Page - 341 - in Intelligent Environments 2019 - Workshop Proceedings of the 15th International Conference on Intelligent Environments

4.2. Decompensated vs. recompensated analysis In this scenario, we aimed at some level of personalization. For 22 out of 51 patients, there is one recording in the decompensated phase (i.e., at the beginning of hospitalization) and one recording in the recompensated phase (i.e., at hospital discharge). Since the number of paired recordings is small and not suitable for ML experiments, we preformed statistical tests to check whether there is a statistically significant difference in the feature values when calculated from the recompensated recordings compared to the decompensated recordings (i.e., before and after a medical intervention). We used the Wilcoxon signed-rank test, which is a non-parametric statistical hypothesis test that tests whether two related paired samples come from the same distribution. In particular, it tests whether the distribution of the differences x - y is symmetric about zero [6]. In this case, x are the values of the features extracted from the decomposition recordings and y are the values of the features extracted from the decomposition. The Wilcoxon test is an alternative to the paired Student's t-test, when the population cannot be assumed to be normally distributed, such as in our case. The statistical tests showed that there are 10 features for which there is a statistically significant change in the distribution. Those features are: 99th percentile of the first derivative of the 4th MFCC coefficient; Quartile deviation of the first derivative of the 4th MFCC coefficient; 1st percentile of the first derivative of the spectral roll-off; 1st percentile of the standard deviation of the spectral roll-off; Percentile range (0-1) of the Spectral roll-off; Flatness of the psychoacoustic sharpness; Percentile range (0-1) of the psychoacoustic sharpness; Percentile range (0-1) of the first derivative of the psychoacoustic sharpness; Standard deviation of the psychoacoustic sharpness; 99th percentile of the spectral entropy. The Mel-frequency Cepstrum Coeffiecients (MFCC) are the coefficients of the MFC representation of the sounds. The MFC is a representation of the short-term power spectrum of the sound [7]. The spectral roll-off is a measure of the amount of the right- skewedness of the power spectrum of the signal. Similarly, the psychoacoustic sharpness and the spectral entropy are two features which quantify the spectral characteristics of the sound. The boxplots of the 10 features are presented in Fig. 3. The boxplots clearly show that there is difference in the values of the features between the recordings of patients in recompensated and decompensated episodes. 5. Discussion In line with our previous research based on the heart sound analysis [1], the stacking- based approach is promising in determining patient’s state of CHF and detecting subclinical signs of threatening CHF deterioration. The accuracy is somewhat lower than what we obtained when comparing healthy controls to CHF patients. However, before, we were dealing only with CHF patients in the fully decompensated state, where the M.Gjoreski etal. /TowardEarlyDetectionandMonitoringofChronicHeartFailure 341