[Gnumed-devel] Extended proposal concerning normcurves

[Christof Meigen]

The name of the game
--------------------

Physicians, especially pediatrians, regulary measure parameters
of a patient in order to check wether or not this patient
develops/is normal in terms of that parameter.

Such parameters include body height, weight respectively BMI,
head circumference, but also blood pressure or peak flow
(for asthma patients).

The results of such measurements are usually viewed in comparison
to the majority of the population, and usually a centile is
calculated, saying that XY% of the population (with that age/sex)
shows a value smaller that the measured one. To calculate
these centiles, normcurves are used.

I think that this idea "normcurves" can be implemented in
GnuMed in a both flexible and standardized manner, aiming for

- a very general implementation of the methods and algorithms used
  to calculate centiles, print nice graphics, mark strange measurements
  etc.

- a standard framework to interchange centile data, easily pick
  the desired norm for a patient etc.


Why all that fuss?
------------------

The view of what is "normal" changes rapidly with time, but there are
also reginal differences.

For example, 14% of the childen from Leipzig have now a BMI over that
97th centile that was used in the late 80's in the GDR.  Similar trends
for height (secular trend) are common.

In an international projekt, one ends easily up with 10 to 20 important
normcurves, and the need to easily add new ones is obvious.


What a normcurve does
---------------------

Ok, let's view the normcurve as a black box. You put in a set
of parameters (1) describing the patient, usually sex and age,
and a parameter (2) the normcurve is about.

The normcurve returns either a specific centile or a code on a scale,
for example "overweight", "obese", "normal" etc.

Parameters of type (1) does not necessarily be sex and age,
for example it is common to check the development of babys
on a "weight for length" chart.

Furthermore, a normcurve can also do the reverse calculation, 
that means to calculate the value of a given centile
or a given code on the scale, which is used for plotting
centile curves on charts etc.

What a normcurve is
-------------------

A normcurve consists of the following data

1) Country: the country in which the study was conducted that
            the norm is based on, may contain "*" for worldwide
            standards
2) Year   : The year the norm was published
3) Name   : A common name, usually the author's of the corresponding 
            paper
4) Reference: the bibliographical entry of where the norm was published
5) A range of ages for which it is applyable
6) Parameter of type (1): Code for the parameter that is used for
            the X-Axis on the chart, usually age
   Parameter of type (2): Code for the parameter that is used for the
            Y-Axis of the chart, for example height, BMI, head 
            circumference
7) For borth sexes "centile" data, which is either
       - a set of functions for some centiles (3rd, 10th, 50th etc.)
         resp. for borders between codes like "normal"/"pathologically high"
       - or two functions, for the mean and the standard deviation 
         (for normally distributed parameters)
       - or three functions, for the median, coefficiant of variation
         and Box-Cox-power (for skew normally distributed parameters
         like BMI)

     a function can be either:
       - a list of fixed points between linear interpolation is used
       - coefficients of splines
       - coefficients of a polynom

annotation for (1): At first sight, it might seem reasonable to
introduce an "ethnical" parameter, which is common in the US, 
where "Hispanic", "Asian", "Afro-american" and "White" or so is used.
This is generally a bad idea, since
- it is, especially in germany, a bit peculiar to push people
  in an ethnical category
- normcurves are usually published with no recommendation for
  which ethnical groups they should be used, but rather a country is
  given (growth in Germany is significantly different from growth in
  Poland, for example, although any statements about ethnical 
  differences between the two populations may be considered akin to
  Nazi-ideology)
- it seems wisest to avoid the whole discussion be leaving the 
  decision, which normcurve can be applied best to a certain 
  patient, to the physician. There are different tastes of the
  physicians and different standards by the health insurences anyway.


How the normcurve is stored
---------------------------

I would propose a XML-file, the scheme or document type definition
can be easily derived from the above definition.

Furthermore, it would be simple to introduce a web service, where
norm curves can be published by everyone (after reviewing by a core
team) and easily downloaded and integrated in GnuMed.

GnuMed would then at startup or on request load the XML-File and
initialize the appropriate objects and GUI-Elements.

How the normcurve is used
-------------------------

In the patient view, you would see the centiles or the text results
("obese" etc) of the normcurve next to the values. You can also
view and print predefined charts, as they are common for height/weight
among physicians now.

Specialities of height and BMI development
------------------------------------------

Well, the following is on the cutting edge of medical research,
but it adresses intuitive problems:

- Height of a child correlates with parentel height, with a 
  coefficient of correlation between 0.2 and 0.7, depending
  on age and sex. Therefor, centiles for children whose parent's
  heights are known, are usually moved up or down (for tall
  parents, taller children are "normal") and they are also
  closer together, since there is less variation in children
  whose parents have the same height. "Parental adjusted centiles"
  can be calculated statistically sound using the correlations
  mentioned above (I calculated an appropriate table from the 
  famous Zurich study by Prader)
- Even more important than the centile itself is the development
  of a child's centile, for example it is remarkable when a child
  keeps "dropping down" the centiles. The check which development
  is normal is usually done by eye, which may be sufficiant for
  a padiatrian with lots of experience. However, there are
  several more or less sophisticated techniques to 
  automatically detect normality in longitudinal data.
  I will give a talk on that in november at the Auxological
  conference of the Deutsche Gesellschaft für Auxologie
  in Glücksburg, Germany. The two most promising techniques
  are based on cluster analysis of normalised SDS developments
  respectively on Principal component analysis of phase-adjusted
  growth curves, both resulting more or less in a definition 
  of common "growth patterns".


What should/could be done
-------------------------

1) Define an DTD for the XML-storage of a normcurve or introduce
  a different storage scheme
---> I would volunteer here

2) Define a python object-tree to handle norm curves, that is
  * Read from the file
  * check measurements of patients, returning a centile or a code
  * return values for given centiles/codes
---> I would volunteer here

3) Define how measurements of various types are stored in GnuMed, 
  including  global definitions of the measurement type (height, 
  bone age etc.atc)
---> I don't feel like having competence here

4) Integration of the results of the normcurve-chacks into the GUI
---> I don't feel like having competence here

5) Set up a web service where people can access a "comprehensive
   GnuMed normcurve database"
---> That should be no problem once point (1) is solved.

What should be done NOW
-----------------------

As soon as people find the whole idea appealing and no-one works
on the yet, and (3) is clear (for example by introducing standard
codes or names for measurement types like "height" "weight" etc.
which would be used both in the XML-file as in the application
to make clear which value is for which type of measurement) 
I would start with point (1) and maybe with point (5).

I would start with point (2), although I have to become
more familiar with python first, but as a perl hacker,
programming in python should be as easy as \pi. Maybe some
GUI guys could then make use of that and integrate the
choice of a normcurve for a given patient and the resulting
centile data into some views.

Mark that this is all still completely independent of how
the charts for the patients will look like, which has to be
discussed with people who will actually use it. For my
personal preliminary approach on growth charts (including
cluster-based growth patterns) you can visit the (german)
website www.willi-will-wachsen.de/cgi-bin/centiles.pl

what do you think about the whole matter?

     Christof