Maternal health is one of the main sustainable development goals of the World Health Organization (WHO). Changes associated with pregnancy can be reflected in the physiological, psychological and behavioral states of women. In this study, we investigated relationships between Adverse Childhood Experiences (ACEs) and mental health symptoms such as depression and anxiety over pregnancy and into the postpartum period, as well as quality of life. Further, we examined the interrelationship between mental health and physical health during and post pregnancy, finding that the efects during pre and postnatal periods tend to difer. We hope that our analysis lays potential groundwork for exploring more complex relationships between mental health symptoms and ACEs during and post pregnancy.
Pregnancy brings extensive psychological and biological changes, significantly impacting
maternal and fetal health. Depression is reported as the most common mental health disorder
during pregnancy [
Our research aims to comprehensively understand the relationships between ACEs and
pregnancy-related symptoms and outcomes. Existing literature primarily explores the
interrelationships between pairs of variables (e.g., ACE and depression, depression and anxiety), leaving
a gap in holistic understanding. The specific objectives of this study are to explore:
• How ACEs afect pregnancy related depression, anxiety, and quality of life
• The intricate relationships among mental health symptoms that occur during and
postpregnancy
• Investigate the relationship between physical and mental health during pregnancy
By illuminating these associations and patterns, our analysis could contribute to a more
nuanced understanding of how ACEs impact maternal mental health. This could raise
awareness of ACE at both individual and community levels so that we could inform more efective
interventions to support individuals who have experienced ACEs [
Numerous studies have indicated a significant association between early childhood adversity and
adverse mental health outcomes in adulthood [
Walker [12] utilized a structural equation modeling (SEM) to explore the relationship between ACEs and mental health symptoms during pregnancy, highlighting potential mechanisms linking early adversity to mental health symptoms during pregnancy. Although the study did not establish causal relationships, it provided strong evidence of a direct association between ACEs and mental health symptoms. These associations were attenuated by positive factors like resilience and social support. The findings presented by Walker [ 12] seem to align with results reported in a similar study conducted by Lydsdottir et al. [13]. Lydsdottir et al. presented SEMs to examine the relationship between adverse experiences in childhood and adulthood, prior history of mental illness, and symptoms of mental disorders during pregnancy. Both adult and childhood adversities were significantly associated with a history of depression and mental disorders during pregnancy. On the other hand, social support was found to have a significant negative association with symptoms of common mental disorders during pregnancy which is similar to what has been reported in [12].
The goal of causal discovery is to identify causal relationships through the analysis of observational data. Granger causality [14] is a popular technique for examining associations among diferent time-series data, assuming covariance stationarity. To handle non-stationary data, windowing techniques are commonly used, but they were not applicable to our dataset due to limited time steps and lack of stationarity. The Inductive Causation (IC) algorithm [15], implemented in the PC algorithm [16], is a well-known constraint-based approach for causal discovery. It starts with a complete, undirected graph and iteratively removes edges based on conditional independence decisions. However, the PC algorithm assumes the absence of hidden confounders, which is challenging to guarantee in systems involving human physiology and psychology. The DAGs with NO TEARS algorithm [17], a score-based approach for learning causal graphs, formulates the structure learning problem as a continuous optimization problem over real matrices. Unlike the PC algorithm, this technique does not assume the absence of hidden confounders. Both the PC and DAGs with NO TEARS algorithms were utilized for causal relationship discovery between ACE and mental health during pregnancy. The construction and performance of both methods are presented and discussed in the following sections.
The dataset used in this study is from the Better Understanding the Metamorphosis of Pregnancy (BUMP) study [18]. The BUMP study is a longitudinal digital health study that aims to gain a deeper understanding of the pregnancy and postpartum individual-level experience through the use of wearable and mobile devices [18]. For the purposes of this study, we use data collected from 4 clinically validated surveys that measured ACE, depression and anxiety symptoms, and quality of life: the ACE survey [19], the participant Health Questionnaire (PHQ-9) [20], the General Anxiety Disorder-7 (GAD-7) survey [21], and the Participant-Reported Outcomes Measurement Information System (PROMIS) survey [22]. PROMIS survey was broken down into PROMIS-physical and PROMIS-mental according to [23].
