Talk:Clinical decision support system

As seems appropriate I added engineering. Robert Tito |  Talk  22:40, 17 May 2007 (CDT)

This article focuses too much on the algorithms, and too little on the concepts. There are probably thousands of analytical platforms out there; the reader probably just needs a conceptual overview of how this technology works, and where it is being used in the clinical enterprise. Let's have sections called "supervised learning" and "unsupervised learning," "applications," "challenges."--Michael Benjamin 02:18, 21 May 2007 (CDT)

technologies

This article lists a number of technologies, but is often very vague about them (for example, it mention artificial neural networks, but doesn't specify whether they are feedforward networks or recurrent, whether backpropogation is used for training, or an unsupervised method such as Hebbian learning, and it doesn't make it clear why connectionism should be listed separately). From an engineering perspective, it just looks like a few terms for technologies, giving no indication of how or why they should be used). Greg Woodhouse 16:04, 23 May 2007 (CDT)

How to take it to the Approval stage?

I had intended this to be an introductory article on CDSS. The various decision support technologies may be discussed in the appropriate sections. If you look at the ANN entry in CZ, that is just a stub till now. Can we make it an approved article without going into details of each and every CDSS mechanics? Supten 02:00, 31 October 2007 (CDT)

Some minor points...

The introduction appears to be a bit too essay-like for me, or a bit too chatty. One important point that is missing so far, however, is issues of responsibility: What concepts for this problem exist for the individual projects, what are the liabilities for the parties involved, be it the company producing the systems or the doctors applying it? This is a general problem of such systems and thus has its place in this overview article. Who gets the blame when a wrong recommendation by the system is followed up on? --Oliver Hauss 14:46, 12 November 2007 (CST)

Approval?

I think this article first needs references converted to cite the inline cite journal format. - Robert Badgett 15:19, 13 November 2007 (CST)

Reorganization needed for Methods of decision support section

Based on my reads of Coiera (ISBN 0-340-76425-2) and Shortliffe (ISBN 0-387-28986-0), I proposed to reorganize "Methods of decision support section" as below. The text below is for sake of discussion and I will flesh out once everyone is ok with the proposal. Please comment.

• Knowledge based systems / expert systems. These systems are created by having experts identify relationships between independent variables (such as signs and symptoms) and dependent variables (such as likely underlying diseases). Per Shortliffe, "instead of modeling the relationships among patient findings and possible diagnoses in purely in terms of statistical associations or mathematical equations, knowledge based systems might represent those relationships in terms of qualitative symbolic structures." Somestimes the inputs may include locally created knowledge (such as local frequency of surgical complications - PMID 11687560)
  • Evidence-adaptive systems (proposed by PMID 11687560)
• Machine learning. In these systems, the relationships between independent variables (such as signs and symptoms) and dependent variables (such as likely underlying diseases) is created by having the system be trained on a "large collection of previously classified examples during a period of supervised learning" (Shortliffe). A classic example is automated electrocardiogram interpretation (PMID 1834940).
  • Artificial Neural Networks is a type of machine learning
    • Bayesian Belief Network is a type of Artificial Neural Networks (do I have this right?)
    • Connectionist expert system is a type of Artificial Neural Networks in which humans can help the system revise weights. Seems this should go here and not be called a hybrid system because although humans help revise the system, literature based knowledge is not directly used to modify weight.
• Hybrid - an example is PMID 9021057.

Should supervised versus unsupervised machine learning would fit in here?

Robert Badgett 03:38, 19 January 2008 (CST)

The proposed sections look good. Supten Sarbadhikari 21:08, 21 January 2008 (CST)

Question

Why is "heuristic reasoning (QMR, DXplain, ILIAD)" call hybrid? - Robert Badgett 07:21, 19 January 2008 (CST)

Because they combine a rational (analytic) approach along with the heuristic (rule of thumb) approach. Supten Sarbadhikari 21:11, 21 January 2008 (CST)

Thanks. This page (http://lcs.mgh.harvard.edu/projects/dxplain.html) makes DXplain look more Bayesian, so I moved dxplain to the Bayesian section. Please correct if I have this wrong. - Robert Badgett 05:21, 22 January 2008 (CST)