Sunday, November 8, 2009

Remote Monitoring: Predictability

One of the most controversial subjects in measurement and analysis is the concept of predictably.  Prediction does not imply causality or a causal relationship.  It is about an earlier event or events indicating the likelihood of another event occurring.  For example, I've run simulation studies of rare events.  If any of my readers have done this, you'll notice that rare events tend to cluster around each other.  This means that if one rare event has occurred, it's likely that the same event will occur again in a relatively short time.  

Interestingly, the clustering does not seem to be an artifact of the simulation system.  There are some real-world examples.  Consider the paths of hurricanes. At any one time, it is rare that a hurricane will make landfall at a particular location.  However, once a hurricane has hit a particular location, it appears that one can predict that the likelihood of the next hurricane hitting in that same general area goes ups.  I can think of a couple of examples in recent history.  In 1996, hurricanes made landfall two times around the area of Wilmington, NC. Furthermore, a third hurricane passed by.  In 2005, New Orleans was hit solidly twice.  If you look at two hurricane seasons - 1996, 2005 - you'll note that they show quite different patterns.  The rare event paradigm suggests that when the patterns for creating rare conditions are established, they will tend to linger. 

In medicine the objective is to find an event or conditions preceding an event before the event of concern occurs.  For example, an event of concern would be a heart attack.  It is true that once one has had a heart attack, another one could soon follow.  The conditions are right for a follow-on event.  However, the objective is to prevent a heart attack - not wait for a heart attack to occur in order to deal with the next one that is likely to soon occur.  Physicians employ a variety of means to attempt to detect possible conditions that may indicate an increased likelihood of a heart attack.  For example, cholesterol levels that are out of balance might signal an increase in likelihood of having a heart attack.  

The problem is that most of the conditional indicators that physicians currently employ are weak indicators of an impending heart attack.  The indicators are suggestive.  Let me provide an example using a slot machine as an example.  Let's assume that hitting the jackpot is equivalent to an heart attack.  Each pull of the lever represents another passing day.  On it's own, with the settings that the machine is initially set to, the slot machine has a possibility of hitting a jackpot with each pull of the lever.  However, the settings on the slot machine can be biased to make it more likely to hit a jackpot.  This is what doctors search for ... the elevated conditions that make a heart attack more likely.  Making hitting a jackpot more likely does not mean that you're ever going to hit one.  It just increases the likelihood that you will hit one.  

To compound the problem, the discovery of biasing conditions that appear to increase the likelihood of events such as heart attacks are often difficult to clearly assess.  One problem is that apparent biasing indicators or biasing conditions generally don't have a clear causal relationship. They are indicators, they have a correlative relationship (that is not always strong), and not a causal relationship.  There are other problems as well.  For one, extending conclusions to an individual from data collected from a group is generally considered suspect.  Yet, that is what's going on with respect to measuring performing assessments on individuals.  Individuals are compared to norms based on data collected from large groups of individuals.  Overtime and with enough data, norms may be considered predictors.  Search out the literature.  You'll note that many times, measurement that once were considered predictive, now no longer are.

The gold standard of prediction is the discovery of predecessor event or events.  It is something that precedes the watched-for event.  In Southern California everyone is waiting for the great earthquake.  Scientists have been attempting to discover a predecessor event to that great earthquake.  Same goes for detecting a heart attack or other important medical events that are threats to ones health.  Two clear problems stand in the way of discovering a clear predecessor event.  The first is finding that event that seems to precede the event of interest.  This not easy.  A review of the literature will inform you of that.  Second, is once you've found what appears to be a predecessor event, what's its relationship to the target event, the event of interest?  Often times that is a very long process and even with effectively predictive predecessor events, the relationship is not always one to one.  In that, one predecessor event may not precede the event of interest.  Several predecessor events could precede the event of interest.  Or, the predecessor event does not always appear before the event of interest.

This ends my discussion of predictability.  Next time ... I'm going to speculate on what may be possible in the near term and how the benefits of remote monitoring and remote programming can be made available relatively inexpensively to a large number of people.

Article update notice

I have updated my article on Digital Plaster.  I have found an image of digital plaster that I have included, plus a link to one of the early news releases from the Imperial College, London, UK.  I shall include Digital Plaster in my next article.

No comments:

Post a Comment