Using an efficient and largely automated approach, we were able to readily identify a selected group of individuals who were seeking information on recent symptoms of TIA/stroke online–sometimes very soon after symptom onset and often before a formal medical evaluation had been completed. Although participants in our study are unlikely to be fully representative of patients with TIA generally, or patients seeking online information on TIA specifically, we did identify a subpopulation with a relatively high burden of cerebrovascular disease that could benefit from a targeted public health intervention. Given the low cost and the substantial efficiency of this approach in terms of cost per participant reached, and since a major goal of acute management of TIA is urgent evaluation to prevent early recurrent stroke, this subpopulation may present an attractive target for interventions designed to reduce delays in presenting to medical attention for TIA/stroke.
A secondary goal of our study was to assess the performance of a self-reported version of ABCD2
score for TIA. The ABCD2
score was originally developed to assess the risk of stroke after TIA using simple elements available to front-line clinicians, 
but this score has also been shown to work in part by distinguishing TIA from TIA mimics. 
Here, we faced the additional challenge from the measurement issues that arise from using self-reported responses from individuals with various levels of health literacy. Accordingly, we found that a modified ABCD2
score by self-report could not reliably rule out a TIA or stroke. Some participants (2 of 12) with the lowest score of 0 had symptoms such as diplopia or amaurosis fugax that are not specifically captured by the score, but would still require urgent evaluation.
With regards to the performance of other self-reported items to predict a cerebrovascular diagnosis, we found a strong association between a higher number of episodes of similar symptoms within the last 12 months and a non-cerebrovascular diagnosis. Having multiple stereotyped episodes of symptoms previously without a resultant stroke may make a given episode less likely to be TIA and may make seizure, migraine or other episodic disorders more likely. 
The predictive value of this finding will require further validation and confirmation of the dose-response relationship that is suggested by our data.
Since such a high proportion of participants in our study had TIA or stroke, we speculate that all of the individuals that participated in our study may have been well served to seek medical attention regardless of risk score, particularly since the online questionnaire was unable to reliably identify low-risk individuals. Some individual participants indicated that they lacked a primary care provider or medical insurance or otherwise had limited access to medical care, while others indicated that they did not understand that TIA was a medical emergency, or were unable to get an urgent appointment, but had already planned on seeking medical attention at some point. The value of some form of risk stratification at the outset may not be to identify those without worrisome symptoms, but to provide some interactivity and engagement in order to enhance behavior change while participants are deciding on next actions. We suspect that if individuals are given feedback and recommendations that are to some extent individualized and interactive rather than presented as static blanket recommendations, they may be more likely to take action after TIA, though a formal assessment of this possibility would require additional study.
Our methods also highlight opportunities to improve the efficiency of clinical research more generally by quickly and efficiently reaching hundreds of potential participants on a daily basis and by enrolling participants over a large geographic area. Direct-to-participant recruitment portals that incorporate initial eligibility screening by self-report may demonstrate a sufficiently high yield for TIA and stroke diagnoses to serve as a primary recruitment method for some types of clinical research. Furthermore, our study frames the potential role for validated and interactive self-assessment online tools address public health problems more generally. A previous study have demonstrated a relationship between internet search queries and influenza epidemics 
and another study showed that the frequency of internet search queries for stroke related search terms by state (“stroke signs”, “stroke symptoms”, “mini stroke”, excluding “heat”) was associated with stroke prevalence. 
There remain potential medicolegal issues including whether providing tailored or personalized information constitutes medical advice, and valid concerns about maintaining privacy and confidentiality of health information online. But as disparities in access 
and adoption of internet technology across demographic and sociodemographic strata narrow, 
internet-based interventions to efficiently target public health problems may become more attractive. Future studies may include questionnaires in the ED regarding internet-based activity prior to presentation in the ED and evaluations of the impact of internet-based tools on the actual the behavior of subjects with possible TIA and ultimately on processes measures of treatment and patient outcomes.
Our study should be interpreted in light of a number of limitations. First, about 1% of visitors to the website enrolled in the study and so the participants in this study were highly selected. Although the motivations for enrollment were not captured in our study, many visitors quickly moved from the website within a few seconds, which suggests that they were not finding what they had expected or what they were looking for. We speculate that individuals with symptoms without a straightforward diagnosis despite prior medical evaluations, those with chronic medical conditions, or those with barriers to accessing care might be overrepresented in our study. Second, the authentication of participants and their responses and the difficulty of reaching some participants by telephone limit the generalizability of our results and illustrate a particular challenge for internet-based research. Third, we attempted to mitigate misclassification of TIA and stroke outcomes by formally assessing interrater reliability and masking the neurologists to questionnaire responses, but since follow-up was conducted by telephone and access to additional clinical information such as neuroimaging studies or results of a neurologic exam was limited to self-report the potential for misclassification remains. For the lay-public, the use of the term, “mini-stroke,” a much more common search term than “TIA”, serves to confuse the distinction between TIA and stroke and the clinical distinction between TIA and minor stroke may be difficult to make even for experienced clinicians. 
Fourth, although the neurologists were masked to responses to the questionnaire, they were free to elicit any information into their final assessments and may have incorporated items that were elicited by the online questionnaire into their assessments.
Individuals seeking information on possible TIA or stroke symptoms online can be efficiently and readily identified, sometimes before they have sought formal medical attention. Although elements of a self-reported ABCD2 score was not able to rule out a cerebrovascular etiology, the burden of true stroke and TIA in this selected subpopulation may be sufficiently high to justify future targeted interventions to encourage urgent medical evaluation.