Automatic Atrial Fibrillation Detection Using Voice Analysis Algorythm

Background: An early identification of atrial fibrillation (AF) has been a long-standing clinical challenge. The ESC guidelines recommend screening patients > 65 years of age using pulse palpation or an ECG rhythm strip. These methods have limited effectivity. Analysis of natural speech signals has been used as a monitoring tool for various medical conditions. Recently it has been reported to enable estimation of heart rate.

Aim: In this study we evaluated the efficacy of vocal features analysis in the detection of AF and in discriminating between sinus rhythm (SR) and episodes of AF.

Methods: Prospective multicenter (2 centers) study has been conducted. Consecutive 86 patients with persistent AF admitted for cardioversion were enrolled. Prior to cardioversion, the patients voiced specific vowels and words according to a pre-specified protocol. An ECG tracing was simultaneously recorded. These recordings were repeated in SR following cardioversion. The recordings of the first 34 patients were used to develop an algorithm of AF detection based on analysis of acoustic features in SR and AF. The algorithm was then validated in all 86 patients: 25% of SR recordings were used to train the algorithm and the remaining 75% of SR and 100% of AF data from every patient were tested with the algorithm to distinguish between AF and SR.

Results: The total of 513 recordings were analyzed. Classification of the recordings as AF or SR was performed using varying cutoff values of the separation parameter. The resulting curve showing the specificity and sensitivity of the developed algorithm for distinguishing AF rom SR is presented in Figure .

Conclusion: This study demonstrates the feasibility of detecting AF and discriminating it from SR using analysis of acoustic features extracted from spoken vowels. The potential use of this method for wider population screening will be further evaluated.