Skillful heat-related mortality forecasting during recent deadly European summers

DRAFT 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 Skillful heat-related mortality forecasting during recent deadly European summers Emma Holmberga,b,c,2, Marcos Quijal-Zamoranod,e,f,g, Joan Ballesterd,1, and Gabriele Messoria,c,h,1 This manuscript was compiled on October 24, 2025 Europe is a heatwave hotspot: numerous temperature records have been broken in recent summers, and roughly 60,000 and 50,000 heat-related deaths occurred in the summers of 2022 and 2023, respectively. With recent summers, like that of 2022, projected to become the new norm, there is a pressing need to further develop heat-health warning systems to help society adapt to a warming climate. Here, we forecast heat-related mortality by applying a statistical epidemiological framework to temperature forecasts extending up to two weeks in advance. Focusing on two recent and exceptional summers in Europe, namely 2022 and 2023, we evaluate the skill of the daily heat-related mortality forecasts, and assess its association with temperature. For most of Europe, milder temperatures, close to the minimum mortality temperature, are associated with more skillful heat-related mortality forecasts. However, some of the hottest regions in Europe instead showed enhanced forecast skill associated with higher temperatures. This suggests that heat-related mortality forecasts can provide valuable information in European regions associated with high levels of heat-related mortality. Consequently, we advocate for local health authorities to include information from forecasts of heat-related mortality in their heat warning systems. Heat-related mortality | early-warning systems | impact forecasting | temperature extremes | Europe R ecent years have repeatedly witnessed unprecedented temperatures, with the summers of 2022 and 2023 being the warmest on record for Europe and globally, respectively ( 1 , 2 ), at the time of occurrence. Europe in particular is emerging as a heatwave hotspot ( 3 ), against the background of a global increase in heatwave frequency, duration and intensity under climate change ( 4 ). Extreme temperatures are linked to multifarious detrimental human health impacts, notably excess mortality ( 5 – 7 ), and increased extreme heat in Europe imposes a heavy burden on society. The European summer of 2022 was associated with over 60,000 heat-related deaths (8), followed by over 47,000 heat-related deaths in 2023 (9). ( 9 ) showed that the 2023 heat-related mortality (HRM) burden would have been 80% higher without adaptation processes that occurred in Europe this century. The importance of adaptation is evidenced by the case-in-point of the summers of 2003 and 2006. The summer of 2003 witnessed over 70,000 excess deaths in Europe ( 10 ), and resulted in the implementation of heat early-warning systems in several European countries ( 11 ). The summer of 2006 again saw parts of Europe experience a severe heatwave. However, in France the observed mortality was markedly less than expected ( 12 ), which was partially attributed to the implementation of early-warning systems after the 2003 heatwave ( 12 ). With summers like the aforementioned projected to become the new norm in the coming decades, the seasonal distribution of deaths will rapidly change. The increase in HRM will outstrip the decrease in cold-related mortality by the second half of the century for most future climate scenarios ( 13 ) unless further adaptations to rising temperatures occur. Heat-health early-warning systems (e.g. 14 ) are one such key adaptation, supporting the broader effort to adapt society to a changing climate. Here, we analyse whether numerical forecasts of temperature enable skillful forecasts of HRM on time scales of several days to two weeks, which forms the core of effective heat-health early-warning systems. We consider the European summers of 2022 and 2023, which are of particular interest for the study of the practical predictability of HRM. Firstly, they exemplify the occurrence of deadly heat in Europe. Secondly, unprecedented temperatures such as those during those two summers are projected to become increasingly common, and are thereby representative of summers to come. Finally, numerical weather prediction (NWP) models are regularly updated to improve model performance, and, on average, forecast capabilities have improved by about 1 day every 10 years ( 15 ). Selecting Significance Statement Heat-related mortality is a growing issue in Europe, and the continued development of heat warning systems is a key aspect of society’s adaptation to a changing climate. To date, many heat warning systems across Europe consider human-health impacts only implicitly. Here, we place the explicit focus on human-health impacts. Specifically, we assess the pan-European skill of heat-related