Skip to content

Proposed Requirements for Cardiovascular Imaging-Related Machine Learning Evaluation (PRIME): A Checklist

Proposed Requirements for Cardiovascular Imaging-Related Machine Learning Evaluation (PRIME): A Checklist

“The greatest benefit of machine learning may ultimately be not what the machines learn but what we learn by teaching them.” 

-Pedro Domingos, author of The Master Algorithm

Cardiology as a subspecialty has significantly increased its imaging-related machine learning projects and publications in the past few years, so this proposed requirements for these projects is a timely reference. This proposal, reviewed by the American College of Cardiology Healthcare Innovation Council, is an attempt to provide a checklist for machine learning models with cardiovascular imaging for all involved in this domain: data scientists, cardiologists, investigators, editors and reviewers. In addition, these suggested requirements put forth by a multidisciplinary panel of clinicians and machine learning experts may reduce the inconsistencies observed in these models.

The proposed 7 steps, each with a separate checklist, are: 1) Designing study plan; 2) Data standardization; 3) Selecting machine learning models; 4) Model assessment; 5) Model evaluation; 6) Model replicability; and 7) Reporting limitations. Important topics discussed under each of these categories include general topics such as data preparation, feature engineering and feature selection. The three main approaches to feature engineering include: hand engineering, dimensionality reduction, and deep learning. Smaller topics that are also critical for the model include: class imbalance, data shift, and data leakage. The authors also suggested freely available data and open-source code but concomitantly fully realize that there are inherent challenges and complexities with machine learning models.

The authors mentioned at the end of the proposal that there are three areas of consideration for the machine learning “best practice”: the use of automated ML (auto-ML); the multiomics approach to data; and the use of GAN for synthesis of data. While the proposed list is a daunting task, this manuscript is a good start towards a machine learning “best practice” for not only for cardiovascular imaging but other projects as well. This proposal should be the beginning of a long journey ahead for such a vision to optimize these projects.

Read the full article here

Show Buttons
Hide Buttons