Seite - 342 - in Intelligent Environments 2019 - Workshop Proceedings of the 15th International Conference on Intelligent Environments

Figure 3. Boxplot and p-values that show a statistically significant difference in the values when calculated from the recompensated recordings compared to the decompensated recordings (i.e., before and after the CHF therapy). effects of increased ventricular filling pressures on heart sounds are much more pronounced. In addition, the dataset for CHF patients was considerably larger and more diverse in the current study, which makes the results more representative. Furthermore, the inclusion of the recording-based features (derived from HRV) expands the feature room upon which we operate. The results of personalized approach in detecting preclinical worsening of CHF condition are encouraging. We have identified at least 10 features that show statistically significant difference for CHF patients before and after the medical intervention. All of the 10 features quantify the spectral representation of the sounds. Thus, the spectral features might be more informative compared to the temporal features for the specific task. To further check whether these features can be used to build personalized models, a larger dataset will be required. Ideally, a pilot study would include CHF patients regularly recording their heart sounds, perhaps using a modified mobile/wearable device instead of a professional stethoscope. However, a labelled dataset for the specific modified mobile/wearable device would be required in order to tune the ML models. In future work, we plan to analyze the effects that different hyperparameters have on the methods accuracy. This analysis includes sampling rates, window size, more advanced feature selection methods, e.g., wrapper method instead ranking methods. Alternatively, dimensionality reduction methods instead of feature selection, e.g., deep autoencoders, can turn out to be efficient. An intrinsic problem in building personalized models lies in the fact that only a small number of recordings will always be available for each individual. Here, the data from healthy patients can be utilized by using multi- task learning techniques, where one task would be healthy vs. non-healthy and the other task would be decompensated vs recompensated. Finally, we plan to test our method on other related datasets, e.g., the dataset used in the PASCAL Classifying Heart Sounds Challenge [8] or the dataset used in the PhysioNet/Computing in Cardiology Challenge [9]. M.Gjoreski etal. /TowardEarlyDetectionandMonitoringofChronicHeartFailure342