Cardiac arrythmias affect millions of individuals worldwide and certain types can lead to serious complications such as stroke or cardiac arrest. Arrhythmias remain sometimes difficult to diagnose with ambulatory electrocardiogram monitors due to its transient nature. Nevertheless, recent advances in wearable single-lead ECG devices are promising for pre-screening of abnormal rhythms in large populations as they are relatively comfortable and can be worn over long periods of time, particularly at night. Herein, we evaluate an embeddable algorithm to detect abnormal rhythms in real time from a single lead ECG signal acquired on 20 patients for 24 hours in daily life conditions. We use a Firstbeat™ Bodyguard 3 as single-lead ECG monitoring device and a 3 lead ECG Holter (Spiderview) with an analysis software for abnormal rhythm detection (Synescope® analysis software) for the reference. The algorithm is composed of a beat detector to extract RR intervals and a classifier for abnormal rhythm detection. The classifier consists of two gated recurrent unit (GRU) layers, takes windows of RR intervals of about 30 seconds as input, and returns 4 rhythm classes: abnormal, bradycardia, tachycardia, and normal. We trained the classifier on five databases including 286'645 30s windows of RR intervals (training: 143'090, test: 94'463, validation: 49'092) from 283 subjects. We achieved the detection of abnormal rhythms with a sensitivity and specificity of 90.7% and 91.1% respectively. The detection of normal rhythm showed a sensitivity of 88%, with 7% detected as abnormal, and a specificity of 99.6%. The performance of our approach, based on training data acquired from a different device and using a distinct R-peak detector, shows promising results for its integration on daily life ECG monitoring device and for performing early detection of cardiac arrhythmia.