Machine Learning-Based Classification of Family Planning Participant Status Using Random Forest and the CRISP-DM Framework

Abstract

The Family Planning (FP) program requires accurate information to support evidence-based decision-making and improve the quality of reproductive health services. Classification of FP participant status can assist health authorities in identifying participant patterns and monitoring program implementation. Previous research using the Support Vector Machine (SVM) algorithm on the same dataset achieved an accuracy of 56.20%, indicating that improvements in classification performance are still required. This study proposes the Random Forest algorithm within the Cross-Industry Standard Process for Data Mining (CRISP-DM) framework to classify FP participant status. The dataset consists of 1,402 FP participant records obtained from SATPEL PPKB, Cilebar District, Karawang Regency. Data preprocessing included data transformation, One Hot Encoding for categorical predictor variables, Label Encoding for the target variable, and Hold-Out Validation with an 80:20 train-test split using stratified sampling. The predictor variables were registration month, wife's birth year, wife's age, and contraceptive method, while the target variable was FP participant status (New, Change Method, and Repeat). Model performance was evaluated using accuracy, precision, recall, F1-score, confusion matrix, classification report, and feature importance analysis. The Random Forest model achieved an accuracy of 59.43%, with weighted precision, recall, and F1-score of 59.00%. However, the macro-average precision, recall, and F1-score were 45.00%, 44.00%, and 44.00%, respectively, indicating performance differences across classes caused by class imbalance. The model achieved the highest F1-score for the New class (0.63), followed by the Repeat class (0.59), whereas the Change Method class obtained the lowest F1-score (0.11). Feature importance analysis identified wife's birth year and wife's age as the most influential predictor variables. Compared with the previous SVM-based model, Random Forest provided a modest improvement in accuracy and enhanced model interpretability through feature importance analysis. Nevertheless, the low macro-average performance indicates that further research should investigate class-balancing techniques and hyperparameter optimization to improve classification performance, particularly for minority classes.