mortality forecasts for 2022 and 2023; two exceptionally hot summers. Our findings demonstrate regionally and temperature dependent skill for heatrelated mortality forecasts. We further note enhanced skill during particularly hot conditions, and thus advocate for the adoption of information from impact-based forecasts in heat warning systems. This information has the potential to be particularly beneficial in helping protect vulnerable members of society. www.pnas.org/cgi/doi/10.1073/pnas.XXXXXXXXXX PNAS — October 24, 2025 — vol. XXX — no. XX — 1–11 DRAFT 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 Author affiliations: a Department of Earth Sciences, Uppsala University, 75236, Uppsala, Sweden; b Centre for Natural Hazards and Disaster Science (CNDS), 75236, Uppsala University, Uppsala, Sweden; c Swedish Centre for Impacts of Climate Extremes (CLIMES), Uppsala University, 75236, Uppsala, Sweden; d ISGlobal, Barcelona, Spain; e Universitat Pompeu Fabra (UPF), 08002, Barcelona, Spain; f Institute of Social and Preventive Medicine (ISPM), University of Bern, Bern, Switzerland; g Oeschger Center for Climate Change Research (OCCR), University of Bern, Bern, Switzerland; h Department of Meteorology and Bolin Centre for Climate Research, Stockholm University, 11418, Stockholm, Sweden All authors contributed to the conceptualisation, methodology, validation and reviewing and editing the manuscript , MQ-Z and JB were responsible for data curation, EH, MQ-Z and JB were responsible for formal analysis and software, GM and JB were responsible for funding acquisition and supervision, EH, GM and JB were responsible for project management, JB was responsible for computing resources, EH was responsible for the visualisation and writing of the original draft. The authors declare no competing interests. 1J.B. and G.M. contributed equally to this work. 2 To whom correspondence should be addressed. Email: emma.holmberg(at)geo.uu.se 2— www.pnas.org/cgi/doi/10.1073/pnas.XXXXXXXXXX Holmberg et al. DRAFT 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 recent summers thus ensures that the analysis is conducted on forecasts which are close to current NWP capabilities. ( 16 ) show that the intensity of temperature extremes can in general be forecasted 1–3 weeks in advance, whilst ( 17 ) showed that temperature-related mortality could in general be forecast more than 8 days in advance based on forecast capabilities prior to 2020. To assess the practical predictability of HRM during recent record-breaking summers we compute HRM forecasts based on temperature forecasts, leveraging the epidemiological association between temperature and HRM. This association is estimated using a non-linear statistical model, meaning that the errors of the HRM forecast may differ to those of the original temperature forecasts. It is thus crucial to investigate the predictability properties of HRM forecasts. To achieve this we first analyse the relationship between temperature and mortality, and the spatial distribution of HRM. We then assess the temporal evolution of temperature and HRM. Finally, we investigate the temporal and spatial distribution of HRM forecast skill for several European regions, ultimately offering a spatiallyand temporally-resolved quantification of the relationship between temperature and HRM forecast skill in Europe. Fig. 1a provides a graphical summary of the connection between the sets of analysis in this study. 1. Results First, we show a visual summary of the analysis conducted in this study (Fig. 1a). We begin the analysis by examining the relationship between temperature and mortality from both continental-average and regional perspectives. The pooled exposure-response association (Fig. 1b) shows a clear nonlinear relationship between temperature and RR, which is especially prominent for heat (temperatures to the right of the curve’s minimum). This minimum point, namely the MMT, shows a marked north-south gradient, with countries bordering the Mediterranean showing the highest values, and more northerly or elevated regions showing lower values (Fig. 1d). The pooled lag-response association at the 99th percentile shows a rapid decay in RR with time after exposure (Fig. 1c). The cumulative RR at the 99th percentile of temperatures (Fig. 1e) again shows a north-south gradient. We now extend our analysis by considering this temperature–mortality relationship from a forecasting perspective. Specifically, we investigate the time series of daily temperature and AF averaged over almost all regions in Europe for the summers of 2022 and 2023 (Fig. 2). These two summers were characterised by consistently aboveclimatology temperatures, with several heatwave periods when temperatures exceeded climatology by 4 ◦ C or more (Fig. 2a,b). As expected, the seven-day lead-time forecasts more closely resembled the observations than forecasts with leadtime 