Meeting Banner
Abstract #0249

Predicting Neurological Outcome in Neonatal Encephalopathy: A Machine Learning and Network Analysis Approach

Etay Ziv1, Olga Tymofiyeva1, Sonia L. Bonifacio2, Patrick S. McQuillen2, Donna M. Ferriero2, A James Barkovich1, Duan Xu1, Christopher P. Hess1

1Department of Radiology & Biomedical Imaging, UCSF, San Francisco, CA, United States; 2Department of Pediatrics, UCSF, San Francisco, CA, United States

Neonatal encephalopathy represents a heterogeneous group of conditions associated with life-long developmental disabilities. The ability to predict outcome early on in the perinatal period could potentially have a significant impact on subsequent treatment. Structural connectivity networks of the brain can be constructed using diffusion MRI. We hypothesize that networks derived from patients who have poor outcome may have different structure than those who have good outcome. Here we present an unbiased approach to enumerate a large set of network properties and using a combination of unsupervised and supervised learning, we demonstrate surprisingly good discrimination between good and poor neurological outcome.

Abstract PDF Presentation Video

Keywords

ability abnormal accuracy affect assembly assessed babies baby biomedical birth born brain brains characteristic class classes clustering coefficient combination complete component components conditions connectivity considered constructed construction cross derived developing developmental diffusion dimensional disabilities discrimination disease early easily edges encephalopathy enumerate equal except excite feature features finally generalized good grants heterogeneous highlight hypothesize impact input kernel label larger learning leave length life linear long machine machines mapped mapping matrix measure measured measures months motivate near neonatal network networks neurological neurologists next notable outcome overlap partitioning path patients pediatric pediatrics perinatal period poor potentially predict predicting prediction preprocessing press principle properties radiology report represents sample scanned scanner score separability smith space sphere structural structure subgraph subgraphs subsequent suggestive supervised support supported surprisingly term traditional trained transitivity treatment typical unbiased univ unsupervised validation various vector