Author + information
- Christopher O’Connor, MD, FACC, Editor-in-Chief, JACC: Heart Failure∗ ()
- ↵∗Address for correspondence:
Dr. Christopher O’Connor, Editor-in-Chief, JACC: Heart Failure, Heart House, 2400 N Street NW, Washington, DC 200347.
Many years ago, I served as a research fellow in an exciting era of the Duke Databank for Cardiovascular Disease prior to the formation of the Duke Clinical Research Institute. At that time, there were only 10 members of this clinical investigative section in the division of cardiology, including Drs. Robert Califf (my primary mentor), David Pryor, Kerry Lee, Frank Harrell, Daniel Mark, and Mark Hlatky. Dr. Eugene Stead, the founder, stated that the collection of data in a databank would be more useful than the human memory in facilitating the care of patients who underwent cardiac catheterization and either coronary artery bypass surgery or medical therapy.
Thus, I became part of an investigation to study the accuracy in which long-term prognosis could be predicted in patients with coronary artery disease, many of whom had reduced left ventricular function, and whether these prognostic predictions obtained from a large, diverse sampling of master clinicians at Duke University would be better than predictions from a multivariable statistical model from the databank. Clearly, my hypothesis was that the esteemed gray-haired clinicians were far superior in their knowledge of prognosis than a cold, unbiased databank. In this unique study, we examined patients who were selected from a large series of medically-treated patients with significant coronary artery disease and ischemic cardiomyopathy. Using detailed clinical summaries, the cardiologists predicted the probability of 3-year survival for these patients. Cox regression models were developed using patients who were not in the test samples but were also used to predict corresponding probabilities for each test patient. Surprisingly, and contrary to my established paradigm, the model estimates of prognosis were 20% better than the master clinicians in predicting the outcome (1). The databank model predictions added significant prognostic information to the master clinicians’ predictions. Although there was significant interphysician variability, neither practice characteristics nor extent of clinical experience significantly affected each physician’s predictive accuracy. Again, to my surprise, master clinicians with over 20 to 30 years of experience caring for thousands of patients each year were inferior to the predictive models.
Today, the Duke Databank for Cardiovascular Disease has over 250,000 patients with over 40 years of follow-up data. The statistical models are very robust in predicting clinical outcomes. At Duke, we changed our practice to not only rely on our clinical estimates but to also use prognostigrams to assist in our counseling of patients on prognosis with varied types of therapies.
To this end, I will say that we must welcome the use of big data to assist in our clinical practice. All of us are utilizing the electronic medical record and other health records with large, actively integrating databases in each and every one of our clinical practices. We need to have clinical decision tools to help us utilize this data to assist us. Today, more than ever, I believe that not only our clinical acumen but also the use of large data will assist us in providing better care for our patients and greater quality and quantity of life. Let's embrace good models.
- American College of Cardiology Foundation