Prediction of cardiac cycle duration for cardiac-gated closed-loop auricular vagus nerve stimulation

Auricular vagus nerve stimulation (aVNS) is a neuromodulation technology that establishes balance in the autonomic nervous system and, in turn, provides therapy for numerous chronic ailments. Personalized aVNS adapts the stimulation parameters in accordance with the time-varying physiological state of the body, and is suggested to improve the therapeutic outcomes and reduce side effects. The physiological state is estimated via recorded biomarkers such as the electrocardiogram (ECG). aVNS can be delivered in synchrony with any phase of the cardiac cycle before and after the R-peak. This paper proposes the prediction of the duration of the next cardiac cycle after the detected R-peak for the realization of the personalized cardiac-gated closed-loop aVNS applied at any time point during the predicted cardiac cycle. We propose and explore the feasibility of four different prediction methods for predicting the duration of the next cardiac cycle. Two methods are respiration-insensitive, last value and averaging, and the other two are respiration-sensitive, extrapolation and interpolation. Offline recorded ECG waveforms were used to evaluate the different methods. Subsequently, three of the four methods (last value, averaging, and extrapolation) were implemented in real-time on a proprietary aVNS hardware setup, with the data acquisition performed across normal and paced deep breathing. Offline evaluation of the methods revealed that extrapolation and interpolation achieved lower prediction errors during deep breathing with the median absolute error (MdAE) of 32.09 ms (interquartile range 16.07-56.61 ms) and 31.71 ms (15.5-54.06 ms), respectively, as compared with the averaging and last-value methods with 88.75 ms (58.73-124.15 ms) and 40.85 ms (19.7-68.4 ms), respectively. During normal breathing, all evaluated methods yielded lower prediction errors relative to the averaging method 28.5 ms (15.2-43.7 ms). Real-time implementation validated these methods for closed-loop cardiac-gated aVNS, with the best performance achieved by the extrapolation method with 31.4 ms (15.17-55.9 ms) during paced deep breathing. During normal breathing, comparable performance across prediction methods favors the computationally simple last-value approach (MdAE: 31.6 ms). Proposed methods establish the potential of ECG-based R-peak prediction in real-time as a reliable and individual biomarker for the personalized cardiac-gated aVNS, creating a foundation for future clinical applications of aVNS.