Accurate, Early Detection of Cerebral Palsy Using Multimodal MRI at Term in Infants Born Preterm

1383 

Abstract Submission 

Nehal Parikh¹, Mekibib Altaye², Armin Allahverdy², Julia Kline³, Karen Harpster², Hailong Li², Junqi Wang², Abiot Yenealem Derbie⁴, Jean Tkach², Beth Kline-Fath², Stephanie Merhar², Weihong Yuan², Lili He²

¹CINCINNATI CHILDREN'S HOSPITAL, Cincinnati, OH, ²CINCINNATI CHIILDREN'S HOSPITAL, Cincinnati, OH, ³National Institutes of Health, Bethesda, MD, ⁴Cincinnati Children's Hospital, CINCINNATI, OH

Nehal Parikh, DO, MS  
CINCINNATI CHILDREN'S HOSPITAL
Cincinnati, OH

Mekibib Altaye, PhD  
CINCINNATI CHIILDREN'S HOSPITAL
Cincinnati, OH

Armin Allahverdy, PhD  
CINCINNATI CHIILDREN'S HOSPITAL
Cincinnati, OH

Julia Kline, PhD  
National Institutes of Health
Bethesda, MD

Karen Harpster, PhD  
CINCINNATI CHIILDREN'S HOSPITAL
Cincinnati, OH

Hailong Li, PhD  
CINCINNATI CHIILDREN'S HOSPITAL
Cincinnati, OH

Junqi Wang, PhD  
CINCINNATI CHIILDREN'S HOSPITAL
Cincinnati, OH

Abiot Yenealem Derbie  
Cincinnati Children's Hospital
CINCINNATI, OH

Jean Tkach, PhD  
CINCINNATI CHIILDREN'S HOSPITAL
Cincinnati, OH

Beth Kline-Fath, MD  
CINCINNATI CHIILDREN'S HOSPITAL
Cincinnati, OH

Stephanie Merhar, MD, MS  
CINCINNATI CHIILDREN'S HOSPITAL
Cincinnati, OH

Weihong Yuan, PhD  
CINCINNATI CHIILDREN'S HOSPITAL
Cincinnati, OH

Lili He, PhD  
CINCINNATI CHIILDREN'S HOSPITAL
Cincinnati, OH

Despite recent advances in early diagnosis of cerebral palsy (CP), accurate and timely detection remains elusive. Advances in quantitative MRI and machine learning technology appear promising to enable early, accurate prediction of CP. Our goal was to improve early CP prediction in preterm infants by exploiting advanced quantitative MRI biomarkers acquired at term-equivalent age.

We recruited a multisite regional cohort of 358 very preterm (VPT) infants born at or below 32 weeks' gestational age from 5 Southwest Ohio Neonatal Intensive Care Units (Cincinnati Infant Neurodevelopment Early Prediction Study [CINEPS]). All infants were imaged at Cincinnati Children's between 39 and 44 weeks postmenstrual age on a single 3T Philips scanner and 32-channel receiver head coil with the following identical sequences: dMRI: TE 88ms, TR 6972ms, FA 90°, resolution 2×2×2 mm3, 36 directions; b-value of 800 s/mm2; MB factor 2; rsfMRI: TE 45, TR 893ms, FA 90°, resolution 2.5×2.5×2.5 mm3; 400 volumes; MB factor 4; Axial T2w: TE 166 ms, TR 8300ms, FA 90°, resolution 1×1×1 mm3. CP was diagnosed at 2 years corrected age (CA) using the Amiel-Tison (1998) standardized neurological exam and the Gross Motor Function Classification System (GMFCS; Palisano RJ, 2000). We used established pre- and post-processing pipelines and neonatal brain atlases from the Developing Human Connectome Project (dHCP) to generate brain morphometry measures (volumes, cortical maturation), structural connectivity (SC) from diffusion MRI (dMRI), and functional connectivity (FC) from resting state functional MRI (rsfMRI) as previously described (Bastiani M, 2019; Kline JE, 2020; Kline JE, 2021). We used CONN to generate six graph theory measures per modality for each of the 81 dHCP atlas regions of interest. We used an unsupervised approach for feature selection/reduction from the nearly 20,000 MRI predictor variables. Specifically, we employed non-negative matrix factorization (NMF) to decompose these predictors into 7 network components each from morphometry, FC, and SC. We elected to include moderate-severe brain injury from conventional MRI (cMRI) and five clinically known predictors of CP a priori (Table 1). Last, we applied a support vector machine to develop a multimodal model that included the above independent variables to predict CP. We created 1,000 bootstrap datasets to evaluate model performance and correct for over-optimism per the TRIPOD guidelines (Moons KGM, 2015). To assess model fit, we calculated optimism-corrected values for sensitivity, specificity, and area under the receiver operating characteristic curve (AUC).

There were no significant differences in the baseline clinical variables between the 307 VPT infants (86%) that returned for standardized CP testing at 2 years CA (shown in Table 1) and the 51 that did not (data not shown). 34 infants (11.1%) developed CP (GMFCS ≥1) of which 11 (3.6%) had moderate-severe CP (GMFCS≥2). The CP vs. non-CP groups differed in several clinical factors (Table 1). We observed moderate-severe injury on cMRI at term in 25 (8.1%) infants. The AUC was 0.657 (95% CI: .581, .730) for the clinical plus cMRI model. For the combined model that included clinical, cMRI, and quantitative MRI modalities, the AUC was 0.870 (.818, .914), sensitivity was 86.0% (73.5, 96.4), and specificity was 88.0% (79.5, 96.0) for the prediction of any CP; the corresponding values for moderate-severe CP were 0.946 (.859, .993), 93.4% (72.2, 100), and 95.8% (90.5, 99.7) (Table 2).

In a regional prospective cohort of VPT infants, we combined multimodal neuroimaging and machine learning to enable early, accurate prediction of CP that outperformed current statistical models. Our internally validated model can be immediately translated to enable targeted risk stratification for research studies following NICU discharge that are designed to prevent or reduce the severity of CP in high-risk VPT infants.

Neurodevelopmental/ Early Life (eg. ADHD, autism) ²

Classification and Predictive Modeling ¹

Diffusion MRI Modeling and Analysis

fMRI Connectivity and Network Modeling

Motor Behavior Other

Development

FUNCTIONAL MRI

Machine Learning

Motor

Movement Disorder

Neurological

PEDIATRIC

STRUCTURAL MRI

WHITE MATTER IMAGING - DTI, HARDI, DSI, ETC