Utilizing Machine Learning for the Identification of Chronic Heart Failure (CHF) from Heart Pulsations

Over 26 million people worldwide suffer from chronic heart failure (CHF), and its prevalence is increasing at a rate of 2% annually. Automated approaches for detecting CHF are limited, even within the academic community, despite the significant burden CHF presents and the ubiquity of sensors in our daily lives. In this study, we introduce a cardiac hypertrophy heart sound detection technique that integrates traditional machine learning (ML) with full-stack deep learning (DL). Traditional ML relies on manually crafted features to learn patterns, whereas DL leverages the signal's spectro-temporal representation to achieve the same goal. Recordings from 947 participants were used to evaluate this approach, sourced from one custom-built CHF dataset and six existing datasets. Using the same evaluation standards as the baseline method employed in a recent PhysioNet challenge, the proposed approach achieved a score of 89.3 out of 100. Although the overall accuracy of the methodology is 92.9%, with an error rate of 7.1%, this is comparable to the percentage of recordings labeled as “unknown” by experts (9.7%). Furthermore, we identified 15 key features with an accuracy of 93.2% that can be utilized to develop ML models capable of distinguishing between the decompensated and recompensated phases of CHF. The proposed technique has demonstrated promising results in differentiating recordings from healthy individuals and those with CHF, as well as in identifying various stages of CHF. This innovation could pave the way for the development of at-home CHF monitoring systems that help reduce hospitalizations and improve patient outcomes.