Seite - 338 - in Intelligent Environments 2019 - Workshop Proceedings of the 15th International Conference on Intelligent Environments

3. Methodology The machine-learning stacking scheme used in this study is presented in Fig. 1. The following subsections describe in detail the key steps of the pipeline. 3.1. Filtering and segmentation According to Choi et al [2], the majority of cardiovascular sounds are most likely to occur in the frequency range below 1 kHz, thus, we filtered the raw audio files using low-pass Butterworth filter with a threshold of 1 kHz. For segmentation of the filtered audio signal, we used a sliding window of 1 s with an overlap of 0.5 s. This provides audio segments with a duration of 1 s. The size of the window was chosen with the reasoning that for an average heart rate of 60 beats per minute, the 1 s window should contain at least one heartbeat, thus containing all the relevant information about the CHF- related sounds. 1. RECORDING (RAW) 2. FILTERING 3. SEGMENTATION 4. FEATURE EXTRACTION 5. FEATURE SELECTION 6. SEGMENT CLASSIFIER 7. SEGMENT AGGREGATION 8. RECORDING CLASSIFIER EVALUATION Figure 1. Machine-learning stacking scheme used for the analysis performed in this study. 3.2. Feature extraction 3.2.1. Audio features We used the OpenSMILE tool [3] for large feature-space extraction. The tool was originally created for acoustic emotion recognition in 2009 but later expanded to more general uses. For example, in addition to affect recognition, it is widely used for music information retrieval (e.g., chord labelling, beat tracking, and onset detection). We extracted 1941 audio features, including statistical features (e.g., variance, skewness, kurtosis), energy-based features (e.g., energy in bands from 250 to 1 kHz), frequency- based features (25 %, 50 %, 75 %, and 90 % spectral roll-off points) and voicing-related features (e.g., jitter, shimmer, Harmonics-to-Noise Ratio). The complete list of features is described by Schuller et al. [4]. M.Gjoreski etal. /TowardEarlyDetectionandMonitoringofChronicHeartFailure338