# Internal KCN: Yupeng Tian

Yupeng will present their work on a R-peak detection algorithm of textile waist-recorded ECG signals using HDIG (History Dependent Inverse Gaussian) model, and textile waist-ECG noise modelling.

IG (Inverse Gaussian) model has great implications for neural activities as well. Neuron firings and heart beats have similar bio-physiological mechanisms. The processes both include polarization and depolarization of membrane potentials, and a threshold. The first passage time distribution of a random walk process with a fixed threshold can be proved to follow IG mathematically. And in this sense, both ISI (inter-spike interval) and RR interval (interval between two R-peaks) distributions can be more appropriately modelled in IG rather than the traditional normal distributions. We’ll show first that IG models can be a more appropriate estimations of ISI, then present the ECG R-peak detection algorithm (with HDIG) and the results.

We assume that RR intervals follow IG. For the algorithm, we use HDIG as a real-time updating version of IG: the current distribution is history dependent, i.e. we modify the parameters of current IG based on the previously observed RR intervals. This will make the current prediction of RR more precise. At each R-peak, we search for the next R-peak in an interval around the right-end point of the predicted RR interval. Since the RR interval distributions are accurate, we are doing an optimal search, i.e. searching around the most probable point that the next R-peak will happen. Thus the R-peak detection prediction accuracy is greatly improved. We can achieve more than 95% in textile waist-recorded ECG jogging data, which are contaminated with serious motion artifacts. Traditional methods like Pan & Tompkins can only achieve ~70% accuracy in such noisy datasets.

Then I’ll present our work on textile waist-ECG noise modelling. The purpose is to (1) find which type of noise affects R-peak detection accuracy most; (2) generate massive textile-like data using the modelled features of noise. In this way, we can address the important noise class for R-peak detection and test the algorithms on much larger ECG datasets.

The following references will be helpful in understanding the rationale of the algorithm: