Mining multi-site clinical data to develop machine learning MRI biomarkers: application to neonatal hypoxic ischemic...

Area of Science:

• Medical imaging
• Neurology
• Biomedical informatics

Background:

• Retrospective analysis of hospital data offers a cost-effective approach for biomarker development compared to prospective trials.
• Neonatal hypoxic-ischemic encephalopathy (HIE) requires accurate diagnostic and prognostic tools.

Purpose of the Study:

• To establish a comprehensive retrospective clinical dataset of neonatal HIE.
• To develop and validate MRI-based analytic algorithms for HIE lesion detection and outcome prediction.
• To compare machine learning algorithm performance against human experts in HIE assessment.

Main Methods:

• Utilizing clinical registries and big data informatics to compile a multi-site dataset.
• Including structural and diffusion MRI, clinical information, and short- and long-term outcomes for at least 300 HIE patients.
• Employing machine learning frameworks to analyze deviations from normative brain atlases.

Main Results:

• The study aims to develop algorithms capable of detecting HIE-related brain abnormalities.
• The developed algorithms are intended to predict patient outcomes with high accuracy.
• Performance of algorithms will be benchmarked against expert human assessments.

Conclusions:

• Retrospective data analysis is a viable strategy for developing advanced HIE diagnostic tools.
• Machine learning applied to MRI data holds promise for improving HIE patient management.
• This research facilitates the creation of objective, data-driven approaches in neonatal neurology.