Inflection point analysis: A machine learning approach for extraction of IEGM active intervals during atrial fibrillation.

OBJECTIVE: In this paper, we propose a novel algorithm to extract the active intervals of intracardiac electrograms during atrial fibrillation.

METHODS: First, we show that the characteristics of the signal waveform at its inflection points are prominent features that are implicitly used by human annotators for distinguishing between active and inactive intervals of IEGMs. Then, we show that the natural logarithm of features corresponding to active and inactive intervals exhibits a mixture of two Gaussian distributions in three dimensional feature space. An Expectation Maximization algorithm for Gaussian mixtures is then applied for automatic clustering of the features into two categories.

RESULTS: The absolute error in onset and offset estimation of active intervals is 6.1ms and 10.7ms, respectively, guaranteeing a high resolution. The true positive rate for the proposed method is also 98.1%, proving the high reliability.

CONCLUSION: The proposed method can extract the active intervals of IEGMs during AF with a high accuracy and resolution close to manually annotated results.

SIGNIFICANCE: In contrast with some of the conventional methods, no windowing technique is required in our approach resulting in significantly higher resolution in estimating the onset and offset of active intervals. Furthermore, since the signal characteristics at inflection points are analyzed instead of signal samples, the computational time is significantly low, ensuring the real-time application of our algorithm. The proposed method is also robust to noise and baseline variations thanks to the Laplacian of Gaussian filter employed for extraction of inflection points.