Predicting endoscopic third ventriculostomy with choroid plexus cauterization success with machine learning: the importance of ventricle size and the irrelevance of etiology

OBJECTIVE: Endoscopic third ventriculostomy with choroid plexus cauterization (ETV/CPC) has decreased rates of shunt dependence in infants with hydrocephalus. The ETV Success Score (ETVSS) is the standard for predicting the 6-month success rate for ETV in children based on age, hydrocephalus etiology, and shunt history. However, the ETVSS does not account for the impact of CPC or preoperative ventricular volume. It also relies on the independence of each contributing variable and does not predict success beyond 6 months. In this study, the authors used a machine learning approach to create a tool specifically for ETV/CPC to predict the likelihood of success at 1 year. METHODS: The records of 206 pediatric patients younger than 2 years of age who received ETV/CPC as a primary, definitive treatment for hydrocephalus at a single institution between 2009 and 2021 were reviewed for patient demographics, presenting characteristics, and medical history. Data on corrected age at surgery, frontal occipital horn ratio (FOHR), hydrocephalus etiology, and whether there was prior CSF diversion were used in developing logistic regression, XGBoost, random forest, and gradient boosting algorithms to predict the percentage likelihood of ETV/CPC failure within postoperative year 1. RESULTS: Unlike the ETVSS, etiology of hydrocephalus did not substantially influence any model's predictions and was removed. A logistic regression model produced the best area under the receiver operating characteristic curve (AUROC) and was used as the final model. Three features selected for the final model (corrected age, prior CSF diversion, and FOHR) proved relevant. The logistic regression predictor had an AUROC of 0.85 (compared to 0.66 for the ETVSS) and an outcome classification accuracy of 76%. The model correctly classified its predictions for success slightly more accurately than its predictions for failure (sensitivity 78%, specificity 75%). CONCLUSIONS: This model can reliably predict the likelihood of ETV/CPC failure at 1 year. Its outperformance of the ETVSS is likely due to the impact of CPC, the novel use of preoperative ventricle size as a predictive parameter, and accounting for variable interdependence. Further validation in additional patient populations is needed.