Machine Learning and End to End Deep Learning for Detection of Chronic Heart Failure from Heart Sounds

Heart failure that does not resolve over time is known as chronic heart failure (CHF). This debilitating ailment occurs when the heart is unable to pump blood to the body's organs and tissues at the pressures necessary for metabolism. The occurrence of CHF is increasing at a rate of 2% each year, reaching epidemic proportions in the population. One to two percent of the general population and ten percent of those 65 and up in the industrialised world are affected by CHF. At present, around 2% of the yearly healthcare expenditure goes towards the cost of diagnosing and treating CHF. The United States spent about $35 billion on CHF treatment in 2018, and that number is projected to grow in the coming decade. The 5-year survival rate of this population is just around 50%, which is a very bad prognosis for congestive heart failure (CHF), even if medical and device-based therapy techniques have improved significantly over the recent decades. In the normal progression of congestive heart failure (CHF), the patient will go through phases where they are feeling fine and showing no indications of fluid overload, and phases where symptoms and signs of systemic fluid overload are more noticeable, such as orthopnea, peripheral edoema, liver congestion, and pulmonary edoema. To successfully achieve a negative fluid balance and return to the compensatory state, patients typically need to be admitted to the hospital during the latter episodes and treated with intravenous drugs (diuretics, inotropes). If a treating physician could quickly identify when a patient's HF is getting worse, they may make adjustments to the patient's outpatient medical therapy and keep them out of the hospital.At present, a skilled doctor can tell whether heart failure is getting worse just by looking at the patient and looking for certain changes in their blood that are called biomarkers for heart failure. It is unfortunate that when a patient's CHF symptoms worsen, it usually indicates that they are already in the midst of a fully established episode of CHF and will likely need to be admitted to the hospital. Also, phonocardiography can detect distinctive alterations in heart sounds that occur in some patients when their heart failure worsens. Hence, this project employs a state-of-the-art end-to-end average aggregate recording model that incorporates features extracted from both machine learning and deep learning to detect chronic heart failure from phonocardiography (PCG) data. Results from both the proposed ChronicNet model and standalone ML/DL models were also evaluated.