RealTime Classification of Atrial Fibrillation using RR Intervals

Real-Time Classification of Atrial Fibrillation using RR Intervals and Transition States Jericho Lawson Faculty advisors: Dr. Yishi Wang & Dr. Cuixian Chen Department of Mathematics at the University of North Carolina at Wilmington 2. 7 million Americans currently have atrial fibrillation, a heart issue described as a “quivering or irregular heartbeat. ” AFib can lead to other heart issues, such as blood clots, stroke, and heart failure [1]. Using AFib data from the MIT-BIH Atrial Fibrillation Database and the 2017 Physio. Net Challenge Dataset, as well as methods from Moody and Mark’s 1983 paper, we explore various classification methods and resampling techniques to detect AFib. Proportions of transition states between RR intervals are used as covariates in logistic regression, LDA, QDA, boosting, and XGBoost models. With 5 -Fold Cross Validation, we can get up to 97. 1% prediction accuracy and 97. 0% sensitivity using the MIT-BIH data. Similar results can be seen with fewer covariates and dimension reduction techniques. Features such as RR interval variance and dimensions from the multi-dimensional scaling of differences in the Kolmogorov-Smirnov tests provide potential usage for AFib classification. Background Results Experimental Design Abstract Covariat es 1. Data Explanation MIT-BIH Atrial Fibrillation Database: 25 ten-hour ECG recordings, 23 of which were used; Locations of R-peaks are already included in dataset; Older, but cleaner and documented better by experts 2017 Physio. Net Challenge Dataset: 8, 528 ECG recordings, ranging from 9 -60 seconds; RR intervals were labeled as normal, AFib, other, or noise; Only documented by one expert; Results in messier data 2. Data Extraction and Cleaning Collect ECG data from Physio. Net Retrieve basic information from R peaks Accuracy Standard Error ______ All Nine Transition States ______ Logistic LDA QDA Boosting Reg. 0. 955 0. 942 0. 922 0. 971 0. 002 0. 003 0. 005 0. 002 Sensitivity 0. 933 0. 890 0. 950 0. 973 0. 925 0. 865 0. 941 Specificity 0. 970 Metrics 0. 976 0. 904 0. 969 Models 0. 982 0. 967 Table 1: Performance of Four Different Methods 0. 970 Using Two Different for AFib Classification Find running mean and transition state for each RR interval Remove any outlying RR intervals 3. Feature Extraction Broken into 30 -second or whole length segments, these features can be determined: Atrial fibrillation (AFib) is described as a "quivering or irregular heartbeat" that can lead to certain health issues, such as blood clots, stroke, and heart failure [1]. Detecting AFib can help warn people of a heart issue before more serious problems occur, especially with the use of real-time heart-rate devices, such as Fitbits and Apple Watches. Nine transition proportions, stating change in transition states between two adjacent RR intervals (e. g. short to long) RR variance, d. RR mean, & d. RR variance Three dimensions from MDS of pairwise Kolmogorov-Smirnov tests for each user 4. Model Testing MIT-BIH data 5 -fold cross validation Figure 1: Normal vs. Afib ECG Recording [4] Using electrocardiograms (ECGs or EKGs), we can identify the heart rate and RR intervals for a subject. However, we want to determine any trends, variables, and patterns that can identify atrial fibrillation for various individuals using RR intervals from the ECG data. The MIT-BIH Atrial Fibrillation Database [2] and the 2017 Physio. Net Challenge Dataset [3] will be used to analyze various ECG data. Data cleaning and feature extraction is done before AFib classification. From there, we will classify AFib using arbitrary segments from both databases with RR interval data. 2017 data Logistic regression Linear discriminant analysis Quadratic discriminant analysis Boosting (gbm) Boosting (XGBoost) All Reg-to-Long Reg-L & S-Reg PCA = 5 Results ____ Regular-to-Long State ____ Logistic LDA QDA Boosting Reg. 0. 952 0. 935 0. 956 0. 957 0. 002 0. 003 0. 004 0. 003 0. 954 0. 960 All 12 Reg-L & RR Var. Accuracy 0. 943 0. 929 Reg-L, PCA = S-Reg, 5 RR/d. R R Var. 0. 940 0. 938 Standard Error Sensitivity Specificity 0. 004 0. 005 0. 725 0. 958 0. 634 0. 947 0. 706 0. 956 0. 687 0. 956 F 1 Score Time (s) Figure 3: Performance Metrics of Four Different Models Using Boosting with the 2017 Physio. Net Data 0. 621 0. 505 0. 602 0. 593 112. 58 65. 340 149. 18 128. 49 4 Metrics of Boosting 4 Using Four 2 Table 2: Performance Different Models for AFib Classification • About a 2% reduction is seen in both accuracy and sensitivity when the regular-tolong covariate is used instead of nine transition proportions for MIT-BIH data • For the 2017 Physio. Net data, while the prediction accuracy is between 92. 994. 3%, the sensitivity and F 1 values are relatively low, showing that our models cannot classify AFib segments well • Using 5 -Fold CV on each MIT-BIH user, at least 12 of 21 saw improvements or equivalencies using MDS dimensions from Kolmogorov-Smirnov tests as additional covariates Main Goal: To correctly detect and classify atrial fibrillation using information from RR intervals given some ECG data. Acknowledgements Figure 4: Performance Metrics of Four Different Models with and without Kolmogorov-Smirnov Test Covariates (for 08215) Conclusions This project was funded by NSF grant DMS-1659288. Special thanks given to Dr. Yishi Wang, Dr. Cuixian Chen, Dr. Yaw Chang, Dr. Rachel Carroll, and Summerlin Thompson for their guidance and support throughout the research project. I would also like to thank Drew Johnston, Georgia Smith, and Christian Austin for being great colleagues while we were working on similar projects. • Our models can achieve up to 97. 1% prediction accuracy and 97. 3% sensitivity using transition states from RR interval data in the MIT-BIH dataset • Usage of Kolmogorov-Smirnov tests to generate differences as covariates can have potential improvement on the models that were generated References [1] American Heart Association. Atrial fibrillation. In Heart. org, June 2019. [2] A. Goldberger, L. Amaral, L. Glass, J. Hausdorff, P. Ivanov, R. Mark, J. Mietus, G. Moody, C. Peng, and H. Stanley. Physiobank, physiotoolkit, and physionet: Components of a new research resource for complex physiologic signals. In Circulation Electronic Pages, volume 23, pages 215– 220, June 2000. [3] Chengyu Liu, David Springer, Qiao Li, Benjamin Moody, Ricardo Abad Juan, Francisco J Chorro, Francisco Castells, Jos e Millet Roig, Ikaro Silva, Alistair E. W. Johnson, Zeeshan Syed, Samuel E. Schmidt, Chrysa D. Papadaniil, Leontios Hadjileontiadis, Hosein Naseri, Ali Moukadem, Alain Dieterlen, Christian Brandt, Hong Tang, Maryam Samieinasab, Moham-mad Reza Samieinasab, Reza Sameni, Roger G. Mark, and Gari D. Clifford. An open access database for the evaluation of heart sound algorithms. volume 37, 2016. [4] S. Steinbaum. A Visual Guide to Atrial Fibrillation. Web. MD, 2017. Figure 2: Performance Metrics of Four Different Methods Using Regular-to-Long as a Covariate • 93. 5 -95. 7% accuracy with only the regular-to-long covariate for AFib classification. • While models performed well on MIT-BIH data, they struggled on the 2017 Physio. Net data, suggesting that signal processing would be needed to identify noisy peaks in the 2017 data • These models provide a basis for real-time AFib detection when implemented and computed simultaneously with ECG data