Aims: A Hidden Markov model (HMM) is a powerful tool used to model a Markovian stochastic process by learning the dynamics of observation sequences (examples) generated by the process itself. However, besides estimating the set of parameters and selecting the appropriate number of states of the model, the selection of the length of the observation sequence is commonly addressed using a priori information of the process to be modeled. In this sense, the aim of this study is to analyze the influence of the duration of continuous univariate observation sequences on HMM for the detection of QRS complexes in single-lead ECG recordings.
Methods: For the characterization of the QRS complexes dynamics, the first ECG channel of the MIT-BIH Arrhythmia database was employed. The length (L) of the observation sequence was varied from 50 ms to 190 ms. For each L, the first 5 min of the recordings were used to train a 2-state HMM and the rest was used to test the detection performance. In the test phase, once the HMM was trained, a detection was signaled when the log-likelihood of observation of the model was above a decision threshold. The area under the curve (AUC) of the receiver operating characteristic plot was used to assess the detection performance for each L.
Results: For observation sequences of L = {50, 70, 90, 100, 110, 130, 150, 170, 190} ms, the performance obtained was AUC = {0.035, 0.008, 0.110, 0.380, 0.053, 0.061, 0.050, 0.030, 0.033}, respectively. The best AUC was 0.380 to characterize observation sequences of 100 ms length.
Conclusion: The best 2-state HMM-based QRS complex detection performance was obtained for observation sequences of 100 ms, which is associated with the average duration of QRS complexes, from a physiological point of view.