Organize all the world's medical information
--------------------------------------------

Fay Chang, Muthian Sivathanu, Rob von Behren, and Deborah A. Wallach


The field of medicine is fundamentally driven by careful observation and
analysis of example occurrences.  Thus, creating more opportunities for
such observation and analysis will enable significant advancements in
medicine.  We believe that large distributed computer systems have the
potential to help achieve this goal in two key ways.  First, they can
provide a comprehensive, reliable, and secure store of all the world's
medical data that is globally accessible and usable.  Second, and more
importantly, massive distributed systems can employ techniques such as
statistical machine translation to automate the "observation and analysis"
part, again potentially resulting in new insights and advancements.

Although a unified repository of all medical information is quite valuable
in itself, analyzing such large amounts of data manually is a daunting
task.  The most interesting application of distributed systems technology
in this scenario would be to mine this information automatically to derive
new truths and insights about correlations between symptoms, geographic
conditions, life patterns, and diseases or health conditions, but even
allowing mining through manual searches would be a key enabler for medical
research.

The structure of such a repository presents interesting challenges.
For example, medical institutions may need to retain ownership of the
data they generate in order to protect the privacy of their patients.
This necessitates a fundamentally distributed solution, since all data
cannot belong to a single domain of administrative control.  Accessing
and analyzing information in this environment requires basic advances
in distributed information security and anonymity techniques.

Achieving knowledge extraction from a large corpus of medical data is
challenging for a variety of reasons.  However, we believe that large
amounts of data coupled with enough computing power can circumvent
fundamentally hard problems by converting them into brute force searches on
data.  Investigating this line of research will require collaboration
between the systems community and the machine learning/data mining
community.  We will also need good abstractions and platform support that
facilitate building such massive distributed applications for statistical
machine learning and brute force search.

This challenge can be broken down into several concrete sub-problems, and
then progressively attacked on several fronts.  Possible steps include:

 - A system that provides data correlation capabilities across a number of
   types of medical information records all taken from the same medical
   center.  Although this sounds simple, records entered by different
   doctors may include different abbreviations for the same symptoms, or
   some records may be input as text while others are transcribed either
   through voice recognition or manually from voice input, etc.

 - A system that provides security guarantees at the proper granularity,
   such that no individual patients can be identified, yet retains enough
   detail to be able to spot geographical or familial trends.

 - A system that would be robust across records from multiple sources, yet
   all still in a single language.

 - A system that could correlate records across multiple languages.

 - A system that, given a set of symptoms, will suggest further symptoms to
   look for, diagnoses and treatments.

 - A system that performs automated knowledge extraction across the corpus,
   with automated processes running in the background attempting to learn
   new information.

 - Although acquiring the medical records themselves is as much of a
   political challenge as a systems one, any steps that can be taken in
   this direction would be useful.