Fig. 1

Data Flow and Feature Engineering Pipeline This pipeline shows the flow of data through our clinical pipeline. This pipeline shows how the data is bootstrapped, subsetted and filtered in order to obtain higher quality notes for use in the proceeding feature selection and classification. If CUIs are to be utilized, then the cTAKES/YTEX pipeline is used to create the initial features through usage of the Collection Processing Engine within the cTAKES suite. This data is output as a sparse matrix which we convert in order to conform to the style of the other feature engineering techniques so each algorithm can be used independently of the data it is being given. If CUIs are not being used, then a stemming process is undergone as this is needed for both BOWs and the inversion method. In the case of using BOWs, punctuation and stop words are additionally removed in order to reduce bias in the dataset. If the inversion method is to be used, then we leverage Word2Vec to create two Word2Vec models which are fine-tuned according to which phenotype they represent. All feature sets are subjected to normalization and feature selection through scikit-learn’s ExtraTreesClassifier’s variable importance to prep for classifier usage [22]