A LASSO Analysis of Maternal, Obstetric, and Perinatal Predictors of Autism

Michal Ron Ben-Gurion University of the Negev Alexa Pohl University of Cambridge Dan Aizenberg Ben-Gurion University of the Negev Tslil Simantov Ben-Gurion University of the Negev Amber Ruigrok Cambridge University: University of Cambridge Paula Smith Cambridge University: University of Cambridge Michael Lombardo Istituto Italiano di Tecnologia Carrie Allison Cambridge University: University of Cambridge Alal Eran Ben-Gurion University of the Negev Simon Baron-Cohen Cambridge University: University of Cambridge Florina Uzefovsky (  orina@bgu.ac.il ) Ben-Gurion University of the Negev https://orcid.org/0000-0003-4571-2514


Introduction
Autism spectrum conditions are a set of neurodevelopmental conditions characterized by di culties in social communication and interaction, and by restricted and repetitive interests and behaviors (1). The prevalence of autism in the United States is 1 in 54, with a male bias of 4:1 (2). The etiology of the condition is thought to be a combination of genetic and non-genetic factors (3,4). However, the speci c factors, both genetic and environmental, are yet mostly unknown (5).
Previous research has shown that for environmental/non-genetic factors the timing of the exposure is critical, and most studies focus on early exposure -in utero or immediately after. Thus, prenatal and obstetric factors, such as maternal medical conditions during pregnancy and obstetric complications (6); and early perinatal factors, such as pre-term labor and infant characteristics (7), seem to be strongly related to autism in the developing child. Even though many of these factors are interconnected or cooccurring, most of the available studies focus on a few selected predictors (e.g., (8-11)), which makes it harder to understand the unique association of each one (beyond the others) with autism. To address this problem, we focus on speci c medical symptoms and conditions that have been previously associated with child's autism, and examine those within a single analysis using least absolute shrinkage and selection operator (12), which allows to choose variables most likely to be associated with the outcome from many potential predictors. We examined predictors of autism in the rst and second-born children (Study 1) and compared siblings discordant for autism (Study 2). Our analyses focused on three groups of diagnoses; (1) Maternal medical factors include any diagnosis, typically chronic, the mother received in her lifetime. These may serve as proximal or distal factors in explaining child's autism. For example, Polycystic Ovary Syndrome (PCOS) is a maternal condition associated with child's autism (8,13).This association has been interpreted as a consequence of maternal testosterone passing to the foetus and inducing neurological changes, which in turn, may trigger elevation of autistic traits and behavior (14,15).
(2) Obstetric factors include any diagnoses, conditions or symptoms that occur during and are related to pregnancy. For example, maternal infection during pregnancy (viral or bacterial) was previously associated with child's autism (16-18). Another such condition is infertility, which has been reported with mixed results (19)(20)(21)(22). These mixed ndings can be the result of an indirect association of fertility treatment and autism, mainly through an increase in the likelihood of obstetric complications, pre-term labor, and low birth-weight; through an association with maternal medical conditions such as PCOS (8); or through parental age (23). All these factors have been independently associated with child's autism(6, 7, 24). (3) Perinatal factors include characteristics of the child immediately after birth, such as gestational age. Gestational age was previously associated with autism (25), but also with various obstetric and maternal medical conditions (26).

The current study
As reviewed, many of the maternal medical conditions, obstetric conditions, and perinatal characteristics have been independently associated with autism, but are also interconnected or co-occurring (7,25,(27)(28)(29). Only by examining a combination of predictors can we assess the relative contribution of each predictor above others and isolate the most relevant predictors. To this end, the current study utilizes information collected from a large group of women regarding their medical conditions and pregnancy course. We compare these factors for children later diagnosed with autism and children who are typically developing. Study 1 uses a data-driven, LASSO, approach to identify which conditions are associated with autism in the rstborn and second-born child, separately. Study 2 examines a sub-sample of mothers who reported on the obstetric course of births resulting in a typically developing child and a sibling later diagnosed with autism. We compare the siblings' obstetric course to isolate pregnancy-course related predictors of autism, while passively controlling for many non-speci c environmental and genetic factors.

