A heart rate sensor and a blood lactate prediction can both appear in an exercise study, but their errors come from different steps. One number comes from comparing two readings taken during the same ride. The other comes from a model learning how several signals tend to change with lactate. In this experiment, the difference shows up as soon as you look past the correlation.
- The heart rate comparison to a chest strap had a mean absolute error of 0.89 beats per minute across 35 people.
- Only 23 of 36 volunteers had complete temperature profiles. The temperature comparator was another ear measurement.
- The best lactate model had a correlation of 0.838 and a mean absolute error of 1.528 mmol/L. Agreement limits ran from −4.52 to +4.55 mmol/L.
- Paper date
- 20 November 2026
- Volunteer sample
- 36 healthy adults
- Exercise setting
- Controlled incremental cycling
- Lactate model evaluation
- Leave one participant out at a time
Original publication: 20 Nov 2026 · The date above refers to this brief.
Lactate prediction error across tested inputs
mmol/L MAE
Three ways to judge a model result
Direction
Correlation says how closely predictions and measurements move together.
Size
Mean absolute error gives the average size of the error in the measurement’s own units.
Spread
Limits of agreement show how widely individual differences can vary in this dataset.
Each comparison used a different reference
Volunteers rode a stationary cycle at increasing workloads. A Polar H10 chest strap supplied the comparison heart rate. The earbud differed from it by 0.89 beats per minute on average when differences were counted without their signs. The study reports 7,568 paired heart rate points, collected from 35 people. The temperature comparison covered 23 complete profiles with a mean absolute difference of 0.35 °C against a Braun thermometer used in the other ear. Wireless data loss during intense exercise reduced the number of complete profiles. Two ear devices agreeing tells us how they compare at that measurement site. The authors identify the absence of a direct core-temperature reference as a limitation.
Source 1 ↗Read error in mmol/L alongside the correlation
For lactate, the team compared model predictions with blood measurements. The final model reported a correlation of 0.838. That shows how closely predicted and measured values tended to move together across the observations. The typical absolute size of the errors averaged 1.528 mmol/L. The reported limits of agreement reached from −4.52 to +4.55 mmol/L, a span of 9.07 mmol/L. Errors also tended to increase at higher lactate levels, with underprediction above the paper’s 4 mmol/L reference. A model can track the rise as a ride gets harder and still give a different value for a particular reading.
Source 1 ↗Source 2 ↗What did adding heart rate contribute?
The smallest mean absolute error among the tested heart-rate-only models was 3.538 mmol/L. For temperature alone it was 1.642. The full five-feature model reached 1.528, 0.114 mmol/L lower than the best temperature-only result. The best models in those comparisons use different algorithms, so that difference cannot isolate the contribution of heart rate. The simplest model using both signals had an error of 1.670 mmol/L, slightly higher than the best temperature-only model.
Source 1 ↗What would convince a reader that it travels?
The researchers left each participant’s data out in turn while fitting models using the other participants. This checks performance on people whose readings were held out of that fit inside the same study. A separate group tested under new conditions would show whether the errors travel outside this lab. A follow-up could report error by workload and the proportion of missing readings, alongside overall correlation and average error.
Source 1 ↗Device agreement and prediction error answer different questions
Publisher full paper. Checked participant and reference device details, comparison tables, model evaluation, limits of agreement, limitations and disclosures.
01How close were the heart rate readings to the comparator?
- What was observed
- The mean absolute error against the chest strap was 0.89 beats/min.
- Where the conclusion stops
- It applies to paired readings in the indoor cycling experiment.
Source 1 · Figure 1 and Table 1
02Did high correlation imply narrow lactate errors?
- What was observed
- Correlation was 0.838 and the limits of agreement were −4.52 to +4.55 mmol/L.
- Where the conclusion stops
- Errors apply to this study cohort and increased at higher measured lactate.
