Using data mining techniques to explore physicians' therapeutic decisions when clinical guidelines do not provide recommendations: methods and example for type 2 diabetes

BMC Med Inform Decis Mak. 2009 Jun 10:9:28. doi: 10.1186/1472-6947-9-28.

Abstract

Background: Clinical guidelines carry medical evidence to the point of practice. As evidence is not always available, many guidelines do not provide recommendations for all clinical situations encountered in practice. We propose an approach for identifying knowledge gaps in guidelines and for exploring physicians' therapeutic decisions with data mining techniques to fill these knowledge gaps. We demonstrate our method by an example in the domain of type 2 diabetes.

Methods: We analyzed the French national guidelines for the management of type 2 diabetes to identify clinical conditions that are not covered or those for which the guidelines do not provide recommendations. We extracted patient records corresponding to each clinical condition from a database of type 2 diabetic patients treated at Avicenne University Hospital of Bobigny, France. We explored physicians' prescriptions for each of these profiles using C5.0 decision-tree learning algorithm. We developed decision-trees for different levels of detail of the therapeutic decision, namely the type of treatment, the pharmaco-therapeutic class, the international non proprietary name, and the dose of each medication. We compared the rules generated with those added to the guidelines in a newer version, to examine their similarity.

Results: We extracted 27 rules from the analysis of a database of 463 patient records. Eleven rules were about the choice of the type of treatment and thirteen rules about the choice of the pharmaco-therapeutic class of each drug. For the choice of the international non proprietary name and the dose, we could extract only a few rules because the number of patient records was too low for these factors. The extracted rules showed similarities with those added to the newer version of the guidelines.

Conclusion: Our method showed its usefulness for completing guidelines recommendations with rules learnt automatically from physicians' prescriptions. It could be used during the development of guidelines as a complementary source from practice-based knowledge. It can also be used as an evaluation tool for comparing a physician's therapeutic decisions with those recommended by a given set of clinical guidelines. The example we described showed that physician practice was in some ways ahead of the guideline.

MeSH terms

  • Algorithms
  • Artificial Intelligence
  • Combined Modality Therapy
  • Decision Support Techniques*
  • Decision Trees
  • Diabetes Mellitus, Type 2 / blood
  • Diabetes Mellitus, Type 2 / diagnosis
  • Diabetes Mellitus, Type 2 / drug therapy*
  • Drug Therapy, Combination
  • Evidence-Based Medicine / statistics & numerical data*
  • Glycated Hemoglobin / analysis
  • Humans
  • Hypoglycemic Agents / therapeutic use*
  • Insulin / therapeutic use
  • Medical Records Systems, Computerized / statistics & numerical data
  • Practice Guidelines as Topic*
  • Retrospective Studies
  • Software

Substances

  • Glycated Hemoglobin A
  • Hypoglycemic Agents
  • Insulin
  • hemoglobin A1c protein, human