Using Machine Learning to Preoperatively Predict ETV/CPC Failure Likelihood Within 1-Year

Introduction: 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 successful hydrocephalus management at six months following ETV based on age, hydrocephalus etiology, and shunting history. The ETVSS does not, however, account for the impact of CPC or quantify likelihood of success beyond six months, and relies on independence of contributing variables. We created a predictive tool to address limitations of the ETVSS by quantifying likelihood of ETV/CPC failure at one year.

Methods: Records for 257 pediatric patients who received ETV/CPC as a primary definitive treatment for hydrocephalus at a single institution between 2009 – 2021 were reviewed for demographics, presenting characteristics, and medical history. Data on corrected age at surgery, hydrocephalus etiology, sex, and prior CSF diversion (such as external ventricular drain or shunt placement) were used to develop a gradient boost classifier paired with a custom function to predict the percentage likelihood of ETV/CPC failure within the first postoperative year.

Results: The classifier produced a receiver operator characteristic curve of 0.95 and an accuracy of 86%. The model predicted failure likelihood within a year with more accurate directionality than the ETVSS in 95% of cases and had an average absolute error of 8.75%, compared to 46% for the ETVSS. The model predicted failure likelihood at one year more accurately than the ETVSS for 93% of patients with myelomeningocele, IVH, or tumor (average error difference: 38.25), 100% of patients with post-infectious etiologies, (average error difference: 30.00) and 96% of patients with aqueduct stenosis, Dandy Walker, or other etiologies (average error difference: 35.63).

Conclusion : This model reliably predicted the likelihood of ETV/CPC failure at one year. It may outperform the ETVSS by accounting for predictive variable interdependence. Further validation on additional patient populations is needed.