676 - Nonlinear Brain Network Disruption in Opioid-Exposed Neonates Revealed by Deep Generative Modeling

Publication Number: 2659.676

Jung-Hoon Kim, Children's National Health System, Washington, DC, United States; Josepheen De Asis-Cruz, Children's National Health System, Bethesda, MD, United States; Kushal Kapse, Developing Brain Institute, Washington, DC, United States; Yao Wu, Children's National Hospital, Washington, DC, United States; Stephanie L. Merhar, Cincinnati Children's Hospital Medical Center, Cincinnati, OH, United States; Carla M. Bann, RTI International, Apex, NC, United States; Jamie E. Newman, RTI International, Research Triangle Park, NC, United States; Nicole Mack, RTI International, Chicago, IL, United States; Sara DeMauro, Children's Hospital of Philadelphia, Philadelphia, PA, United States; Namasivayam Ambalavanan, University of Alabama School of Medicine, Birmingham, AL, United States; Scott A. Lorch, The Children's Hospital of Philadelphia, University of Pennsylvania, Philadelphia, PA, United States; Deanne Wilson-Costello, Case Western Reserve University School of Medicine, Cleveland, OH, United States; Brenda Poindexter, Children's Healthcare of Atlanta, Atlanta, GA, United States; Myriam Peralta-Carcelen, University of Alabama at Birmingham, Birmingham, AL, United States; Jonathan M. Davis, Tufts University School of Medicine, Boston, MA, United States; Catherine Limperopoulos, Children's National Health System, Washington DC, DC, United States

Background: Antenatal opioid exposure (AOE) has been linked to disrupted brain growth and neurobehavioral impairments. Our Outcomes of Babies with Opioid Exposure (OBOE) study, and other groups have revealed altered brain networks in opioid-exposed neonates. Prior work, however, has relied on linear methods that may fail to capture nonlinear temporal fluctuations on newborn resting-state fMRI (rs-fMRI), which are highly plausible given the exponential brain growth in neonates. To address this, we applied a variational autoencoder (VAE), a type of deep generative model, to newborn rs-fMRI, enabling detection of nonlinear temporal functional modes (TFMs) to elucidate dynamic shifts in brain networks in AOE neonates.

Objective: To determine whether nonlinear modeling of rs-fMRI reveals functional brain alterations associated with AOE neonates.

Design/Methods: We analyzed 158 AOE newborns and 90 age-matched controls (CTL) enrolled in our OBOE study. Variational autoencoder (VAE), which has been extensively validated in healthy fetuses, newborns, and adults, and trained on an independent large neonatal dataset (dHCP), was applied to rs-fMRI to learn nonlinear latent representations (Fig. 1A left). Temporal ICA on VAE-derived latent time series defined nonlinear TFMs (Fig. 1A right). We evaluated classification performance using a cross-validated SVM and identified opioid-sensitive TFMs via leave-one-out analysis.

Results: Baseline study participant characteristics are presented in table 1. Using VAE-derived TFMs, the classifier distinguished AOE from CTL neonates with a statistically significant accuracy, reaching 59.7 ± 0.4% at 20 TFMs (Fig. 1C; p< 0.05, permutation test). Four opioid-sensitive TFMs were identified (Fig. 2A; Bonferroni-corrected p< 0.05), which also predicted postmenstrual age (Fig. 2B; r=0.72, 0.56, 0.52, and 0.62 for TFM1-4, respectively; p< 10-12 for all). Abnormal between-TFM interactions were found in AOE newborns (Fig. 2C). Opioid-sensitive TFMs were localized in the primary sensorimotor (TFM2-3) and visual regions (TFM4), with TFM1 reflecting global synchronization (Fig. 2D).

Conclusion(s): We report that nonlinear TFMs revealed significant alterations in functional brain organization in AOE newborns. Notably, opioid-sensitive TFMs were predominantly localized in the primary sensorimotor and visual regions, with global synchronization also affected. VAE-based modeling offers a powerful framework for detecting dynamic disruptions in neonatal brain networks, laying the foundation for early identification of neurodevelopmental risk.

Figure 1. Workflow for deriving temporal functional modes (TFMs) using a variational autoencoder (VAE) and assessing their sensitivity to antenatal opioid exposure.
(A) Cortical resting-state fMRI patterns are encoded into a latent space via a VAE, generating latent trajectories that capture nonlinear temporal dynamics. Temporal independent component analysis (tICA) is then applied within this latent space to extract TFMs, which are subsequently projected back onto the cortical surface for interpretability. (B) VAE-derived TFMs are used to train a support vector machine (SVM) classifier to distinguish opioid-exposed neonates from controls, using 10-fold cross-validation. (C) Classification accuracy as a function of the number of TFMs included. The boxplot represents the null distribution from permutation testing, and the red error bar indicates observed accuracy. Statistical significance was assessed via permutation testing (p < 0.05).

Table 1. Characteristics of studied cohort.


Figure 2. Opioid-sensitive temporal functional modes (TFMs) and their functional relevance.
(A) Sensitivity analysis of individual TFMs to opioid exposure, assessed by excluding each TFM from the classification model. Red error bars indicate classification accuracy after exclusion; the gray shaded area represents baseline accuracy using all TFMs. (B) Relationship between each TFM's age-prediction performance and its sensitivity to opioid exposure, highlighting developmental relevance. (C) Connectivity between TFMs showing significant group differences. Each dot represents an individual subject; red error bars denote the standard error of the mean. OE: opioid-exposed group. * indicates Bonferroni-corrected p < 0.05 (two-sample t-test). (D) Full set of 20 temporal functional modes (TFMs) of healthy newborns derived using a variational autoencoder (VAE). TFMs defined in the VAE latent space are back-projected into a cortical space using the VAE decoder. The order of TFMs is set by its difference in classification accuracy between leave-one-TFM-out model vs. whole TFMs.