Title

Development of Machine Learning-Based Predictor Algorithm for Conversion of an Ommaya Reservoir to a Permanent Cerebrospinal Fluid Shunt in Preterm Posthemorrhagic Hydrocephalus.

Publication/Presentation Date

6-1-2022

Abstract

BACKGROUND: An Ommaya reservoir can be used to treat posthemorrhagic hydrocephalus secondary to intraventricular hemorrhage of prematurity until an acceptable weight can be obtained to place a permanent shunt. Identifying newborns at higher risk of developing shunt conversion may improve the management of these patients. This study aimed to develop a predictive algorithm for conversion of an Ommaya reservoir to a permanent shunt using artificial intelligence techniques and classical statistics.

METHODS: A database of 43 preterm patients weighing ≤1500 g with posthemorrhagic hydrocephalus (Papile grades III and IV with Levene ventricular index >4 mm above the 97th percentile) managed with an Ommaya reservoir at our institution between 2002 and 2017 was used to train a k-nearest neighbor algorithm. Validation of results was done with cross-validation technique. Three scenarios were calculated: 1) considering all features regardless whether or not they are correlated with the output variable; 2) considering the features as predictors if they have a correlation >30% with the output variable; 3) considering the output of the previous analysis.

RESULTS: When considering the outputs of a previous multivariate analysis, the algorithm reached 86% of cross-validation accuracy.

CONCLUSIONS: The use of machine learning-based algorithms can help in early identification of patients with permanent need of a shunt. We present a predictive algorithm for a permanent shunt with an accuracy of 86%; accuracy of the algorithm can be improved with larger volume of data and previous analysis.

Volume

162

First Page

264

Last Page

264

ISSN

1878-8769

Disciplines

Medicine and Health Sciences

PubMedID

35259501

Department(s)

Department of Surgery, Fellows and Residents, Department of Surgery Residents

Document Type

Article

COinS