Predicting endoscopic third ventriculostomy success in childhood hydrocephalus: an artificial neural network analysis

Abstract

Object: Artificial neural networks (ANNs) can be used as a measure for the clinical decision-making process. The aim of this study was to develop an ANN model to predict endoscopic third ventriculostomy (ETV) success at 6 months and to compare the findings with those obtained using traditional predictive measures in childhood hydrocephalus.

Methods: The ANN, ETV Success Score (ETVSS), CURE Children's Hospital of Uganda (CCHU) ETV (CCHU ETV) Success Score, and logistic regression models were applied to predict outcomes. The cause of hydrocephalus, patient age, whether choroid plexus cauterization (CPC) was performed, previous shunt surgery, sex, type of hydrocephalus, and body weight were considered as input variables for an established ANN model. Data from hydrocephalic children who underwent ETV were applied, and the computer program that analyzes the data was trained to predict successful ETV by using several input variables. Successful ETV outcome was defined as the absence of ETV failure within 6 months of follow-up. Then, sensitivity analysis was performed for the established ANN model to identify the most important variables that predict outcome. The area under a receiver operating characteristic curve, accuracy rate of the prediction, and Hosmer-Lemeshow statistics were measured to test different prediction models.

Results: Data for 168 patients (80 males and 88 females; mean age 1.4 ± 2.6 years) were analyzed. Data from patients were divided into 3 groups: a training group (n = 84), a testing group (n = 42), and a validation group (n = 42). The successful ETV outcome rate, defined as the absence of ETV failure within 6 months of follow-up, was 47%. Etiology, age, CPC status, type of hydrocephalus, and previous shunt placement were the most important variables that were indicated by the ANN analysis. Compared with the ETVSS, CCHU ETV Success Score, and the logistic regression models, the ANN model showed better results, with an accuracy rate of 95.1%, a Hosmer-Lemeshow statistic of 41.2, and an area under the curve of 0.87.

Conclusions: The findings show that ANNs can predict ETV success at 6 months with a high level of accuracy in childhood hydrocephalus. The authors' results will need to be confirmed with further prospective studies.