Let denote the set of participants in the form of unique user identification numbers, with || = 254 individuals. Let = {ACE, GAD-7, PHQ-9, PROMIS-physical, PROMIS-mental} denote the set of surveys in the study. Let survey denote the set of questions of survey. For each participant ∈ , the answers to each survey ∈ ∖ {ACE} were collected at each check-in time point ∈ , which contains 10 check-in time point before delivery prenatal, and 3 check-in time after delivery ∈ postnatal. The time interval between two consecutive check-in () time points is approximately two weeks. ,,, where ∈ , ∈ ∖ {ACE}, ∈ , ∈ , represents the answer to question of survey given by participant at check-in time point . Survey ACE was only taken once for each participants. (AC)E,, where ∈ , ∈ ACE, represent the answer to question in the ACE survey given by participant . Lastly, missingness in the dataset occurred due to non-compliance of participants in filling out surveys at various time points , and thus, for each participant , we denote the available data time points as () ⊂ .
The collected survey data was preprocessed for deploying causal structure discovery algorithms. We utilized two methods: survey-to-survey and question-to-question analysis, with surveys or questions treated as nodes respectively (Figure 1). We performed analyses both across all time points and separately for prenatal and postnatal data, to allow us to gain further insights into the data and explore the efects of time on our results.
Survey-to-survey analysis This analysis was performed to investigate the causal relationships between attributes that surveys measured on a population level. An overview of the analysis is presented in Appendix 2. We aggregated the question scores in each survey into survey scores. For each participant and survey ∈ at time point ∈ (), denote (,) = ∑︀∈ ,, as the sum of the question scores for survey of participant at time . The () sum was then mapped using a survey-specific mapping ℱ : R → Z≥ 0, where the values and numbers of mapping thresholds 1, · · · , are survey-specific, sourcing from corresponding literature[20][21][22][19]. A general example of a mapping is shown as follows: ℱ(, ) = ⎧⎪0, ⎪ ⎪⎨⎪1,
() if 0 < , ≤ 1
() if 1 < , ≤ 2 ⎪⎪· · · ⎪⎩⎪ − 1, if − 1 < , ≤ () .
Since the ACE survey was only administered once by each participant at the beginning of their study enrolment, the mapping function for ACE does not consider time domain. However, the survey scores of the rest of the surveys for each individual were aggregated across time. For each survey ∈ ∖ {ACE} and participant ∈ , let = AGGREGATE∈ () (ℱ(, )) be the aggregated result over all the time points. Examples of AGGREGATE functions are mean, median, max, min. We used mean as our AGGREGATE function. Therefore, = | 1()| ∑︀∈ () (ℱ(, )). The scores of the ACE survey were unchanged. Eventually, data matrix D ∈ R||×|| was generated. The distributions of the aggregated survey score were summarized in Appendix: 1.3.
In temporal analysis, we separated the time points into prenatal and postnatal periods and calculated separate scores for each period. Participants with both pre- and postnatal check-in data available (in ) were included in the analysis, totaling 42 participants. Out of the 42 participants, 38 had no ACE experience, while 4 did. The prenatal scores were given by: ,,prenatal = AGGREGATE∈p(ren)atal ((), , ). The postnatal scores were given by: ,,postnatal = AGGREGATE∈p(ost)natal ((), , ). Two data matrices were generated based on these scores, Dprenatal and Dpostnatal. The data matrices were computed on the same participant cohort who had both pre- and postnatal data available. The distributions of the aggregated survey score of both pre- and postnatal check-in times were summarized in Appendix 1.4. Question-to-question analysis The preprocessing is similar to survey-to-survey analysis, but we directly aggregated each single question score in the surveys for an individual across time. For each survey ∈ ∖ {ACE}, question ∈ , and participant ∈ , denote , as the aggregated result over all the time points, , = | 1()| ∑︀∈ () ((,),). We kept the score of ACE survey unchanged. Eventually, data matrix C ∈ R||× was generated, where = ∑︀∈ ||. Similar procedures were also followed for the temporal analysis.
