Ten steps towards improving prognosis research

Many thanks to Hemingway et al for this long needed paper although I regret that in the current online version, part
of the Purpose section was apparently truncated.

The authors are sadly to the point when they mention that
prognosis studies are often improvised as the “cherry
on the cake” of a work designed for quite another
purpose. This is true of marker studies that often use
available biological material, literally “in the
freezer”, and of multivariate risk prediction models
which are often a kind of fishing expedition across all
available data items, no matter the nature of their
putative relationship with the outcome.

Another point, that the authors did not raise, is the need
for appropriate statistical methods. Published prognosis
results (I dare not say studies) are too often based on
statistical inference methods, using the “p value” as a measure of predictive value,
although a full range of statistical methods are available
for predictive analysis. This is likely a consequence of a
lack of planning and thus thinking about these analyses.

Last, as somewhat indirectly pointed by the authors, in
medical matters, prognosis is not an aim per se.
The clinical usefulness of a marker or a combination of
thereof should be considered at design time. Should this
marker be used for therapeutic decisions? Might it be used
as a proxy outcome in clinical trials? Such questions are
central to medical prognosis research.

Competing interests:
None declared