跳到正文
原文
Nature 健康护理主题·· 9 小时前AI 评分35

MOMENTUM Sport 耳塞在运动生理参数监测中的表现

Assessment of the MOMENTUM sport earbud for multi-parametric physiological tracking and exploratory lactate modeling during incremental exercise

AI 导读

该研究评估了 MOMENTUM Sport 耳塞在运动过程中的心率(HR)和核心温度(\(T_c\))监测准确性,并探讨了基于这些信号预测乳酸水平的可能性。结果显示,MOMENTUM Sport 在心率监测方面的平均绝对误差(MAE)为 0.89 bpm,平均绝对百分比误差(MAPE)为 0.78%,表明其测量结果与 Polar H10 参考设备高度一致。

正文

Introduction

Heart rate (HR) measurements in sports have been a topic of extensive discussion for many years, as they provide a cost-effective, quick, and non-invasive way to assess the condition of the autonomic nervous system and cardiovascular fitness1. While it has been shown that training significantly impacts changes in autonomic nervous system status, suggesting that HR measurements could effectively reflect an athlete’s training condition1,2. Additionally, during exercise, increased metabolic activity raises core temperature (\(T_c\)), which in turn elevates heart rate since the muscles need more oxygen to supply the metabolic rate demands, to help in heat dissipation and to stabilize the internal environment3. Conversely, exposure to cold can lower \(T_c\), leading to a decrease in heart rate as the body conserves energy and reduces heat loss through mechanisms like vasoconstriction and shivering4.

While this reduces cardiovascular strain, cold environments may increase metabolic strain by elevating blood lactate levels5,6. This is primarily due to a ’compression of the recruitment order,’ where cold-impaired muscle function forces the nervous system to recruit glycolytic Fast-Twitch (Type II) fibers at much lower intensities7,8. Combined with reduced oxygen delivery from peripheral vasoconstriction, this shift leads to a higher reliance on anaerobic metabolism and greater lactate accumulation compared to temperate conditions5,9.

Despite the physiological link between these variables, there remains a significant gap in the availability of a valid device that can simultaneously track key variables such as HR, \(T_c\), and blood lactate levels. Understanding how these measurements correlate during various exercise intensities is essential for optimizing training regimens and preventing heat-related illnesses10,11. This study seeks to investigate a novel earbud device, the MOMENTUM Sport (MSPORT), designed for monitoring HR and \(T_c\) within the ear canal. The importance of this site lies in its unique proximity to the carotid arterial supply and the brain12.

The elevation of core temperature is recognized as a contributing factor to fatigue11,13 and a limitation in endurance exercise performance14. Intense exercise can increase the risk of exertional heat illnesses, particularly when \(T_c\) surpasses 40 \(^\circ\)C15. However, training at an appropriate core temperature can enhance various performance metrics, including maximum oxygen consumption (VO\(_2\)max). A study by Bergh et al. found that participants could not achieve their VO\(_2\)max when body temperatures fell below 37.5\(^\circ\)C and 38.0\(^\circ\)C, indicating that lower core temperatures negatively affect peak performance. Additionally, after the initial two minutes of exercise, the increase in VO\(_2\) became dependent on core temperature, with a significant compromise in the ability to enhance oxygen consumption as core temperature continued to decrease16.

Furthermore, tracking blood lactate accumulation is vital, as it serves as a key indicator of metabolic stress and fatigue during exercise17. The theoretical basis for predicting lactate via \(T_c\) and HR lies in the “Lactate Shuttle Theory,” which identifies lactate as a major oxidative fuel source and signaling molecule that directly influences systemic thermogenesis18. During high-intensity exertion, the exothermic nature of rapid glycolysis creates a distinct thermal flux; however, a potential “temporal decoupling” may exist between the metabolic production of lactate and its subsequent manifestation as a measurable rise in cranial temperature. By investigating the interactions between HR, \(T_c\), and lactate levels, this study provides valuable insights into optimizing training strategies.

Existing methods for monitoring HR, \(T_c\), and lactate include traditional devices such as wristwatches19, chest straps, ingestible telemetric capsules20, and blood lactate analyzers. While these methods have proven reliable in controlled settings, they often come with limitations, such as discomfort, invasiveness, and logistical challenges during competition20. Recent advancements in wearable technology have emerged as a promising alternative; however, the reliability of these newer devices in accurately measuring critical parameters like \(T_c\) during exercise remains under-researched, particularly regarding earbud-based trackers21,22.

This study aims to provide empirical data on the accuracy of the MSPORT in measuring HR and \(T_c\) during incremental exercise and to examine the potential for predicting lactate levels based on these signals using a Leave-One-Subject-Out (LOSO) cross-validation framework. Specifically, we address: (1) how MSPORT vital signs compare to references like the Polar H10 and Braun ThermoScan 7; (2) whether integrated HR and \(T_c\) data can provide a robust non-invasive estimation of blood lactate; and (3) what temporal kinetic relationship, specifically the lead or lag time, exists between systemic lactate accumulation and cranial thermal flux.

Results

Heart rate validation of the MSPORT compared to the polar H10

The Fig. 1a shows a Bland-Altman plot in all the participants where the Mean Absolute Error (MAE) of 0.89 bpm indicates that, on average, the MSPORT device deviates from the Polar H10 by less than one bpm, suggesting a high level of accuracy in heart rate measurement. The Mean Absolute Percentage Error (MAPE) of 0.78% demonstrates that the discrepancies between the two devices are minimal relative to the measured values. The Root Mean Square Error (RMSE) of 1.10 bpm further supports the reliability of the MSPORT device, as it reflects the standard deviation of the differences, indicating that most measurements are closely aligned with the reference device. Additionally, the bias of 0.79 bpm suggests a slight tendency for the MSPORT device to report heart rates that are consistently lower than those of the Polar H10.

Overall, the results from the Bland-Altman plot indicate that the MSPORT is a reliable alternative for heart rate tracking, with minimal systematic error and high accuracy compared to the Polar H10. The Bland-Altman analysis comprised 7,568 data points across 35 participants. A weak proportional bias was identified (\(r = 0.32\), p < 0.001), suggesting that measurement differences slightly increase at higher heart rates, which should be considered when interpreting MSPORT accuracy during maximal intensity exercise.

Figure 1b presents data from a single representative participant (selected as the individual whose MAE most closely matched the group median MAE of 0.89 bpm), comparing HR measurements taken by the MSPORT tracker with those from the Polar H10 chest strap. The MAE is 0.59 bpm, indicating strong agreement between the two devices. Additionally, the MAPE is 0.54%, and the RMSE is 0.79 bpm, both reflecting minimal variability and high precision. A bias of 0.38 bpm suggests that the MSPORT tracker slightly overestimates HR compared to the Polar H10, though this difference is minor.

Fig. 1

a) Bland-Altman plot demonstrating the comparison of heart rate measurements between the MSPORT device and the Polar H10 across all participants, red dashed lines represent the upper and lower limits of agreement and the blue line indicates the mean difference. Blue represent male and red female participants. b) Example HR readings of one participant over time with the MSPORT (blue) and Polar H10 (red).

Table 1 Bland-Altman comparison of heart rate measurements between male and female participants.

Full size table

Although the mean difference in measurements between the MSPORT and the Polar H10 is statistically significant, as shown in Table 1, there is no significant difference in the mean difference between male and female participants. The p-value of 0.06 suggests that the difference in measurements between the two devices is similar across biological sexes, with no substantial disparity observed between male and female participants.

Ear-canal temperature agreement with braun ThermoScan 7

The results are illustrated in two plots. Figure 2a presents a Bland-Altman analysis, revealing a mean bias of \(0.12~^{\circ }\text {C}\). Due to wireless telemetry data loss during high-intensity movement, complete temperature profiles were obtained for 23 out of the 36 participants. A comparative analysis confirmed that the remaining cohort showed no systematic differences in sex distribution, age, weight, or \(\dot{V}\text {O}_{2\text {max}}\) relative to the excluded individuals, indicating that data missingness occurred at random and did not introduce selection bias. The analysis yielded a MAE of 0.35 \(^\circ\)C, indicating that, on average, the temperature readings differ by this amount between the MSPORT tracker and the Braun 7 reference device. This relatively small MAE suggests good accuracy in the MSPORT device’s measurements. Additionally, the MAPE of 0.92% indicates that the average error is less than 1% of the actual temperature readings, further supporting the reliability of the MSPORT tracker. The RMSE of \(0.40~^{\circ }\text {C}\) reflects the overall precision of the device, with lower values indicating better performance. The mean difference between devices was \(0.12~^{\circ }\text {C}\) (95% CI: 0.11 to \(0.13~^{\circ }\text {C}\)), with limits of agreement from \(-0.72\) to \(0.95~^{\circ }\text {C}\).

Figure 2b presents the same representative participant (median HR MAE), comparing the temperature readings from the MSPORT tracker with those from the Braun ThermoScan 7. In this case, the MAE is 0.18 \(^\circ\)C, demonstrating a close alignment between the two devices. The MAPE is 0.48%, and the RMSE is 0.23 \(^\circ\)C, both of which indicate minimal variability and high accuracy. The bias of 0.05 \(^\circ\)C suggests that the MSPORT tracker tends to report slightly higher temperatures than the Braun 7, but this difference is small.

Overall, the small mean differences, low MAE and MAPE values, and reasonable RMSE highlight the MSPORT device’s reliability and accuracy as a ear-canal temperature monitoring device, reinforcing its effectiveness in sports and personal use (Table 2).

Fig. 2

a) Bland-Altman plot demonstrating the comparison of body temperature measurements between the MSPORT device and the Braun 7 across 23 participants, red dashed lines represent the upper and lower limits of agreement and the blue line indicates the mean difference. Blue represent male and red female participants. b) Example Tc readings of one participant over time with the MSPORT (blue) and Braun 7 (red).

Table 2 Bland-Altman comparison of body temperature measurements between male and female participants.

Full size table

While the mean difference in temperature measurements between the MSPORT and the Braun 7 is significant, as shown in Table 2, there is no significant difference in measurement discrepancy between biological sexes. With a p-value of 0.19, the differences in temperature readings between the devices are similar for male and female participants.

Blood lactate prediction and kinetic coupling (Fig. 3)

Fig. 3

a) Scatter plot of predicted versus actual blood lactate levels. The red dashed line indicates perfect agreement (\(y = x\)); the blue line shows the regression fit (\(y = 0.69x + 1.46\)), indicating a calibration slope of 0.69 and a positive intercept of 1.46 mmol/L, consistent with overestimation at low concentrations and underestimation at higher concentrations. The green dashed vertical line marks the conventional 4 mmol/L reference value. b) Residual plot showing differences between predicted and actual values; the red dashed line indicates zero difference and the green dashed line marks 4 mmol/L. c) Bland-Altman plot of predicted versus measured lactate across all participants; dashed lines represent the 95% limits of agreement and the blue line the mean difference. Detailed performance metrics are reported in the Results.

Lactate levels were predicted using a combination of HR and \(T_c\), with model performance assessed through a Leave-One-Subject-Out (LOSO) cross-validation framework23. Five features were derived from the raw signals for model input. The cumulative core temperature delta (\(\Delta T_c\)) was computed as the rise in \(T_c\) relative to each participant’s resting baseline. The normalised heart rate (\(HR_{norm}\)) was expressed as a ratio of resting HR to remove between-person baseline differences. The rate of change of temperature (\(\dot{T}_c\)) and rate of change of heart rate (\(\dot{HR}\)) were computed as the first temporal difference of each calibrated signal at each 10-second window, capturing the velocity of each physiological response. An interaction term (\(HR_{norm} \times \Delta T_c\)) was included to capture the combined cardiovascular-thermal signal. A baseline model using only \(\Delta T_c\) and \(HR_{norm}\) was compared against an extended model using all five features. Among the models evaluated, a Gradient Boosting Regressor (GBR) with the extended feature set achieved the best performance by MAE under LOSO cross-validation and was selected as the final model. The scatter plot of predicted versus measured blood lactate (Fig. 3a) demonstrates a Pearson correlation of \(R = 0.84\) (\(p < 0.001\)), this corresponds to a coefficient of determination of \(R^2 = 0.699\), indicating that approximately 70% of the variance in measured lactate is explained by the model at the pooled observation level. Dispersion increases at concentrations exceeding 4 mmol/L, where the model exhibits a systematic underestimation bias, as further illustrated in the residual plot (Fig. 3b).

To quantify the independent contribution of each signal, a systematic ablation analysis was conducted using LOSO cross-validation across all three model architectures. Heart rate alone was a poor predictor of blood lactate (best model: Neural Network, \(r = 0.233\), MAE \(= 3.538\) mmol/L, MAPE \(= 162.2\%\)), indicating that cardiovascular load does not carry sufficient metabolic specificity to distinguish glycolytic state during incremental exercise. Ear-canal temperature alone was substantially more informative (best model: Neural Network, \(r = 0.809\), MAE \(= 1.642\) mmol/L, MAPE \(= 53.3\%\)), confirming that \(\Delta T_c\) is the primary metabolic signal in this framework. Combining HR and \(\Delta T_c\) (basic feature set) produced minimal additional improvement over \(\Delta T_c\) alone (\(r = 0.817\), MAE \(= 1.670\) mmol/L for Linear Regression), while the full five-feature Gradient Boosting model, incorporating rate-of-change and interaction features, achieved the best absolute performance (\(r = 0.838\), MAE \(= 1.528\) mmol/L, MAPE \(= 48.8\%\)). These results demonstrate that the improvement from multi-signal fusion is driven primarily by \(\Delta T_c\) rather than by HR, consistent with the thermodynamic hypothesis that glycolytic heat production is encoded in the thermal signal independently of cardiovascular demand.

To establish the physiological significance of the predictors, a Linear Mixed Model (LMM)24 was employed to account for repeated measures within participants. The LMM revealed that ear canal temperature (\(T_c\)) was a highly significant driver of the model (\(z = 11.65\), \(p < 0.001\)), confirming a robust physiological link between ear canal thermal flux and metabolic stress. Heart rate was also a significant predictor (\(z = -6.78\), \(p < 0.001\)), while the quadratic temperature term was significant (\(p = 0.004\)), suggesting a nonlinear relationship at higher intensities.

