########## Response to Reviewer #1 Feedback: ##########
* Three barcharts depicting of relative influence of each predictor variable in the highest-performing models with respect to F1-score are included (one for each of the three binary classification tasks).
* The following changes have been made with respect to figures:
1) Font sizes of axis ticks, axis labels, and plot titles were adjusted to improve aesthetics.
2) Thicker arrows are used in Fig 3 (the flowchart).
3) A figure depicting the persistence diagrams of two ECG signals - one with sinus rhythm and one with atrial fibrillation - before and after including the isoelectric baseline is included to illustrate how the isoelectric baseline creates topological features corresponding to the P,Q,S,T-waves (if they exist to begin with). This is Fig 4.
4) A figure depicting the tranformations applied to preprocess an ECG signal prior to computing its topological features is included. This is Fig 5.
5) A figure depicting the computation of the effective centroid coordinates is included. This is Fig 6.
6) The line thicknesses of the ROC curves are increased to improve aesthetics.
* A figure is included which illustrates how the isoelectric baseline affects the persistence diagrams of ECG signals with sinus rhythm and atrial fibrillation. This is Fig. 4.
* More detail is included in the "Construction of Predictor Variables" and "Results/Discussion" sections to hopefully improve readability and to convey the methods and results more clearly. The new figures may also help explain the methods behind constructing predictor variables more clearly.
* The standard deviations of the evaluation metrics are now included in Tables 2-4 to help portray their distribution across the five folds. Since these standard deviations are small, the evaluation metrics across each of the five folds are very similar and are not presented.

########## Response to Reviewer #2 Feedback: ##########
1) These papers along with several others are briefly discussed and cited. The previous methods of applying TDA to ECG analysis have utilized statistics derived from the birth and death radii of topological features, and one distinct aspect of the approach presented here is that it utilizes information derived from area-minimal cycle representatives of topological features as predictor variables as well. Hopefully this is made more clear in the Introduction.
2) This is corrected.
3) The prevalence of arrhythmia and of atrial fibrillation are now reported in the Introduction.
4) This is updated.
5) The difference between homology and homotopy is dense and beyond the scope of this article. We briefly mention this difference because intuitively, we conceptualize topological features as homotopy groups, although for more efficient computations, topological features are treated as homology groups. This is made more clear in the manuscript. Since we use the Vietoris Rips Complex as the choice of simplicial complex, homology and homotopy are equivalent in the context of TDA as applied here. The classic textbook in Algebraic Topology by Hatcher covers these details in great depth and is cited.
6) The database of 12-lead ECG signals contains 10646 signals, of which 10605 signals have a non-empty lead 2 signal. It is not clear in the documentation of the database why 41 signals do not have a recording for lead 2. Additionally, in order to compute the effective time coordinate of the centroid of each area-minimal cycle representative, there must be a QRS-complex following the area-minimal cycle representative, so we do not consider the data following the onset of the last QRS-complex. So, we exclude one small subset of the ECG signal following the last large amplitude peak (i.e. the last QRS-complex). The frequency is only considered in computing the QRS-complexes.
7) We removed the equations that define the evaluation metrics and referred the reader to a paper which describes these evaluation metrics (although they can probably be found in most statistics or machine learning textbooks). Three figures depicting the relative influence of each predictor variable in the top-performing models for each classification task are now included.
8) Perhaps the Gradient Boosted Decision Tree models outperformed the others due to heterogeneity of the data. The tree-based models 'learned' the rules used to diagnose arrhythmias better than the other modeling approaches taken.
9) The standard deviation of the evaluation metrics across the five folds are now reported in Tables 2-4, and the evaluation metrics of the top-performing model for each evaluation metric are presented in bold text.
10) The motivation behind presenting the ROC curve for Classification of Arrhythmia vs. Sinus Rhythm is that it can be compared to the ROC curve of Arrhythmias with Morphological Changes vs. Sinus Rhythm with Bradycardia and Tachycardia Treated as Non-Arrhythmia. TDA does well as quantifying 'the shape of the data', so comparing these ROC curves illustrates how when we use TDA to classify only arrhythmias with morphological changes vs all other rhythms, the results are much better than when we use TDA to classify arrhythmia vs non-arrhythmia when the arrhythmia group contains the two arrhythmias with non-morphological changes (i.e tachycardia and bradycardia). This is made more clear in the manuscript.
11) We are hesitant to include a table presenting our results alongside the results from other studies due to (1) different datasets with vastly different sample size, ECG signal duration, and sampling frequency and (2) different rhythms annotated for the ECG signals between studies. We mentioned this information and more detail with respect to comparisons with other studies in the Results/Discussion.
12) This is updated.
13) The new figure Fig. 4 depicts the persistence diagrams of two real ECG signals: one with normal sinus rhythm and one with atrial fibrillation. The effect of the isoelectric baseline on the persistence diagrams and the corresponding area-minimal cycle representatives it spawns is illustrated.

########## Miscellaneous ##########
* An acknowledgement that the computations of (1) topological features and (2) the statistical models were performed on Wayne State University's High-Performance Computing Grid is now included.
* The scripts to create the new figures and compute the relative influence of predictor variables in the top-performing models have been added to the GitHub repository.