Thursday, May 3, 2012

Proposal for a N=1 Experimentation Online Community




INTRODUCTION

There is a paradox when it comes to optimizing ones health through diet.  All indications are that individual characteristics, circumstances and history are tremendously important as far as fine-tuning choices of nutrition and exercise.  This leads many people to try N=1 experimentation in order to evaluate exactly how their body will respond to changes and find what's right for them.   The trouble with this approach is that these experiments are difficult to to conduct, interpret and use for future decision making.  What to do?  

An idea has been floating around to develop an online community -- the larger the better -- of N=1 experimenters to
  • provide support, new ideas and discussion
  • structure and conduct experiments across a variety of nutritional (and other factors)
  • share results and allow analysis of both pooled and specifically relevant community members
  • develop tools allowing one to interpret the community results in their individual context, make predictions and forward looking decisions
This community could even ultimately become a useful scientific complement to the traditional approaches of large scale epidemiological studies and small targeted controlled trials.

WHY IS N=1 DIFFICULT?

Truly informative N=1 nutritional studies are difficult because of 
* The self-discipline, consistency and time (and possibly cost) required
* The intrinsic 'high dimensionality' of the problem.  (By high dimensionality we mean that there are many, certainly tens maybe hundreds of variables necessary to describe ones food intake and relevant circumstances.  Of course not all of these are equally important but from a statistical modeling standpoint, if one reduces this dimensionality in an ad-hoc fashion -- e.g. by just ignoring some of them -- it will almost always introduces systematic biases in the results.)  These attributes of food intake might include total calories, macronutrient breakdown, types of fats, vitamins, glycemic index, etc.  Timing of feedings, combinations of macro-nutrients eaten, are other obvious potential considerations.  And when one considers the background factors: observable individual characteristics like age, weight, body fat, activity level and type, sleep, supplementation and history thereof, the picture is further complicated.
So a) there are many many compelling experiments to try out and only one lifetime to do it in, b) it may be  tempting to prematurely move to the next experiment because something seems 'not to be working', c) interpreting results even after a lot of data has been collected can be difficult because the other variables are not adequately controlled or accounted for, d) extrapolating from other's N=1 results is exceedingly problematic.
* It is difficult to know what can be objectively measured and tracked and with what averaging and what lags.  For instance, if I add some starches on top of a   low carb diet, conjectured to accelerate weight loss, how long until it starts working, should I look at instantaneous weight or averaged over some period?
* There are few available to help predict outcomes and analyze realizations against these to figure out what went wrong (or right).  This can emotionally be quite important and for certain types of individuals will help them persevere with the programs.  For example one would like to be able to say:  given my current state of health and personal history, if I eat a 15%/25%/60% carb/prot/fat diet for 1 month, I would expect might weight change over that period to be in the following range.   

What are we proposing?

The vision is to have a way to explore many hypotheses with large populations that cover the high-dimensional space of characteristics  with the necessary emotional and intellectual support of the participants, effective monitoring and reporting tools and ongoing analysis and interpretation.

The community would consist of
* A handful of interested researchers willing to design/program and oversee experiments
* A large number of individual N=1 experimenters who would participate in these and report their results to some level of granularity
* Possibly, funding permitting, some number of self-identified but screened guinea-pigs willing to follow a baseline diet with proscribed tweaks and larger (instantaneous and secular) changes and put up with continual tests and diagnostics.
* Statistical modelers and tool builders who would create open source analytical and predictive tools that would be continually updated using the results.  Machine learning techniques are ideal for coping with the large dimensionality and (hopefully) large data -- inferring the right conditional distribution of outcomes for an individual and not losing sight of the error-bars 

A website would  provide background on the experiments (including summaries of relevant traditional studies), forums for discussion and Q&A, tools for reporting/tracking, statistics on the trials underway and summary results from completed ones, analytic and modeling tools that would provide statistics and machine-learning based predictive capability.

Please contact us if you have interest in participating either as an organizer, researcher or participant.

No comments:

Post a Comment