This study utilized the Yahoo Answers platform to examine queries of parents suspecting their child has ASD. On the Yahoo Answers platform, queries are posted using natural language, answers are submitted by users, and a community forms around this interaction. A query can elicit multiple answers; one of them is rated the best answer, either by the asker or by the community. Yahoo Answers queries—rather than search-engine queries—were selected for the purpose of this study, as they consist of anonymously posted queries on a public platform and hence are more likely to represent parents’ true need for an answer.

We extracted all English-language queries from Yahoo Answers that were submitted between 6 June, 2006 and 12 December, 2013 and contained the words autism, Asperger, ASD, or PDD. A total of 8681 queries met these criteria. We used crowdsourcing [35] (using CrowdFlower) to differentiate between queries posted by parents of a child diagnosed with autism (n=2412), queries posted by parents of a child diagnosed with autism whose sibling they suspected might have autism (n=41), queries posted by parents who suspected their child might have autism (n=1081), and queries posted by parents who did not match any of the above descriptions (n=5147).

Of the 1081 queries of suspecting parents, 195 were randomly selected and analyzed for this study. Among these, in 96.4% (188/195) of the queries that mentioned the age of the child, it was, on average, 3.2 years (SD 2.9; range 1.25 months-18 years; 60.3% [114/189] below 3 years of age). Boys were the subject of 84.7% (160/189) of the queries, girls were the subject of 15.3% (29/189), and the gender of the child was not specified in the rest.

In 4.6% (9/195) of the queries, parents reported a family history of ASD. In 30.8% (60/195) of the queries, parents did not mention reporting their concern to a health care provider.

The content analysis of Yahoo Answers queries was conducted using the NVIVO software. Two types of content analysis processes were applied:

First, one type of content analysis was used to rate a child's risk of ASD as either low, medium, or high. To this end, a set of rules was devised for defining levels of ASD-risk from text. Medium-risk was defined by concerns related to one type of ASD-specific sign, general description of developmental delay, non-ASD–specific concern requiring evaluation and mentioning a risk factor for ASD such as a family member with ASD but no ASD-specific sign. For example:

High-risk was defined as concerns related to at least two types of ASD-specific sign, 1 from the RRBI domain and another from the Social and Communication domains. High-risk rating also considered the severity of the described signs and urgency expressed by the parent:

Low-risk was defined, by default, as queries that did not meet the above criteria, for example:

Then, 2 clinical experts in ASD separately rated the risk level of children described in 38 (19.49%) queries and reached kappa of .72 in their differentiation of at risk queries. Finally, ASD risk in the remaining queries was rated by 1 clinical expert.

Second, another process of content analysis was conducted to identify the types of warning signs noted by parents. This process, which involved deductive and inductive analysis methods, was conducted by a different clinical expert. The deductive method implied coding the warning signs according to domains and subdomains that match DSM-V criteria for ASD. An inductive method was used to identify concerning signs that did not correspond to the DSM-V diagnosing criteria such as describing cognitive impairment or language delay. The resulting taxonomy of the concerning signs mentioned in parents’ queries consisted of 12 domains and 72 subdomains (Multimedia Appendix 1). The first author and clinical expert coded 38 queries and obtained an inter-rater agreement as measured by significant kappa values between .54 and 1 across subdomains, with .71-1 values for the 12 domains. Due to challenges in agreeing on the number of manifestations pertaining to a single type of warning sign, each subdomain of sign (ie, the lowest hierarchy level within a subdomain) was coded for its presence in the query, regardless of the number of times that type was mentioned in the query (eg, “rocks, swings and sways body” is coded once for rocking). While coding, the clinical expert and the first author discussed coding dilemmas and refined the coding rules accordingly. To summarize, each query received an ASD global risk score and was coded for either presence or absence of each sign domain and its subdomains.

Data regarding the age and gender of the child were extracted using a combination of automated text analyses, followed by crowdsourcing rating to correct for errors. The sample with available age data was divided into 4 age groups: 0-2 (n=62), 2-3 (n=52), 3-6 (n=58), and ≥6 years (n=18).

Only signs with an occurrence of more than 5% were included in the analyses. Alpha level (Type 1 error) was corrected using Bonferroni for multiple comparisons, such that the threshold P value (.05) was divided by the number of comparisons. The sign domains differentiating the 3 ASD risk groups were analyzed using chi-square tests, given their dichotomous nature and Fisher’s exact tests for pairwise comparisons. The null hypothesis was that the distribution of sign domains across groups is equal. The association between the child’s gender and age group and the ASD-risk level was determined using chi-square tests. The length of a query differed significantly between ASD-risk groups (F_2,192=10.93, P<.001; mean number of words associated with each risk level was as follows: low=132.57, medium=225.32, high=281.62). Bonferroni post-hoc tests indicated that the number of words was significantly (P<.05) higher in the high-risk group relative to the low-risk and medium-risk groups.

