Climate change projections indicate that the world's most populated regions will experience more frequent, intense and longer-lasting heatwave periods over the coming decades. Such events are likely to result in widespread overheating in the built environment, with a consequential increase in heat-related morbidity and mortality. In order to warn the population of such risks, Heat-Health Warning Systems (HHWSs) are being progressively adopted world-wide. Current HHWSs are, however, based solely on weather observations and forecasts and are unable to identify precisely where, when, or to what extent individual buildings (and their occupants) will be affected. In contrast, AutoRegressive models with eXogenous inputs (ARX) have been demonstrated to reliably forecast indoor temperatures in individual rooms using minimal data. Thus, the large-scale deployment of forecasting models could theoretically enable the development of a high-resolution indoor HHWS (iHHWS). In this study, ARX models were tested over the long-lasting UK heatwave of 2018 using hourly monitored dry-bulb temperature data from 25 rooms (12 living rooms and 13 bedrooms) in 12 dwellings, located within the London Urban Heat Island (UHI). The study investigates different approaches to improving the reliability of room-based heat exposure predictions at longer forecasting horizons. The effectiveness of the iHHWS system was assessed by evaluating the accuracy of predictions (using fixed and adaptive temperature thresholds) at different lead times (1, 3, 6, 12, 24, 48 and 72 h ahead). Compared to forecasted indoor temperatures, a Cumulative Heat Index (CHI) metric was shown to increase the reliability of heat-health warnings up to 24 h ahead.