14 days. The continental-mean AF (Fig. 2c–f) broadly follows the peaks of the temperature time series, although the non-linear epidemiological relationship amplifies AF during periods of elevated temperatures. Thus, even a moderate temperature forecast spread at elevated temperatures leads to a large AF forecast spread (Fig. 2c,d). Furthermore, the uncertainty due to the epidemiological fit is also largest for the most elevated temperatures (Fig. 2e,f). We show the spatial distributions of seasonal-average temperature and AF for the summers of 2022 and 2023 in Figs. 1,2,3 of the appendix. Next we evaluate the forecasts to assess how many days in advance the forecasts typically provide ’useful’ information. For both temperature and AF, the predictability horizon varies between several days and almost two weeks, with variations in the AF horizon often but not always matching those in the temperature horizon (Fig. 3a–d, appendix: Fig. 4). During summer 2022, south-western Europe shows mostly enhanced temperature predictability relative to the other regions (Fig. 3a). AF shows a predictability peak for south-western Europe during July 2022 (Fig. 3c), which matches the temperature predictability peak. This corresponds to record-breaking temperatures across south-western Europe at that time. However, AF also displays a prolonged period of heightened predictability in south-eastern Europe starting at roughly the same time, which is not clearly visible in temperature. In 2023, western Europe shows a predictability peak in June, while southwestern and south-eastern Europe show prominent peaks in July (Fig. 3b). During summer 2023, all macro-regions display a peak in AF predictability horizon between early and late July (Fig. 3d), even though western Europe and Europe as a whole do not show particularly prominent peaks in temperature predictability during the same time. As evidenced by the differences between the four analysed macro-regions, the predictability horizons also display a marked spatial variability. In 2022, the predictability horizon for temperature varies between approximately 4–8 days in northern and north-western Europe, whilst reaching up to 12 days in parts of southern Europe (Fig. 3e). 2023 shows a similar north-south gradient, albeit less pronounced than for 2022 (Fig. 3f). During 2022, southern Europe showed a reduced predictability horizon for AF relative to temperature, with differences often in the range of 1–4 days (Fig. 3g, appendix: Fig 5a). Northern Europe, including the British Isles, instead showed an increased predictability horizon. In 2023 the AF predictability horizon is generally higher than for 2022 (Fig. 3h), and is longer than the temperature horizon in many regions, notably in eastern and northern Europe, and Portugal (appendix: Fig 5b). To better understand the varying relationship between the predictability of temperature and AF, we perform regressions between several combinations of temperature, temperature predictability horizon and AF predictability horizon. Summer 2022 displays significant positive regression-coefficients between temperature and temperature predictability, and positive coefficients for temperature predictability versus AF predictability in some parts of France, Spain and Italy bordering the Mediterranean (Fig. 4a,c). Thus, higher temperatures generally benefit temperature predictability, which in turn benefits AF predictability for some parts of the Mediterranean coastline. Positive coefficients for temperature versus temperature predictability are also visible for summer 2023, with southern and south-eastern Europe showing a more pronounced signal than in 2022 (Fig. 4b). However, the signal is weakened when considering temperature predictability versus AF predictability (Fig. 4d). The differences between the two summers become evident when regressing temperature against AF predictability (Figs. 4e,f). In summer 2022 there Holmberg et al. PNAS — October 24, 2025 — vol. XXX — no. XX — 3 DRAFT 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 Fig. 1. Summary graphic of the analysis in this study (a), curves showing the pooled, population-weighted European average of both the association between temperature and the relative risk of mortality (RR) (b) and the lag response at the 99th percentile of temperature (c) where the 95%confidence interval is shaded. The dashed horizontal lines in (b, c) denote RR = 1. Maps of the minimum mortality temperature (MMT) for each region (d) and the RR at the 99th percentile for each region (e). Missing data is in white (d,e). 