« Forget Clinical Involvement | Main | Podcare: Care and iPod »

Bayesian Machine Learning: dealing with complexity

bayes-small.jpgWhat do a spam filter and a Nonconformist minister who lived 300 years ago have in common?

About 300 hundred years ago Thomas Bayes invented Bayes Theorem. Though it was given some recognition in his lifetime, the field of Artificial Intelligence adopted it centuries later to create Bayesian Networks (BN) which became the basis of Bayesian Machine Learning.

BNs help us to manage complexity and uncertainty by inferring cause and effect, rather than having to be told it.
They modify an initial theory by applying evidence to a fixed set of rules based on Bayes Theorem. Using BN's machines learn about the real world, where the relationships between events are often uncertain.

Interrelationships in BNs can achieve mind-boggling complexity, but increases in processing power have made them practicable. They are used in data mining, operating systems, fault diagnosis and fraud detection. And, as the number of intelligent devices grows, they will become more common.

My spam filter Popfile uses Bayesian techniques. Since installing it, Popfile has become more than 90 percent accurate and helps to keep my Inbox clear. Bayesian filters identify spam from keywords such as "free", "new" and "enlargement". They also learn from my actions by associating the way I deal with an email with its characteristics.

BNs also mimic the way in which we think. Experience that is consistent with our beliefs tend to reinforce them; whereas inconsistent experience may force us to modify them or form new ones. Unlike Neural Networks (more in future FHIT entries) the workings of BNs are visible and so easy to consider and discuss.

It's easy to see how such techniques could be applied to decision and diagnostic support in healthcare. Also, clinical trials can be modified while they are in progress using Bayesian techniques; for example, dosages can be changed in the light of findings to allow more data to be collected closer to the optimal dose. This is inimical to traditional statistics whose supporters--sometimes called frequentists--say that it risks contaminating data and introducing bias.

Nonetheless, Bayesian techniques are likely to become common, allowing machines to anticipate our actions and help us to make decisions and deal with complexity--which seems perfect for the healthcare world.

If you are interested in reading more about machine learning, please see my article on the E-Health-Insider website: Cyber Care.


TrackBack URL for this entry:

Post a comment

(If you haven't left a comment here before, you may need to be approved by the site owner before your comment will appear. Until then, it won't appear on the entry. Thanks for waiting.)