Nik Noor Syamimi Ismail1*, Muhd Zulfadli Hafiz Ismail1, Evi Diana Omar1, Zamtira Seman1, Ridwan Sanaudi1, Anis Syakira Jailani2, Hasnah Mat1, Yusrina Mohd Yusoff1, Najjah Tohar1, Asyraf Syahmi Mohd Noor1, Ainil Zafirah Abd Karim1, Nabilah Hanis Zainuddin1, Munerah Mohamad Zaman1, Mohd Azahadi Omar1
1. Sector for Biostatistics & Data Repository, Office of NIH Manager, National Institutes of Health, Ministry of Health, Malaysia, Setia Alam, 40170 Shah Alam, Selangor.
2. Research and Grant Management Unit, Office of NIH Manager, National Institutes of Health, Ministry of Health, Malaysia, Setia Alam, 40170 Shah Alam, Selangor.
*Corresponding author: Nik Noor Syamimi Ismail.
ABSTRACT
INTRODUCTION: Patient socio-demographics significantly influence outpatient healthcare utilisation, impacting service delivery and resource planning. Traditional methods may overlook complex predictor interactions. Machine learning offers enhanced predictive capabilities, especially with large-scale datasets, allowing for more accurate classification of outpatient care types and informing targeted healthcare policies.
OBJECTIVE(S): This study evaluates random forest (RF) and XGBoost model performance in forecasting utilisation of outpatient facilities among patients, and identifying key features in predicting the utilisation type.
MATERIALS & METHODS: A total of 1,604 records from the National Health and Morbidity Survey (NHMS) 2019 were analysed using 19 socio-demographic features. The target variable was the binary outcome of outpatient utilisation type, either private or government health clinics. The dataset was split into training (70%) and testing (30%) sets. RF and XGBoost models were used to predict outpatient utilisation type. Model performance was evaluated using sensitivity, specificity, positive predictive value, balanced accuracy, F1-score, and area under the curve (AUC). Feature importance was identified using SHAP (SHapley Additive exPlanations).
RESULTS: RF slightly outperformed XGBoost in most metrics: sensitivity (0.923 vs 0.920), specificity (0.410 vs 0.375), precision (0.785 vs 0.774), balanced accuracy (0.666 vs 0.647), and F1-score (0.845 vs 0.841). XGBoost showed a marginally higher AUC (0.738 vs 0.730), indicating slightly better discrimination. The top five key features from the overall models identified were respondents’ age, financial health coverage (employer), total household monthly income, respondents’ highest education level, and respondents’ ethnicity.
CONCLUSION: Both RF and XGBoost demonstrated promising predictive performance for the outpatient utilisation type. The key features highlight the importance of socio-economic and age-related factors in determining healthcare facility choice, reinforcing the value of machine learning in healthcare planning.
KEYWORDS: Machine Learning, Random Forests, XGBoost, Outpatient, Healthcare


