Seasonal prediction systems (SPSs) have the potential to inform forecasts of heat-related risks, but their application to human heat stress remains limited by the availability of radiation data required to compute thermal stress indices and by the generally low predictive skill for European summer climate. Here, we develop a methodological framework to derive the Universal Thermal Climate Index (UTCI) and Wet Bulb Globe Temperature (WBGT) from SPS outputs by reconstructing the radiation components required for their computation. Validation against ERA5 shows that the proposed methodology provides accurate estimates of both indices, with root-mean-square error (RMSE) generally below 1.2 K for UTCI and 0.4 K for WBGT. We then investigate whether teleconnection-based subsampling can improve seasonal predictions of human heat stress by enhancing the representation of the summer North Atlantic Oscillation (NAO) in the European Centre for Medium-Range Weather Forecasts (ECMWF) SPS SEAS5. The subsampling is implemented within a "fair-sliding" framework designed to emulate real-time forecasting conditions over 2003–2024. Results show that subsampling enhances NAO prediction skill, increasing the correlation between the SEAS5 and ERA5 summer NAO indices from 0.40 to 0.63, while reducing the RMSE from 1.10 to 0.91. The enhanced NAO skill is accompanied by better predictions of summer-mean UTCI and WBGT, as well as the number of summer days exceeding heat stress thresholds, particularly over western, central, and northern Europe. The gains in thermal stress prediction are primarily driven by better skill for temperature- and radiation-related components, while improvements in humidity and wind are weaker or more regionally confined. Overall, these results demonstrate a pathway to extend SPS capabilities toward seasonal forecasting of heat stress conditions, supporting health-oriented climate services.