Data-efficient machine learning methods in the ME-TIME study: Rationale and design of a longitudinal study to detect atrial fibrillation and heart failure from wearables

Figures

• 

  Figure 1. Data analysis pipeline for the ME-TIME study. Included participants (image 1) are given a smartwatch (image 2), which is connected to our data acquisition and storage platform (image 3). The resulting data are then preprocessed (image 4) and put into the data-efficient machine learning model (image 5). AF = atrial fibrillation group; HF = heart failure group; Ref = reference group.

• 

  Figure 2. Pipeline for the study’s proposed approach. a: Segmentation of the time series of each subject (1 healthy and 2 atrial fibrillation) using a sliding window. Only the label of the entire subject is available, instead of each individual window. b: The windows are inputs to an autoencoder and are compressed to a smaller (2-dimensional for illustrative purposes) representation. c: The compressed representation is used to train a multiple-instance classifier that can distinguish between healthy, atrial fibrillation, or heart failure). AF = atrial fibrillation.

• 

  Figure 3. Illustration of multiple-instance learning. The red and blue lines indicate bags of heart disease patients and reference subjects. Even though the labels for each instance are not known, for the sake of this example, the plus and minus signs depict time windows where heart disease is present or absent, respectively. The decision boundary is depicted by a dotted circle, where instances within the circle are classified as heart disease, and instances outside the circle are classified as healthy.

• 

  Figure 4. Our leave-p-subjects-out cross-validation strategy consists of the following: On the left, cross-validation folds are illustrated using 3 subjects with corresponding subject number. Green and blue represent training and validation subjects, respectively. On the right, a single fold is expanded and additionally illustrates the internal and external test sets. Each row corresponds to a subject, and the dotted squares within each row represent windows. The yellow external test block encompasses entire subjects and their corresponding data points that have not been encountered by the model. Meanwhile, the dark yellow internal test block consists of unobserved data points, representing the final 20% measurements of the time series from subjects already encountered during model development.

• 

  Figure 5. Visualization of heart rate windows (upper image) and step windows (lower image) of 6 subjects. Each window has data of 1 hour (720 time points for heart rate, 60 for steps, respectively). Each high-dimensional window is mapped to a 2-dimensional (2D) location using UMAP.11 The contour curves illustrate the distribution of the 2D UMAP samples for each subject, with each color representing 1 of the 6 subjects.

• 

  Figure 6. Acceleration-deceleration curves during light activity. Red and blue represent data from subjects with persistent atrial fibrillation (n=7) and no heart disease (n=15), respectively. The mean and standard deviation are shown per time point (5 second intervals) for both groups.

• 

  Figure 7. Cross-correlation matrices between windowed heart rate data and number of steps for healthy and the persistent atrial fibrillation (AF) group. Rows are window sizes and columns lag between heart rate and step counter.