The overall MAE was 1.53 mmol/L, with a MAPE of 48.8% and RMSE of 2.309 mmol/L. To contextualise the MAPE, results were stratified at the conventional 4 mmol/L reference value (noting that this threshold may differ considerably from an individually determined lactate threshold): below this threshold, MAPE was 64.4%, MAE was 0.92 mmol/L, and RMSE was 1.69 mmol/L; above it MAPE was 28.1%, MAE was 2.34 mmol/L, and RMSE was 2.94 mmol/L. Participant-level out-of-fold MAE, derived from the LOSO framework by treating each held-out participant as the unit of analysis, was 1.539 mmol/L (95% CI: 1.218–1.861 mmol/L; t-distribution, \(n = 23\) participants), with individual values ranging from 0.534 to 3.447 mmol/L. Participant-level RMSE was 2.129 mmol/L (95% CI: 1.717–2.542 mmol/L) and mean bias was \(-0.069\) mmol/L (95% CI: \(-0.715\) to 0.577 mmol/L), indicating no systematic directional error at the participant level. Sex-stratified participant-level MAE was 1.424 ± 0.591 mmol/L for males (\(n = 13\); range: 0.660–2.313 mmol/L) and 1.689 ± 0.917 mmol/L for females (\(n = 10\); range: 0.534–3.447 mmol/L). This pattern reflects the mathematical amplification of percentage errors at low absolute lactate values, where even small absolute errors produce large percentage deviations. The model therefore performs more accurately at the clinically relevant high-intensity range. Bland-Altman analysis (Fig. 3c) confirmed a negligible systematic bias of 0.02 mmol/L with 95% Limits of Agreement of \(-4.52\) to 4.55 mmol/L. Statistical analysis revealed a significant sex-based difference in prediction bias (\(p < 0.05\)). Males exhibited a positive mean difference of 0.48 mmol/L, whereas females showed a negative mean difference of \(-0.81\) mmol/L (Table 3).

Table 3 Bland-Altman comparison of lactate prediction by biological sex.

Full size table

Fig. 4

Exploratory cross-correlation analysis: (a) mean temporal offset between lactate and ear-canal temperature per sex group, and (b) mean cross-correlation curves by sex. Given the three-minute lactate sampling interval, these results should be interpreted as preliminary directional observations only.ics across a 300-second lag window.

To assess the contribution of each feature to model performance, permutation importance was calculated by randomly shuffling each feature 30 times (sufficient to produce stable mean importance estimates as evidenced by the low SD values reported) and measuring the resulting mean increase in MAE. Gradient Boosting built-in importance was also reported alongside for comparison. Both analyses were performed on the final GBR model trained on all data, serving as a feature ranking tool rather than for accuracy reporting.

To investigate the temporal relationship between metabolic stress and thermal response, an exploratory cross-correlation was performed25. Cross-correlation analysis identifies the optimal time offset between two signals, where a negative lag indicates that one signal precedes the other. Given that blood lactate was sampled only every 3 minutes, whereas ear-canal temperature was recorded continuously at 1 Hz, any second-scale lag estimate is highly sensitive to interpolation assumptions and the shared monotonic rise of both signals during the ramp protocol. Accordingly, we do not interpret the apparent lag as a precise physiological constant.

The analysis indicated that blood lactate accumulation precedes the rise in ear canal temperature by an average of 16.5 seconds (95% CI: 6.7 to 26.3 seconds; SD: 24.0 s across subjects), meaning that the metabolic signal is detectable in the blood before it manifests as a measurable thermal change in the ear canal. This temporal decoupling likely reflects the time required for lactate-driven heat production to propagate through the circulatory system to the cranial thermal compartment. The male mean lead time was \(-20.0\) s (SD: 26.5 s) and the female mean lead time was \(-12.0\) s (SD: 21.0 s), a difference that was not statistically significant (\(p = 0.4418\)), suggesting that the metabolic-to-thermal signaling mechanism is conserved across biological sexes (Fig. 4a). However, because of the coarse lactate sampling, the precise lag magnitude cannot be reliably estimated from the present data. This observation should therefore be regarded as hypothesis-generating and requires confirmation with higher-frequency lactate measurements.

Feature importance analysis

Permutation importance analysis identified the cumulative core temperature delta (\(\Delta T_c\)) as the strongest individual predictor, with a mean MAE increase of 1.46 mmol/L (SD: 0.10) when shuffled, while the rate of change of heart rate (\(\dot{HR}\)) was the second strongest predictor (1.02 mmol/L, SD: 0.08). Additionally, GBR built-in importance ranked \(\Delta T_c\) first (0.68), followed by \(\dot{HR}\) (0.16) (Table 4). Temperature-related features combined were 1.5 times more important than HR-related features in permutation importance. These results are exploratory given the post-hoc nature of the analysis.

The cumulative ear-canal temperature delta (\(\Delta T_c\)) emerged as the dominant predictor in both analyses (Fig. 5), with a MAE increase of 1.46 mmol/L, followed by the rate of change of heart rate (\(\dot{HR}\): 1.02 mmol/L) and the rate of change of temperature (\(\dot{T}_c\): 0.42 mmol/L). Combined, temperature-related features substantially outweighed HR-related features in predictive contribution. This finding extends the LMM result (\(z = 11.65\), \(p < 0.001\)) by showing that it is specifically the velocity of thermal change, rather than its absolute magnitude, that carries the strongest metabolic signal. This is physiologically consistent with the 16.5-second kinetic lead time: a rapidly rising core temperature reflects the systemic propagation of lactate-induced heat production.

Table 4 Permutation feature importance for the Gradient Boosting lactate prediction model.

Full size table

Fig. 5

Feature importance for the lactate prediction model. (a) GBR built-in importance showing the relative contribution of each feature; the grey dashed vertical line at 0.2 denotes the equal-importance baseline (\(\frac{1}{5}\) features), indicating that any feature exceeding this threshold contributes more than an average uniform share to the model’s splits. (b) Permutation importance showing the mean increase in Mean Absolute Error (MAE) in \(\text {mmol/L}\) when each feature column is randomly shuffled (error bars represent the standard deviation across 30 permutations to ensure statistical stability based on the stability of estimates across repetitions). Red bars indicate temperature-related features; blue bars indicate HR-related features. \(\Delta T_c\) (rate of change of ear canal temperature) was the dominant predictor in both analyses.

Methods

Participants

Thirty-six healthy adults, 19 males and 17 females, volunteered to participate in this study (see Table 5). One participant was excluded from the heart rate analysis due to persistent Bluetooth signal dropout throughout the entire recording, yielding \(n = 35\) for HR analyses. No formal a priori power calculation was conducted as this was a validation study; prior estimates of agreement variance for this novel device-protocol combination were unavailable. A precision-based justification is provided: with \(n = 35\) for HR, the 95% CI for the mean bias is approximately ±0.26 bpm, providing sufficient precision to detect a clinically meaningful systematic error. For temperature and lactate (\(n = 23\)), precision is more limited and is acknowledged as a study limitation. An additional 12 participants were excluded from temperature and lactate analyses due to incomplete MSPORT data records caused by intermittent wireless connectivity issues and movement-induced sensor displacement during the strenuous exercise phases, yielding \(n = 23\) for temperature and lactate analyses. A participant flow diagram is provided in the Supplementary Material. Baseline characteristics (age, sex, weight, VO\(_2\)max) did not differ significantly between included and excluded participants. Before the tests, participants were informed about the study details, potential risks and aims of the study and signed a written informed consent form. Participants without any disability in the lower limbs were included in the study. Exclusion criteria comprised current or previous lower extremity/back injuries or any condition affecting leg muscle activation; known cardiovascular, metabolic, or systemic diseases; medications altering heart rate or thermoregulation; acute infections or fever; and ear-related conditions (e.g., active infections, severe cerumen impaction, or ear canal anomalies) that would compromise the fit of the in-ear sensor or safe exercise testing. The study was conducted in accordance with the Declaration of Helsinki principles and approved by the ethics committee of the Technical University of Munich (Approval Number: 2025-216-S-SB).

Table 5 Participant characteristics (mean ± SD). Height, weight and VO\(_2\)max were measured on the day of the laboratory test.

Full size table

Recording devices

MOMENTUM sport (MSPORT)

The MSPORT device was provided by Sennheiser (Sonova Consumer Hearing GmbH) free of charge for research purposes. The authors received no financial compensation and declare no conflict of interest. The MSPORT wireless earbuds, measuring 23.3 \(\times\) 20.9 \(\times\) 26.1 mm and weighing 6.4 grams, feature integrated biometric in-ear sensors in the left earbud, to track heart rate using photoplethysmography and body temperature. They are designed to operate within a temperature range of 0\(^\circ\)C to 45\(^\circ\)C, making them suitable for use in both cold and hot conditions. The device is compatible with most fitness trackers, including smartwatches and apps, via Bluetooth 5.2. Because the biometric sensors are present only in the left earbud, all MSPORT heart rate and temperature measurements were obtained from the left ear. Reference ear-canal temperature measurements were taken in the contralateral right ear using the Braun ThermoScan 7 to avoid disturbing the MSPORT sensor placement during exercise. Cycling was not paused for Braun readings; measurements were taken during the rest periods between stages. The MSPORT earbud remained in the left ear throughout and was not removed. Three consecutive Braun readings were taken from the right ear at each time point and averaged. All Braun measurements were performed by the same trained researcher to minimise operator variability. Braun readings were aligned to the nearest 10-second sensor window timestamp for analysis.

Polar H10 chest strap

The Polar H10 (Polar Electro, Kempele, Finland) is a widely validated ECG-based chest-strap heart rate monitor26, used here as a reference comparator for MSPORT HR measurements. It is not a clinical diagnostic ECG device but has been independently validated for accuracy in exercise settings and serves as the established reference standard for wearable HR comparison studies. The Polar H10 monitors HR based on electrocardiography (ECG) measurements and has been previously validated against the ECG Holter gold standard27.

Braun ThermoScan 7 - IRT6520

The Braun-IRT7 is a widely used infrared tympanic thermometer that measures heat generated by the eardrum and surrounding tissue within the ear canal, providing a reliable proxy for core body temperature. The device employs infrared technology, allowing for fast temperature readings. The Braun has shown a mean difference of \(-0.044\)\(^{\circ }\)C with the gold standard SureTemp Plus in the study of Mah et. all (2021)28. Because of its accuracy and given the extensive usage and reliability of this device, we have included it as a reference for body temperature measurements.

Metalyzer 3B

The Metalyzer 3B (Cortex Biophysik GmbH, Leipzig, Germany) is a widely used cardiopulmonary function testing device29 for collecting vital sign parameters such as respiratory rate, heart rate, and respiratory exchange rate during real-time assessments. Utilizing mixed gas or heart-to-heart testing, it supports comparisons of cardiopulmonary function across various populations and conditions30,31, and operates as a breath-by-breath indirect calorimetry system. It has been validated as a reliable instrument for performing exercise testing in both routine sports medicine practice and research settings29.

EKF Biosen C-Line

Blood lactate concentration was determined using the Biosen C-Line stationary analyzer (EKF Diagnostic GmbH, Barleben, Germany). The device quantifies L-lactate in whole blood via an enzymatic-amperometric method, using an immobilized lactate oxidase chip sensor, with a required sample volume of 20 \(\upmu\)l and a precision of less than 3% coefficient of variation32. The Biosen C-Line is an established laboratory reference standard for blood lactate measurement in exercise science research32.

Study protocol

Participants initially wore a Metalyzer 3B spirometer mask, a Polar H10 chest band, and the MSPORT (in the left ear) for five minutes while seated. Reference ear-canal temperature measurements were obtained from the contralateral right ear using the Braun ThermoScan 7. This bilateral approach was adopted to avoid disturbing the MSPORT sensor placement during exercise; both devices measure heat generated in the ear canal and surrounding tympanic tissue. While contralateral ear measurements are a recognised limitation, bilateral agreement studies have reported minimal systematic differences at rest and during exercise in healthy adults. The use of a tympanic reference rather than an esophageal or rectal thermometer means the present temperature analysis characterises inter-device agreement at the same anatomical region rather than absolute core temperature accuracy. They then rested on the cycle ergometer for three minutes before beginning the warm-up cycling phase. Incremental ergospirometry ramp tests were conducted on a cycle ergometer to determine volitional exhaustion and metabolic thresholds. The spirometer test, a widely used method for measuring an individual’s VO\(_2\)max, determines the maximum amount of oxygen (O\(_2\)) a person can utilize during intense exercise. During the test, participants breathed into the Metalyzer 3B device, which measures oxygen (O\(_2\)) consumption and carbon dioxide (CO\(_2\)) production. After completing a three-minute warm-up at 50 W, participants began the slow ramp test. The test protocol started with an initial load of 50 W, increasing by 10 W every minute (slow ramp: 50/10/1). This load increment was achieved by increasing the resistance by 1 W every six seconds. Participants were instructed to maintain a pedal cadence between 70 and 100 rpm. \(\dot{V}\)O\(_2\)max was considered attained when at least two of the following criteria were met: (i) a respiratory exchange ratio \(\ge 1.10\), (ii) heart rate \(\ge 90\%\) of age-predicted maximum (\(220 - \text {age}\)), and (iii) a plateau in \(\dot{V}\)O\(_2\) of less than 2.1 mL\(\cdot\)kg\(^{-1}\cdot\)min\(^{-1}\) despite an increase in workload33. Gas exchange variables were measured continuously using the MetaMax 3B metabolic analyser (Cortex Biophysik GmbH, Leipzig, Germany). Prior to testing, participants were instructed to avoid vigorous exercise for 24 h, caffeine consumption for 12 h, and heavy meals for 3 h, and to arrive well-hydrated. Heart rate, breathing, and lactate resting values were measured during the warm-up phase. Blood samples were collected from the right earlobe every three minutes to measure lactate levels. Prior to sampling, the sample site was cleaned and dried with a swab. Samples were obtained using heparinized capillary tubes through standard capillary punctures performed with a blood lancet (Safe-T-Pro Plus, Roche Diagnostics, Basel, Germany)34. The test concluded with a 10-minute recovery phase, during which post-exercise blood samples were taken at one-, three-, and five-minute intervals.

Data acquisition

Throughout the protocol, the Polar H10 chest strap measured HR based on ECG and transmitted the HR data automatically to the Metalyzer 3B via Bluetooth. VO\(_2\) and VCO\(_2\) were continuously measured using the MetaLyzer 3B. The MSPORT earbud was connected to a mobile app and a gateway, a single-board computer designed for data processing and transmission, via Bluetooth 4.2 and measured the HR and Tc in real time with a sampling rate of approximately 1 Hz. In addition, ear-canal temperature was measured every two minutes in the right ear using a Braun ThermoScan 7. Three consecutive readings were taken at each time point and averaged to obtain the reference ear-canal temperature.

Data analysis & statistics

Data preprocessing and synchronization were conducted offline using Python (v3.12). MSPORT heart rate (HR) and ear-canal temperature (\(T_c\)) data were synchronized with metabolic data from the Metalyzer 3B and blood lactate reference values. To minimize signal noise while preserving physiological trends, data were aggregated into ten-second non-overlapping time windows to compute mean values.

