Prediction of 6 months endoscopic third ventriculostomy success rate in patients with hydrocephalus using a multi-layer perceptron network

Objective: Discrimination between patients most likely to benefit from endoscopic third ventriculostomy (ETV) and those at higher risk of failure is challenging. Compared to other standard models, we have tried to develop a prognostic multi-layer perceptron model based on potentially high-impact new variables for predicting the ETV success score (ETVSS).

Methods: Clinical and radiological data of 128 patients have been collected, and ETV outcomes were evaluated. The success of ETV was defined as remission of symptoms and not requiring VPS for six months after surgery. Several clinical and radiological features have been used to construct the model. Then the Binary Gravitational Search algorithm was applied to extract the best set of features. Finally, two models were created based on these features, multi-layer perceptron, and logistic regression.

Results: Eight variables have been selected (age, callosal angle, bifrontal angle, bicaudate index, subdural hygroma, temporal horn width, third ventricle width, frontal horn width). The neural network model was constructed upon the selected features. The result was AUC:0.913 and accuracy:0.859. Then the BGSA algorithm removed half of the features, and the remaining (Age, Temporal horn width, Bifrontal angle, Frontal horn width) were applied to construct models. The ANN could reach an accuracy of 0.84, AUC:0.858 and Positive Predictive Value (PPV): 0.92, which was higher than the logistic regression model (accuracy:0.80, AUC: 0.819, PPV: 0.89).

Conclusion: The research findings have shown that the MLP model is more effective than the classic logistic regression tools in predicting ETV success rate. In this model, two newly added features, the width of the lateral ventricle's temporal horn and the lateral ventricle's frontal horn, yield a relatively high inter-observer reliability.