With a large amount of data collected from studies of sleep quality and based on the physiological signals (PS) that are collected, it is possible to use mechanisms that intelligently detect sleep disorders such as arousals (ARS). In this detection, the triggers can be present in any of the PS or can occur from their combinations. Thus, with the characterization of the PS and with a considerable number of examples, it is possible to generate a model that recognizes ARS zones in new samples. In this way, by segmenting the signals and decomposing them into variable frequency bands, thanks to the application of discrete wavelet transform (DWT), it is possible to characterize the contributions of each PS in time and frequency. The features that are extracted give information about the contributions in frequency and time of each PS. Then these characteristics feed a neural network model that iteratively learns the best non-linear function that approximates the input to its corresponding label. Once the methodology was tested, with less than 3% of the training data, it was possible to reach an Area Under Precision-Recall Curve (AUPRC) of 0.261.
It is possible to exploit the predictive capacity of data collected in intensive care units (ICU) with a high ratio of missing values. Combining several sources of information, a considerable number of missing values are generated. In this manuscript, an alternative approach to impute this type of data, together with the use of deep learning techniques to improve the early detection of sepsis in ICU is proposed. Initially, laboratory tests are separated and summarized. Then, their most representative information is extracted by taking codes from an autoencoder. This information is combined with the rest of the variables and used to exploit temporal dependencies through long short-term memory recurrent neural networks. With the proposed approach our team, WIN-UAB, was ranked in the position 38/78 with a utility score (defined in the the PhysioNet/Computing in Cardiology Challenge 2019) of 0.241 on the full test set. The predictive capacity of the proposed solution demonstrated the potential of integrating an alternative approach for imputing variables with a high ratio of missing values. In terms of dimensionality reduction, it is possible to reduce 27% of features through the codes of autoencoders.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.