Methods
Participants-A total of N = 1,230 women aged 15-77 years (mean = 38.42, SD = 12.4) were recruited for a study on health through two websites managed by the Autism Research Centre (ARC) at Cambridge University, UK; 721 women through a website that targets individuals diagnosed with autism and their family members (https://autismresearchcentre.net/) and 509 women through a website targeting the general population (a liation to ARC is not mentioned; https://cambridgepsychology.com). Of all the participants enrolled in the study, only mothers of biological children were included in the current analyses. Participants were provided with information regarding the study and gave their consent before gaining access to the questionnaire. The study was approved by the Psychology Research Ethical Committee (PREC) at Cambridge University. In current analyses only women who reported regarding at least one pregnancy resulting in a live birth were included (N = 572). Births of multiples and children whose diagnosis status was missing were excluded from analyses. Study 1. Two separate analyses were conducted, for the rst and second live-born child of each mother. The nal sample included: (1) N = 557 mothers and their rst-born children (four mothers were removed due to missing data regarding child's autism diagnosis; and 11 due to multiple fetuses). (2) N = 374 mothers and their second-born children (31 mothers were removed due to missing data regarding child's diagnosis; and 13 due to multiple fetuses). Study 2. Analyses were conducted on a sub-sample of 132 mothers (264 pregnancies) who reported on at least one child diagnosed with autism and at least one typically developing child (TD). For each mother we selected the rst reported single birth of a TD child and the rst reported single birth of a child later diagnosed with autism. The siblings share a biological father, as determined by the maternal report regarding father's date of birth (see Table 1 for details). Measures-Mothers lled out the following self-report measures: (1) Demographic questionnaire -including mother's and father's date of birth, BMI before pregnancy, number of pregnancies, and autism diagnoses in the family.
(2) Health and pregnancy questionnaire (see Supplementary Information A) -maternal medical conditions (i.e. not speci c to the pregnancy); conditions and complications during pregnancy; and infant characteristics. Only items relating to maternal medical conditions, obstetric course and perinatal factors were analyzed. See full list in Table 2 and Supplementary Information A. Child sex X X X Notes: a) The prevalence of each condition in the sample was examined separately for each analysis. Each analysis included only variables with prevalence of more than 5 valid cases for binary variables and less than 40% missing value for quantitative variables. For each sample the variables included in the analysis are marked in X. b) In the sub sample, only obstetric factors and perinatal factors were included, as maternal characteristics are shared between children.

Data Analysis
Most of the predictor variables were treated as binary variables (yes/no), and few as ordinal or quantitative variables (see Supplementary Information A for more information).
Missing data imputation-Variables with less than 40% missing data (quantitative variables) or less than 5 cases (binary variables) were included in the analyses, and the missing data was imputed using "MICE" (Multivariate Imputation via Chained Equations) package in R, and the Predictive Mean Matching (PMM) technique, which is appropriate for numeric variables. MICE assumes missing at random, meaning that the probability that a value is missing depends only on other observed values, and therefore observed values can be used to predict the missing value. We used the MICE procedure to create 1,000 imputed data sets (30), and averaged out the imputed variables across the data sets.

Study 1
The main aim of the study was to examine the association between maternal medical conditions, obstetric factors and perinatal factors, with child's autism. To select the predictors which show a unique contribution to autism we analyzed the data using the procedure LASSO (in R, "glmnet" package; (31)).