Bland-Altman analyses were performed on the full time-series data (HR: \(n = 7{,}568\) observations from 35 participants; temperature: n observations from 23 participants). Because each participant contributed multiple correlated time points, observations within participants are not independent. The reported Limits of Agreement should therefore be interpreted as descriptive estimates of the overall measurement spread rather than as participant-level agreement bounds; participant-level Bland-Altman analysis is presented in the sex-stratified tables. Proportional bias was examined by computing the Pearson correlation between the differences and the means. Confidence intervals for the mean difference (bias) were computed using a normal approximation. Confidence intervals for the limits of agreement themselves were not computed. To account for inter-individual physiological baselines, HR was normalized to the resting baseline (\(HR_{cal} = HR_{actual} / HR_{rest}\)) and \(T_c\) was expressed as a delta change from baseline (\(\Delta T_c = T_{actual} - T_{rest}\)). The validity of the MSPORT device was assessed using Pearson’s correlation (R), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and Standard Error of the Estimate (SEE). Agreement between reference devices (Polar H10 and Braun ThermoScan 7) and the MSPORT was evaluated via Bland-Altman analysis35,36, with limits of agreement defined as the mean difference ±1.96 standard deviation.

Predictive modeling for blood lactate was performed using a Leave-One-Subject-Out (LOSO) cross-validation framework implemented in Scikit-learn23, ensuring the model’s generalizability across participants. The modelling dataset comprised 289 lactate observations from 23 participants (range: 10–16 per participant, mean: 12.6). A Gradient Boosting Regressor (GBR) was used as the primary model with the following hyperparameters: n_estimators \(= 200\), max_depth \(= 2\), learning_rate \(= 0.05\), min_samples_leaf \(= 5\), subsample \(= 0.8\), loss \(=\) squared error, random_state \(= 42\). These hyperparameters were fixed prior to LOSO cross-validation and were not re-tuned within individual folds to prevent information leakage. Feature selection was also fixed prior to cross-validation. Each blood lactate measurement was paired with one 10-second window immediately preceding the blood draw; features were computed over the rolling window ending at that measurement time point, yielding one prediction per blood sample (not multiple predictions per sample). The physiological contribution of each predictor was quantified using a Linear Mixed Model (LMM) in the Statsmodels library24 to account for the nested structure of repeated measures within individuals. The model was specified as:

$$\text {Lactate} \sim \text {HR\_norm} + \Delta T_{\text {c}} + I(\Delta T_{\text {c}}^2) + C(\text {sex}) + (1 \mid \text {participant}),$$

estimated by maximum likelihood, with participant as a random intercept. Fixed-effect estimates with 95% confidence intervals and z-statistics are reported.

To quantify participant-level prediction uncertainty, the LOSO framework was used to derive one out-of-fold MAE value per participant (treating each held-out participant as the unit of analysis). A 95% confidence interval for the mean participant-level MAE was computed using the t-distribution across \(n = 23\) participants. Sex-stratified participant-level summaries (mean, SD, and range of MAE) are reported in the Results to characterise between-participant variability in prediction error.

To explore the temporal dynamics between metabolic stress and thermal response, the cumulative physiological load was quantified using Area Under the Curve via the trapezoidal rule in SciPy25. Furthermore, a cross-correlation analysis was employed to determine the optimal temporal lag between lactate accumulation and \(T_c\) rise. Sex-based differences in prediction bias and kinetic lags were assessed using independent t-tests. For all analyses, the significance threshold was set at \(p < 0.05\).

To ensure computational efficiency and code optimization, a Large Language Model (Claude; Anthropic, San Francisco, CA) was utilized as a coding assistant for script refinement.

AI assisted copy editing

Generative AI tools were employed during the manuscript preparation phase to enhance grammatical precision, linguistic clarity, and overall readability. All final content and analytical outputs were reviewed and verified by the authors to ensure scientific accuracy.

Discussion

The results of this study provide compelling evidence for the accuracy and reliability of the MSPORT device in measuring heart rate and body temperature when compared to established reference devices, namely the Polar H10 for heart rate and the Braun 7 for body temperature. Additionally, in this study, lactate levels were predicted using HR and Tc data collected from the MSPPORT. The performance of the prediction model was assessed through regression analysis.

Heart rate measurement

The analysis of heart rate measurements, as depicted in the Bland-Altman plot (Fig. 1a), indicates a MAE of 0.89 bpm, suggesting that the MSPORT device deviates from the Polar H10 by slightly more than one beat per minute on average. This level of performance aligns with findings from previous studies investigating the efficacy of heart rate monitors integrated into earbuds using photoplethysmography3738. Such accuracy is particularly important in sports contexts, where precise heart rate monitoring is essential.

To contextualize the MSPORT’s performance, comparisons were made with previously validated earbud-based HR monitors tested under similar exercise conditions. The Bose SoundSport Pulse Headphones, evaluated during a comparable graded exercise test, reported a MAPE of 15.42% at 200 W and an average MAPE of 7.44%39, whereas the MSPORT achieved a MAPE of 0.73% at 200 W and an average MAPE of 0.97% under the same intensity. Similarly, a clinical-grade in-ear photoplethysmography device (Cosinuss, Munich, Germany)40, validated in a clinical rather than exercise setting, reported limits of agreement of LLoA = 3.95 bpm and ULoA = 5.53 bpm with an average MAPE of 2.43%, compared to LLoA = −3.25 bpm and ULoA = 5.07 bpm and a MAPE of 0.97% for the MSPORT. It should be noted that these comparisons are indicative rather than conclusive, as differences in participant populations, exercise protocols, and measurement settings limit direct equivalence between studies.

The RMSE of 1.10 bpm and a mean difference of 0.79 bpm, which indicates a slight tendency for the MSPORT device to report higher heart rates, are also noteworthy. While the bias suggests a systematic difference, it remains within an acceptable range less than 1 bpm, reinforcing the notion that the MSPORT device is a reliable alternative for heart rate monitoring. The overall findings suggest that the MSPORT device can serve as a practical tool for both fitness enthusiasts and professional athletes, providing accurate heart rate data that aligns closely with established standards.

Ear-canal temperature agreement

Similarly, the validation of body temperature measurements against the Braun 7 device demonstrates good inter-device agreement across all participants. Importantly, both devices measure ear-canal (tympanic) temperature; this analysis characterises agreement between two measurement approaches at the same anatomical site rather than validation against a gold-standard core temperature measure such as rectal or oesophageal thermometry. The MAE values, indicate that the MSPORT device consistently performs within acceptable limits. The MAPE values, which are below 1%, further support the reliability of the MSPORT device in temperature assessment.

Notably, the RMSE observed in this study is lower than that reported in a systematic review, which found that 17 out of 18 published algorithms achieved a clinical acceptance threshold of RMSE less than or equal to 0.5 \(^\circ\)C41 supporting the conclusion that the MSPORT device performs well in terms of precision.

The limits of agreement for the MSPORT (−0.72 \(^\circ\)C to 0.95 \(^\circ\)C) were compared to those of the CORE sensor (−0.62 \(^\circ\)C to 0.59 \(^\circ\)C), a validated wearable core temperature monitor tested during cycling exercise15. The CORE sensor demonstrated slightly narrower limits of agreement, suggesting marginally higher temperature precision than the MSPORT. Nevertheless, both devices fall within clinically acceptable thresholds, and the MSPORT offers the additional advantage of simultaneous heart rate monitoring within a single earbud device.

