OBJECTIVES: In the event of a disaster, the unmet need for healthcare access prevails among disaster victims, which is an imminent public health risk. It is essential to predict the vulnerable population and their different levels of need for efficient disaster management. Given this, we aimed to investigate the level of unmet healthcare needs of disaster victims and build a better unmet needs prediction model using machine learning.

METHODS: We used the 2017 Disaster Victims Panel Survey, a nationally representative sample survey of disaster victims designed for the relief policy and technology in South Korea. Three machine learning approaches (logistic regression, C5.0 tree-base-model, and random forest) with 80% of the dataset were used for prediction. The 10-fold cross-validation was performed with the remaining 25% of the dataset. Features were selected based on Andersen and Newman's health service utilization framework and disaster-related factors such as type of disaster, income decrease and debt increase after the disaster, and disaster-related disease or injury. AUC-ROC, specificity, and sensitivity tested the predictive performance of models.

RESULTS: The data contained 1,659 subjects, and 31.5% of subjects experienced unmet healthcare access needs. The random forest model had the best predictive performance with AUC-ROC, specificity, and sensitivity of 0.83, 74.0%, and 79.2%, respectively. Perceived health status, residence area and disaster-related disease or injury were the most important features in prediction.

CONCLUSIONS: This study presents the significantly higher unmet need for healthcare access among disaster victims even within the universal coverage of health insurance in South Korea and current public policy for disaster victims. It supports the need for special post-disaster care of healthcare service access for disaster victims.