Use of Supervised Learning for the Characterization of Pediatric Postinfectious Hydrocephalus
Restricted (Penn State Only)
- Author:
- Parra Rodriguez, Flor
- Graduate Program:
- Engineering Science and Mechanics (MS)
- Degree:
- Master of Science
- Document Type:
- Master Thesis
- Date of Defense:
- June 22, 2023
- Committee Members:
- Steven Schiff, Thesis Advisor/Co-Advisor
Corina Stefania Drapaca, Committee Member
Albert Segall, Program Head/Chair
Andrea P. Arguelles, Committee Member - Keywords:
- hydrocephalus
pediatric hydrocephalus
brain calcifications - Abstract:
- Postinfectious hydrocephalus (PIH) is the condition of excess fluid accumulation in the ventricular system as result of a preceeding acquired bacterial or viral infection. Though hydrocephalus remains the leading cause of pediatric neurosurgery worldwide with a particularly elevated incidence of PIH in sub-Saharan countries, the current treatment approaches are remedial and invasive in nature. As part of the effort to identify early biomarkers that are distinct to PIH, this work proposes a supervised learning pipeline to investigate whether volumetric estimates of brain tissue, cerebrospinal fluid (CSF), and calcified tissue, can be used individually or collectively as predictor biomarkers that allow for the distinction between postinfectious and non-postinfectious hydrocephalus (NPIH). The volumetric data was obtained from a cohort of 134 pediatric hydrocephalus patients (mean age = 49.8 days, SD = 36) via image segmentation of computed tomography (CT) scans at the admission and a follow up time point. The pipeline incorporated six trained classification models with repeated K-fold (K = 10) and Leave-One-Out cross-validation methods and 100 train/test split resampling iterations. Classification analysis of each volumetric feature determined CSF to be the best predictor feature of PIH/NPIH, yielding the highest average performance on test data (F-1 score = 0.868 , Error rate = 0.198), and intracranial calcifications as the weakest predictor (F-1 score =0.793 , Error rate =0.347). Multivariate classification using the volumetric data from the initial scan provided a statistically validated decision boundary in R3, providing preliminary evidence that volumetric information at pre-treatment is useful to non-invasively characterize the early stages of PIH in pediatric patients.