Determination of Relevant Feature Combinations for Detection Stunting Status of Toddlers

Early detection of stunting status in toddlers must be done by recognizing important factors that affect the stunting status of toddlers to be precise and accurate. The stunting status of this study was normal and stunting. A two-stage filter method has been used to determine the subset of dataset features capable of producing a stunting status detection system with the expected accuracy. The first stage was carried out to discover the important features related to the stunting status of toddlers. These important features are obtained by measuring the level of relevance of dataset features to stunting status (Target Class). The level of relevance was measured using univariate and mutual information filter methods. The univariate types used were ANOVA and Chi-Square, while the mutual information types used were Information Gain and minimal-Redundancy-Maximal-Relevance (mRMR). The second stage aims to determine the subset of dataset features that differentiate the target class. The feature subset is formed using a forward selection technique starting from features with a high level of relevance to the target class. The addition of feature subset members is based on the ranking order obtained. Naive Bayes, K-Nearest Neighbour, and Support Vector Machine are applied to determine classification accuracy using a combination of features provided by the filter method. The experimental results show that the best accuracy is produced by SVM without using a filter. Prediction of stunting status will decrease its performance if one or more features are removed. The results of this study can contribute to the preprocessing step to determine the best features in the early detection of stunting status in toddlers.