Revisiting the Endoscopic Third Ventriculostomy Success Score using machine learning: can we do better?

Area of Science:

• Neurosurgery
• Medical Informatics
• Biostatistics

Background:

• The Endoscopic Third Ventriculostomy Success Score (ETVSS) aids in predicting surgical outcomes for hydrocephalus.
• The original logistic regression (LR) model for ETVSS had moderate predictive performance (AUROC 0.68).
• A larger dataset is needed to develop and validate improved predictive models for ETV success.

Purpose of the Study:

• To develop and validate more accurate machine learning (ML) models for predicting endoscopic third ventriculostomy (ETV) success.
• To perform the largest external validation of the ETVSS to date.
• To compare the performance of various ML algorithms against the established ETVSS.

Main Methods:

• Utilized the MarketScan database (2005-2022) to identify pediatric patients (<18 years) undergoing first-time ETV.
• Collected data on ETVSS predictors: age, hydrocephalus etiology, and prior shunt history.
• Applied six ML algorithms (LR, SVM, Random Forest, k-NN, XGBoost, Naive Bayes) and nested cross-validation for model assessment.

Main Results:

• 2047 patients were included; 61.6% had successful ETV.
• The original ETVSS achieved an AUROC of 0.693 on the validation set and 0.661 on the test set.
• New LR and XGBoost models showed comparable performance to the original ETVSS, with AUROCs around 0.67-0.69.

Conclusions:

• This large-scale validation confirms the modest predictive performance of the ETVSS.
• Sophisticated ML algorithms did not substantially enhance prediction accuracy compared to the ETVSS.
• Future improvements require novel and more dimensional input data, not just advanced modeling techniques.