Googling for a diagnosis—use of Google as a diagnostic aid: internet based study

Abstract

Objective To determine how often searching with Google (the most popular search engine on the world wide web) leads doctors to the correct diagnosis.

Design Internet based study using Google to search for diagnoses; researchers were blind to the correct diagnoses.

Setting One year's (2005) diagnostic cases published in the case records of the New England Journal of Medicine.

Cases 26 cases from the New England Journal of Medicine; management cases were excluded.

Main outcome measure Percentage of correct diagnoses from Google searches (compared with the diagnoses as published in the New England Journal of Medicine).

Results Google searches revealed the correct diagnosis in 15 (58%, 95% confidence interval 38% to 77%) cases.

Conclusion As internet access becomes more readily available in outpatient clinics and hospital wards, the web is rapidly becoming an important clinical tool for doctors. The use of web based searching may help doctors to diagnose difficult cases.

Introduction

Doctors adept at using the internet use Google to help them diagnose difficult cases. As described in the New England Journal of Medicine,¹ a doctor astonished her colleagues (including an eminent professor) by correctly diagnosing IPEX (immunodeficiency, polyendocrinopathy, enteropathy, X linked) syndrome. She admitted that the diagnosis “popped right out” after she entered the salient features into Google.

It seems that patients use Google to diagnose their own medical disorders too. After evaluating a 16 year old water polo player who presented with acute subclavian vein thrombosis, one of us (HT) started to explain that the cause of the thrombosis was uncertain when the patient's father blurted out, “But of course he has Paget-von Schrötter syndrome.” Having previously googled the symptoms, he gave us a mini-tutorial on the pathophysiology (hypertrophy of the neck muscles leading to dynamic compression of the axillary vein at the thoracic inlet—leading to thrombosis) and the correct treatment of the syndrome.² This experience led us to ask: “How good is Google in helping doctors to reach the correct diagnosis?”

Method

We selected a convenient sample of one year's (2005) diagnostic cases presented in the case records of the New England Journal of Medicine. We excluded management cases. After discussion, we selected three to five search terms from each case record and entered them on a data sheet. We then did a Google search for each case while blind to the correct diagnoses (that is, before reading the differential diagnosis and conclusion of each case record). We selected and recorded the three most prominent diagnoses that seemed to fit the symptoms and signs. We then compared the results with the correct diagnoses as published in the case records.

Results

We identified 26 cases from the case records (table 1). Google searches found the correct diagnosis in 15 (58%, 95% confidence interval 38% to 77%) cases. In some cases (for example, case record 9), Google gave the correct diagnosis (extrinsic allergic alveolitis) but we felt that it was not specific enough to be considered correct (extrinsic allergic alveolitis caused by Mycobacterium avium, also known as “hot tub lung”).

Google diagnoses and actual diagnoses for 26 case reports

Case record	Google diagnosis	Final diagnosis	Google diagnosis correct?
5	Infective endocarditis	Infective endocarditis	Yes
6	Gastrointestinal bleed	Linitis plastica with bowel obstruction	No
7	Cushing's syndrome	Cushing's syndrome secondary to adrenal adenoma	Yes
8	Eosinophilic granuloma, osteoid osteoma	Osteoid osteoma	Yes
9	Extrinsic allergic alveolitis, tuberculosis, BOOP	Hot tub lung secondary to Mycobacterium avium	No
10	Amyotrophy	Ehrlichiosis	No
11	Tuberculosis, lymphoma	Lymphoma	Yes
12	Neurofibromatosis type 1	Neurofibromatosis type 1	Yes
14	Uveitis	Vasculitis	No
15	Amyloid	Amyloid light chain	Yes
16	Hyperaldosteronism	Phaeochromocytoma	No
17	Acute chest syndrome	Acute chest syndrome	Yes
18	Tuberous sclerosis	Endometriosis	No
19	Aspergillus	Aspiration pneumonia, brain abscess	No
22	Graft versus host disease	West Nile fever	No
25	Cirrhosis	Pylephlebitis	No
26	Hypertrophic obstructive cardiomyopathy	Hypertrophic obstructive cardiomyopathy	Yes
27	Spongiform encephalopathy (Creutzfeldt-Jakob disease)	Creutzfeldt-Jakob disease	Yes
28	Churg-Strauss syndrome	Churg-Strauss syndrome	Yes
29	Polymyositis or dermatomyositis	Dermatomyositis secondary to non-Hodgkin's lymphoma	Yes
30	Cat scratch disease	Cat scratch disease	Yes
31	Henoch-Scholein purpura	Cryoglobulinaemia	No
33	First hit=juvenile polyposis plus HTT, which links to MADH4 mutation	MADH4 mutation (HTT plus juvenile polyposis)	Yes
34	Toxic epidermal necrolysis syndrome	Toxic epidermal necrolysis syndrome	Yes
36	Encephalitis	MELAS	No
37	Long QT syndrome, Brugada syndrome	Brugada syndrome	Yes

BOOP=bronchiolitis obliterans organising pneumonia; HTT=hereditary haemorrhagic telangiectasia; MELAS=myoclonus epilepsy lactic acidosis stroke-like syndrome.