Overall, the findings from this study highlight that the MSPORT tracker shows good agreement with a co-located tympanic reference device for ear-canal temperature measurement during exercise. The small mean differences, low MAE and MAPE values, and reasonable RMSE underscore its effectiveness for both sports and personal use. As technology continues to advance, the integration of such reliable devices into everyday sport monitoring could enhance user experience and promote better performance outcomes. Future studies could expand on these findings by including a larger and more diverse participant pool, as well as exploring the device’s performance under varying environmental conditions and activity levels.

Blood lactate prediction and kinetic coupling

To our knowledge, this is the first study to predict blood lactate concentrations using a non-invasive ear-worn sensor combining HR and \(T_c\). Direct comparisons with existing research are therefore limited. While prior studies have investigated near-infrared spectroscopy inflection points and blood lactate thresholds42,43, predictive modeling of lactate from thermal and cardiac signals remains largely unexplored. Among the few comparable studies, Huang et al. (2019) employed a mathematical model during cycling and reported an overall correlation of \(R = 0.64\), with a higher correlation of \(R = 0.96\) at low-to-moderate intensities44. Our model achieved an overall correlation of \(R = 0.84\), exceeding Huang et al.’s overall figure, although differences in modeling approach, predictors, and protocol preclude direct equivalence.

A systematic ablation analysis (see Results) demonstrated that ear-canal temperature alone carries most of the predictive signal for blood lactate, whereas heart rate alone has near-chance performance. Combining HR and Tc yields only a modest additional improvement, with the best performance achieved by the full 5-feature Gradient Boosting model. These findings indicate that the metabolic signal is primarily thermal, with HR contributing limited additional information under the present protocol.

However, it is essential to interpret these results in the context of absolute prediction accuracy. The overall MAE of 1.53 mmol/L and MAPE of 48.8% represent a meaningful improvement over the original linear model (MAE: 1.72 mmol/L, MAPE: 61.8%), achieved by incorporating rate-of-change and interaction features into a Gradient Boosting framework. The limits of agreement (\(-4.52\) to 4.55 mmol/L) remain wide, and the model is not yet suitable for individual-level lactate quantification in clinical or performance settings. The stratified MAPE analysis reveals an important nuance: MAPE was 64.4% below 4 mmol/L but only 28.1% above it. This apparent paradox reflects the mathematical amplification of percentage errors at low absolute values rather than genuine model failure at low intensities; the absolute error is similar across the lactate range. The model therefore performs with enhanced reliability at the high-intensity range where lactate monitoring is most useful.

These limitations are primarily driven by the ramp test protocol, in which HR and \(T_c\) exhibit near-linear progression while lactate increases exponentially beyond the anaerobic threshold. Consequently, the current model should be interpreted as suitable primarily for group-level trend monitoring during incremental loading, and its generalizability to steady-state, interval, or outdoor exercise remains to be established. Future enhancements should incorporate additional physiological parameters such as oxygen saturation, respiration rate, and heart rate variability, and evaluate step or constant work rate protocols where the linear-exponential mismatch is less pronounced.

The Role of Core Temperature and Feature Importance A critical finding was the high statistical significance of cumulative ear-canal temperature delta (\(\Delta T_c\)) as an independent predictor in the Linear Mixed Model (\(z = 11.65\), \(p < 0.001\)). This demonstrates that \(\Delta T_c\) is not merely a correlate of HR but captures a distinct biological signal, likely reflecting the systemic thermal flux generated during glycolysis. To account for inter-individual baselines and the nested structure of repeated measures, a random-intercept model was employed, with participant included as a random effect, considered the standard approach for confirmatory hypothesis testing in physiological data45.

As an exploratory analysis, permutation feature importance was assessed to identify the relative contribution of each predictor in the Gradient Boosting model. The cumulative ear-canal temperature delta (\(\Delta T_c\)) was the strongest individual predictor, with a mean MAE increase of 1.46 mmol/L (SD: 0.10) when randomly shuffled. The rate of change of heart rate (\(\dot{HR}\)) was the second strongest predictor, (1.02 mmol/L, SD: 0.08), suggesting that HR acceleration carries additional metabolic information beyond absolute HR during incremental exercise. The rate of change of temperature (\(\dot{T}_c\): 0.42 mmol/L) and absolute normalised HR (\(HR_{norm}\): 0.20 mmol/L) were weaker contributors. Combined, temperature-related features were 1.5 times more important than HR-related features, a modest but consistent difference. These findings should be interpreted as exploratory given the post-hoc nature of the feature importance analysis and the relatively small sample of 23 participants. Confirmation in prospective studies with pre-specified feature selection is warranted.

Exploratory Temporal Analysis A novel contribution of this work is the examination of the temporal coupling between systemic lactate and ear canal thermogenesis. Cross-correlationcan alysis indicated that blood lactate tends to rise before \(T_c\) by an average of 16.5 seconds during incremental exercise, suggesting a preliminary temporal association between the metabolic and thermal signals. However, because blood lactate was sampled only every 3 minutes, whereas ear-canal temperature was recorded continuously at 1 Hz, any second-scale lag estimate is highly sensitive to interpolation assumptions and the shared monotonic increase of both signals during the ramp protocol. As such, we do not interpret the apparent lag as a precise physiological constant.

The observed temporal pattern is consistent with the notion that lactate-driven heat production to propagate through the circulatory system to the cranial thermal compartment, consistent with the Lactate Shuttle Theory which identifies lactate as both a fuel and a signaling molecule18. The proximity to highly perfused tissues and major vessels like the carotid artery likely explains why the ear canal is able to capture such rapid thermal flux12. Nevertheless, given the 3-minute lactate sampling interval, the exact magnitude of any lag remains uncertain and may vary substantially depending on the interpolation method and analysis choices. Therefore, these temporal findings should be regarded as hypothesis-generating and require confirmation with higher-frequency lactate measurements and more appropriate time-series methods.

Sex-Based Differences The significant sex-based difference in lactate magnitude (\(p < 0.05\)) contrasts with Rascon et al. (2020), who reported no sex differences during exercise46, but aligns with Deschenes et al. (2006)47. This discrepancy is likely attributed to fitness level: the \(\dot{V}O_{2\text {max}}\) values of our cohort (Males: 51.93; Females: 39.00 mL/kg/min) are comparable to the trained subjects in Deschenes et al., whereas Rascon’s participants had substantially lower aerobic capacities. Beyond fitness, estrogen promotes fat oxidation and glycogen sparing in women, typically resulting in lower lactate accumulation at equivalent relative submaximal workloads48,49. Men additionally show higher reliance on carbohydrate oxidation during incremental exercise49. Importantly, while biological sex influenced the magnitude of lactate response, the kinetic lead time remained stable across sexes (\(p = 0.4418\)), suggesting that the metabolic-to-thermal coupling mechanism is a conserved physiological process independent of sex.

These sex-based comparisons are based on relatively small groups and should be interpreted as exploratory; participant-level summaries are presented in the sex-stratified tables, and repeated observations within participants are not treated as independent.

Implications and limitations

The findings from this study support the use of the MSPORT device as a reliable tool for continuous HR and \(T_c\) tracking during laboratory cycling. The HR validation results, in particular, are strong and comparable to established earbud-based monitors. However, it is essential to emphasize that all results were obtained under strictly controlled indoor laboratory conditions and should not be generalised without further validation. Furthermore, the reference device (Braun ThermoScan 7) is an infrared tympanic thermometer and not a gold-standard core temperature measure (e.g., esophageal, rectal, or ingestible telemetric capsules). Consequently, our temperature analysis characterises inter-device agreement at the ear canal rather than absolute core temperature accuracy, and bilateral ear measurements (MSPORT in the left ear, Braun in the contralateral right ear) may be affected by side-to-side anatomical and perfusion differences.

