Novel imputation method using average code from autoencoders in clinical data

It is possible to improve the reconstruction of clinical data combining codes from autoencoders (AE). The extracted information can be used for enhancing existing imputation methods in this type of data. In the proposed approach, initially, encoder and decoder functions from trained autoencoder are extracted. Then, imputers equally spaced from normalized distribution of the variables generate codes that are combined in the average one that is finally used to reconstruct the original information. The proposed method is compared imputing by mean values of variables and using a single AE for reconstruction. The proposed approach has an outstanding performance recovering original information. It is even better with missing values in more than one variable. The error is at least 70% less than the other methods imputing one variable, and also the proposed approach is highly recommended with missing values in more than one variable.