Project Details

Development of neuroimaging biomarkers for 2-year neurocognitive outcomes in mild neonatal brain injury

Rina Bao, PhD

Summary

One in two hundred newborns suffers from brain injuries related to the lack of oxygen and blood supply, a neonatal brain dysfunction known as hypoxic-ischemic encephalopathy (HIE). We aim to use image analysis and artificial intelligence (AI) to address the challenge that a sixth of HIE patients still develop neurocognitive impairments by 2 years of age, even though their brain Magnetic Resonance Imaging (MRI) scans present little or no explicit injuries in the newborn stage. To address this, we investigated whether computer-extracted whole-brain patterns could establish imaging biomarkers associated with 2-year outcomes and support individualized outcome prediction. Among 237 patients with HIE from Massachusetts General Hospital and Boston Children’s Hospital, we analyzed 33 infants with mild HIE, including 24 with normal and 9 with adverse 2-year outcomes. From apparent diffusion coefficient (ADC) and anatomy-normalized ZADC maps, we extracted 5,782 whole-brain, regional, and fiber-tract features. Among them, 362 showed nominal associations with outcome (p < 0.05). In an exact-p-value sensitivity analysis, three left thalamic ZADC features remained significant after within-family false-discovery-rate correction. Infants with adverse outcomes had lower kurtosis and uniformity and higher entropy (q = 0.047 for each; absolute rank-biserial r = 0.824 to 0.833). We next evaluated four machine-learning classifiers using leave-one-patient-out analysis with feature selection. The four classifiers selected unique feature sets containing 40-50 features for logistic regression, linear support vector machine, Gaussian-kernel support vector machine, and decision tree, respectively. The Gaussian-kernel support vector machine performed best, achieving 0.788 accuracy, 0.715 balanced accuracy, 0.588 F1 score, 0.556 sensitivity, and 0.875 specificity for 2-year outcome prediction in mild HIE cases. These exploratory findings suggest that quantitative neonatal MRI may help reveal subtle abnormalities associated with 2-year neurocognitive outcome and may support individualized prognosis. We will next validate the entire analytical pipeline in a larger multisite cohort of infants with mild HIE from 23 sites, including feature extraction, feature selection, identification of imaging patterns associated with neurocognitive outcomes, and outcome prediction, to determine whether imaging biomarkers can improve prediction of 2-year neurocognitive outcomes and assess their robustness and generalizability.

Publications:

Bao R, Song Y, Bates SV, Weiss RJ, Foster AN, Jaimes C, Sotardi S, Zhang Y, Hirschtick RL, Grant PE, Ou Y. BOston Neonatal Brain Injury Data for Hypoxic Ischemic Encephalopathy (BONBID-HIE): I. MRI and Lesion Labeling. Sci Data. 2025 Jan 11;12(1):53. doi: 10.1038/s41597-024-03986-7. PMID: 39799120; PMCID: PMC11724925.

Bao R, Foster AN, Song Y, Vyas R, Kesri A, Eddine Toubal I, Soltani Kazemi E, Rahmon G, Kucukpinar T, Almansour M, Ho ML, Palaniappan K, Ninalga D, Prasad Koirala C, Mohapatra S, Schlaug G, Wodzinski M, Muller H, Ellis DG, Aizenberg MR, Arda Aydin M, Abdinli E, Unal G, Tahmasebi N, Punithakumar K, Song T, Peng Y, Bates SV, Hirschtick R, Ellen Grant P, Ou Y. BONBID-HIE 2023: Lesion Segmentation Challenge in BOston Neonatal Brain Injury Data for Hypoxic Ischemic Encephalopathy. IEEE Trans Med Imaging. 2026 Apr;45(4):1711-1725. doi: 10.1109/TMI.2025.3638977. PMID: 41379890; PMCID: PMC12954739.

Bao R, Weiss RJ, Bates SV, Song Y, He S, Li J, Bjornerud A, Hirschtick RL, Grant PE, Ou Y. PARADISE: Personalized and regional adaptation for HIE disease identification and segmentation. Med Image Anal. 2025 May;102:103419. doi: 10.1016/j.media.2024.103419. Epub 2025 Feb 1. PMID: 40147073; PMCID: PMC12777737.