Detection of Respiratory Effort-Related Arousals Using a Hidden Markov Model and Random Decision Forest

The efficient detection of respiratory effort-related arousals requires enormous amount of data and a suitable learning model. Using a dataset taken from PhysioNet.org, windows of 20 seconds were extracted with their median aligned with the starting point of the arousals. The same amount of data was selected from non-arousal regions. Features derived using these windows were reduced to 38 by using various feature selection methods. A cross-validated Random Forest (RF) was used for the evaluation. The training data was processed with a 20-second sliding window and a 1 second resolution. Windows were labelled according to their temporal location in the data. This was used to train three separate RFs on different parts of the data, which provided a probability emission model. The probability values used in a Hidden Markov Model and established the most probable path with the Viterbi algorithm. Probability values were aggregated based on the Viterbi path, then smoothed and resampled to match the original sample rate. This method achieved an 0.29 score of AUPRC.