Data analysis was then conducted to assess whether children at risk of ASD could be detected from the queries using the text-coding method devised. The goal of this analysis was to differentiate between low-risk, medium-risk, and high-risk queries. To this end, each query was analyzed to determine (1) the number of times each word or word pair (bigram) appeared in the text, and, separately, (2) the warning signs, coded according to domains and subdomains, as previously explained.

Additional attributes analyzed included the child’s age and gender, the number of words in the query, as well as the length of the query, as measured by the number of characters in it. We trained a linear classifier [31] (with a least-squares criterion) and estimated its performance in terms of the receiver operating curve (ROC) using Leave-One-Out, that is, for each query i, we trained a classifier using all other queries, and tested on the i-th query [36].

Our study examined the nature of online queries of parents who were concerned that their child might have ASD. The Internet offers parents a venue for expressing and verifying their concerns anonymously at any time and place. The analysis of these narratives highlights signs that alert parents, in the general public, of the possibility of their children having ASD. These concerns mirror parental developmental knowledge, awareness, and expectations, as well as levels of parenting anxiety. Most of the Yahoo Answers queries were judged by clinical experts as reflecting high-risk for ASD or as medium-risk, validating parents’ concerns. Differentiating ASD early in life is important, so that the child with ASD can gain the most from targeted interventions [37]. Parental concerns expressed online offer a new method for facilitating earlier screening and referral for evaluation. In the long run, an online tool which provides a gross estimate of ASD risk based on textual descriptions of warning signs and guided questions could prompt parents to approach a professional. Such a tool can harness social media to support worried parents and minimize the risk of acting on nonprofessional advice or disregarding worries.

Our findings that parents with mostly no family history of ASD (95.4% [186/195]) associated a broad range of ASD-specific signs with the disorder were encouraging. The prevailing signs that parents found worrisome were within the DSM-V [12] ASD core domain of RRBI, followed by the social-communication domain. Nevertheless, more than a third of the parents sampled, expressed concerns related to domains of language, cognitive, and emotional, which according to the DSM-V are non-ASD–specific; however, they are highly prevalent in ASD. Other non-ASD–specific concerns mentioned in some of the queries related to motor development, ADL, and medical conditions. Interestingly, concerns related to motor problems entered as a meaningful domain in the decision-tree classifier. The mix of concerns from ASD and non-ASD–specific domains is consistent with previous studies documenting the types of first parental concerns of children later diagnosed with ASD [9,10,15].

The average child age at which an online concern was raised was 38.03 months, close to the average age of ASD diagnosis [38,39]. Nonetheless, this age is greater than the average age of the first concerns that was reported in ASD research, which is 10-18 months [5,9,19,21,24]. The majority of queries described sons, although analysis of queries did not reveal different types of signs for sons versus daughters, and the rate of ASD risk was not significantly higher for boys than for girls. Comparisons between age groups showed that language concerns differed in prevalence across age groups. Language problems were mentioned most frequently in queries regarding children 0-3-year-old and less in the oldest group. It is likely that parents of a nonverbal 7-year-old have reached the stage beyond suspicion and hence would be less likely to query this. Note that the peak of emotional concerns was in the 3-6 years old age group, particularly descriptions of outbursts or extreme shyness. Social concerns increased with highest prevalence in the older group. These differences between age groups can be explained in the light of the changes in developmental expectations of parents from children at different ages.

At the extremes, there were parents who raised concerns too early to determine risk, for example:

At the other end of the spectrum, there were parents of older children who were either questioning a non-ASD diagnosis their child received or were never evaluated but always felt something was different, for example:

When designing an online screening tool, the age of the child must be considered and, based on the findings, the need to continue monitoring such a child is warranted.

