Medical Conditions in the First Years of Life Associated with Future Diagnosis of ASD in Children

The associations of 38 medical conditions with future ASD diagnosis reported here strengthen the evidence presented in previous studies of the higher prevalence of gastrointestinal conditions, developmental delays, mental health disorders, musculoskeletal abnormalities, neurological conditions, and sleep disorders among children with ASD. In contrast to the few previous studies that relied on parent reports of medical conditions in the years before ASD diagnosis—which is highly subject to recall bias—the key innovation of our study is that we relied on diagnoses made by clinicians and recorded in the child’s medical record prospectively before the ASD diagnosis was made. Our findings provide new evidence indicating that children with ASD have a disproportionate burden of certain medical conditions in the years preceding ASD diagnosis.

A high prevalence of GI conditions among children with ASD has been reported in several studies with no comparison to control children (Molloy and Manning-Courtney 2003; Nikolov et al. 2009; Ming et al. 2008). However, a recent consensus report evaluating the diagnosis and treatment of GI conditions among individuals with ASD concluded that: “The prevalence of gastrointestinal abnormalities in individuals with ASDs is incompletely understood,” and went on to highlight the methodological limitations of previous studies and the possible influence of referral bias (Buie et al. 2010). For example, one recent study using descriptive reports from structured interviews estimated the prevalence of GI symptoms among children with ASD to be 70% compared to a prevalence of only 28% among control children (Valicenti-McDermott et al. 2006). In stark contrast, a small retrospective case–control study of the medical records of 96 ASD children and 449 control children reported that children with ASD did not have higher odds of a history of GI disorders before ASD diagnosis compared to children without ASD (OR 1.0, 95% CI 0.5–2.2) (Black et al. 2002). This study is the most similar in design to ours, although much smaller, and the CI bounds are consistent with our results (adjusted OR estimates ranging from 1.0 to 2.2 across GI conditions). Overall, our findings support our hypothesis that children with ASD are more likely than control children to have a documented comorbid GI condition in the time before ASD diagnosis, but the difference is quite modest, with effects smaller than we expected. Our findings suggest that descriptive reporting of gastrointestinal conditions may substantially over-estimate the difference in prevalence between children with ASD and controls.

We found that developmental delays and mental health conditions showed the strongest associations with future ASD diagnosis. The strong association between language delay and ASD risk is not surprising given that language delay is a component that characterizes ASD diagnosis (Rapin 1991; American Psychiatric Association 2013). Notably, we found a moderately strong association of motor delay with future ASD diagnosis. Motor delays are not considered prodromal to ASD development, but some studies have reported increased motor impairment among children already diagnosed with ASD as compared to control children (Noterdaeme et al. 2002; Mari et al. 2003). To our knowledge, our finding of an association between ASD and a motor delay diagnosis that precedes ASD diagnosis has not previously been reported.

For 41 other medical conditions, we did not find statistically significant evidence of an association with ASD risk in this highly-powered study. We found that children who are later diagnosed with ASD were not more likely than controls to be diagnosed with allergy, asthma, infections, autoimmune conditions, hematology, genitourinary, accidental injury, endocrine disorders or cardiovascular disorders in the years preceding ASD diagnosis. In contrast, some previous studies have reported that ASD is associated with comorbid medical conditions of allergies (Yaghmaie et al. 2013; Theije et al. 2014), asthma (Kotey et al. 2014), infections (Atladottir et al. 2010), autoimmune conditions (Gesundheit et al. 2013), and endocrine disorders (Chen et al. 2009).

We previously reported that both allergies and autoimmune diseases were diagnosed more often among children already diagnosed with autism than among controls (Zerbo et al. 2015). One possible explanation for why results during the time period prior to ASD diagnosis differ from results when studying children already diagnosed with ASD is that some medical conditions may have a later age of onset than the age of ASD diagnosis. Alternatively, increased health care utilization among children already diagnosed with ASD could lead to increased opportunity for diagnoses of other medical conditions relative to children without ASD who have fewer encounters with the health care system. Further research is needed to elucidate how medical conditions that co-occur with ASD may differ before and after ASD diagnosis and the reasons driving those differences.

Our study also demonstrated the ability of medical condition clusters to differentiate relative risks of ASD. We identified 14 risk strata and illustrated a gradient of increasing ASD risk across these strata, with ORs ranging from 1.5 to 100 in comparison to the lowest risk group of children. Risk stratification of children into groups who have lower and higher risks of ASD could be a first step toward earlier ASD diagnosis by identifying groups of children for whom increased monitoring or earlier screening protocols could be developed. Earlier diagnosis of ASD is critical because earlier treatment can reduce the degree of impairment and improve function (Lord et al. 2005). Our analysis of the latency time between diagnosis of all conditions in each cluster and subsequent diagnosis of ASD found that the mean latency time was greater than 12 months for every cluster, and as high as 50 months for some, illustrating that there is a substantial window of time between identification of higher risk children and the diagnosis of ASD. Not surprisingly, the high risk clusters were dominated by developmental delays and mental health conditions. However, there were interesting features regarding how these delays clustered together in different combinations. Interestingly, both cluster analyses identified several distinct clusters within the high risk group of children with language delays, where language delays in combination with certain other comorbid conditions showed heterogeneous relative risks for ASD that were all higher than the relative risk in the cluster of language delay alone. These results underscore the importance of considering combinations of conditions. Global delay sometimes incorporates language delay because global developmental delay is defined as significant delay in two or more areas of development (American Psychiatric Association 2013). Children diagnosed with a global delay who do not have a separate language delay diagnosis may have a language delay as a component. Thus, it is difficult to interpret the higher relative risk for the cluster of global delay with language delay compared to the cluster of global delay without language delay. Our findings of an association between motor delay and ASD are consistent with other recent studies suggesting an increased prevalence of motor impairment among those with ASD as compared to controls (Ming et al. 2007). Notably, our cluster analyses showed an association of motor delay with ASD risk even among children who had no co-occurring language delay diagnosis.