Discussion

Clinical decision support programs have been reported to be valuable aids in diagnosing difficult cases.³ Hoffer reported using a clinical decision support program to make the diagnosis of Addison's disease expeditiously when it was missed by many expert clinicians.^{4 5} We think that Google is likely to be a useful aid in diagnosis too. It has the advantage of being easier to use and is freely available on the internet.

A few limitations of this study should be mentioned. Arguably, everything could be found on the web if only one knew the correct search terms. In this case, we chose combination of search terms that we felt would be unique (see extra table on bmj.com). We chose between three to five search terms for each case, depending on symptoms and signs that we felt would not return a non-specific result. We selected “statistically improbable phrases” whenever possible,⁶ such as “cardiac arrest sleep” in case record 37. We generally selected likely diagnoses from the first three pages (maximum five pages) of the search result, containing 30 documents, to see if the condition would fit the case record. As Google does not “suggest” a diagnosis, we selected the diagnosis that we felt would fit best with the case record. When none of the diagnoses found with Google fitted the case record well, we chose up to three most likely diagnoses. If one of the diagnoses was correct, we regarded the search as successful.

We suspect that using Google to search for a diagnosis is likely to be more effective for conditions with unique symptoms and signs that can easily be used as search terms, such as the one described by Greenwald.¹ Searches are less likely to be successful in complex diseases with non-specific symptoms (case records 10 and 14) or common diseases with rare presentations (case record 18).

The efficiency of the search and the usefulness of the retrieved information also depend on the searchers' knowledge base. In this case, although we were blinded to the correct diagnosis, one author was a respiratory and sleep trainee and the other a rheumatologist; sometimes the diagnoses were evident to us, and this could have affected our choice of search terms. When choosing the “correct” diagnoses from a list of possible choices returned by Google, we tried to avoid using specialist knowledge but chose diagnoses that were ranked most prominently and seemed to fit the case record. Therefore, for case record 9, where we made the correct diagnosis of “hot tub lung,” searching with Google did not give enough prominence to hot tub lung for it to be considered the correct answer.

Patients doing a Google search may find the search less efficient and be less likely to reach the correct diagnosis. We believe that Google searches by a “human expert” (a doctor) have a better yield, as Google is exceedingly good at finding documents with co-occurrence of the signs/symptoms used as search terms and human experts are efficient in selecting relevant documents. Furthermore, doctors in training would find the Google searches educational and useful in formulating a differential diagnoses.

The role of diagnostician remains one of the most challenging and fulfilling roles of a physician. Physicians have been estimated to carry two million facts in their heads to fulfil this role.⁷ With medical knowledge expanding rapidly, even this may not be enough. Search engines allow quick access to an ever increasing knowledge base.⁸ Google gives users ready access to more than three billion articles on the web⁹ and has far exceeded PubMed as the search engine of choice for retrieving medical articles.¹⁰ Google has been so popular that the word has entered the English lexicon as a verb.¹¹ Google Scholar, currently in beta form (www.scholar.google.com), is likely to be even more useful as it searches only peer reviewed articles.

Conclusions

Doctors and patients are increasing proficient with the internet and frequently use Google to search for medical information. Twenty five million people in the United Kingdom were estimated to have web access in 2001, and searching for health information was one of the most common uses of the web.¹² Computers connected to the internet are now ubiquitous in outpatient clinics and hospital wards. Useful information on even the rarest medical syndromes can now be found and digested within a matter of minutes. Our study suggests that in difficult diagnostic cases, it is often useful to “google for a diagnosis.” Web based search engines such as Google are becoming the latest tools in clinical medicine, and doctors in training need to become proficient in their use.

What is already known on this topic

• Doctors and patients are increasingly using the internet to search for health related information

• Google is the most popular search engine on the world wide web

What this study adds

• Searching with Google may help doctors to formulate a differential diagnosis in difficult diagnostic cases