Artificial intelligence based platform for the automatic and simultaneous explainable detection of apnoea, oxygen desaturation, and artefacts in paediatric polygraphy exams (REST).

The gold standard for the diagnosis of sleep apnoea (SA) is polysomnography, consisting of overnight in-lab tests, which are expensive for both patients and healthcare systems. Airflow and pulse/oximetry signals contain most of the necessary information for detecting SA and widely simplify the data acquisition process, hence holding the promise to increase the availability of SA diagnosis and reduce waitlists. Deep learning has recently shown some interesting steps forward in analysing these signals in paediatric patients. Here we introduce a novel platform, REST, that is able to simultaneously detect the presence of apnoea, desaturation, and artefacts in input signals. To achieve this goal, we developed a novel 1D deep neural network architecture that leverages prior knowledge of the information distribution across signals, allowing for the concurrent detection and interpretation of target events. The platform was trained, validated, and tested on data from 86 paediatric patients. We show that our approach outperforms other three approaches from the literature, reaching 92.50% (1.10%), 98.30% (0.43%), and 97.59% (0.28%) balanced classification accuracies for apnoea, desaturation, and artefact, respectively (mean and standard deviation, in brackets). Notably, the REST platform also gives a confidence score as output, highlighting to the doctor the samples that need to be reviewed and further boosting the performances of the other samples. Lastly, based on gradient-weighted class activation mapping (grad-CAM) heatmaps, our platform allows the explanation of the decision process, pointing out the regions of the input signals in which events occur, increasing the reliability of the whole process for a human user.