Ambreen Shafqat, Eunsik Park
The rising frequency of extreme temperature events due to climate change presents an urgent public health challenge, particularly regarding their complex impact on psychiatric emergencies and suicide. This study investigated the complex relationship between multidimensional apparent temperature (AT) anomalies and suicide mortality in Seoul, South Korea (2014-2021), utilizing a dual inferential and predictive framework. Analyzing 17,710 suicide deaths, we applied Distributed Lag Non-linear Models (DLNM) to estimate exposure-response associations across three AT indices (Humidex, Thermal Workload Parameter (TWP), and Thermal Workload Parameter with Radiation (TWPR)) and multidimensional anomaly classifications. To evaluate the operational utility for public health mobilization, we implemented an Extreme Gradient Boosting (eXGBoost) machine learning classifier with a chronological 70/30 train-test split to predict high-mortality days. DLNMs revealed a temporal divergence in risk: AT indices triggered immediate mortality increases (RR ≈ 1.28 at lag 0), whereas complex multidimensional anomalies exhibited robust delayed effects (RR = 1.002 at lag 3), reflecting cumulative psychological exhaustion. The eXGBoost models demonstrated moderate discrimination for suicide mortality (AUC ≈ 0.70-0.72) with high sensitivity (0.714-0.750) and moderate specificity (∼0.52-0.58) across all thermal indices. Anomaly indicators achieved comparable performance (AUC ≈ 0.67-0.70) across lag structures. The predictive performance was higher and more stable for females (AUC ≈ 0.61-0.64) than for males (AUC ≈ 0.47-0.50) and highest in the age 35-64 years group (AUC ≈ 0.63-0.66). These findings indicate that environment-based suicide early warning systems should prioritize high sensitivity to identify high-risk periods, supporting proactive public health awareness, even at the cost of moderate false-positive alerts.