The prevalence of prenatal maternal stress (PNMS) has increased alarmingly; a recent large-scale study reported approximately 30% of pregnant women reported one or more types of stressors [Loomans et al., 2013; reviewed in Van den Bergh et al. (2020)]. Further, climate change events, in the form of hurricanes, tropical storms, wildfires, flooding, and droughts, are increasing in frequency and becoming more extreme in nature. These weather events cause serious disruptions in people’s lives and pose important threats to the mental health of individuals, especially among vulnerable populations such as pregnant women (Clemens et al., 2020; Zakrison et al., 2020). Considering the prevalence rate of PNMS, it is imperative to investigate its impact on early neurobehavioral development in their offspring (Monk et al., 2019).

An accumulating number of studies have established that various forms of PNMS serve as a risk factor for future child psychopathology (Monk et al., 2019; Lautarescu et al., 2020; Van den Bergh et al., 2020; Nomura et al., 2022). This included in utero maternal stress due to exposure to weather-related disasters, such as the Quebec Ice Storm, which led to greater externalizing and internalizing clinical behaviors (King et al., 2012) and Superstorm Sandy (SS), which led to greater clinical and adaptive behaviors (Nomura et al., 2022). Other studies have investigated the biophysiological consequences of suboptimal neurobehavioral development using magnetic resonance imaging (MRI) (Buss et al., 2010; El Marroun et al., 2016; Lebel et al., 2016; Davis et al., 2017; Wen et al., 2017; Mareckova et al., 2022, reviewed in Lautarescu et al., 2020). MRI serves as an unparalleled technology to pinpoint the structural and functional brain changes in various brain regions of the limbic system, such as the amygdala and hippocampus, and frontal lobe, such as the prefrontal cortex, among offspring exposed to PNMS and how those changes underlie the consequent neurobehavioral, emotional, and cognitive changes observed. Machine learning (ML), a rapidly developing technique in data science, is defined as a robust data-driven approach to automatically detect underlying patterns in high-dimensional data with high accuracy (Bi et al., 2019). Several studies have adopted ML algorithms and identified an enlarged amygdala as an important risk factor for early childhood anxiety (Qin et al., 2014), while other studies have discovered a wide range of biopsychosocial causal features using a predictive classification model (Saxe et al., 2017). However, research on the application of ML onto both neurobehavioral and brain volumetric data has been very limited in pediatric populations (Oskar and Stingone, 2020).

To date, there are a small number of ML studies pertaining to the association between prenatal stress or trauma exposure and altered neurobehavioral development and brain volumetric changes of offspring. One study demonstrated detection and classification of prenatal alcohol exposure (Rodriguez et al., 2021) and another identified the most affected regions of the brain (volumes) in children exposed to alcohol prenatally (Little and Beaulieu, 2020). However, a gold-standard of randomization is not possible among human populations, nor had any other work investigated disaster related PNMS in a quasi-experiment model in a human population. Quasi-experimental designs make it possible to pseudo-randomize prenatal stress independent of confounding personal attributes, such as genetic makeups and maternal psychological disorders status (Lafortune et al., 2021). SS, which hit a wide region in New York City (NYC), randomly “assigned” stressful conditions to pregnant women and their offspring, constructing an objective measure of stress independent of the mothers’ genetic background, psychopathology, and socioeconomic status. Thus, the Stress in Pregnancy (SIP) study, with its uniform exposure to a stressor, SS, enabled us to address the inherent bias that traditionally plagues studies assessing the impact of stress in a human population. As such, by leveraging the technique of explainable ML, this study attempts to identify important child neurobehavioral, and brain volumetric features associated with natural disaster-related PNMS; explainable ML makes it possible to uncover patterns of high-dimensional neurobehavioral and brain data of children exposed to such disasters.

The study capitalized on a longitudinal study that followed mother/child dyads from in utero to age 11, who were exposed, or not exposed, to Superstorm Sandy (SS) that hit metropolitan New York in 2012. An ML classifier was built to identify neurobehavioral and brain volumetric features that set a child with in utero exposure to SS apart from those without. The binary in utero exposure to a natural disaster was formulated as the target of the classification problem. ML can facilitate identification of features that dominate the classification process leading to high prediction accuracy. These features can be understood as the phenotypes most affected by natural disaster-related PNMS and studying them helps understand which behaviors and brain volumetric data in the selected brain regions are most affected as a result of SS exposure in utero. To this end, we trained a classification ML algorithm for features associated with SS exposure on a given participant and applied explainable artificial intelligence (AI) to identify features that can more accurately contribute to predicting structural changes in brain volume from the exposure.

