Abstract
Celiac disease (CD) is a common immune-based condition with diverse clinical manifestations. This study utilized Bing search engine data from the United States to identify symptoms and conditions that precede a celiac-related query (CRQ). By comparing query ratios (QR) before and after an initial CRQ, researchers identified key antecedent symptoms such as diarrhea, headache, and anxiety. While specific symptoms showed increased frequency, predictive modeling (AUC 0.53) highlighted the difficulty of identifying CD based solely on search patterns due to the disease's protean nature.
Introduction
Celiac disease affects approximately 1% of the population, yet the average diagnostic delay in the US remains around 11 years. This delay is largely due to the shift from "classical" presentations (malabsorption, weight loss) to "non-classical" symptoms (anemia, neurological issues). Clinical records often fail to capture the "pre-clinical" phase—the period when individuals experience symptoms but have not yet sought medical advice. This study leverages the "digital breadcrumbs" left in search engine logs to characterize the symptoms that prompt individuals to investigate celiac disease or a gluten-free diet.
Methods
The researchers analyzed anonymized English-language queries from the Bing search engine in the US between January and October 2017.
Cohort Selection
The study identified 90,142 users who performed a "Celiac-Related Query" (CRQ) in October 2017, having made no such queries in the previous nine months. Users were further categorized by "sustained interest"—defined as making a CRQ on more than one day—which served as a proxy for a potential diagnosis or serious health concern (n=6,528).
Analytical Framework
The primary metric was the Query Ratio (QR): the ratio of searches for a specific symptom or condition in the 14 days preceding the first CRQ versus the 14 days following it. A QR > 1 indicates a symptom likely driving the initial interest in CD. Predictive models, including Random Forests, were tested to determine if antecedent searches could identify users who would eventually show sustained interest in CD.
Results and Insights
Symptom Signatures
The study found that both gastrointestinal and extraintestinal symptoms preceded the first CRQ. For the sustained interest group, the top precursor symptoms included:
- Diarrhea: QR 1.34
- Cough: QR 1.28
- Headache: QR 1.27
- Bloating: QR 1.26
Notably, neuropsychiatric symptoms like anxiety (QR 1.10), depression (QR 1.03), and ADHD (QR 1.64) also showed increased frequency before the CRQ, suggesting that "brain fog" and mental health changes are significant triggers for patient self-investigation.
Temporal Trends
The data revealed a distinct temporal pattern. Interest in precursor symptoms typically remained stable until approximately two weeks before the first CRQ, at which point search frequency spiked significantly.
Behavioral Shifts
Following the initial CRQ, users exhibited an immediate pivot in dietary searches. Queries for standard ingredients (e.g., "all-purpose flour") were rapidly replaced by gluten-free alternatives (e.g., "xanthan gum," "gluten-free pie crust"), indicating that the internet search often marks the start of a dietary intervention.
Predictive Modeling Challenges
Despite identifying these trends, the predictive power of antecedent searches was low. The maximum Area Under the Curve (AUC) achieved was 0.53, barely exceeding random chance. This underscores the "protean" nature of celiac disease—its symptoms are so common and diffuse that they do not form a unique digital signature easily distinguishable from other conditions.
Conclusion
Search engine data provides a unique window into the patient journey before clinical diagnosis. While certain symptoms like diarrhea and anxiety are clear precursors to interest in celiac disease, the diversity of these symptoms makes automated case-finding a significant challenge. Future research linking search logs with electronic health records (EHR) may refine these digital phenotypes and help bridge the gap between symptom onset and formal diagnosis.
Summary updated