In 2011 I posted:

“An answer engine.  For commercial reasons I have to be vague on this for now (something I’m not comfortable with) but it’ll take shape over 2012.“

I have revisited the idea many times in the last five years, always trying to solve the same problem: a search engine doesn’t answer questions, it gives you 10-20 results to articles that may answer your question.  I borrowed the term ‘answer engine’ from Steve Wozniak (co-founder of Apple) who used it in relation to the release of Siri.  He said people want and need an answer engine.

Until recently the problem has been scalability – how to extract the answers from the great content held in Trip.  But, due to my involvement in various machine learning initiatives, the missing link(s) has been found and we’re busy developing an ‘answer engine’.  Our initiative works on a number of connected problems areas:

  • From a set of search terms can you infer the likely question?
  • Can we find an answer to that likely question?
  • Given there will be potentially multiple answers in Trip, can we surface the best answer?

We have made great strides in all three areas.  So much so we are currently working on the second iteration of our answer engine.  For the test it has two search boxes based on PICO.  The first box (P) represents the disease and the second box is the (I) intervention of interest. To illustrate what it does is using the P of ‘sore throat’ and I of ‘antibiotics’ – the answer engine returned:

“Antibiotics appear to have no benefit in treating acute laryngitis. Erythromycin could reduce voice disturbance at one week and cough at two weeks when measured subjectively. We consider that these outcomes are not relevant in clinical practice. The implications for practice are that prescribing antibiotics should not be done in the first instance as they will not objectively improve symptoms.“

The above is a great answer.  But we can’t rely on a single test so we’ve been extensively testing it and the current version gets the following results:

  • Fail – 19%
  • Partial pass – 32%
  • Pass – 49%

So virtually a 50% success rate, which impressed me!  But the biggest reason for the partial passes is our system not pulling through the answer (a relatively simple fix) and the biggest reasons for the failures was the inability to exclude additional terms which confused the answer (again, this should be a relatively simple fix).

Our system should be great for a number of reasons:

  • It can integrate seamlessly with the existing Trip interface but also act as a standalone product/app!
  • It can easily be integrated with our multi-lingual systems to users will be able to search and obtain answers in languages such as French, German and soon to come languages such as Spanish.  No English will be needed!
  • It will always be as up to date as the answer are based on all the evidence in Trip.
  • It is modular and will launch with a focus on therapeutics.  It will then expand to include medicines information (eg side-effects, interactions, contraindications) and from their I’d like to tackle clinical guidelines or lab tests (depends on the resource availability).
  • A user can get an answer in less than 5 seconds, when previously they would have had to scan the first 10-20 results to see which result was most likely to answer their question.  Assuming they get the best article they still need to read/scan it for the answer.  So, considerably longer.

 This is brilliant, it really is…!