The lactate prediction model, a Gradient Boosting Regressor with extended features, achieved a correlation of \(R = 0.84\) with MAE of 1.53 mmol/L, MAPE of 48.8%, and limits of agreement of \(-4.52\) to 4.55 mmol/L. These figures indicate the model is currently suitable only for group-level trend monitoring and not for individual-level lactate quantification. In a ramp protocol, elapsed time and workload increase monotonically together with HR and temperature, so simple time- or workload-based models would also correlate with lactate. Our ablation (HR-only, \(T_c\)-only, combined) shows that ear-canal temperature alone carries most of the predictive information, but we did not implement explicit time-only or workload-only benchmark models. This remains a limitation and a priority for future work. These findings are specific to an incremental cycling ramp protocol at a stable room temperature (\(23.21 \pm 0.49^{\circ }\)C) and humidity (\(35.77 \pm 6.05\%\)). The model’s generalizability to other contexts, including outdoor cycling, running, or thermally challenging environments, cannot be assured. In outdoor settings, variables such as wind-induced cooling of the ear canal or varying solar radiation could alter the relationship between ear canal thermal flux and systemic metabolic heat.

A further limitation of this study is the data attrition encountered during the temperature and lactate analyses. Complete validation datasets for both parameters were available for 23 out of the 36 participants (Table 3). These exclusions were primarily necessitated by intermittent wireless connectivity constraints and movement-induced sensor displacement during the strenuous exercise protocols, rather than systematic participant characteristics. Consequently, statistical power and precision for temperature- and lactate-related analyses are limited, and the relatively small sample constrains the generalisability of the lactate prediction model. Furthermore, the concentration of data loss among participants during high-intensity exercise phases should be interpreted as a device usability finding: the MSPORT sensor is susceptible to signal dropout under conditions of vigorous movement, which represents an important practical limitation for future field deployment and should be addressed in device development.

Future research

To bridge the gap between laboratory validation and practical application, future research must move beyond the constraints of the fast ramp test protocol used in this study. Specifically:

Future studies should employ different exercise tests, such as steady-state constant work rate tests or step-incremental protocols, to assess whether the6.5-second lag is a physiological constant or a protocol-dependent variable. Investigating how the metabolic-to-thermal coupling behaves in cold or hot environments is necessary to determine if the “kinetic delay” remains stable across different thermoregulatory demands. Future research should assess if different movement patterns (e.g., vertical oscillation in running) impact the photoplethysmogram based heart rate accuracy and the stability of the ear canal temperature sensor. Future iterations of predictive models could explore temporal shifting of the \(T_c\) signal if higher-frequency metabolic sampling confirms a physiological lag between lactate production and ear-canal temperature response. Such studies should include continuous or near-continuous lactate monitoring, detrended cross-correlation analysis, and permutation-based significance testing to provide a robust estimate of any temporal coupling.lood lactate, potentially improving accuracy during rapid, non-steady-state intensity changes.

Conclusion

In conclusion, this study demonstrates that the MSPORT device provides a reliable platform for monitoring heart rate and core temperature during controlled laboratory cycling, with HR accuracy comparable to established earbud-based monitors. Beyond simple tracking, we have established a preliminary framework for estimating metabolic strain through the integration of cardiovascular and thermal signals. An exploratory cross-correlation analysis suggested a preliminary temporal association between lactate and ear-canal temperature signals; however, given the three-minute lactate sampling interval, the precise magnitude of any lag cannot be reliably estimated from the present data and this observation requires confirmation with higher-frequency metabolic sampling.

The current lactate prediction model demonstrates promising group-level correlation (\(R = 0.84\)) with improved accuracy following the integration of rate-of-change and interaction features into a Gradient Boosting framework (MAE: 1.53 mmol/L, MAPE: 48.8%, LoA: \(-4.52\) to 4.55 mmol/L). The stratified analysis revealed that MAPE above the 4 mmol/L conventional reference value was 28.1%, substantially lower than below it (64.4%), suggesting the model performs most reliably at high-intensity range. Feature importance analysis identified the cumulative ear-canal temperature delta (\(\Delta T_c\)) as the dominant predictor (mean MAE increase: 1.46 mmol/L), a finding consistent with the statistical significance of \(T_c\) in the LMM and supporting the physiological relevance of ear-canal temperature for metabolic estimation. Despite these improvements, individual-level error remains substantial, limiting immediate clinical applicability. The model is best interpreted as a proof-of-concept for thermal-metabolic coupling rather than a validated predictive tool. Further research across diverse exercise protocols, environmental conditions, and athletic populations is necessary before these algorithms can be refined for real-world competition monitoring. Ultimately, the ability to non-invasively monitor both HR and thermal flux from a single ear-worn device represents a meaningful step toward integrated physiological monitoring in sport.

References

  1. Buchheit, M. Monitoring training status with hr measures: do all roads lead to rome?. Front. Physiol. 5, 71297 (2014).

    Article  Google Scholar 

  2. Lamberts, R. P., Rietjens, G. J., Tijdink, H. H., Noakes, T. D. & Lambert, M. I. Measuring submaximal performance parameters to monitor fatigue and predict cycling performance: A case study of a world-class cyclo-cross cyclist. Eur. J. Appl. Physiol. 108, 183–190 (2010).

    Article  PubMed  Google Scholar 

  3. Hara, S., Kawabata, T. & Nakamura, H. Real-time sensing, transmission and analysis for vital signs of persons during exercises. In 37th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC) 4017–4020 (IEEE, 2015).

  4. Tansey, E. A. & Johnson, C. D. Recent advances in thermoregulation. Adv. Physiol. Educ. 39 (3), 139–148 (2015).

    Article  PubMed  Google Scholar 

  5. No, M. & Kwak, H.-B. Effects of environmental temperature on physiological responses during submaximal and maximal exercises in soccer players. Integr. Med. Res. 5(3), 216–222 (2016).

    Article  PubMed  Google Scholar 

  6. Murtuza, F. et al. Is there an association between body temperature and serum lactate levels in hip fracture patients?. Ann. R. Coll. Surg. Engl. 97(7), 513–518 (2015).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  7. Wakabayashi, H., Wijayanto, T. & Tochihara, Y. Neuromuscular function during knee extension exercise after cold water immersion. J. Physiol. Anthropol. 36(1), 28 (2017).

    Article  PubMed  PubMed Central  Google Scholar 

  8. Rome, L. C. Influence of temperature on muscle recruitment and muscle function in vivo. Am. J. Physiol. - Regul. Integr. Comp. Physiol. 259(2), 210–222 (1990).

    Article  Google Scholar 

  9. Beelen, A. & Sargeant, A. Effect of lowered muscle temperature on the physiological response to exercise in men. Eur. J. Appl. Physiol. Occup. Physiol. 63(5), 387–392 (1991).

    Article  CAS  PubMed  Google Scholar 

  10. Armstrong, L. E. et al. Exertional heat illness during training and competition. Med. Sci. Sports Exerc. 39(3), 556–572 (2007).

    Article  PubMed  Google Scholar 

  11. Nybo, L. & Nielsen, B. Hyperthermia and central fatigue during prolonged exercise in humans. J. Appl. Physiol. 91(3), 1055–1060 (2001).

    Article  CAS  PubMed  Google Scholar 

  12. Yeoh, W. K. et al. Re-visiting the tympanic membrane vicinity as core body temperature measurement site. PLoS One 12(4), e0174120 (2017).

