Reliability of Self-Applied Smartphone Mechanocardiography for Atrial Fibrillation Detection

Abstract

Smartphone mechanocardiography (sMCG) is a new technique that allows patients to record a cardiac rhythm strip using a smartphone built-in tri-axial accelerometer and gyroscope. In this study, we investigated how a self-applied sMCG can reliably contribute to the differentiation of atrial fibrillation (AFib) from the sinus rhythm (SR). A study sample of 300 elderly adults including 150 AFib cases with persistent and paroxysmal AFib was recruited. Among the 300 subjects, 182 subjects (82 AFib) completed two recordings, one physician-applied and one self-applied. The remaining patients (n = 118) were nervous, in a quite poor condition, and not interested in or capable of concentrating on using a smartphone in the acute setting of a hospital in order to perform the self-applied recording. Two data processing frameworks were used, knowledge-based learning (KL) which is a rule-based algorithm and machine learning (ML) which is an automated classification technique. For the ML approach, we considered four classifiers, namely random forest (RF), extreme gradient boosting (XGB), support vector machines (SVM), and artificial neural network (NN). For evaluation, a leave-one-subject-out cross-validation was adopted for the ML approach. Compared to physician-interpreted ECG-derived labels, the KL approach predicted AFib with sensitivity values of 0.963 and 0.976, specificity values of 0.980 and 0.929, and F-measure values of 0.972 and 0.952 for the physician- and self-applied measurements, respectively. Similarly, NN which was the best ML classifier according to the F-measure values, demonstrated, on average, sensitivity values of 0.976 and 0.938, specificity values of 0.962 and 0.936, and F-measure values of 0.969 and 0.937, respectively. All other classifiers delivered quite similar results. The sMCG technology for AFib detection, supported by the KL and ML approaches, can accurately differentiate AFib from SR in both physician- and self-applied recording scenarios. This new technology can help to screen patients with episodic or undiagnosed AFib and also be used as a home-based self-applied monitoring technique.

Keywords

• Accelerometers
• biomedical engineering
• cardiography
• gyroscopes
• machine learning
• signal processing