5:20 PM - 5:50 PM
[II-IL-09] 2 Recent Advances in Pediatric Pulmonary Hypertens
Listening to the voice of the heart: Acoustic diagnosis of pulmonary hypertension using automated speech and language recognition inspired algorithms
The diagnosis of pulmonary hypertension is delayed or missed during the early stages of the disease, when it is most treatable, because the clinical signs may be difficult to discern. There is, therefore, a pressing and unmet need to explore diagnostic screening methods that are cost effective, non-invasive and result in a timely diagnosis. We hypothesized that an automated speech- recognition-inspired classification algorithm could differentiate between the heart sounds in subjects with and without pulmonary hypertension and would outperform trained clinicians. Heart sounds, electrocardiograms, and pulmonary artery pressures were recorded simultaneously during catheterization of the pulmonary artery subjects undergoing right heart cardiac catheterization. Digitised heart sound recordings were used to train and test speech-recognition-inspired classification algorithms. We used mel-frequency cepstral coefficients to extract features from the heart sounds, and built Gaussian-mixture models to classify the features as pulmonary hypertension or non-pulmonary hypertension. Physicians, blinded to all clinical data and patient identity listened to heart sound recordings and attempted a diagnosis. We studied 164 subjects: 86 subjects with pulmonary hypertension (mean pulmonary artery pressure 40.0 ± 11.5 mmHg) and 78 without pulmonary hypertension (mean pulmonary artery pressure 16.5 ± 4.5 mmHg) (p <0.005). The correct diagnostic rate of the automated speech-recognition-inspired algorithm was 74% compared to a 56% by physician diagnosis (p < 0.05). The false positive rate of diagnosis for the algorithm was 34% versus 50% (p<0.05) for physicians and false negative diagnostic rate for the algorithm was 23% versus 68% (p<0.05) for clinicians. Our results suggest that changes in the heart sounds are directly related to increased pulmonary artery pressure and that computer-aided auscultation could be used as a screening tool in the early diagnosis of pulmonary hypertension. We have developed an automated speech-recognition-machine learning-inspired classification algorithm for the acoustic diagnosis of pulmonary hypertension that outperforms clinicians.
The diagnosis of pulmonary hypertension is delayed or missed during the early stages of the disease, when it is most treatable, because the clinical signs may be difficult to discern. There is, therefore, a pressing and unmet need to explore diagnostic screening methods that are cost effective, non-invasive and result in a timely diagnosis. We hypothesized that an automated speech- recognition-inspired classification algorithm could differentiate between the heart sounds in subjects with and without pulmonary hypertension and would outperform trained clinicians. Heart sounds, electrocardiograms, and pulmonary artery pressures were recorded simultaneously during catheterization of the pulmonary artery subjects undergoing right heart cardiac catheterization. Digitised heart sound recordings were used to train and test speech-recognition-inspired classification algorithms. We used mel-frequency cepstral coefficients to extract features from the heart sounds, and built Gaussian-mixture models to classify the features as pulmonary hypertension or non-pulmonary hypertension. Physicians, blinded to all clinical data and patient identity listened to heart sound recordings and attempted a diagnosis. We studied 164 subjects: 86 subjects with pulmonary hypertension (mean pulmonary artery pressure 40.0 ± 11.5 mmHg) and 78 without pulmonary hypertension (mean pulmonary artery pressure 16.5 ± 4.5 mmHg) (p <0.005). The correct diagnostic rate of the automated speech-recognition-inspired algorithm was 74% compared to a 56% by physician diagnosis (p < 0.05). The false positive rate of diagnosis for the algorithm was 34% versus 50% (p<0.05) for physicians and false negative diagnostic rate for the algorithm was 23% versus 68% (p<0.05) for clinicians. Our results suggest that changes in the heart sounds are directly related to increased pulmonary artery pressure and that computer-aided auscultation could be used as a screening tool in the early diagnosis of pulmonary hypertension. We have developed an automated speech-recognition-machine learning-inspired classification algorithm for the acoustic diagnosis of pulmonary hypertension that outperforms clinicians.