Machine Learning Analysis of Infant Brain Odor Responses in an fMRI Study
Poster No:
411
Submission Type:
Abstract Submission
Authors:
Dahyeon Kang1, Sarah Reeser1, Allegra Johnson1, Stephen Dager1, Mary Larimer1, Natalia Kleinhans1
Institutions:
1University of Washington, Seattle, WA
First Author:
Co-Author(s):
Introduction:
With the legalization of cannabis, pregnant women have increasingly turned to cannabis products to alleviate symptoms such as nausea, anxiety, and pain during pregnancy.¹ Yet, there remains a significant gap in research regarding the impact of prenatal cannabis exposure on infant brain development. Infant brains pose unique challenges for study due to their rapid developmental trajectories.² While functional Magnetic Resonance Imaging (fMRI) has provided a means to study infant brain development, traditional fMRI statistical methods (e.g., general linear models) encounter limitations in precisely identifying task-dependent brain activation. This is due to the heterochronicity in the maturation process across brain regions, potentially influencing the hemodynamic response function. Therefore, integrating models capable of handling multiple variables, i.e., machine learning, becomes crucial in understanding these complexities. Our study, using fMRI data from infants aged 6 to 9 months with prenatal cannabis exposure (PCE) and a control group (CON), aims to uncover nuanced insights into how such exposure might affect neural processing and responses to olfactory stimuli during this early developmental period.
Methods:
Twenty-eight infants (14 PCE, 14 CON) provided valid fMRI data (57.1% male) under natural sleep. Olfactometer tubes were positioned toward the infant's nose to present phenylethyl alcohol, a rose-like odorant which was presented in a block-design, with 'odor+air' for 9s and 'air only' for 18s, repeated 4 times. Olfactometer equipment setup was as previously described.³ Quiet BOLD fMRI scans were obtained on a Philips Ingenia Elition 3T with a 32-channel head coil (TR/TE=1500/30ms, 2.5 mm³ isotropic, MB 3, SENSE factor=2, 72 dynamics). Preprocessing was performed using FMRIB's Software Library and included motion correction, brain extraction, detrending, band pass filtering, and registration to the 7.5 month infant template.² The regions of interest (ROIs) included in the analyses covered both primary and secondary olfactory cortex (Fig. 1).⁴ Python with Scikit-Learn was used for data analysis. A Random Forest Regression was chosen for its capability in handling high-dimensional data and managing multicollinearity among features. SHAP (SHapley Additive exPlanations) values were computed for interpreting feature importance within the defined ROIs.
Results:
The Random Forest model demonstrated strong performance in distinguishing between odor and air sensory processing, accounting for 43% of the observed response variance, R²=0.43, MSE=0.125. SHAP analysis highlighted specific ROIs, such as the pars orbitalis, entorhinal, pallidum, and insula, among others, showing substantial influence in predicting hemodynamic responses to odor (Fig. 1). Further, using multi-level modeling, group- and individual-level differences in time to peak activation and SHAP-values were examined. A significant main effect of Group (b=-0.004, t=-2.131, p=0.03) and an interaction between time-to-peak-activation and Group (b=0.002, t=5.685, p<0.01) were found in the medial orbitofrontal cortex (Med OFC). Specifically, PCE infants exhibited lower SHAP values and longer time-to-peak-activation (10.5s vs. 6s post-stimulus in CON) in the Med OFC (Fig. 3). Lastly, among PCE infants, significant associations were discovered between tetrahydrocannabinol (THC) exposure levels and SHAP values in the left pallidum (b=0.0001, t=2.334, p=.04), such that higher levels of THC correlated with lower SHAP values.
Conclusions:
This study offers initial evidence supporting the potential of machine learning in delineating precise neural responses to sensory stimuli in infant brains. By revealing key contributors and their temporal dynamics, it highlights the intricate interplay within the developing brain during sensory processing tasks, thereby opening avenues to develop more precise models of the hemodynamic response that will allow for deeper investigations into early neurodevelopmental processes.
Disorders of the Nervous System:
Neurodevelopmental/ Early Life (eg. ADHD, autism) 1
Lifespan Development:
Normal Brain Development: Fetus to Adolescence
Modeling and Analysis Methods:
Activation (eg. BOLD task-fMRI) 2
Methods Development
Perception, Attention and Motor Behavior:
Chemical Senses: Olfaction, Taste
Keywords:
Addictions
Data analysis
Development
Machine Learning
Smell
1|2Indicates the priority used for review
Provide references using author date format
2. (Sanchez et al., 2012) Sanchez, C. E., Richards, J. E., & Almli, C. R. Neurodevelopmental MRI brain templates for children from 2 weeks to 4 years of age. Dev. Psychobiol. 54, 77–91.
3. (Kleinhans et al., 2019) Kleinhans, N. M., et al. FMRI correlates of olfactory processing in typically-developing school-aged children. Psychiatry Res. Neuroimaging 283, 67–76.
4. (Zhou et al., 2019) Zhou, G., Lane, G., Cooper, S. L., Kahnt, T., & Zelano, C. Characterizing functional pathways of the human olfactory system. eLife 8, e47177.