Tuesday, March 28, 2006

Isabel versus Dxplain in a case of giant cell arteritis with atypical manifestations

A while back I mentioned two web based differential diagnosis resources, Isabel and Dxplain. I have access to Dxplain through Merck Medicus and signed up for a free trial of Isabel, so I decided to plug a challenging case into both programs and compare their performance.

The CPC in the February 9 2006 NEJM featured an uncommon presentation of giant cell arteritis dominated by chest pain and respiratory complaints. Myalgias were also present, mistakenly attributed to statin drugs. The following findings were entered into both programs: chest wall pain; ear pain; shoulder pain; myalgias; hoarseness; exertional dyspnea; fatigue; sore throat. Age (elderly) and gender (female) were entered. Although both programs encourage entry of laboratory results you can enter as much or as little data as you wish, and I decided to see how the programs did with only the initial presenting clinical manifestations. I opted not to enter the patient’s sedimentation rate of 90. That would have made it too easy.

Isabel missed the boat entirely with a long list of diagnoses across multiple specialties, with the only rheumatologic diagnosis offered being relapsing polychondritis. Dxplain fared better by coming up with polymyalgia rheumatica. Neither program diagnosed giant cell arteritis or temporal arteritis. So, Dxplain wins the first round. As time permits before my free Isabel trial expires I’ll try more cases and report back.

1 comment:

Anonymous said...

RW, I lead the team that developed Isabel. The ISABEL-Dxplain comparison [NEJM CPC case of Giant Cell arteritis ] is a good idea but you need to understand fully how Isabel works and how it is different from Dxplain.

Dxplain is a rules-based system [this was the prevalent methodology in the eighties when Dxplain was developed - clinical features were manually linked to diseases and given evoking strengths]. Isabel, on the other hand, uses natural language processing software to search pages of textbooks through a taxonomy [of over 11000 diagnoses and 4000 drugs] - a bit like how Google searches web-pages. Like Google, Isabel is fast and easy to use. The more clinical features you enter into Isabel the ‘tighter the spread’ [more useful] of the differential diagnosis. There is no Oracle sitting behind the screen and therefore holding back a clinical feature like elevated ESR yields a different search & result on Isabel. If you were looking for John Smith MD from Portland, Oregon on Google and you held back on entering Portland, Oregon – the result will be different and less useful.

Mark Graber MD, Chief of Internal Medicine at the VA did a more elaborate study in November 2005 using 50 CPC cases from the NEJM. Mark found that when they cut and paste the CPC case [excluding the discussion, of course] into Isabel, Isabel came up with the final diagnosis in 74 % cases and when they extracted the clinical features Isabel came up with the diagnosis in 96%. For best results we recommend that clinical features are extracted, avoid repetition and abbreviation and interpret numeric values.

I extracted all the clinical features and entered them into Isabel [each on a separate line- myalgia|fatigue|chronic pain in the morning|sore throat|increased proximal muscle weakness|tenderness with abduction of shoulders|pain over lateral aspect of chest wall extending to rib cage|high ESR|inspiratory crackles at both lung bases|pitting edema of the lower limbs|patchy bibasilar atelectasis|hoarseness|exertional dyspnea]– Giant Cell Arteritis [GCA] was one of the differentials. I also cut and pasted the case – except for the discussion into Isabel and Isabel came up with the diagnosis of GCA.

Given the key differences of how they work I'm up for you testing Isabel out again. Any interest RW?