The search task within the TREC 2011 Medical Records track 
was an ad hoc search task that modeled the clinical
task of finding cohorts for comparative effectiveness research.
The document set used in the track was based on a set
of de-identified clinical reports made available by the
University of Pittsburgh's BLULab NLP Repository.  This document
set is not available to non-participants.   There
is a many-to-one mapping between reports and ''visits'',
where a visit is an individual patient's single stay
at a hospital.  The visit was used as the unit of retrieval
in the track (so a document for the purposes of the track
was the union of the content in all records mapped to that visit).
This mapping of report ids to visit ids is posted
on the web site as "UnivOfPittReportMappingToVisit.txt".

Thirty-five topics, numbers 101--135, were created for
the track (file topics101-135.txt).  However,
the 2011 evaluation was over 34 topics: topic 130
was dropped because it has no relevant visits.
The qrels file contains three-way judgments of
   0:  not relevant
   1:  partially relevant
   2:  relevant
The judgment sets for the track were built using
a sampling technique compatible with computing
extended inferred measures (see "A Simple
and Efficient Sampling Method for Estimating
AP and NDCG" by Yilmaz, Kanoulas, and Aslam
in SIGIR 2008).  A judgment file that includes
the strata at which each document was sampled
is posted as fullsample-qrels.gz.  A qrels
file compatible for trec_eval is posted as
qrels.txt.