Figure 2 shows the plots of ranked feature importance scores for different categories of features: neurobehavioral and brain volumetric data. The most relevant neurobehavioral feature that distinguished SS exposure was aggression. Moreover, the left amygdala, left hippocampus, and the right parahippocampal gyrus (PHG) were the three brain regions that distinguished SS exposure.

To the best of our knowledge, the current study is the first to show that the implementation of a machine learning model was predictive in identifying children exposed in utero to natural disasters based on various neurobehavioral phenotypes and brain regions, shown to be sensitive to stress and trauma exposure (Lautarescu et al., 2020; National Scientific Council on the Developing Child, 2020; Van den Bergh et al., 2020).

Our results showed distinctly higher aggression and adaptability scores among our SS exposed group, which, consistent with the work of King et al. (2012) and Nomura et al. (2022), suggests in utero maternal stress, from exposure to a weather-related disaster, is strongly associated with child neurobehavioral phenotypes. Our results demonstrate the stress a child experiences in utero increases the risk for future psychopathology but may also impact a child’s potential to acquire and express certain adaptive skills in their postnatal environment (Nomura et al., 2022).

In agreement with the findings from numerous earlier studies, our results suggest in utero stress exposure may be associated with a larger amygdala volume (Buss et al., 2012; Wen et al., 2017; Acosta et al., 2019; Jones et al., 2019) and a larger hippocampal volume (Cao-Lei et al., 2021), although less supported. It has been further suggested that prenatal stress exposure appears to accelerate the development of these regions, particularly the amygdala, as a means of allowing more rapid detection of incoming danger and threat (Lautarescu et al., 2020). Our results are consistent with the evolutional model of stress that suggests the exaggeration of automatic fear detection programmed prenatally to better prepare for the likely at stake environment postnatally (Glover, 2011).

Our results offer further, more novel, insights as in utero SS exposure was found to be linked with both a larger caudate and a smaller PHG. Enlargements in the caudate have been suggested to underlie the pathophysiology of various neurobehavioral and emotional disorders, characterized by repetitive and ritualistic tendencies, such as obsessive-compulsive disorder (OCD) and autism (Ring and Serra-Mestres, 2002). A reduced PHG was found to be associated with elevated prenatal maternal anxiety (Buss et al., 2010; Acosta et al., 2019). The caudate and PHG are two brain regions shown to influence emotion regulation and the formation of emotional memories (Bhatt et al., 2012; Zhu et al., 2019; Driscoll et al., 2022). Our findings on a larger caudate and a smaller PHG may extend the current literature and support the association between prenatal natural disaster-related maternal stress and altered child volumetric brain development, which may have long-term implications on child neurodevelopment and emotion regulation (Talge et al., 2007; Wu et al., 2022).

Of note, no notable sex differences were observed between the assessed clinical and adaptive neurobehaviors nor between the GMV of the SS sensitive brain regions. In efforts to explore the association between the GMV of the SS sensitive brain regions and aggression scores, a post hoc analysis was conducted. Upon stratification of the sample by the SS exposure status, higher aggression scores were negatively associated with left amygdala GMV [r(19) = −0.45, p = 0.04] among the SS unexposed and marginally associated with left amygdala GMV among the SS exposed [r(7) = −0.60, p = 0.09]. No other notable associations between aggression and the GMV of SS sensitive brain regions were detected. We anticipated seeing a strong correlation between amygdala GMV and aggression scores as previous work among a sample of children, aged 6–9, found an association between smaller amygdala GMV and higher aggression scores (Thijssen et al., 2015). As validated by the aggression diathesis model, we speculate this strong correlation may be due to increased aggression leading to an imbalance between top down and bottom up control systems (Siever, 2008). Greater acts of aggression may lead top down systems, which are modulated by brain regions that are highly interconnected with the amygdala, such as the orbitofrontal cortex (OFC) and the anterior cingulate cortex (ACC), to fail to suppress aggressive acts in the presence of anger inducing stimuli (Siever, 2008). Moreover, increased aggression may induce hyperresponsivity of limbic brain regions, such as the amygdala, which provide the “drive” to respond to incoming danger and threat (Siever, 2008).

