ECG R-R Interval Estimation

Magnetic resonance imaging (MRI) is a biomedical imaging modality that provides high resolution images and excellent soft tissue contrast [1]. This makes MRI a preferable imaging modality for many applications. One limitation of MRI is slow data acquisition. MRI acquires data in the Fourier domain one point at a time, usually until enough data has been collected to approximately satisfy Nyquist sampling conditions. Thus MRI scans are on the order of minutes to tens of minutes and this makes MRI particularly sensitive to motion. Existing methods can robustly correct small amounts of motion but correcting large motions is an active area of research [2]. In cardiac MRI, motion comes from cardiac and respiratory motion. Respiratory motion is relatively small and can be corrected. Cardiac motion during diastole is also relatively small and correctable and thus data is usually collected during this period (Figure 1). Current implementations have a fixed data acquisition duration as shown in Figure 1 and are triggered by the R-peak. Using the results of this project, we aim to modify this scheme to create a dynamic-length data acquisition sequence to maximize signal-to-noise ratio (SNR) efficiency (SNR efficiency ∝ SNR/ √ T ). If diastole is shorter than the acquisition length, we will end up collecting data during systole. This data will be too corrupted by motion to recover and all the data collected during the previous interval must be discarded. If diastole is longer than acquisition length, we will wasting time during diastole in which we could have acquired data, reducing SNR efficiency. We would like to estimate diastole length but this is not an easy problem. Instead, since we are already collecting ECG data during the exam and diastole length is correlated to the R-R interval (the time between two consecutive R-peaks) [3, 4], we estimate the R-R interval. This project first implements several classification approaches as a proof of concept. The first approach is a series of classification approaches using logistic regression and support vector machines (SVM) to predict if the next R-R interval will be sufficiently long to accommodate the data acquisition. These methods use an n-gram model of previous R-R intervals as input and output a prediction of the length of the next R-R interval. We then try a softmax classification approach to predict the largest integer number of imaging blocks (IMG, Figure 1) that will fit in the interval. We again use an n-gram model of previous R-R intervals for input and output a predicted maximum integer. A nonlinear autoregressive neural network (NARNN) is then implemented to estimate the length of the next R-R interval. The input to the NARNN is the ECG time series.