1354195 - Prediction of Aortic Hemodynamics Using Convolutional Neural Networks (CNN) and Time-Frequency Transformation of Chest Vibrations Measured by Seismocardiography (SCG)

4D flow MRI provides a comprehensive assessment of cardiovascular hemodynamics. However, because of utilizing advanced MRI sequences, 4D flow MRI can be costly, time-consuming, and not readily available in many clinical environments. Seismocardiography (SCG) measures the chest vibrations with an inexpensive wearable device through an easy-to-perform 2-min surface recording of the chest. We hypothesize that SCG can accurately predict 4D flow MRI measures, such as the systolic aortic peak velocity (V_max). We demonstrate that time-frequency representations of the SCG data and the demographic attributes of the subjects can serve as input data to deep neural networks (DNN) in order to predict the V_max within good accuracy compared to the reference standard 4D flow MRI.

Methods:

We recruited twelve patients with aortic valve diseases (3 females, age: 56.2 ± 13.5 years) and forty-six healthy control subjects (20 females, age: 45.9 ± 17.2 years) with no known history of cardiovascular diseases to receive 4D flow MRI (spatial resolution: 1-3 mm³, time resolution: 30-40 ms, venc = 150-375 cm/s, 1.5T or 3T). We also used a custom-designed, wearable cardiac sensor incorporating a MEMS accelerometer to acquire the SCG measurements on the same day.  Continuous wavelet transform (CWT) was used to obtain the time-frequency representation of each measured SCG pulse. We used the images of scalograms (N = 3919) to train a convolutional neural network (CNN). Each CNN unit was composed of a sequence of a Conv2D layer, a ReLU activation layer, a Batch Normalization layer, and a MaxPooling2D layer. We also created a multi-layer perceptron (MLP) to incorporate the demographic information of the study participants. The outcome layers of the MLP and CNN models were combined to predict the systolic V_max in the aorta. 4D flow MRI was used to calculate the ground-truth V_max along the ascending aorta (AAo). We statistically compare the results obtained by the DNN model with 4D flow MRI assessment using correlation and Bland-Altman analyses.



Results:

Peak systolic velocities predicted by deep learning were in excellent agreement with 4D flow derived V_max, as demonstrated by low absolute difference of 3.9 ± 1.6% between methods over five random trials. There was also a strong linear correlation between DNN-estimated and measured V_max (y = 0.78x + 30.6, r² = 0.86, p< 0.01). In addition, the Bland-Altman plots of all V_max for DNN vs. 4D flow MRI indicated low non-significant bias (p = 0.41) and small limits of agreement.   

Conclusion:

We demonstrated that the scalograms of the SCG recordings of chest vibrations can predict the aortic V_max using a DNN. This study suggests that cardiac SCG obtained using easy-to-use and low-cost wearable electronics can be utilized as a supplement to CMR exams to screen patients for aortic valve abnormalities or when advanced imaging such as 4D flow MRI is not available.