LASSO (Least Absolute Shrinkage and Selection
Operator) is a technique for regression analysis, which performs both variable selection and regularization; and can be used for binary outcomes (12). The technique uses L1 penalty, which minimizes the absolute value of each predictor by shrinking all estimates toward zero (31,32). In the context of this research, the technique assigns positive and negative weights to variables associated with child's autism. Variables unrelated to the diagnosis are assigned a weight of zero, which effectively excludes them from the nal model. LASSO yields a regularization parameter (a penalty; i.e "λ") that creates a parsimonious model which contains the maximum number of parameters and minimum cross-validation errors. The nal model contains the selected predictive variables that were associated with the outcome. Despite some limitations (31,33), LASSO has high accuracy, and the chosen variables are highly likely to represent true contributions (33), and therefore is a reliable method for predictors selection. Although the LASSO yields coe cient value for each predictive variables, one cannot interpret the relative magnitude of the coe cients, because LASSO creates bias in the estimation of the parameters (due the shrinking toward zero), and the coe cients are not necessarily accurate. Therefore, an estimator was calculated for every coe cient by training LASSO classi ers on 1,000 bootstrapped samples of the data using the glmnet R package. To ensure a coe cient of zero for unrelated variables, the median of the coe cients for every variable, rather than the mean, was calculated as the nal estimator. Additionally, a 95% con dence interval for the medians was calculated using scipy.stats.binom (34). To calculate the P value for every estimator, 10,000 equallydistributed random datasets were generated by independently bootstrapping each column of the original dataset. Then, a one-sided p-value was calculated for each estimator as the number of times the respective random coe cients were higher than or equal to the actual estimator (for positive estimators), or lower than or equal to the actual estimator for negative estimators, divided by the total number of random datasets (10,000). P values were not calculated for coe cients estimated to be 0.

Study 2
Ten predictors, which could differ between pregnancies, were included, negating the need for predictor selection. We conducted bootstrapped (5,000 repetition) paired t-test analysis to compare the rate of each condition for TD vs autistic siblings. A Bonferroni correction was applied to control for multiple testing (.05/10 = .005).

Study 1:
A LASSO analysis was conducted for the rst and second-born children separately. See descriptive statistics for each sample in Table 3. See summary of the predictors of child's autism in Table 4. For the rst-born, the nal model included 15 predictors associated with increased likelihood for autism diagnosis in the child For the second-born, the nal model included 20 predictors of child's autism.

Study 2:
132 parents and their 264 children were included in the analysis. See results in Table 5. In this analysis, only child-related variables were included (obstetric and perinatal factors), as maternal factors are shared by both children.