4— www.pnas.org/cgi/doi/10.1073/pnas.XXXXXXXXXX Holmberg et al. DRAFT 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 Fig. 2. Time series of 2m temperature (temperature, K )(a,b), and Attributable Fraction (AF,%)(c–f) for the summers (June-August) of 2022 (a,c,e) and 2023 (b,d,f). Dashed black denotes climatology, solid black denotes observations, red denotes forecast values (median) at lead time 7 days and blue denotes forecast values (median) at lead time 14 days. For panels a–d we show the median of the forecast ensemble, whilst in panels e,f we show the median from the Monte-Carlo simulations. The shading shows the forecast ensemble spread for temperature (a,b), and AF (c,d) and the 95%confidence interval from the epidemiological fit (e,f). Values are population-weighted averages of all analysed European regions. Holmberg et al. PNAS — October 24, 2025 — vol. XXX — no. XX — 5 DRAFT 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 Fig. 3. Time series’ of 2m temperature (temperature; a,b) and Attributable Fraction (AF; c,d) predictability horizons, for Europe (black), south-western Europe (orange), western Europe (blue) and south-eastern Europe (red) for 2022 (a,c) and 2023 (b,d). Shading denotes the 95%confidence interval determined by bootstrapping the regions of the dataset. Values have been population-weighted prior to averaging. Composites of the predictability horizon of temperature (e,f) and AF (g,h) for 2022 (e,g) and 2023 (f,h). Missing data is shown in white (e–h). 6— www.pnas.org/cgi/doi/10.1073/pnas.XXXXXXXXXX Holmberg et al. DRAFT 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 Fig. 4. Coefficients estimated from negative binomial regression fits for 2m temperature (temperature) vs the predictability horizon of temperature (a,b), the predictability horizon of temperature vs the predictability horizon of Attributable Fraction (AF) (c,d) and temperature vs the predictability horizon of AF (e,f) in 2022 (a,c,e) and 2023 (b,d,f). The significance of the regression coefficients was corrected for multiple tests following (18). Non-significant values are shaded light grey; missing data is in white. Holmberg et al. PNAS — October 24, 2025 — vol. XXX — no. XX — 7 DRAFT 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 Summary of predictability for temperature and AF Temperature AF Region north-south gradient; north-south gradient; higher predictability in southern regions higher predictability in northern regions Year Higher predictability Higher predictability on average for 2022 than 2023 for 2023 than 2022 (climatologically cooler places) Region and year High predictability during the High predictability during the hottest part of 2022 for southern Europe hottest part of 2022 for southern Europe Table 1. Summary of the results varying by region, year, and the combination of region and year. is a significant positive relationship in some regions bordering Mediterranean, and either no significant relationship or a negative relationship elsewhere in the rest of Europe. Summer 2023 displays a widespread significant negative relationship, which is particularly pronounced along the Atlantic coast but also present along the Mediterranean coastline. This indicates that milder temperatures lead to an increased AF predictability horizon. We summarise these results in Table 1, which provides an overview of the predictability of temperature and AF by year and region. 2. Discussion Our analysis highlights the non-linear, and spatio-temporally varying relationship between temperature and AF in a predictability context. For most of Europe, AF forecasts show reduced skill during the hottest days of summer due to the epidemiological transformation. This means that results from the literature on the evaluation of temperature forecasts cannot be easily translated to AF forecasts. Nonetheless, the two are not entirely decoupled, and particularly large forecast errors for temperature (so-called ”forecast busts” ( 19 , 20 )) may also detract from AF predictability, especially if such busts occur during a heatwave. Summertime temperature extremes in Europe result from multiple processes occurring across different spatial scales ( 21 ). There are nonetheless specific large-scale circulation patterns associated with elevated temperatures across the continent, notably persistent high-pressure features termed atmospheric blocking (e.g. 22 ). The mature phase of blocking has been associated with enhanced operational predictability, whilst blocking onset and decay remain challenging to forecast ( 23 , 24 ), potentially leading to large errors in temperature forecasts. This may in turn have direct consequences for AF forecasts. Summer 2022 was characterised by a blocked flow pattern over western and north-western Europe during June and July ( 25 ), which then decayed in late July. The imperfect representation of such flow patterns could explain the comparatively poor 14-day temperature forecast for late July 2022 (Fig. 2a), which did not capture the temporary decrease in temperatures seen in observations. We posit that improvements in NWP models’ capabilities of capturing the onset and decay of blocking events would be of direct relevance to AF forecasting. Indeed, forecast errors at high temperatures are amplified during the transformation to AF, due to the non-linear relationship between the two (Fig. 1b). This non-linear relationship between temperature and AF is also reflected in the transformation of temperature forecast spread to its AF counterpart. Unlike temperature, AF is bounded from below at 0, naturally giving an asymmetry in forecast spread. This asymmetry is also evident for forecasts of larger AF, where ensemble members with higher values for temperature correspond to amplified AF values, whilst temperature values closer to the minimum mortality temperature show a compressed spread in the AF forecast (Fig. 2c,d). We argue that improvements in NWP model performance associated specifically with hot temperatures would be particularly valuable for improving the accuracy of HRM forecasts. Furthermore, the use of data-driven forecast models could represent a fruitful line of inquiry if such models could be specifically tailored either to perform well for hot temperatures when health impacts are greatest, or to directly forecast heat-health impacts (26). Assuming that each ensemble member of any given temperature forecast represents a physically plausible evolution of temperature over Europe, it is possible to attain temperatures corresponding to AF values which are much larger than those we estimated here for the summers of 2022 and 2023. Furthermore, we suggest that recent studies considering worst-case heatwave scenarios (e.g. 27 , 28 ) could be examined using the methodology employed here to help guide preparedness plans. However, we acknowledge the caveat that the epidemiological model employed here is only trained on temperatures that have been observed, meaning that estimations of AF associated with record-breaking temperatures should be considered cautiously. ( 29 ) find that heat warnings based on fixed-temperature thresholds underestimate the number of days when heat affects the population, and call for better alignment between the issuing of heat warnings and societal responses to heat. Impact-based forecasts have the advantage of providing relevant information specifically tailored to the societal response. In order to inform the issuing of vulnerability-specific warnings by heat-health warning systems, epidemiological models could be extended to consider physiological factors (e.g. sex, age, comorbidities), or even socio-economic vulnerabilities. Our results evidence a regionally-varying temperaturedependent skill of AF forecasts, and we highlight the parts of France, Spain and Italy bordering Mediterranean as a particularly interesting case. In 2022, a coherent, statistically significant relationship between temperature and the predictability horizon of AF is visible there (Fig. 4e), which also broadly corresponds to the regions highlighted in 8— www.pnas.org/cgi/doi/10.1073/pnas.XXXXXXXXXX Holmberg et al. DRAFT 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 Fig.1a of the appendix as having the largest AF values, and where the seasonal-mean temperature is well above the MMT (appendix: Fig. 6a). We conjecture that these extreme temperatures seen along parts of the Mediterranean coastline correspond to the far right end of the temperature-RR curve (Fig. 1b), where the relationship is strongly non-linear. Consequently, climatology – which we take as a baseline against which we evaluate our forecasts– would provide a poor estimate of AF. On the other hand, an inverse association between temperature and the predictability horizon of AF can be seen along Europe’s Atlantic coast (Fig. 4e,f). This region is climatologically cooler than the Mediterranean basin, with mean temperatures for the season only slightly above the MMT (appendix: Fig. 6). We thus conjecture that here the relationship is dominated by comparatively good AF forecast performance corresponding to cooler temperatures, close to the MMT. For temperatures close to the MMT the RR curve is relatively flat, meaning that temperature errors in this vicinity would be of little importance to the AF, which would have a value close to 0. We acknowledge a number of limitations of this study. First and foremost we consider only the summers of 2022 and 2023. These years were chosen so as to focus on recent and particularly hot summers. Whilst we argue that they exemplify projected future summers, the small sample size curtails making reliable generalisations, and further verification of the findings presented here is imperative. This should include the use of observed mortality data, where available. Indeed, the verification approach we adopted here is akin to the perfect model approach ( 30 ), as AF forecasts are verified against AF estimations from reanalysis temperature data. A further limitation is the correction of systematic errors in temperature forecasts. Here we adopted a computationally advantageous method, which is however not tailored to AF forecasts. In order to generate optimal AF forecasts, we suggest that temperature forecasts should undergo a tailored bias correction accounting for the non-linear relationship between temperature and AF. At the time of this study, data for some regions was not accessible, particularly in north-eastern Europe. Continued efforts to improve the accessibility of health data are vital for the development of the heat-health field, and to address biases in the focus regions of studies ( 31 ). In particular, the methodology presented here could be expanded to different geographical regions, and we highlight the need for further studies targeting regions in the global south. Our results suggest that such forecasts could be useful for summers hotter than 2022 and 2023, although it is not possible to verify this until such summers occur. The use of climate models could provide a useful avenue of research to simulate potential future record-breaking events, whilst the use of reforecast data from weather models could prove informative for a systematic verification of current forecast capabilities, albeit for hypothetical scenarios. Finally, our findings have important implications in the context of a warming climate. In the absence of meaningful adaptation measures or major advances in NWP modelling, AF forecasts in many European regions will be associated with larger forecast spread and errors for future, unprecedentedly high temperatures. This is particularly pertinent for the Mediterranean basin, which is both climatologically warmer than the European average, and has been identified as a climate change hotspot ( 32 ). The study of future scenarios presents an important avenue for further research, and a key advantage of the methodology used here is that it could be generalised to such set-ups. Furthermore, dynamical changes in the atmospheric circulation can reinforce the thermodynamic warming occurring at the surface ( 33 ). Different circulation patterns are associated with differing levels of predictability ( 23 ), and it is as yet uncertain how changes in their relative occurrence could influence the predictability of temperature and hence AF. Future studies could fruitfully investigate the link between flow patterns driving temperature extremes, the predictability of these flow patterns and how this could inform AF skill. 3. Conclusions This study has considered the predictability of Europe-wide heat-related mortality in the context of numerical weather forecasts of temperature during the exceptionally warm summers of 2022 and 2023. The non-linear relationship between temperature and heat-related mortality has a marked effect on the propagation of forecast spread and error between the temperature and mortality forecasts, with the latter being particularly sensitive to errors in temperature forecasts at high temperatures. Nonetheless, in some of the hottest regions in Europe, mortality is more skillfully predicted for higher temperatures than low ones. This suggests that mortality forecasts can provide valuable information. However, further interdisciplinary efforts are needed to overcome the challenges in forecasting mortality for the hottest and most deadly days. In the context of a warming climate, the continued development of heat-related mortality forecasts for early-warning systems is of the utmost importance in order to protect society from the increasing burden of extreme heat. Data Archival. The meteorological data used in this study can be downloaded from the Copernicus Climate Data Store https://cds.climate.copernicus.eu/. The TIGGE forecast data is available at https://apps.ecmwf.int/datasets/data/ tigge/levtype=sfc/type=cf/ and the ERA5-Land reanalysis data is available at https://cds.climate.copernicus.eu/datasets/ reanalysis-era5-land?tab=download. The health data used in this study cannot be made publicly available due to confidentiality agreements with data providers. Materials and Methods 4. Data and Methods A. Mortality data. We used the spatio-temporally-homogeneous daily regional EARLY-ADAPT mortality database ( 34 ), which in the case of this study contains over 153 million death-counts for 31 European countries. These represent the countries’ entire urban and rural population of over 504 million people. The data was reported for 580 contiguous regions as follows: Austria (35 regions), Belgium (40), Bosnia and Herzegovina (1), Bulgaria (6), Croatia (1), Cyprus (1), Czechia (14), Denmark (1), England (9), Estonia (1), France (96), Finland (4), Germany (16), Greece (46), Hungary (8), Ireland (1), Italy (99), Latvia (6), Lithuania (2), Luxembourg (1), Montenegro (1), the Netherlands (1), Norway (11), Portugal (7), Romania (42), Scotland (23), Serbia (24), Slovenia (12), Holmberg et al. PNAS — October 24, 2025 — vol. XXX — no. XX — 9