We used causal structure discovery algorithms to find a graph = (, ), where is the set of defined nodes and is the set of edges representing causal relationships between variables. The graph should be a directed acyclic graph (DAG), where all the edges have orientations without directed cycles or directed loops. We applied two causal structure discovery algorithms to our dataset: PC algorithm [24][25] and DAG with NO TEARS [26] to output a DAG.
The PC algorithm [24] is a constraint-based method that tests for conditional independence between d-separation sets to determine the skeleton of the graph. This skeleton can be extended to a completed partially directed acyclic graph (CPDAG) within the equivalence class of the underlying DAG, where ambiguous causal directions are represented by arrow-to-arrow connections. However, the CPDAG is not a DAG since it cannot distinguish certain observational equivalence classes. To accommodate the independence test’s limitation with continuous values, we rounded obtained scores from previous step. The implementation of the PC algorithm utilized the pcalg package [25]. DAG with NO TEARS[26] is a score-based method. Given the data matrix X ∈ R× , it optimizes a loss function with the constraint of a DAG: min ( ) subject to ( ) ∈ , where ∈ R× is the weighted adjacency matrix inducing ( ). Since a sparse DAG is desired, the ( ) = 21 ||X − X ||2 + || ||1. The discrete constraint can be replaced by a smooth constraint ℎ( ) = ( ⊙ ) − = 0 as proposed by [26]. Solving the constraint optimization problem yields a DAG, which is what we desired.
This section presents the results obtained from using the PC and DAG with NO TEARS methods to investigate the relationship between ACEs and mental health symptoms during pregnancy and postpartum. Due to the measured and unmeasured confounders, we cannot attribute causality to the associations we uncovered; however they provide valuable insights into the dynamic relationship between ACEs and mental health symptoms during pregnancy and postpartum.
The survey-to-survey results are shown in Figure 3. Figure 3a shows the results of the PC algorithm, where it can be observed that ACE has no connections with other nodes. This is likely due to the ACE data’s highly unbalanced nature in the dataset, with only a few individuals experiencing significant levels of childhood adverse events. Therefore, we hypothesize that this unbalanced data distribution leads to the ACE variable being independent of all other variables, the skewness of the data afecting the results of the independence tests. The second result in 3b is from the DAGs with NO TEARS algorithm. Based on the consultation with domain experts, it was determined that the DAG with NO TEARS algorithm yields more reasonable results. Therefore, we utilized this algorithm for further analyses. For better interpretation, we focused on strong connections (weight > 0.7, average of edge weights in both prental and postnatal graphs). Though the learned graph is fully connected, we see that the strongest relations are from ACE to Anxiety (0.9) to Depression (0.8) to Mental Well-being (0.5) to Physical Well-being (0.9). These findings are consistent with previous studies [12, 13].
As for the results of the question-to-question level analysis, the weighted adjacency matrix generated from the DAG with NO TEARS is shown in Figure 4. Notably, question 9 from the PHQ9 survey exhibits stronger edge weights than other questions, which can be observed as dark red regions (weight ≈ 2) in Figure 4. This question pertains to suicide risk, asking about thoughts of self-harm or feeling better of dead in the past two weeks. It reflects an exceptionally severe symptom if answered afirmatively compared to other survey questions. Detailed descriptions of each survey question can be found in Appendix 1.2. Based on Figure 4, PHQ9-q9 (the suicide risk question) is particularly strongly positively associated with ACE-q4 (feeling that no one in your family loved you), PHQ9-q6 (feeling bad or like a failure), PROMIS_GLOBAL10_PHYSICAL-q1 (rating physical health) and PROMIS_GLOBAL10_PHYSICAL-q4 ( rating your pain on average) and negatively associated with PHQ9-q4 (trouble falling or staying asleep, or sleeping too much). The association between increased risk of suicide (PHQ9-q9) and childhood adversity has been previously presented in [27], particularly the positive correlation between neglect which is captured in ACE-q4 and risk of suicide attempts [28].
