Page - 339 - in Intelligent Environments 2019 - Workshop Proceedings of the 15th International Conference on Intelligent Environments

3.2.2. Heart rate variability features Heart rate variability (HRV) describes the variation in the time interval between the heartbeats (R-R intervals). Fig. 2 presents an example of filtered audio data with marked R-R intervals that were detected automatically using a peak detection algorithm that finds the maximum of the first heart tone, S1. Once the peaks were detected, the R-R intervals simply represent the distance between these peaks. We calculated the following HRV features: mean of the R-R intervals, the standard deviation of the R-R intervals, the square root of the mean of the squares of differences between the adjacent R-R intervals, the percentage of differences between adjacent R-R intervals that are greater than 50 ms, and the Poincaré plot indexes of HRV (SD1 and SD2). Each of the segments within the same recording was then attributed the HRV- based feature value of the entire recording [5]. Figure 2. Representative filtered segment with detected R-R intervals 3.3. Feature selection Since the number of available features (6 HRV and 1941 audio features) is several times larger than the number of available recordings, we selected only the best features to avoid overfitting. We ranked the features using Mutual information values between the features and the class values on the training data. Mutual information is a measure that estimates the dependency between two random variables. In the training phase, we used only the top-ranked n features, where n was set to be equal to the number of training samples. 3.4. Machine learning classifiers The ML pipeline contains two ML classifiers, a segment-based and a recording- based classifier. The segment-based classifier takes as input 1 s segments (represented via the extracted features), and outputs the estimated class probabilities for each segment. The recording-based classifier takes as input a recording, and outputs estimated class probabilities for the recording. The motivation here is the fact that all the segments in a chosen recording belong to the same class, therefore the aggregation of segment-based forecasts should improve the overall classification. M.Gjoreski etal. /TowardEarlyDetectionandMonitoringofChronicHeartFailure 339