Suggestions were formulated to reduce bias and error related to the use of machine learning (ML) approaches in cardiology research, and published in the Journal of American College of Cardiologists: Cardiovascular Imaging.
The use of ML approaches for cardiovascular research has recently increased, as the technology offers approaches to automatically discover relevant patterns among datasets. This review authored by members of the American College of Cardiology Healthcare Innovation Council, points to the fact that many studies using ML approaches may have uncertain real-world data sources, inconsistent outcomes, possible measurement inaccuracies, or lack of validation and reproducibility.
The authors provide here a framework to guide cardiovascular research in the form of a checklist.
When considering employing a ML approach for their research work, investigators should initially determine whether it would be applicable for the specific study aim. An important caveat of ML is that it requires large sample sizes. Therefore, if collecting and labeling fewer than hundreds of samples per class is not feasible, overfitting is likely be a relevant concern. When sufficient samples are available, ML approaches are best suited for unstructured data, exploratory study objectives, or for feature selection purposes.
Next, data should be standardized, if necessary. During this process, redundant features are normalized, duplicates are removed, outliers removed or corrected for, and missing data removed or imputed. As a general rule, the ratio of observations to measurements should be ≥5. In cases in which this ratio is too large, dimension reduction may be considered.
Many ML approaches are available to researchers, and the choice of which model to implement is critical. Some models are preferable for high-dimensional data (regression or instance-based learning) or imaging data (convolutional neural networks). The authors recommend selecting the simplest algorithm that is appropriate for one’s dataset.
Several methods are available to assess and evaluate models. Model assessment should always be performed through random division of the data into training, testing, and validation sets. Cross-validation and bootstrapping methods are best suited for big data, and jack-knifing methods for smaller datasets. Model evaluation should include appropriate plots (Bland-Altman). In addition, inter-observational variability should be reported, and misclassification risk be made clear.
To maintain a level of reproducibility across studies, the authors encourage researchers to release the code and data used, when possible. All chosen variables and parameters, as well as specific versions of software and libraries should be clearly indicated.
The authors acknowledge that these methods are complex, and while they have the opportunity to advance the field of cardiology, especially personalized medicine, many concerns remain when translating these findings into clinical practice. This checklist should assist researchers in reducing bias or error when designing and carrying out future studies.
Sengupta P P, Shrestha S, Berthon B, et al. Proposed Requirements for Cardiovascular Imaging-Related Machine Learning Evaluation (PRIME): A Checklist. JACC Cardiovasc Imaging. 2020;13(9):2017-2035.