Acoustic diagnosis of pulmonary hypertension: automated speech- recognition-inspired classification algorithm...

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence in Medicine

Background:

• Pulmonary hypertension (PH) diagnosis can be challenging.
• Accurate and timely diagnosis of PH is crucial for patient outcomes.
• Current diagnostic methods may have limitations in accessibility and accuracy.

Purpose of the Study:

• To develop and evaluate an automated classification algorithm inspired by speech recognition for diagnosing PH using heart sounds.
• To compare the diagnostic performance of this algorithm against experienced physicians.

Main Methods:

• Simultaneous recording of heart sounds, electrocardiograms, and mean pulmonary artery pressures (mPAp).
• Digitization of heart sound recordings for algorithm training and testing.
• Feature extraction using mel-frequency cepstral coefficients and classification with Gaussian-mixture models.
• Physician diagnosis based on listening to heart sounds, blinded to patient data.

Main Results:

• The automated algorithm achieved a correct diagnostic rate of 74%, significantly outperforming physicians at 56% (p=0.005).
• The algorithm demonstrated a lower false positive rate (34% vs. 50%, p=0.04) and a substantially lower false negative rate (23% vs. 68%, p=0.0002) compared to physicians.
• The study included 164 subjects, with 86 diagnosed with PH (mPAp ≥ 25 mmHg) and 78 without (mPAp < 25 mmHg).

Conclusions:

• An automated speech-recognition-inspired algorithm can accurately diagnose pulmonary hypertension from heart sounds.
• This algorithm surpasses physician performance in correctly identifying PH, with fewer false positives and negatives.
• The developed algorithm shows potential for widespread screening of PH, facilitating earlier specialist referral and intervention.