Discussion
Studies on maternal medical factors predicting child's autism tend to focus on few factors within the same analysis. This can be problematic as some of these factors are etiologically or phenomenologically dependent, which makes it harder to determine which factors can have the largest contribution to predicting child's diagnosis. In Study 1, a case-control approach revealed 15 predictors of autism for the rst-born, and eight of those were replicated in the second-born along with 12 additional predictors. In Study 2, we compared a subset of predictors, relating to pregnancy course and perinatal factors for siblings discordant for diagnosis. In this analysis two factors were signi cant, replicating those found in Study 1. Our analyses broadly showed that obstetric factors and maternal medical conditions were most predictive of child's autism. Below we will discuss the main ndings.
Gestational bleeding during the second or third trimesters was associated with child's autism in all analyses (yet the effect is small in the rst-born), replicating previous studies (6). Gestational bleeding is a non-speci c condition, and can indicate a host of other pregnancy complications, such as placental insu ciency and preeclampsia, which have also been associated with autism (26) (but were of low prevalence in the current study and should be studied further). Interestingly, a cumulative factor of pregnancy complications was signi cant only for the rst, but not second-born, which may re ect an indirect association through maternal age, as pregnancy complications are associated with older maternal age (29).
The current ndings suggest that infertility, rather than infertility treatments and other obstetric conditions, is related to autism, as we nd this effect after controlling for infertility treatments and multiple obstetric complications, some of which are independently related to autism. Previous ndings regarding the association between autism and infertility and fertility treatments were mixed. One study (27) found a higher incidence of autism among children born after infertility treatments, as compared to natural conception. The association remained statistically signi cant even after adjusting for parity, infant gender, and parental age, all factors that are independently associated with both infertility treatments and autism. In contrast, Hvidtjørn and colleagues (2010) found no association between autism in children and infertility treatments, in a large cohort study. Other studies (22,35) also didn't nd any association between infertility treatments and autism diagnosis, in two separate studies conducted in California, USA. One hypothesis to explain the mixed ndings is that infertility treatments are indirectly related to autism through an association with obstetric complications, pre-term labor, and low birthweight, which are all factors that have been independently associated with an child's autism (6, 7, 24), and with older parental age (21,36,37). Parner and colleagues (2012) analyzed a cohort of all singleton births between 1980 to 2003 in Denmark and found that older parental age is associated with child's autism(38). Similarly, a cohort study of singleton births between 1989 to 2002 conducted in California, USA, also found that advanced parental age was associated with child's autism (39). Older maternal age may be associated with autism not only because of the increased risk of chromosomal abnormalities in the ova (6), but also due to the relationship between mature age and increased risk for pregnancy complications (29).
The current ndings suggest that gestational diabetes and type II diabetes are both related to autism, but not BMI prior to pregnancy or weight gain during pregnancy. Previous studies found a relationship between various metabolic conditions such as maternal obesity prior to pregnancy, increased weight gain during pregnancy and gestational diabetes (de ned as glucose intolerance with onset or rst recognition during pregnancy) with autism and with general poor neurodevelopmental outcomes (40,41). Krakowiak and colleagues found that mothers of children diagnosed with autism had a higher rate of gestational diabetes and obesity (de ned as BMI > 30), compared to mothers of TD children (28). In addition, Xiang and colleagues found higher rates of child's autism among mothers diagnosed with type II diabetes before pregnancy, and among women diagnosed with gestational diabetes by 26 weeks of pregnancy (42). The current ndings emphasize the importance of these metabolic conditions in autism.
The non-signi cant ndings also shed light on the relative importance of various factors. Here we found no evidence for an association between child's gestational age and autism. Pre-term labor and low birthweight were previously found to be non-speci c risk-factors for neurodevelopmental and intellectual disabilities such as learning disabilities, attention problems, and poor executive function (25). According to one study, low birth-weight infants (< 2500 g) have a 60% increased likelihood of autism, and pre-term infants (< 32 weeks) had twice the chance of developing autism, for both sexes (31). In addition, Burstyn and colleagues (2010) found that low birth-weight (< 2500 g) increased the likelihood for autism diagnosis by 33%. However, pre-term labor and low birth-weight often can be the consequents of obstetric complications (26), and the current ndings emphasize the importance of obstetric complications over gestational age in predicting child's autism.
Immune system activation during pregnancy is another suspected environmental mechanism for child's autism. Animal studies show that immune activation during pregnancy is associated with abnormal brain development (16, 43). Speci cally, Shi and colleagues found a reduction in social behavior and elevated anxiety behaviors among mice born to mothers infected by a virus during pregnancy (44). Studies in humans nd similar associations. Lee and colleagues (2015) found that maternal infection (viral or bacterial) which leads to hospitalization, increased the likelihood for autism by 37%, in a cohort study of all live births born 1984-2007, in Sweden (17). However, another study found an association only for hospitalization due to bacterial infections(16, 45), and another found no association between infection and autism in a Danish population (46). We also do not nd evidence for this effect in the current study.
Importantly, some maternal medical conditions which were found in previous research as associated with autism, were not found to be signi cant or did not replicate across the analyses. For example, maternal PCOS was found to be related to child's autism in multiple large-scale analyses (8, 13), but in our study was not signi cant.

Limitations
The main limitation of the study is the use of self-report, recollection data, as opposed to the use of medical records. It is possible that some conditions and diagnoses were misremembered or missed entirely. Therefore, the ndings of the current study should be replicated using medical records in order to validate the ndings. Another limitation of the study is the sample size, which did not allow to examine the relationship between child's autism and relatively rare disorders and conditions. A larger sample, or a sample biased to include reports of mothers who suffered speci c serious health and obstetric complications would allow to focus on these effects.

Conclusions
Our ndings emphasize the role of infertility, including miscarriages, and obstetric complications as predictive factors of autism, beyond typically correlated factors such parental age and vaginal bleeding.
Thus, these ndings are important in guiding further research into the involvement of infertility and obstetric complications in the etiology of autism.

Declarations
Ethics approval and consent to participate