DAEimp: Denoising autoencoder-based imputation of sleep heart health study for identification of cardiovascular diseases

Since it has been recognized that the disordered breathing during sleep is related to cardiovascular diseases, it is possible to predict cardiovascular diseases from sleep breathing data, which however is usually inevitable to have missing data, resulted probability from the loss to follow-up, failure to attend medical appointments, lack of measurements, failure to send or retrieve questionnaires, and inaccurate data transfer. In this paper, we propose a denoising autoencoder-based imputation (DAEimp) algorithm to impute the missing values in the sleep heart health study (SHHS) dataset for the predication of cardiovascular diseases. This algorithm consists of three major steps: (1) based on the missing completely at random assumption, the random uniform noise is added to the positions of missing values to convert missing data imputation into a denoising problem, (2) feed the noisy data and a missing position indicator matrix into an autoencoder model and use the reconstruction error, divided into observation positions reconstruction error and missing positions error, for denoising, and (3) the logistic regression is applied to the generated complete dataset for the identification of cardiovascular diseases. Our results on the SHHS dataset indicate that the proposed DAEimp algorithm achieves state-of-the-art performance in missing data imputation and sleep breathing data-based identification of cardiovascular diseases.