Though in some ways replaced by ultrasound technology, cardiac auscultation–using a stethoscope to listen to a patient’s heart–remains an important screening modality for recognizing heart disease. Auscultation serves as a cost-effective screening tool for heart disease and is of particular importance in several clinical scenarios. Less emphasis has been placed on training US clinicians in auscultation, however, making this something of a “lost art.” This may delay a patient’s diagnosis of heart disease.
Using a digital stethoscope for input, a trained machine learning algorithm could help support the listener with an interpretation of heart sounds. Much of the work I performed was in the important phases of training called normalization and feature selection. Normalization involves making small optimizations on the original sounds that help reduce uninformative variation. Feature selection then determines which characteristics are most important for distinguishing normal sounds from abnormal sounds.
- Obtaining standardized patient heart sounds
- Normalizing heart sounds of different amplitudes and rates
- Dividing heart sound signals into individual beats
- Developing a strategy for selecting optimal features to train a neural network
- Dupont Award for Research in Medicine and Health
- New York Academy of Sciences Francine Salom Memorial Award
- Bausch and Lomb Science Award