Automated detection of atrial fibrillation based on vocal features analysis

Introduction: Early detection of atrial fibrillation (AF) is desirable but challenging due to the often-asymptomatic nature of AF. Known screening methods are limited and most of them depend of electrocardiography or other techniques with direct contact with the skin. Analysis of voice signals from natural speech has been reported for several applications in medicine. The study goal was to evaluate the usefulness of vocal features analysis for the detection of AF. Methods: This prospective study was performed in two medical centers. Patients with persistent AF admitted for cardioversion were enrolled. The patients pronounced the vowels “Ahh” and “Ohh” were recorded synchronously with an ECG tracing. An algorithm was developed to provide an “AF indicator” for detection of AF from the speech signal. Results: A total of 158 patients were recruited. The final analysis of “Ahh” and “Ohh” syllables was performed on 143 and 142 patients, respectively. The mean age was 71.4 ± 9.3 and 43% of patients were females. The developed AF indicator was reliable. Its numerical value decreased significantly in sinus rhythm (SR) after the cardioversion (“Ahh”: from 13.98 ± 3.10 to 7.49 ± 1.58; “Ohh”: from 11.39 ± 2.99 to 2.99 ± 1.61). The values at SR were significantly more homogenous compared to AF as indicated by a lower standard deviation. The area under the receiver operating characteristic curve was >0.98 and >0.89 (“Ahh” and “Ohh,” respectively, p <.001). The AF indicator sensitivity is 95% with 82% specificity. Conclusion: This study is the first report to demonstrate feasibility and reliability of the identification of AF episodes using voice analysis with acceptable accuracy, within the identified limitations of our study methods. The developed AF indicator has higher accuracy using the “Ahh” syllable versus “Ohh.”.