Overview

Schema Version

https://atlas.rsna.org/schemas/2025-11/model.json

Name

Outcome Prediction in Neonates with Encephalopathy (OPiNE)

Link

https://github.com/chris-lew/neonatal_HIE_outcome_prediction

Indexing

Keywords: neonatal encephalopathy, hypoxic-ischemic encephalopathy, prognosis, deep learning, MRI, 2-year outcome, neurodevelopmental impairment

Content: MR, NR, PD

RadLex: RID12698, RID10312, RID5056, RID38778, RID49531, RID10796

Author(s)

Christopher O. Lew

Evan Calabrese

Joshua V. Chen

Felicia Tang

Gunvant Chaudhari

Amanda Lee

John Faro

Sandra Juul

Amit Mathur

Robert C. McKinstry

Jessica L. Wisnowski

Andreas Rauschecker

Yvonne W. Wu

Yi Li

Organization(s)

Duke University Medical Center

University of California, San Francisco

University of Washington

Saint Louis University

Washington University School of Medicine in St Louis

Children’s Hospital Los Angeles

University of Southern California

Version

1.0

Contact

ude.ekud@eserbalac.nave

Funding

Authors declared no funding for this work.

Ethical review

Duke Health Institutional Review Board–approved, retrospective analysis of HEAL trial data with written informed parental consent.

Date

Published: 2024-07-10

References

[1] Lew CO, Calabrese E, Chen JV, Tang F, Chaudhari G, Lee A, Faro J, Juul S, Mathur A, McKinstry RC, Wisnowski JL, Rauschecker A, Wu YW, Li Y. "Artificial Intelligence Outcome Prediction in Neonates with Encephalopathy (AI-OPiNE)". Radiology: Artificial Intelligence. 2024 Sep;6(5):e240076. Published online 2024 Jul 10.. 2024-07-10. doi:10.1148/ryai.240076. PMID: 38984984. PMCID: PMC11427921.

Model

Architecture

3D convolutional neural network encoder with four convolutional blocks followed by two linear layers with 30% dropout and final sigmoid; AdamW optimizer, binary cross-entropy loss; Kaiming initialization; image features concatenated with tabular data at first linear layer.

Availability

Open-source code: https://github.com/chris-lew/neonatal_HIE_outcome_prediction

Clinical benefit

Prognostic prediction of death or neurodevelopmental impairment at 2 years in neonates with hypoxic-ischemic encephalopathy using early MRI and basic clinical data.

Clinical workflow phase

Clinical decision support systems; neuroprognostication following neonatal encephalopathy.

Decision threshold

Binary classification threshold selected by Youden index on each test set.

Degree of automation

Automated image-based prediction using MRI and basic tabular variables to generate a risk probability; intended as decision support.

Indications for use

Term neonates (≥36 weeks gestation) with moderate to severe encephalopathy enrolled within 1–6 hours of age; MRI obtained 4–6 days after birth; intended for use in NICU/tertiary pediatric care settings to aid prognostication.

Input

Preprocessed multisequence neonatal brain MRI (T1-weighted, T2-weighted, diffusion trace, ADC) and tabular variables (sex, gestational age, erythropoietin administration, ADC-defined brain injury volume).

Instructions

Provide T1-, T2-weighted, and DTI data (for trace and ADC) processed as in study: brain extraction, coregistration, N4 bias correction (T1/T2), intensity normalization to [0,1], crop to 96×112×96 at 1.0-mm3; tabular inputs scaled (0–1) and z-normalized. Use model to output probability of outcomes; apply threshold per Youden index if binary classification is desired.

Limitations

Developed and validated on HEAL trial cohort of neonates ≥36 weeks with moderate/severe encephalopathy; not generalizable to other populations (e.g., mild encephalopathy, outside 6 hours after birth). Sample size relatively small for deep learning; potential model variability; small number of death cases; performance depends on threshold selection; further site-specific external validation recommended before clinical use.

Output

Description: Predicted probability of 2-year outcomes, including primary endpoint (death or any neurodevelopmental impairment) and secondary endpoints (NDI severity strata, severe NDI or death, death alone).

Recommendation

Use as a prognostic aid alongside clinical judgment; perform site-specific validation and select operating threshold appropriate to local clinical needs before deployment.

Regulatory information

Authorization status: Not a regulated/cleared medical device; research model described in a peer-reviewed study.

Reproducibility

Evaluation included in- and out-of-distribution test sets; uncertainty estimated using Monte Carlo dropout (30% dropout, 30 samples). Code repository provided for reproducibility.

Sustainability

Training performed in a Linux Docker environment on a workstation with two NVIDIA RTX A6000 GPUs; no energy consumption metrics reported.

Use

Intended: Prognosis

Out-of-scope: Prognosis

Excluded: Prognosis

User

Intended: Physician, Subspecialist diagnostic radiologist, Researcher

Out-of-scope: General diagnostic radiologist

Excluded: Patient, Layperson, Caregiver