Introduction: Atrial fibrillation (AF) is a very common cardiac rhythm abnormality, present in approximately two percent of all people, i.e. in approximately 140 million people globally. Due to failure of the atrium to contract effectively, a blood clot may form within, which may enter blood flow at a later time and in some cases block a blood vessel in the brain, resulting in a stroke. Up to 25 percent of all strokes are caused by AF. AF sometimes begins as asymptomatic (“silent AF”), in which case an otherwise healthy patient remains unaware of the condition. However, the risk of suffering a stroke due to periodical silent AF is similar to recurrent or persistent AF; hence, when silent AF is detected, similar anticoagulant medication is prescribed as in the case of symptomatic continuous AF. We propose detecting AF with IMU (Inertial Measurement Unit) sensors available in modern smartphones and wearable devices.

Purpose: Our objective is to provide simple and cost-effective means for detecting AF (including the “silent AF”). Since many existing mobile devices, such as smartphones, are today equipped with accelerometers and gyroscopes (IMU), they can be used for monitoring the operation of the heart. The purpose is to prevent the consequences of undetected AF with early prevention.

Methods: In order to detect AF, a mobile/wearable device is placed on the chest of the patient, and an accelerometer and gyroscope measurement recording is taken from the subject. The patient is advised to lie in a prone/supine position when the measurement is taken. The procedure is non-invasive and can be taken without support from medical staff (or other persons). An automated algorithm extracts features such as autocorrelation and spectral entropy from the pre-processed (FFT filtering and potential noise exclusion) IMU data. A machine learning algorithm, which operates based on the extracted features, is used for classification. We used state-of-the-art machine learning methods such as SVM (Support Vector Machine), Kernel SVM (KSVM) and Random Forest (RF) classifier. The used data was captured in controlled conditions from total of 16 AF patients and additional 20 recordings from healthy volunteers. The data from each person was divided into 10 second non-overlapping segments for feature extraction and classification. The learning part was implemented in desktop environment and the critical data acquisition was implemented on a smartphone.

Results: We used 10-fold cross validation to verify classifier performance. The best performing classifier was KSVM with sensitivity of 98.5% and specificity of 95.2%. Out-of-bag classification error for RF was 4.95% (200 grown trees).

Conclusions: We proposed a method for detecting AF using smartphone only solution without any add-on hardware. It can be implemented as a software solution for the existing smartphones globally available.