Optimizing the short- and long term regression for QRS detection in presence of missing data

Recently presented new QRS detection algorithm uses a detection function based on the value of angle between two regression segments adjacent in a given point. The optimization of these segments, representing long-term and short-term acceleration of heart activity, is presented in this paper. The detection algorithm is based on four simple steps: (a) iterative linear regression based on samples selected in two windows of different length Δt_L and Δt_M (b) calculation of series of angle values between regression segments adjacent to each point, (3) expressing the synchronicity in both series as a detection function and (4) adaptive thresholding of the detection function. The method was tested with original records from MlTDB for diferent combinations of thresholding parameters and for Δt_L and Δt_M varying in ranges of 41.775 ms and 8.3-41.7 ms respectively. The quality of the detection was measured with false detection rates and represented by the area under the ROC curve. For the normal ECG, the values yielding the highest area of ROC for sensitivity and positive predictive value of QRS detection are respectively: Δt_L= 58.3 ms and Δt_M = 19.5 ms.