Sequential Probability Assignment for Outlier Detection in Heartbeat Timings

Artifacts and noise as well as outlier beats from physiological causes can lead to inaccurate estimates of heart rate and heart rate variability. Especially with the increased popularity of wearables, there is an increased need to be able to identify when motion artifacts and poor sensor contact lead to incorrect detection of heartbeats. In this paper, we propose a sequential probability assignment procedure to detect outlier heartbeats. The procedure uses a time-varying point process model that estimates a two-parameter exponential family distribution per time index. By allowing both parameters of the distribution to vary with time, this model has more flexibility than many previous models and is able to capture changes in both the mean and variance of the intervals. We formulate a maximum likelihood problem with a Kullback-Leibler regularizer at each time step. The usage of an exponential family parametrization makes the estimation at each time point a convex optimization problem, guaranteeing that the solution we find is optimal. We test three different distributions: inverse Gaussian, gamma, and log-normal. We find that the inverse Gaussian fits the distribution of interbeat intervals from clinical electrocardiogram data the best when evaluated using the Kolmogorov-Smirnov statistic. We then show in simulations as well as in clinical data the model's ability to successfully detect outliers.Clinical relevance-Identification of heartbeat timings can be difficult in noisy settings. In addition, ectopic beats and arrhythmic events can produce irregular timings. This outlier detection is one method to help identify timings that are statistically unlikely.