It would be nice, wouldn’t it, if we could work out which patients would not benefit from an intervention, in order to a) not use it and b) use something (probably more toxic) instead? It’s a frequent thought of mine, as an oncologist, when I sign off another chemotherapy chart with multiple agents on it.
I know that other have the problem too – for instance, those deciding how to treat patients with Kawasaki disease. For some patients the usual treatment of high dose immunoglobulin is ineffective at preventing cardiac artery aneurysm formation. There have been clinical prediction rules developed for this, and in Japanese populations, the Kobayashi score is reputed to be effective. The disease does appear to differ across the world though, and it’s always worth confirming that prediction models do work in different areas.
A group from Imperial College, London, took the opportunity to do this by examining a cohort of patients with Kawasaki disease and reviewed their presenting scores, and demonstrated an absolute lack of predictive ability, with 11/37 (29%) of the ‘high risk’ group being treatment resistant, and 8/22 (36%) of the low risk group failing to respond.
This dismal lack of validation is not uncommon, being seeing repeatedly in many different rules and specialities, and demands that we need to remember that just deriving a predictive or diagnostic rule is not enough – it’s just the first step on a path to decent prognostication.