We developed an ensemble ML model to find brain and behavioral features that were associated with SS exposure on given participants. Our main goal was to see whether certain features contribute accurately to predicting structural changes in brain morphometry from exposure using ML. We achieved the best accuracy by using the feature selection method, RFE. We also examined the effect of oversampling by using SMOTE on all features. For high dimensional data, SMOTE generally performs better by avoiding overfitting (Dessie et al., 2021; Park et al., 2021). However, for data with a very small sample size (n = 30), involving augmented (simulation) data provides little benefit if we reduce feature dimensionality because overfitting is no longer a problem and oversampling can lead to underperformance due to a deteriorating sampling (Elreedy and Atiya, 2019). From our best results, we see that with a small data set, with very few positive cases, reducing the dimensionality, rather than introducing synthetic data, leads to higher model performance. Considering achieving greater generalizability is one of our aims, reducing the dimensionality would be one of the critical solutions. Our finding is consistent with prior studies using small sample sizes. Crippa et al. (2015) identified children with autism spectrum disorder (ASD) from a sample size of N = 30. Güven et al. (2020) predicted attention-deficit hyperactivity disorder (ADHD) children from a sample size of N = 44. Both studies implemented feature selection either manually or algorithmically.

Within the summary SHAP plot, most of the important features that affect our target value came from the brain volumetric data (i.e., the amygdala, hippocampus, and caudate) suggesting the brain’s particular sensitivity to stress exposure. Further, the dependence plots showed distinctly larger amygdala and smaller PHG brain volumes, respectively, increase the likelihood of SS exposure.

The study has various limitations. First, despite our prior research indicating the binary SS index (exposed vs. not exposed in utero) to be a significant and critical indicator of impairment, we acknowledge the use of binary, objective indices prevent the analysis of important, fine-grained measures, such as post-traumatic psychosocial reactions to SS. Second, the BASC-2P questionnaire used to assess child behavioral problems was based on parental report, which may lead to subjective bias. However, our study incorporated brain imaging to corroborate and enhance the objectivity and robustness of the findings. Third, information on the handedness of the children was not collected and fourth, the sample size was relatively small with mostly girls. Thus, the results should be interpreted with caution and warrant replication with bigger sample sizes in the future.

There are several strengths in this study. First, our study utilized a quasi-experimental design and rendered the pseudo-randomization of a natural disaster-related stress in utero. Currently, most of the studies that have investigated stress during pregnancy relied on the measurement of maternal psychopathology, trauma history or nominal pregnancy stress—which are likely to be confounded by factors, such as genetics, and heterogeneity of level of stress exerted. Second, the current study is part of a larger longitudinal project prospectively tracking and assessing the neurodevelopment of children, preventing the need to rely on retrospective parental reports. This eliminates the problem due to recall bias, which is common in cross-sectional research using information based on retrospective reports in part. Third, our study is the very first to deploy the technique of ML on neurobehavioral and brain imaging data in the research of in utero maternal stress. Future studies will aim to expand on our initial findings by tracking changes from childhood into adolescence.

Exploiting ML, we developed multiple models for classification and chose the RFE algorithm to extract the subset of features that contributed the most optimally to detection of natural disaster exposure in utero. The use of SHAP values assisted in model interpretation. SHAP plots provided critical insights into behavioral phenotypes and brain volumetric changes associated with natural disaster exposure in utero among offspring, especially within brain regions implicated in emotion regulation (the amygdala, caudate, and PHG). The goal of this study was to leverage advanced computational methodology to understand how a significant stressor connects brain and behavioral development among high-risk populations. Our preliminary results suggest that targeted intervention on behavioral phenotypes, such as aggression, and structural morphological alterations in the emotional centers of the brain can potentially play a key role in buffering the adverse impact of prenatal maternal stress on brain development and improve subsequent developmental outcomes.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

The studies involving human participants were reviewed and approved by the City University of New York University Integrated Institutional Review Board (protocol number: 2018-1305-QC). Written informed consent to participate in this study was provided by the participants’ legal guardian/next of kin.

GD and C-LT led the study and oversaw the integrity of the ML implementation in the study. AS supervised the interpretation of the brain structure findings. YN supervised the interpretation of the behavioral findings. All authors contributed to the drafting, revising, and interpreting the data.