Please note that we ensured that semantic similarities were not confounding our analysis. To that extent, domain experts assisted in determining semantically similar questions, as shown in Appendix 1.5. For instance, question 5 in GAD-7 inquires about restlessness, while question 8 in PHQ-9 asks about excessive fidgeting or restlessness. Correlations between semantically similar questions were removed from the analysis and not considered.
The results on the comparison between prenatal and postnatal survey-to-survey analysis are
shown in Figure 5. First, we note that the associations difer between prenatal and postnatal
stages. We also found that ACE is associated with depression and anxiety both pre- and
postnataly (edge weights 0.9 and 0.7 in the pre and postnatal periods, respectively). These findings
are consistent with [
In this paper we focused on understanding relations between measurements, specifically ACE and mental health, during pregnancy as well as post-nataly. Our findings indicate that the way ACE afects mental health may difer during the pre and postnatal stages. The caveat of any such study is in the frequency of the extreme measures. Since some of the values, especially extreme ones, tend to be rare, their presence/absence can skew the independence tests that causal analysis is based on quite severely. In addition, while we have administered some of the expected surveys to identify depression and anxiety as well as other factors, there are a myriad of unmeasured factors that may influence the association and dependency between reported variables that may result in a direct dependency link in the graph, obscuring the potential unmeasured confounding. Thus, while our results indicate that the relation efect of ACE on mental health difers during the pre- and postnatal periods, further in depth studies are recommended to facilitate understanding of the direct vs indirect efects in these scenarios. We hope that ultimately this work is a first step in ofering insights to developing clinical interventions that can improve maternal mental health during pregnancy and reduce the risk of serious prenatal and postpartum mental health conditions in addition to overall well-being.
*This work is done based on equal contributions from JY, MA, RZ,and PE. JY, MA, RZ, PE are generously supported by 4YouandMe. AG is further supported by CIFAR AI chair as well as the Varma Family Chair. [10] R. E. Norman, M. Byambaa, R. De, A. Butchart, J. Scott, T. Vos, The long-term health consequences of child physical abuse, emotional abuse, and neglect: A systematic review and meta-analysis, PLoS Medicine 9 (2012) e1001349. URL: https://doi.org/10.1371/journal. pmed.1001349. doi:10.1371/journal.pmed.1001349. [11] M. V. Smith, N. Gotman, K. A. Yonkers, Early childhood adversity and pregnancy outcomes, Maternal and Child Health Journal 20 (2016) 790–798. URL: https://doi.org/10.1007/ s10995-015-1909-5. doi:10.1007/s10995-015-1909-5. [12] H. A. Walker, Maternal Adverse Childhood Experiences and Mental Health Symptoms in
Pregnancy: Behavioural and Social Mediators, Ph.D. thesis, University of Victoria, 2019. [13] L. B. Lydsdottir, L. M. Howard, H. Olafsdottir, H. Einarsson, T. Steingrimsdottir, J. F.
