leads to variable and difficult-to-predict postoperative outcomes. This study aimed to identify prognostic factors by developing a machine learning (ML) model on a single-center cohort and to contextualize the findings through a parallel comprehensive literature review, thereby testing the feasibility of this dual approach for prognostic discovery.
Material and Methods
We conducted a retrospective analysis of 86 consecutive patients surgically treated for SWMs. An XGBoost model was trained to predict unfavorable outcomes (functional decline), with feature importance interpreted using SHAP (SHapley Additive exPlanations) values. A review was performed following search strategy principles of the PRISMA guidelines to benchmark our cohort against the global literature.
Results
The XGBoost model achieved a predictive accuracy of 61.1%. SHAP analysis identified pre-operative visual loss (mean SHAP = 0.145) and internal carotid artery involvement (mean SHAP = 0.069) as the most significant predictors of unfavorable outcomes. Conversely, a longer symptom duration was strongly associated with a favorable prognosis (mean SHAP = -0.234). Our cohort (n=86) showed demographic consistency with the literature review data (n=3,176), but notable differences in surgical metrics and visual outcomes were observed.
Conclusion
The machine learning analysis identified pre-operative visual loss and internal carotid artery involvement as robust predictors of functional decline, while unexpectedly revealing that longer symptom duration is associated with a more favorable prognosis. Although the predictive accuracy was constrained by the disease\'s inherent heterogeneity, the parallel literature review confirmed the external validity of our cohort, demonstrating that these local findings align with global epidemiological patterns. Ultimately, this study serves as a methodological blueprint for using machine learning as a hypothesis-generating engine to extract actionable insights from single-center neurosurgical datasets.
The surgical complexity of sphenoid wing meningiomas (SWMs) leads to variable and difficult-to-predict postoperative outcomes. This study aimed to identify prognostic factors by developing a machine learning (ML) model on a single-center cohort and to contextualize the findings through a parallel comprehensive literature review, thereby testing the feasibility of this dual approach for prognostic discovery.
Material and Methods
We conducted a retrospective analysis of 86 consecutive patients surgically treated for SWMs. An XGBoost model was trained to predict unfavorable outcomes (functional decline), with feature importance interpreted using SHAP (SHapley Additive exPlanations) values. A review was performed following search strategy principles of the PRISMA guidelines to benchmark our cohort against the global literature.
Results
The XGBoost model achieved a predictive accuracy of 61.1%. SHAP analysis identified pre-operative visual loss (mean SHAP = 0.145) and internal carotid artery involvement (mean SHAP = 0.069) as the most significant predictors of unfavorable outcomes. Conversely, a longer symptom duration was strongly associated with a favorable prognosis (mean SHAP = -0.234). Our cohort (n=86) showed demographic consistency with the literature review data (n=3,176), but notable differences in surgical metrics and visual outcomes were observed.
Conclusion
The machine learning analysis identified pre-operative visual loss and internal carotid artery involvement as robust predictors of functional decline, while unexpectedly revealing that longer symptom duration is associated with a more favorable prognosis. Although the predictive accuracy was constrained by the disease\'s inherent heterogeneity, the parallel literature review confirmed the external validity of our cohort, demonstrating that these local findings align with global epidemiological patterns. Ultimately, this study serves as a methodological blueprint for using machine learning as a hypothesis-generating engine to extract actionable insights from single-center neurosurgical datasets.
Keywords
Meningioma
surgical outcomes
machine learning
literature review
prognosis.