Using Rough Sets, Neural Networks, and Logistic Regression to Predict Compliance with Cholesterol Guidelines Goals in Patients with Coronary Artery Disease

Coronary artery disease is a leading cause of death and disability in the United States and throughout the developed world. Results from large randomized, blinded, placebo-controlled trials have demonstrated clearly the benefit of lowering LDL cholesterol in lowering the risk for coronary artery disease. Unfortunately, despite the quantity of evidence, and the availability of medications that can efficiently lower LDL cholesterol with few side effects, not everyone who could benefit from cholesterol lowering interventions actually receives them. Despite the dissemination of national care guidelines for the evaluation and treatment of cholesterol levels (NCEP - National Cholesterol Education Program), compliance with such guidelines is suboptimal. There clearly is room for improvement in narrowing the gap between evidence based guidelines and actual clinical practice. The ability to classify those patients who are or will likely to be noncompliant on the basis of patient data routinely collected during patient care could be potentially useful by enabling the focusing of limited health care resources to those who are or will be at high risk of being under treated. In order to explore this possibility further, we attempted to create such classifiers of cholesterol guideline compliance. To do this, we obtained data from an ambulatory electronic medical record system at use at the MGH adult primary care practices for over 20 years. We obtained the data from this hierarchically-structured EMR using its own native query language, called MQL (Medical Query Language). Next, we applied to the collected data the machine learning techniques of rough set theory, neural networks (feed forward backpropagation nets), and logistic regression. We did this by using commonly available software that for the most part is freely available via the internet. We then compared the accuracy of the classifier models using the receiver operating characteristic (ROC) area and C-index summary metrics.