Reproducibility of Machine Learning Models for Paroxysmal Atrial Fibrillation Onset Prediction

• Main Authors: Gilon, Cedric, Gregoire, Jean-Marie, Hellinckx, Jerome, Carlier, Stephane, Bersini, Hugues
• Format: Conference Proceeding
• Language: English
• Subjects: Atrial fibrillation; Heart; Machine learning; Prediction algorithms; Predictive models; Reproducibility of results
• ISSN: 2325-887X
• DOI: 10.22489/CinC.2022.171

Atrial fibrillation (AF) is the most common heart arrhythmia. Paroxysmal AF onset prediction is a more complex task than screening AF. Published methods using the AFPDB database show excellent results, suggesting that paroxysmal AF onset prediction is possible with machine learning (ML) models using heart rate variability (HRV) parameters. Aims To understand if AF onset prediction is possible using previously published methods. Reproduce results of published studies using the Physionet database. Methods We searched the literature for all articles on paroxysmal AF onset prediction. We analysed in depth 3 methodology using ML methods to replicate their results. Results With the information available in the publication, we were unable to reproduce the results presented by the authors with differences up to 20%. For each publication, we explored different scenarios with multiple splits and parameters choice for the model. Conclusion Reproducibility of the models and results is becoming a key aspect of ML research and authors must describe and make available the whole methods required to achieve their results.