To the Editor: Dr Fonarow and colleagues1 estimate a risk model for in-hospital mortality in acutely decompensated heart failure using a large data set (>33 000 patients) and validate it on an equally large independent data set. However, their conclusion that a classification and regression tree (CART) model performs better than a linear logistic model in predicting the risk of death in patients with acutely decompensated heart failure is surprising. In the independent data set, the CART model yielded an area under the receiver operating characteristic curve (AUROC) of 0.668 compared with 0.757 for the logistic model. Although apparently modest, this represents a large difference in performance. A useless model would on average have an AUROC of 0.5. Subtracting 0.5 from the AUROC and considering a scale from 0 to 0.5 (the upper limit), the relative improvement offered by the logistic model is (0.257-0.168)/0.168 or 53%, which is considerable.
The authors argue that their CART model is easier for clinical use than the presumably complex logistic model. In fact, the latter has only 4 predictors compared with 3 for the CART model. It is straightforward to scale the weights of a logistic model to whole numbers to make the arithmetic in calculating the risk score easy. A simple table can provide data to convert the risk score into a probability of dying.
The dichotomization of the continuous predictors in the CART model (a necessary consequence of CART) has led to a substantial loss of information. The ability of the CART model to discriminate between patients with a good and a poor prognosis is significantly reduced compared with the logistic model. Use of the CART model would lead to an unnecessary increase in the chance of suboptimal clinical management of patients with acutely decompensated heart failure compared with use of the logistic regression model.
In general, we recommend the use of continuous predictors in risk models without dichotomization or categorization.2 The creation of risk groups should be the final stage after the model-building process is complete. It is particularly undesirable to dichotomize continuous predictors before building a model.
Country-Specific Mortality and Growth Failure in Infancy and Yound Children and Association With Material Stature
Use interactive graphics and maps to view and sort country-specific infant and early dhildhood mortality and growth failure data and their association with maternal
Instructions
Comments are moderated and will appear on the site at the discretion of the Journal of American Medical Association editors. Comments should not exceed 500 words of text and 10 references.
Do not submit personal medical questions or information that could identify a specific patient, questions about a particular case, or general inquiries to an author. Only content that has not been published, posted, or submitted elsewhere should be submitted. By submitting this Comment, you and any coauthors transfer copyright to the journal if your Comment is posted.
* = Required Field
Disclosure of Any Conflicts of Interest* Indicate all relevant conflicts of interest of each author below, including all relevant financial interests, activities, and relationships within the past 3 years including, but not limited to, employment, affiliation, grants or funding, consultancies, honoraria or payment, speakers’ bureaus, stock ownership or options, expert testimony, royalties, donation of medical equipment, or patents planned, pending, or issued. If all authors have none, check "No potential conflicts or relevant financial interests" in the box below. Please also indicate any funding received in support of this work. The information will be posted with your response.
Register and get free email Table of Contents alerts, saved searches, PowerPoint downloads, CME quizzes, and more
Subscribe for full-text access to content from 1998 forward and a host of useful features
Activate your current subscription (AMA members and current subscribers)
Some tools below are only available to our subscribers or users with an online account.
Download citation file:
Customize your page view by dragging & repositioning the boxes below.
and access these and other features:
Register Now
Enter your username and email address. We'll send you a reminder to the email address on record.
Athens and Shibboleth are access management services that provide single sign-on to protected resources. They replace the multiple user names and passwords necessary to access subscription-based content with a single user name and password that can be entered once per session. It operates independently of a user's location or IP address. If your institution uses Athens or Shibboleth authentication, please contact your site administrator to receive your user name and password.