Core temperature estimated from heart rate and skin temperature missed a rectal probe by more once algorithms were re-run on new data than their papers claimed: the standard heart-rate Kalman filter, reported at about 0.21-0.41 °C RMSE, gave 0.41 °C in treadmill walking (16 people) and 0.66 °C in cycling in the heat (13 men), caught only 33% of samples at or above 38.0 °C in the cycling data, and no method of 14 re-implemented from a 38-study systematic review was best in both datasets
Main caveat: The review did no risk-of-bias or quality appraisal and did not pool; its reported-performance figures are as each study printed them.
Zhao, Li and Bergmann reviewed studies that estimate core body temperature during physical activity from signals a wearable can collect (mostly heart rate and skin temperature), then tested whether the published accuracy holds up.
Of 38 included studies, 14 used only heart rate, skin temperature and standard environmental inputs and described their equations well enough to rebuild.
The authors re-implemented these, re-estimated every parameter on training folds of each dataset and ran them causally (no future samples) on two earlier laboratory datasets: treadmill walking in heat in 16 adults, and indoor cycling at 35 °C in 13 men, with rectal temperature as the reference.
The heart-rate-only extended Kalman filter, adopted in six earlier papers that reported RMSEs of about 0.21-0.41 °C, gave 0.41 °C on the walking data and 0.66 °C on the cycling data. Adding a chest skin temperature helped on the walking data, and the benefit of other skin sites flipped between datasets.
Specificity for core temperature at or above 38.0 °C stayed fairly high while sensitivity dropped in the cycling data, to zero for every single-site skin regression. The authors conclude that no approach won everywhere, that heart-rate-only filters are a stable baseline when sensing is limited, and that published accuracy mostly reflects each study's own conditions.
Participants
Review: 38 studies; 13 enrolled 10 or fewer participants, 12 enrolled 11-20, 10 enrolled 21-50 and 3 more than 50. Activities: walking or running in 24, military or load carriage in 9, cycling in 8; hot conditions (30 °C or more) in 27. Reference: gastrointestinal pill in 19, rectal in 15, oesophageal in 2, rectal and pill in 2.
Dataset 1: 16 participants (8 M / 8 F), 25.1 ± 5.6 years, body fat 17.1 ± 6.1%; six settings of 25-40 °C, 20-50% relative humidity, with or without 530 W/m² of simulated sun, in permeable or impermeable coveralls; rest, 40-60 min treadmill walking, cooling until rectal temperature fell 0.4 °C, then 40 min more; 1-min sampling.
Dataset 2: 13 men, 30.9 ± 5.4 years, body fat 13.1 ± 4.3%; 35 °C and 57% humidity; 20-60 min of cycling at 75% of maximal heart rate in light clothing (stage 1) or 50% in firefighter gear (stage 2), stopped at a 1.5 °C rise, 38.5 °C rectal, exhaustion or 75 min; 10-s sampling.
Training status: not reported for either dataset.
Protocol
Search to June 2025; study selection by two reviewers; extraction of RMSE, MAE, Bland-Altman mean difference and SD, and R², taking the best variant per model category per study.
Benchmark inputs: heart rate and skin temperatures (11 sites in dataset 1; the text names 10 sites in dataset 2 while its table lists 6); rectal temperature only as the criterion.
All models re-fitted on each dataset (published coefficients not used); three-fold trial-based cross-validation; a common evaluation window after the longest warm-up.
Detection of 38.0 °C or above: sensitivity, specificity, accuracy and AUC.
Key Results
| Algorithm (inputs) | Originally reported RMSE | Benchmark RMSE, walking (dataset 1) | Benchmark RMSE, cycling (dataset 2) | Sensitivity / specificity at ≥38.0 °C, walking | Same, cycling |
|---|---|---|---|---|---|
| Extended Kalman filter, heart rate only (Buller and five later papers) | 0.21-0.41 °C | 0.41 °C (bias −0.05, R² 0.47) | 0.66 °C (bias +0.14, R² −0.40) | 0.71 / 0.92 | 0.33 / 0.85 |
| Extended Kalman filter, heart rate + chest skin (Seng) | 0.29 °C | 0.33 °C | 0.52 °C | 0.77 / 0.87 | 0.81 / 0.79 |
| Phase-dependent filter, heart rate (Rizvi) | 0.16 °C | 0.40 °C | 0.54 °C | 0.75 / 0.88 | 0.72 / 0.83 |
| Biphasic filter, heart rate (Falcone) | 0.28 °C | 0.37 °C | 0.49 °C | 0.74 / 0.91 | 0.46 / 0.92 |
| Linear filter, heart rate (Buller) | 0.30 °C | 0.41 °C | 0.47 °C | 0.97 / 0.46 | 0.84 / 0.98 |
| Linear filter, heart rate + thigh / + forehead skin (Welles) | 0.28-0.30 °C | 0.36 / 0.45 °C | 0.68 / 0.53 °C | 0.80 / 0.85 and 0.68 / 0.89 | 0.77 / 0.75 and 0.44 / 0.85 |
| Single-site skin-temperature regression (Kim) | R² only | 0.42-0.57 °C by site | 0.51-0.55 °C by site | 0.22-0.82 / 0.62-0.78 by site | 0.00 at every site / ≥0.997 |
| Kalman filter + deep-learning forecaster, heart rate only (Han) | 0.28 °C | 0.41 °C | 0.34 °C | 0.81 / 0.61 | 0.00 / 1.00 |
Across the review, reported RMSEs ran from about 0.02 °C (a random forest using sweat sodium and potassium) to 0.67 °C (a population mixed-effects model); Kalman-filter studies reported 0.13-0.41 °C. The authors note most reported values fell below a commonly cited 0.5 °C clinical acceptability line.
24 of 38 studies could not be benchmarked because they needed heat flux, insulated skin temperature, metabolic rate or clothing data; among them a dual heat-flux sensor study reported 0.29 °C RMSE at the forehead and 0.40 °C at the wrist (reported values, not re-tested).
Errors were larger on the cycling dataset for most filters and regressions; the authors call sensitivity at a fixed threshold fragile, because a small under-reading near 38.0 °C removes true detections, and recommend RMSE as the main metric.
Practical Application
A wearable's core temperature is an estimate whose error grew when the algorithms met new people and a new task. Re-run on fresh laboratory data, the standard heart-rate method missed rectal temperature by 0.41-0.66 °C RMSE and, in cycling in the heat, flagged only a third of the samples at or above 38.0 °C: read it as a rough direction, never as a heat-safety threshold.
This fits [SYNTH-91]: heart rate and skin temperature are what the sensor measures; core temperature is what the algorithm estimates from them, so it belongs in the guide's context class. The single-device test already in the base, CORE against a pill in elite hockey players, found the same under-reading pattern as core temperature rose (REF-1258).
Do not dose or end a hot session on a wearable's core-temperature number: SYNTH-04 already says a skin sensor cannot confirm its core-temperature target (REF-1258); in this benchmark's cycling data the estimates often stayed below 38.0 °C while the rectal probe was above it (sensitivity 0.00-0.84 across methods).
Inference, not tested here: a skin-site or heart-rate estimate that held for one sport and one clothing condition may not hold for another; the same filter's error rose from walking to cycling, and the best skin site changed between datasets.
Limitations
The review did no risk-of-bias or quality appraisal and did not pool; its reported-performance figures are as each study printed them.
Benchmark datasets are small (16 and 13), laboratory-based, young and healthy, with training status not reported; the authors give population-level RMSEs and not per-person errors, which may hide large individual misses.
The algorithms were re-fitted to each dataset; a commercial device ships fixed coefficients, so its error in a new user could differ in either direction. No commercial device or firmware was tested.
Heat-flux methods were excluded from the benchmark, so the paper says nothing new about sensors that use heat flux.
Inference, not tested here: rectal temperature was the benchmark criterion while the benchmarked papers mostly used a gastrointestinal pill (Table 4); that difference may account for part of the gap between reported and benchmark errors, which the authors attribute to each study's own conditions.
Dataset 2's skin sites are listed differently in the text (10) and its table (6).
Funding: UK Biotechnology and Biological Sciences Research Council and NIHR; no competing interests declared.
How this entry was checked and changed
Why this level
the review part has no risk-of-bias appraisal and no pooled estimate, so it has no route to A; the new numbers come from re-running algorithms on two laboratory datasets of 16 and 13 people, a validation-study design (B); no marker step, because the estimated core temperature is itself the subject