Most heat-health studies identified thresholds just outside human comfort zones, which are often too low to be used in heat-warning systems for reducing climate-related health risks. We refined a generalized additive model for selecting thresholds with substantial health risk enhancement, based on Taiwan population records of 2000-2017, considering lag effects and different spatial scales. Reference-adjusted risk ratio (RaRR) is proposed, defined as the ratio between the relative risk of an essential health outcome for a threshold candidate against that for a reference; the threshold with the highest RaRR is potentially the optimal one. It was found that the wet-bulb globe temperature (WBGT) is a more sensitive heat-health indicator than temperature. At lag 0, the highest RaRR (1.66) with WBGT occurred in emergency visits of children, while that in hospital visits occurred for the working-age group (1.19), presumably due to high exposure while engaging in outdoor activities. For most sex, age, and sub-region categories, the RaRRs of emergency visits were higher than those of hospital visits and all-cause mortality; thus, emergency visits should be employed (if available) to select heat-warning thresholds. This work demonstrates the applicability of this method to facilitate the establishment of heat-warning systems at city or country scales by authorities worldwide.