Page - 337 - in Intelligent Environments 2019 - Workshop Proceedings of the 15th International Conference on Intelligent Environments

can accompany heart failure worsening and can be heard using phonocardiography. One of these “typical” CHF-associated sounds is the third sound (S3) that appears 0.1-0.2 s after the second sound S2. Unfortunately, clinical worsening of the CHF patient likely means that we are already dealing with a fully developed CHF episode. Recently, it has been demonstrated that some physiological parameters (such as appearance of additional heart sounds or increased blood pressure in the pulmonary circulation) already start to show a several weeks before CHF patient develops a clinically evident heart failure deterioration. This is also a window where outpatient-based treatment interventions can reverse CHF deterioration and return the patient to the compensated state without the need for hospital admission. In this paper, we focus on the detection of state of CHF (compensated vs. decompensated) based on the analysis of heart sound recordings. Our work builds upon the initial studies where we demonstrated that it is possible to distinguish between healthy individuals and patients in the decompensated CHF episode using a stack of machine-learning classifiers, showing promising results on a limited dataset [1]. We expand upon this approach using a considerably larger patient dataset. Furthermore, we investigate the difference in heart sounds during the transition, between decompensated and recompensated state of CHF, with the aim of developing personalized monitoring models. Early detection of worsening of CHF has the potential to lead towards reduction of hospitalizations due to worsening CHF, with both improving the quality of life of patients and decreasing the financial and logistic burden on the patient and the health system. 2. Dataset Heart sound recordings were obtained using a professional digital stethoscope 3MTM Littmann Electronic Stethoscope Model 3200. Our dataset (Table 1) contains recordings of 110 “healthy” people (meaning that they had no medical condition that would manifest itself in abnormal heart sound) and 51 people diagnosed with CHF. For 22 CHF patients, recordings were obtained both during the decompensation episode (when hospitalized) and during the compensated phase (when discharged from the hospital). The recordings were always obtained at Erb’s point and each recording was up to 30 s long (stethoscope’s limit). For some healthy people, more than one recording was obtained to increase the amount of data in the study (recordings of patients were obtained in clinical settings which limited the available time). The study was approved by the medical ethics committee beforehand. Table 1. Overview of the experimental data recorded on healthy individuals and on patients in decompensated and recompensated CHF episodes Decompensated Recompensated Healthy Overall # Subjects 51 22 110 183 # Recordings 52 22 159 233 # Segments 2017 865 6272 9154 Duration (min) 17 7 52 76 M.Gjoreski etal. /TowardEarlyDetectionandMonitoringofChronicHeartFailure 337