Had a chance yesterday to share the last developments in the quest to improve differential diagnosis in physical therapy through the use of causal models and Bayesian networks.

The presentation is online here (I plan to add voice over and break it up into smaller online webinar based sessions soon):

http://www.knowledgebasedpractice.com/differential_diagnosis/

All the files for it – including the example networks built with samiam and full conditional probability tables (estimates, but at least these assumptions are explicit and shared) are on GitHub:

https://github.com/scollinspt/differential_diagnosis

The most amazing part of the presentation to me was that with 41 therapists in the room; no one challenged my repeated statements that all of their decision making is founded on probabilistic causal inferences within complex causal networks. In practice they simply do not consider the underlying assumptions and probabilities, but when they make a decision they are making a statement about what they believe, and therefore about a probability. There seemed to be general ascent on this claim.

Thank you to all of you that attended – I look forward to more work and interaction with collaborators as the project continues to unfold; and to offering additional continuing education and even training opportunities for anyone interested.

The general premises of the project are:

– Clinical practice situations are complex systems (see this post)

– Complex systems include high dimensional, multi – variable causal networks.

– Reasoning with such systems is challenging and often includes several hidden assumptions.

– Both for teaching and for research aimed to improve the decision making from such reasoning, it is valuable to make as many of the previously hidden assumptions explicit.

– Existing methods to attempt to improve differential diagnosis (use of diagnostic accuracy such as sensitivity, specificity, likelihood ratios) don’t go far enough. They are far too simple compared to actual complexity of actual clinical practice situations.

– Bayesian networks offer – for clinical educational and clinical research purposes – a bridge between the highly complex system of actual clinical practice and existing methods that attempt to improve the process.

– Clinical education and clinical research support clinical practice and are necessary components to the system attempting to improve practice, which is something that providers, clients and payors are all seeking.

Screen shot of a sample samiam model for back pain with several clinical signs observed and instantiated leading to adjusted posterior probabilities: