Preprocessing of predictor data and Handling Sample Clauses

Preprocessing of predictor data and Handling of missing values‌ The dataset is screened for highly redundant variables, i.e. variables that carry the same or nearly the same information, that have derived after transformation. For example, if A is a categorical variable of N levels (e.g. ‘marital status’), a redundant variable may be the one that derives after the merging of the N levels into two (e.g. ‘alone’). Only one relevant variable is kept. Furthermore, competing predictors are combined to form new variables. The dataset is subsequently checked for the existence of very high correlations among predictors (>0.8). Certain machine learning algorithms cannot operate on categorical data directly. They require all input and output variables to be numeric (e.g. XGBoost, svm etc). An integer encoding had already been applied to the categorical variables of the dataset. However, for categorical variables with more than two categories, such an action imposes an ordinal relationship where no such relationship exist. To overcome this problem, we apply the dummy encoding technique, which represents N categories/levels with N-1 binary variables (the method avoids redundancy). Predictors that only have a single unique value (“zero-variance predictor”) or a number of unique values but with the second more frequent value occurring with a very low frequency (“near-zero-variance predictor”) are removed. The reason is to avoid unstable fitting of the classifier, or an undue influence that few samples may have on the final classification model. Pre-processing is performed in R using the caret package. Missing values may occur due to skip pattern in survey and data collection design, e.g. certain questions are only asked to respondents who have given a certain answer to a previous question. This type of missing is treated before applying any imputation technique. Patients and predictors with high percentage of missing values (>30%) are removed. For the results presented here single imputation has been applied using missForest package of R (Nonparametric Missing Value Imputation using Random Forest). The developed framework also supports multiple imputation with mice and kml packages of R. Next step planning includes the examination of the effect of different/multiple imputations on the classification results. Multiple imputation allows for missing rate of this magnitude [I1]. Furthermore, missForest can successfully handle missing values up to 30%, in datasets including different types of variables [I2].