Developing and comparing machine learning models to detect sleep apnoea using single-lead electrocardiogram (ECG) monitoring

BackgroundSleep apnoea has a high disease burden but remains underdiagnosed, in part due to the expensive and resource intensive nature of polysomnography, its definitive investigation. Emerging literature suggests that it may be possible to detect sleep apnoea using single-lead ECG signals, such as those obtained from smartwatches. In this study, we use two forms of recurrent neural networks (RNNs) to detect sleep apnoea events from single-lead ECG signals. MethodsWe use single-lead ECG data from the PhysioNet Apnea-ECG database, which contains data from 70 patients. We train a bidirectional gated recurrent unit (GRU) model and a bidirectional long short-term memory (LSTM) model on labelled ECG signals from 35 patients and test the models on the remaining 35 patients in the dataset. ResultsBoth models achieved 97.1% accuracy, sensitivity and specificity to detect whether the ECG recordings belonged to a patient diagnosed with sleep apnoea. This corresponds to 34/35 patients in the dataset. At detecting individual apnoea events, the GRU and LSTM models achieved 90.4% and 91.7% accuracies respectively. DiscussionThe models achieved high levels of accuracy, specificity and sensitivity. Bidirectional RNNs are strengthened by the ability of the models to be informed by both past and future states when analysing sequential data, such as ECGs. The models also require minimal human intervention as they automatically extract features from the data. If single-lead ECGs prove a suitable tool for sleep apnoea detection, this may enhance the diagnosis of sleep apnoea and potentially allow widespread screening for the condition. ConclusionsWe note that using models such as bidirectional RNNs has the potential to augment model performance. However, more research and validation is required in order to test whether these may be applicable to other datasets and in clinical practice.