Source 1 · Figure 3 and Table 3
Numbers you can inspect
| Measure | Value & unit | Origin & method |
|---|---|---|
| Exercise volunteers | 36 people | Reported Count of healthy adults in the cycling study. Source 1 · Methods, Participants |
| Participants with paired heart rate data | 35 people | Reported Paired device comparison. Source 1 · Results, Heart rate validity |
| Complete temperature profiles | 23 people | Reported Complete profiles used in the temperature comparison. Source 1 · Results, Temperature validity |
| Heart rate mean absolute error | 0.89 beats/min | Reported Compared with Polar H10 chest strap. Source 1 · Figure 1 |
| Ear temperature mean absolute error | 0.35 °C | Reported Compared with Braun ThermoScan 7 ear thermometer. Source 1 · Figure 2 |
| Best heart-rate-only model error | 3.538 mmol/L MAE | Reported Best mean absolute error for the heart-rate-only feature set. Source 1 · Lactate prediction, ablation analysis |
| Best temperature-only model error | 1.642 mmol/L MAE | Reported Best mean absolute error for the temperature-only feature set. Source 1 · Lactate prediction, ablation analysis |
| Best two-feature model error | 1.67 mmol/L MAE | Reported Linear regression using only the two basic features. Source 1 · Lactate prediction, ablation analysis |
| Full model error | 1.528 mmol/L MAE | Reported Gradient boosting with five features. Source 1 · Lactate prediction, ablation analysis |
| Difference in full vs temperature-only mean absolute error | 0.114 mmol/L | Calculated Subtract 1.528 from 1.642. This compares separately selected models. Source 1 · Lactate prediction, ablation analysis |
| Full model correlation | 0.838 Pearson r | Reported Association of predictions and measurements. Source 1 · Lactate prediction, ablation analysis |
| Span between agreement limits | 9.07 mmol/L | Calculated Subtract the lower limit of −4.52 from the upper limit of +4.55. Source 1 · Figure 3 and Table 3 |
Compare the actual experiments
These studies answer different questions. Read the unit and endpoint before comparing results.
| Study | Unit & setting | Readout | Interpretation boundary |
|---|---|---|---|
| Two-device agreement in cycling and methodological agreement analysis Source 2 · Bland and Altman, publication record | Paired measurements | Mean absolute error, bias and limits of agreement | Read the limits relative to the size of an error that matters in the intended setting. |
| Within-cohort and external model checks Source 1 · Methods, Leave-One-Subject-Out cross-validation | Held-out individuals and a separate cohort | Error on unseen participants in the same study | A separate group can also challenge the collection conditions and population. |
The export includes claims, available numbers, methods and source locations. It contains our reading notes and published summaries; it is not raw participant data or an independent reanalysis.
Evidence update · 22 Nov 2026
First publication. Any calculations in this brief use reported figures. The reading notes identify possible follow-up questions. AI source check; no human editorial or clinical review.
What the earbud measured and what the model estimated
| Measurement | Reported average absolute error | Reference and coverage |
|---|---|---|
| Heart rate | 0.89 beats/min mean absolute error | Polar H10 chest strap; paired data in 35 adults. |
| Ear temperature | 0.35 °C mean absolute error | Braun ear thermometer; 23 complete profiles. |
| Blood lactate prediction | 1.528 mmol/L mean absolute error | Prediction against sampled blood lactate; limits of agreement −4.52 to +4.55 mmol/L. |
The three rows describe different measurements. The size of the numbers cannot be compared across units.
Your questions, answered
Did the earbud measure blood lactate directly?
The researchers measured blood lactate separately. A model estimated lactate using the wearable signals.
Why did the number of people differ by metric?
The study reports 36 volunteers, 35 with paired heart rate data and 23 with complete temperature profiles. Some temperature readings were lost during the test.
Does 0.838 correlation mean 84% accuracy?
It is a correlation coefficient. Mean absolute error and limits of agreement tell you how closely individual values matched in mmol/L.
Was the test done on a new outside group?
The leave-one-subject-out procedure used held-out participants from the same cohort. A study in a separately recruited group would test how well the model transfers.
Limits of this interpretation
- A controlled cycling test in healthy adults covers a limited setting and population.
- The ear temperature comparison used a second ear device. Complete profiles were available for 23 of 36 volunteers.
- The lactate model was exploratory. Performance was evaluated within the study cohort.
- The authors reported no external research funding and declared no competing interests.
Sources & transparency
- Camargo, Zellhuber, Weber and Schönfelder (2026) — Assessment of the MOMENTUM sport earbud for multi-parametric physiological tracking and exploratory lactate modeling during incremental exercise
Publisher full paper. Checked participant and reference device details, comparison tables, model evaluation, limits of agreement, limitations and disclosures. · Accessed 22 Nov 2026
DOI: 10.1038/s41598-026-80803-8 - Bland and Altman (1986) — Statistical methods for assessing agreement between two methods of clinical measurement
Original methods article, indexed publication record. The earbud paper supplies the application and numerical results. · Accessed 22 Nov 2026
DOI: 10.1016/S0140-6736(86)90837-8
Prepared and source-checked with AI for Press-news Team. The source notes identify the material available during preparation. The article received no human editorial or clinical review. The source list supports the reported results and the calculations shown.
Source check: AI source check — primary research and stated source limits
Clinical review: Not applicable to this educational guide
Suggest a correction