Development of a Disaster Survivor Interview Dataset in South Korea and a Comparative Study of Feature Extraction Techniques

This study constructed the Korean Disaster Survivor Interview (KDSI) dataset and utilized it to compare and analyze the results of training models based on different audio feature groups. Using audio data labeled with PTSD, anxiety, and depression, four deep learning models were trained and evaluated across five feature groups. While there were slight differences in performance metrics depending on the feature group, no significant differences were observed, and overall, high AUC and F1 Scores were achieved. Additionally, the KDSI dataset was found to be effective for the development of disaster psychological diagnosis models, with the ensemble approach particularly demonstrating its effectiveness.