Sigurdsson, Adverse life experiences and common mental health problems in pregnancy: a causal pathway analysis, Archives of women’s mental health (2019). [14] A. Shojaie, E. B. Fox, Granger causality: A review and recent advances, 2021. URL: https: //arxiv.org/abs/2105.02675. doi:10.48550/ARXIV.2105.02675. [15] J. P. Verma, Emotions understanding model from spoken language using deep neural networks and mel-frequency cepstral coeficients, in: Proceedings of the Sixth Conference on Uncertainty in Artificial Intelligence, 1990, pp. 220–227. [16] C. G. Peter Spirtes, R. Scheines, Causation, Prediction, and Search, MIT Press, 2000. [17] X. Zheng, B. Aragam, P. Ravikumar, E. P. Xing, Dags with no tears: Continuous optimization for structure learning, 2018. URL: https://arxiv.org/abs/1803.01422. [18] S. M. Goodday, E. Karlin, A. Brooks, C. Chapman, D. R. Karlin, L. Foschini, E. Kipping, M. Wildman, M. Francis, H. Greenman, L. Li, E. Schadt, M. Ghassemi, A. Goldenberg, F. Cormack, N. Taptiklis, C. Centen, S. Smith, S. Friend, Better understanding of the metamorphosis of pregnancy (bump): protocol for a digital feasibility study in women from preconception to postpartum, npj Digital Medicine 5 (2022) 40. URL: https://doi.org/ 10.1038/s41746-022-00579-9. doi:10.1038/s41746-022-00579-9. [19] V. J. Felitti, R. F. Anda, D. Nordenberg, D. F. Williamson, A. M. Spitz, V. Edwards, M. P. Koss, J. S. Marks, Relationship of childhood abuse and household dysfunction to many of the leading causes of death in adults: The adverse childhood experiences (ace) study, American Journal of Preventive Medicine 14 (1998) 245–258. URL: http://www.ajpmonline.org/article/S0749379798000178/fulltexthttp: //www.ajpmonline.org/article/S0749379798000178/abstracthttps://www.ajpmonline.org/ article/S0749-3797(98)00017-8/abstract. doi:10.1016/S0749-3797(98)00017-8. [20] K. Kroenke, R. L. Spitzer, J. B. Williams, The phq-9: Validity of a brief depression severity measure, Journal of General Internal Medicine 16 (2001) 606. URL: /pmc/articles/PMC1495268//pmc/articles/PMC1495268/?report=abstracthttps: //www.ncbi.nlm.nih.gov/pmc/articles/PMC1495268/. doi:10.1046/J.1525-1497.2001. 016009606.X. [21] R. L. Spitzer, K. Kroenke, J. B. Williams, B. Löwe, A brief measure for assessing generalized anxiety disorder: the gad-7, Archives of internal medicine 166 (2006) 1092–1097. URL: https://pubmed.ncbi.nlm.nih.gov/16717171/. doi:10.1001/ARCHINTE.166.10.1092. [22] R. D. Hays, B. D. Schalet, K. L. Spritzer, D. Cella, Two-item promis® global physical and mental health scales, Journal of Patient-Reported Outcomes 1 (2017). URL: /pmc/articles/PMC5934936//pmc/articles/PMC5934936/?report=abstracthttps://www. ncbi.nlm.nih.gov/pmc/articles/PMC5934936/. doi:10.1186/S41687-017-0003-8. [23] V. Slavin, J. Gamble, D. e. a. Creedy, Measuring physical and mental health during pregnancy and postpartum in an australian childbearing population - validation of the promis global short form, BMC Pregnancy Childbirth (2019). [24] P. Spirtes, C. Glymour, R. Scheines, Causation, Prediction, and Search, 2nd ed., MIT press, 2000. [25] M. Kalisch, B. M. E. Ch, Estimating high-dimensional directed acyclic graphs with the pc-algorithm peter b ¨ uhlmann, Journal of Machine Learning Research 8 (2007) 613–636. [26] X. Zheng, B. Aragam, P. Ravikumar, E. P. Xing, Dags with no tears: Continuous optimization for structure learning (2018). URL: https://github.com/xunzheng/notears. [27] S. R. Dube, R. F. Anda, V. J. Felitti, D. P. Chapman, D. F. Williamson, W. H. Giles, Childhood abuse, household dysfunction, and the risk of attempted suicide throughout the life span, JAMA 286 (2001) 3089. URL: https://doi.org/10.1001/jama.286.24.3089. doi:10.1001/jama. 286.24.3089. [28] J. Kingree, M. P. Thompson, N. J. Kaslow, Brief report risk factors for suicide attempts among low-income women with a history of alcohol problems, Addictive Behaviors 24 (1999) 583–587. URL: https://www.sciencedirect.com/science/article/pii/ S0306460398001099. doi:https://doi.org/10.1016/S0306-4603(98)00109-9.
