ABSTRACTEvaluating long‐term changes in thermal comfort can be useful for considering measures against thermal‐related health risks. In this study, spatio‐temporal changes in thermal comfort, using the Universal Thermal Climate Index (UTCI), were calculated from observations at 140 weather stations across Japan for the 1980–2020 period. To derive the mean radiant temperature (MRT) values not readily measured at the stations but required in the estimation of UTCI, a machine‐learning model (XGBoost) was developed. The model uses the station observations, coarse‐resolution radiation data from the ERA‐5 reanalyses dataset, and available globe temperature measurements as predictors. The trend of UTCI throughout Japan in summer was found to be significantly positive. Meanwhile, negative trends were found in stations located in northern areas during the winter. This suggests that not only heat stress risks but also cold stress risks should be given careful attention in colder regions. Lastly, a comparison of the estimated UTCI with prefecture‐level daily summertime heat‐stroke data reveals that the UTCI threshold above which heat‐stroke cases rise drastically varies distinctly between warm and cold regions, with the latter having a lower threshold. These findings could contribute to the estimation of risks attributable to climate change and to better planning of climate‐change‐resilient cities.