Heart failure, a slow, progressive condition, where heart muscles are unable to pump sufficient blood to different parts of the body, can progress to a heart attack. To reduce the risk of heart attack, we need to regularly monitor the symptoms of heart failure.
Weight gain is one such symptom that can be monitored at home. But weight gain generally becomes apparent only in later stages of heart failure.
Another clue, related to heart disorders, is speech. When we speak, vibrations produced by the vocal tract are modified by the glottis, a valve in the throat that controls airflow. The glottis lies between two vocal folds which can swell in heart conditions. This affects phonation and respiration when speaking.
So Madhu Keerthana from IIT Kharagpur collaborated with M Kiran Reddy and others in Finland to propose automated heart failure detection by monitoring speech signals. They extracted vocal tract and glottal excitation features from a database of speech signals of healthy people and heart patients.
They found that people with heart failure tend to speak using a softer glottal flow pulse, richer in low-frequencies than observed in healthy controls. The glottal flow of healthy speakers showed a clear closed phase and a short open phase. In people with heart failure, the closed phase was practically absent from the glottal flow and the shape of the glottal pulse was more rounded.
The spectrogram of healthy speakers showed a clear harmonic structure, especially at low frequencies. In contrast, the spectrogram of those with heart failure showed a blurred harmonic structure and contained more noise-like components.
The researchers used all significant features identified as inputs to train four different machine learning classifier algorithms to identify the speech signals of heart failure patients.
The feedforward neural network algorithm, which learns non-linear relations between input and output training data, showed better classification accuracy of more than 80 per cent when trained using both vocal tract and glottal features.
“We could get such high accuracy even though we were working with a limited dataset,” says Madhu Keerthana, IIT Kharagpur.
“Using speech signals as biomarkers for heart failure is cost-effective. And comfortable for patients since it is non-invasive,” says M Kiran Reddy, her collaborator from Finland.
The findings can be used to design a simple device for epidemiological studies on the prevalence of heart conditions in a population and a computer application for monitoring individual patients in hospitals, clinics and, perhaps, even at home. Time for medical technology companies to step in.
Induja M S,
Freelance writer, Kochi
STEAMindiaReports: straight from the heart of Indian science
Science news that news media in India can use free of cost