Although there were differences in rates of RRBI concerns across risk groups, they were mentioned in at least 75.4% [147/195] of the queries. The RRBI domain included the largest number of hierarchy levels of subdomain coding as well as individual signs (the subdomains are: repetitive movements, stereotypes and repetitive use of objects, sensory issues, unusual and narrow interests, repetitive speech, eating, difficulty with change, rigid thinking, and rituals). This reflects the diversity in types of symptoms described in the DSM-V criteria. The most prevalent RRBIs mentioned were repetitive movements, repetitive use of objects, and sensory abnormalities. The RRBIs characterizing queries of medium- or high-risk were repetitive speech, sensory issues (particularly tactile over-reactivity), unusual use of objects, repetitive speech (particularly idiosyncratic language), and repetitive interest. Interestingly, research shows that RRBIs are not the most prevalent first concerns of parents of children later diagnosed with ASD [18,22]. It may be the most reported domain in online queries, as parent’s attention is more easily drawn to atypical behaviors or socially inappropriate behaviors, compared with their ability to recognize a delay in attaining a milestone. Signs within this domain can be intense and can interfere with play and participation, and thus are noticeable. The presence of some RRBIs in typical development (eg, head banging, noise sensitivity) presents a further challenge for relying on this domain to verify risk status. There is a need to increase parental awareness of the typical manifestations of RRBI during the first 3 years of life, to help parents understand when such concerns may be warranted. Online concerns reflect a parent’s call for help regardless of whether a child has ASD. Therefore, there is clearly a need for interactive parenting education materials aimed at interpreting and coping with RRBIs.

Our results indicated that it is possible to predict risk from the text using machine learning methods once the text is classified into sign domains, whereas using text alone provided insufficient information (at least in our corpus) for accurate identification of children at risk. Developing an ASD-specific flowchart into which parents could insert their narratives related to certain types of concerns may provide a basis for a more accurate prediction of ASD risk from text. An automated screening tool in online forums will benefit from starting with a social concern question, and if not present then ask about communication while if present ask about the presence of RRBIs. While ASD specific questions will need to dominate such a tool, probing about cognitive and motor markers, which are not ASD-specific, is also important. Results from the decision tree indicate that the combination of signs from the social, RRBI, and motor domains predicted the highest likelihood of ASD risk from the coded text. This is in line with the evidence showing that parental concerns pertaining to a combination of several domains predicted an ASD diagnosis [3]. The prediction of high risk from text or text combined with coded signs was better for the high ASD risk group alone rather than predicting for both medium- and high-risk groups. Future research relying on a larger corpus could robustly test different combinations of concerns in predicting ASD risk. As parents seek first-level support online, developing Internet- and mobile-health tools to automate ASD screening relying on the decision-tree classification described in this study may reduce screening time, increase response to screening, and increase accuracy.

The distribution of the types of signs mentioned in online queries did not fully correspond to the signs most commonly reported in previous studies. For instance, in an ASD study, parental language and communication concerns were found to be the most prevalent early concerns, followed by social, RRBI, medical, and emotional domains [15]. Communication signs were highly prevalent among parental concerns noted in other studies as well; however, they did not consistently differentiate those later diagnosed with ASD [3,9,10,17,19,20,22]. These differences in the most common ASD concerns of parents may be explained by the fact that previous studies relied on retrospective reports and samples characterized by high genetic risk for ASD [3,15,19,22,25], as opposed to a sample with little familiarity with ASD symptomatology and a lower likelihood for the child to have ASD.

The limitation of the current exploratory study is the lack of clinical testing for the actual ASD-risk status of the child as opposed to other neurodevelopmental disorders. The next step would be to study the external validity of risk status using standardized developmental measures, develop a structured format for parents to enter their concerns and, test our algorithm for predicting risk from text with a new corpus. Extracting signs from text has its own limitations, as in some cases it requires the clinical inference of meaning, as opposed to the ability to probe a parent and thus extract the meaning behind the concern. We attempted to minimize such bias by developing coding rules and testing coding reliability. The anonymous nature of online queries makes online parental concerns a unique resource that offers an authentic snapshot of parental ASD-related concerns, unaffected by issues of social desirability or other emotional biases.

Early parental concerns constitute a valuable component of early childhood screening. There is accumulating evidence that early parental concerns regarding specific ASD markers are associated with a higher likelihood of an eventual ASD diagnosis [3,15,40]. Our study is the first to investigate online queries describing the types of signs that lead parents to suspect ASD in their child. The fact that the clinical experts found that the majority of queries corresponded to either medium- or high-risk for ASD validated the need to facilitate the parents’ earlier consultation with a professional. We showed the potential of utilizing machine learning methods for ASD screening based on parental concerns. Findings also highlight the need for designing parent-education tools regarding behaviors that are age appropriate, particularly those pertaining to the RRBI domain. Finally, it is important to empower parents’ confidence in their concerns and increase their awareness of the disadvantages of relying solely on an online community for determining ASD-risk status. This study’s findings support the call for health care providers to closely listen to parental ASD-related concerns, as recommended by screening guidelines [13]. Results also demonstrate the need for Internet-based screening systems that utilize parents’ narratives combined with a hierarchical screening questioning. Worried parents approach online communities, comprised mostly lay person answerers, to obtain opinions regarding their child’s likelihood of having an ASD diagnosis. A more efficient mechanism for supporting worried parents online is important for prompting a clinical evaluation when needed and reducing parental anxiety.