Mental health conditions in the pre-diagnostic period showed very strong associations with subsequent ASD diagnosis in the main analysis. The all-age cluster analysis showed that mental health diagnoses were important in both the absence and the presence of developmental delays. These mental health condition clusters were not identified in the age 3 cluster analysis which was restricted to medical conditions identified before age 3. This finding is not surprising given that typical age-of-onset is 4–11 years for ADHD and 5–15 years for conduct disorders (Kessler et al. 2007). Increased monitoring or screening of children with mental health conditions may be important considering the evidence of (i) the strength and timing of these associations with mental health conditions, (ii) more than 10% of ASD cases were diagnosed with a mental health condition prior to their first ASD diagnosis, and (iii) the cluster of children diagnosed with a language delay and a mental health condition who had no diagnosed global delay had the longest mean latency time in Table 3 (mean = 32.7 months, SD = 25.5). These findings are supported by a recent study of children with ASD, ages 2–17, in which approximately 20% of ASD cases were first diagnosed with ADHD, and on average those children were not diagnosed with ASD until 3 years after their ADHD diagnosis (Miodovnik et al. 2015).

Outside of the domains of mental health and developmental delay, the conditions of ophthalmology; ear, nose and throat; and nutrition defined other condition clusters. Individually, each condition was also associated with a moderate increase in ASD risk. The conditions of constipation, gastroesophageal reflux disease, musculoskeletal abnormalities, and seizures contributed to defining the clusters only when studying conditions diagnosed before age 3. Several conditions reached Bonferroni significance in individual analyses but not cluster analyses, including metabolic, gastrointestinal, genetic, pulmonary, sleep, and neurologic conditions; these conditions tended to have either low prevalence or a weak association with ASD risk (OR < 2).

Several recent studies of ASD have utilized a clustering approach to identify risk factors, but the focus of those studies differs from ours. Doshi-Velez et al. only examined children diagnosed with ASD and sought to cluster underlying patterns that may represent heterogeneous etiologic pathways for ASD (Doshi-Velez et al. 2014). Brennan et al. focused on symptoms identified in screening protocols and used hierarchical cluster analyses to detect underlying clinically distinct subgroups (Brennan et al. 2015). Chawarska et al. studied younger siblings of children with ASD, a high risk population, and used a tree-based method to identify clusters of behavioral features associated with subsequent ASD diagnosis or atypical development (Chawarska et al. 2014). In contrast, our study included children with ASD and age-matched controls and identified condition clusters that indicated increased risk of ASD in the pre-diagnostic period.

Our study has a number of strengths and limitations. Our large and diverse study population across three health systems in different parts of the U.S. contributes to the generalizability of these results. Physician documentation of conditions was recorded prospectively across a broad range of medical conditions prior to ASD diagnosis, preventing recall bias and diagnostic bias. However, the ascertainment of medical diagnoses relies on the documentation in each record, so differential utilization of health services that leads to a differential reporting of conditions by case–control status could bias our results. In addition, there is potential for some misclassification in ASD phenotypic assessment since not all diagnoses were made by experts.

A statistical strength of this study is the Bonferroni control for multiple testing in individual and cluster analyses. A concern in tree-based supervised clustering is model overfitting by creating too many clusters that are not meaningful and cannot be replicated (Strobl et al. 2009). Pruning, a cross-validation procedure, can combat overfitting, but a strength of our study is using a Bonferroni test for the stopping criteria which provides a strong degree of control for overfitting without loss of predictive performance (Hothorn et al. 2006). The main statistical limitation of our analyses is that the matched case–control design prohibits us from calculating an estimate and confidence interval for the absolute risk, thus limiting our statistical inference to ORs (Agresti and Kateri 2011). We cannot directly evaluate the predictive characteristics of this clustering model in terms of the absolute risk of ASD for each cluster; this is a topic for future work. This future work could extend to quantifying the sensitivity and specificity of different screening protocols that might be targeted to particular high risk clusters. Although we have demonstrated risk stratification with a gradient of increasing relative risk across clusters, the clinical utility of integrating the risk stratification with screening protocols would depend on a number of factors including the strength of the association of each risk cluster with ASD and the proportions of cases and controls in each risk cluster.