Assessment of the Frequency of Online Searches for Symptoms Before Diagnosis: Analysis of Archival Data

Irit Hochberg, Raviv Allon, Elad Yom-TovJournal of Medical Internet Research 2020 · OpenReview

TL;DR

While "Dr. Google" is a household name, only about 15.5% of patients actually search for their specific symptoms online before receiving a formal medical diagnosis. This rate varies significantly by condition—ranging from 1.8% for cellulitis to over 30% for angina—suggesting that the effectiveness of search engines as early-warning screening tools depends heavily on how "worrying" or unfamiliar the symptoms are.

The Problem: Quantifying the "Dr. Google" Effect

Public health researchers have long eyed internet search data as a potential "early warning system" for disease outbreaks or individual diagnoses. However, most existing data comes from surveys, which are notoriously prone to recall bias. We know people use the internet for health information, but we don't know the true rate at which they do so in the critical window between symptom onset and a clinical diagnosis.

The authors sought to quantify this behavior using objective archival data. By understanding who searches for what and when, developers can better design "digital phenotypes" or screening systems that might one day nudge a user toward a doctor before a condition worsens.

Methodology

The researchers analyzed anonymized logs from Microsoft Bing (US users) over a one-year period (2017–2018). The study focused on 20 conditions, such as Gout, Lymphoma, and Bell’s palsy, which present clear symptoms but are not typically identified through routine screenings like mammograms.

  1. Diagnosis Ascertainment Queries (DAQs): The team identified users who self-identified their diagnosis by searching for phrases like "I was diagnosed with COPD" or "I have Crohn’s disease."
  2. Back-tracing: Once a user made a DAQ, researchers looked backward in that user's search history to see if they had queried for symptoms associated with that specific condition before the diagnosis.
  3. Severity and Familiarity: They used the Medical Severity Rank (MSR) to quantify symptom urgency and conducted a crowdsourced survey to measure how well laypeople recognize which symptoms belong to which diseases.

Key Insights and Validation

The Drivers of Search Behavior

  • The "Unfamiliarity" Driver: There is a negative correlation (rho = -0.47) between symptom recognition and search frequency. If a symptom is mysterious or doesn't "fit" a known disease in the user's mind, they are more likely to turn to a search engine.
  • Duration Matters: Conditions with a longer "lead time" between the first symptom search and the eventual diagnosis had higher search rates (rho = 0.6). A slow-burning condition gives a user more opportunities to consult the internet.
  • The Worry Factor: Symptoms ranked as more "worrying" (lower MSR scores) were 14% more likely to be searched for. Pain and visible distress drive digital engagement more than vague, mild discomfort.

Timing Validation

To ensure "I have X" queries actually occurred near the time of a real-world diagnosis, the authors checked for searches for "hospitals" or "clinics." They found a clear peak: 67% of hospital searches occurred within two weeks of the diagnosis query, validating the search log as a proxy for clinical events.

Figure 6: Distribution of hospital and clinic searches relative to the day of the self-reported diagnosis query (day 0).

Key Results

  • Average Search Rate: On average, 15.49% of people searched for symptoms before their diagnosis.
  • High Variability: The range was massive. Only 1.8% of cellulitis patients searched for symptoms (likely because it is acute and painful, leading straight to the ER), while 30.8% of those with angina or coronary heart disease searched for their symptoms beforehand.
  • Recognition: Laypeople recognized the symptoms of the 20 conditions about 86% of the time, but some conditions like plantar fasciitis (40%) and COPD (29%) were poorly understood, leading to different search behaviors.

Limitations and Future Work

The study relies on "self-declaration," meaning it only captures users who explicitly type their diagnosis into a search bar. This creates a biased sample—likely younger, more tech-savvy, and more likely to be female. Furthermore, the one-year observation window is relatively short; users who had symptoms for years before the study began might have been missed.

Future work could involve integrating these findings into search engine algorithms to provide "low-intensity" screening. If a user's search pattern matches the "digital signature" of a condition like Lymphoma or Crohn's, the engine could potentially provide more authoritative medical resources or suggest a professional consultation earlier in the process.

Summary updated

Abstract

Background: Surveys suggest that a large proportion of people use the internet to search for information on medical symptoms they experience and that around one-third of the people in the United States self-diagnose using online information. However, surveys are known to be biased, and the true rates at which people search for information on their medical symptoms before receiving a formal medical diagnosis are unknown.Objective: This study aimed to estimate the rate at which people search for information on their medical symptoms before receiving a formal medical diagnosis by a health professional.Methods: We collected queries made on a general-purpose internet search engine by people in the United States who self-identified their diagnosis from 1 of 20 medical conditions. We focused on conditions that have evident symptoms and are neither screened systematically nor a part of usual medical care. Thus, they are generally diagnosed after the investigation of specific symptoms. We evaluated how many of these people queried for symptoms associated with their medical condition before their formal diagnosis. In addition, we used a survey questionnaire to assess the familiarity of laypeople with the symptoms associated with these conditions.Results: On average, 15.49% (1792/12,367, SD 8.4%) of people queried about symptoms associated with their medical condition before receiving a medical diagnosis. A longer duration between the first query for a symptom and the corresponding diagnosis was correlated with an increased likelihood of people querying about those symptoms (rho=0.6; P=.005); similarly, unfamiliarity with the association between a condition and its symptom was correlated with an increased likelihood of people querying about those symptoms (rho=−0.47; P=.08). In addition, worrying symptoms were 14% more likely to be queried about.Conclusions: Our results indicate that there is large variability in the percentage of people who query the internet for their symptoms before a formal medical diagnosis is made. This finding has important implications for systems that attempt to screen for medical conditions.J Med Internet Res 2020;22(3):e15065doi:10.2196/15065
BibTeX
@article{hochberg2020assessment,
  author = {Irit Hochberg and Raviv Allon and Elad Yom-Tov},
  title = {{Assessment of the Frequency of Online Searches for Symptoms Before Diagnosis: Analysis of Archival Data}},
  journal = {Journal of Medical Internet Research},
  year = {2020},
  url = {https://openreview.net/forum?id=SB0RvGIVtQ},
}
Figure