The code for the project can be accessed with the following link: https://github.com/ smilejennyyu/BUMP_causality 1. Did a parent or other adult in the household often . . . Swear at you, insult you, put you down, or humiliate you? or Act in a way that made you afraid that you might be physically hurt? Yes/No 2. Did a parent or other adult in the household often . . . Push, grab, slap, or throw something at you? or Ever hit you so hard that you had marks or were injured? Yes/No 3. Did an adult or person at least 5 years older than you ever. . . Touch or fondle you or have you touch their body in a sexual way? or Try to or actually have oral, anal, or vaginal sex with you? Yes/No 4. Did you often feel that... No one in your family loved you or thought you were important or special? or Your family didn’t look out for each other, feel close to each other, or support each other? Yes/No 5. Did you often feel that . . . You didn’t have enough to eat, had to wear dirty clothes, and had no one to protect you? or Your parents were too drunk or high to take care of you or take you to the doctor if you needed it? Yes/No 6. Were your parents ever separated or divorced? Yes/No 7. Was your mother or stepmother: Often pushed, grabbed, slapped, or had something thrown at her? or Sometimes or often kicked, bitten, hit with a fist, or hit with something hard? or Ever repeatedly hit over at least a few minutes or threatened with a gun or knife? Yes/No 8. Did you live with anyone who was a problem drinker or alcoholic or who used street drugs? Yes/No 9. Was a household member depressed or mentally ill or did a household member attempt suicide? Yes/No 10. Did a household member go to prison? Yes/No
Over the last 2 weeks, how often have you been bothered by the following problems? (Not at all, Several days, More than half the days, Nearly every day) 1. Little interest or pleasure in doing things 2. Feeling down, depressed, or hopeless 3. Feeling tired or having little energy. 4. Trouble falling or staying asleep, or sleeping too much. 5. Poor appetite or overeating. 6. Feeling bad about yourself or that you are a failure or have let yourself or your family down 7. Trouble concentrating on things, such as reading the newspaper or watching television. 8. Moving or speaking so slowly that other people could have noticed or the opposite - being so fidgety or restless that you have been moving around a lot more than usual. 9. Thoughts that you would be better of dead or of hurting yourself in some way. 10. If you checked of any problems, how dificult have these problems made it for you to do your work, take care of things at home, or get along with other people?
Over the last 2 weeks, how often have you been bothered by the following problems? (Not at all, Several days, More than half the days, Nearly every day) 1. Feeling nervous, anxious, or on edge 2. Not being able to stop or control worrying 3. Worrying too much about diferent things 4. Trouble relaxing 5. Being so restless that it’s hard to sit still 6. Becoming easily annoyed or irritable 7. Feeling afraid as if something awful might happen 8. If you checked of any problems, how dificult have these made it for you to do your work, take If you checked of any problems, how dificult have these made it for you to do your work, take care of things at home, or get along with other people?
You are about to start a survey that asks about your overall health. It is 10 questions and will take about 1 minute to complete. Question 3, 7, 9, and 10 belong to PROMIS-physical. Questions 2, 4, 5, and 8 belong to PROMIS-mental.
1. In general, would you say your health is: Excellent, Very good, Good, Fair, Poor 2. In general, would you say your quality of life is: Excellent, Very good, Good, Fair, Poor (Mental) 3. In general, how would you rate your physical health? Excellent, Very good, Good, Fair,
Poor (Physical) 4. In general, how would you rate your mental health, including your mood and your ability to think? Excellent, Very good, Good, Fair, Poor (Mental) 5. In general, how would you rate your satisfaction with your social activities and relationships? Excellent, Very good, Good, Fair, Poor (Mental) 6. In general, please rate how well you carry out your usual social activities and roles. (This includes activities at home, at work and in your community, and responsibilities as a parent, child, spouse, employee, friend, etc.) Excellent, Very good, Good, Fair, Poor 7. To what extent are you able to carry out your everyday physical activities such as walking, climbing stairs, carrying groceries, or moving a chair? Completely, Mostly, Moderately, A little, Not at all (Physical) 8. In the past 7 days, how often have you been bothered by emotional problems such as feeling anxious, depressed or irritable? Never, Rarely, Sometimes, Often, Always (Mental) 9. In the past 7 days, how would you rate your fatigue on average? None, Mild, Moderate,
Severe, Very severe (Physical) 10. In the past 7 days, how would you rate your pain on average? Put it on a 10-point scale from No pain to Worst pain imaginable (Physical)