    Article  MathSciNet  PubMed  PubMed Central  Google Scholar 

  13. González-Alonso, J. et al. Influence of body temperature on the development of fatigue during prolonged exercise in the heat. J. Appl. Physiol. 86(3), 1032–1039 (1999).

    Article  PubMed  Google Scholar 

  14. Nielsen, B. et al. Human circulatory and thermoregulatory adaptations with heat acclimation and exercise in a hot, dry environment. J. Physiol. 460(1), 467–485 (1993).

    Article  MathSciNet  CAS  PubMed  PubMed Central  Google Scholar 

  15. Verdel, N. et al. Reliability and validity of the core sensor to assess core body temperature during cycling exercise. Sensors 21(17), 5932 (2021).

    Article  PubMed  PubMed Central  Google Scholar 

  16. Bergh, U. & Ekblom, B. Physical performance and peak aerobic power at different body temperatures. J. Appl. Physiol. 46(5), 885–889 (1979).

    Article  CAS  PubMed  Google Scholar 

  17. Beneke, R., Leithäuser, R. M. & Ochentel, O. Blood lactate diagnostics in exercise testing and training. Int. J. Sports Physiol. Perform. 6(1), 8–24 (2011).

    Article  PubMed  Google Scholar 

  18. Brooks, G. A. The science and translation of lactate shuttle theory. Cell Metab. 27(4), 757–785 (2018).

    Article  CAS  PubMed  Google Scholar 

  19. Hajj-Boutros, G., Landry-Duval, M.-A., Comtois, A. S., Gouspillou, G. & Karelis, A. D. Wrist-worn devices for the measurement of heart rate and energy expenditure: A validation study for the Apple Watch 6, Polar Vantage V and Fitbit Sense. Eur. J. Sport Sci. 23(2), 165–177 (2023).

    Article  PubMed  Google Scholar 

  20. Byrne, C. & Lim, C. L. The ingestible telemetric body core temperature sensor: a review of validity and exercise applications. Br. J. Sports Med. 41(3), 126–133 (2007).

    Article  PubMed  Google Scholar 

  21. Van Remoortel, H. et al. Validity of activity monitors in health and chronic disease: A systematic review. Int. J. Behav. Nutr. Phys. Act.https://doi.org/10.1186/1479-5868-9-84 (2012).

    Article  PubMed  PubMed Central  Google Scholar 

  22. Lyons, E. J., Lewis, Z. H., Mayrsohn, B. G. & Rowland, J. L. Behavior change techniques implemented in electronic lifestyle activity monitors: A systematic content analysis. J. Med. Internet Res. 16(8), 192. (2014).

    Article  Google Scholar 

  23. Pedregosa, F. et al. Scikit-learn: Machine learning in Python. J. Mach. Learn. Res. 12, 2825–2830 (2011).

    MathSciNet  Google Scholar 

  24. Seabold, S. & Perktold, J. statsmodels: Econometric and statistical modeling with python. In 9th Python in Science Conference (2010).

  25. Virtanen, P. et al. SciPy 1.0: Fundamental algorithms for scientific computing in python. Nat. Methods 17, 261–272. https://doi.org/10.1038/s41592-019-0686-2 (2020).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  26. Gillinov, S. et al. Variable accuracy of wearable heart rate monitors during aerobic exercise. Med. Sci. Sports Exerc. 49(8), 1697–1703 (2017).

    Article  PubMed  Google Scholar 

  27. Gilgen-Ammann, R., Schweizer, T. & Wyss, T. Rr interval signal quality of a heart rate monitor and an ecg holter at rest and during exercise. European J. Appl. Physiol. 119(7), 1525–1532 (2019).

    Article  Google Scholar 

  28. Mah, A. J. et al. Studying the accuracy and function of different thermometry techniques for measuring body temperature. Biology 10(12), 1327 (2021).

    Article  PubMed  PubMed Central  Google Scholar 

  29. Meyer, T., Georg, T., Becker, C. & Kindermann, W. Reliability of gas exchange measurements from two different spiroergometry systems. Int. J. Sports Med. 22(08), 593–597 (2001).

    Article  CAS  PubMed  Google Scholar 

  30. Xiong, K.-Y., He, H. & Ni, G.-X. Effect of skill level on cardiorespiratory and metabolic responses during tai chi training. Eur. J. Sport Sci. 13(4), 386–391 (2013).

    Article  PubMed  Google Scholar 

  31. Wang, Z. et al. Enhancing instantaneous oxygen uptake estimation by non-linear model using cardio-pulmonary physiological and motion signals. Front. Physiol. 13, 897412 (2022).

    Article  PubMed  PubMed Central  Google Scholar 

  32. Mentzoni, F. et al. Precision and accuracy of four handheld blood lactate analyzers across low to high exercise intensities. Eur. J. Appl. Physiol. 124(12), 3781–3788 (2024).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  33. Midgley, A. W., McNaughton, L. R. & Wilkinson, M. Is there an optimal training intensity for enhancing the maximal oxygen uptake of distance runners? empirical research findings, current opinions, physiological rationale and practical recommendations. Sports Med. 36(2), 117–132 (2006).

    Article  PubMed  Google Scholar 

  34. Moran, P., Prichard, J. G., Ansley, L. & Howatson, G. The influence of blood lactate sample site on exercise prescription. J. Strength Cond. Res. 26(2), 563–567 (2012).

    Article  PubMed  Google Scholar 

  35. Bland, J. M. & Altman, D. Statistical methods for assessing agreement between two methods of clinical measurement. The lancet 327(8476), 307–310 (1986).

    Article  Google Scholar 

  36. Bland, J. M. & Altman, D. G. Agreement between methods of measurement with multiple observations per individual. J. Biopharm. Stat. 17(4), 571–582 (2007).

    Article  MathSciNet  PubMed  Google Scholar 

  37. Poh, M.-Z., Kim, K., Goessling, A., Swenson, N. & Picard, R. Cardiovascular monitoring using earphones and a mobile device. IEEE Pervasive Comput. 11(4), 18–26 (2010).

    Article  Google Scholar 

  38. Bunn, J., Wells, E., Manor, J. & Webster, M. Evaluation of earbud and wristwatch heart rate monitors during aerobic and resistance training. Int. J. Exerc. Sci. 12(4), 374 (2019).

    Article  PubMed  PubMed Central  Google Scholar 

  39. Boudreaux, B. D. et al. Validity of wearable activity monitors during cycling and resistance exercise. Med. Sci. Sports Exerc. 50(3), 624–633 (2018).

    Article  PubMed  Google Scholar 

  40. Adams, T. et al. Accurate detection of heart rate using in-ear photoplethysmography in a clinical setting. Front. Digit. Health 4, 909519 (2022).

    Article  PubMed  PubMed Central  Google Scholar 

  41. Dolson, C. M. et al. Wearable sensor technology to predict core body temperature: A systematic review. Sensors 22(19), 7639 (2022).

    Article  PubMed  PubMed Central  Google Scholar 

  42. Bellotti, C. et al. Determination of maximal lactate steady state in healthy adults: Can nirs help. Med. Sci. Sports Exerc. 45(6), 1208–1216 (2013).

    Article  CAS  PubMed  Google Scholar 

  43. Borges, N. R. & Driller, M. W. Wearable lactate threshold predicting device is valid and reliable in runners. J. Strength Cond. Res. 30(8), 2212–2218 (2016).

    Article  PubMed  Google Scholar 

  44. Huang, S.-C. et al. Noninvasive prediction of blood lactate through a machine learning-based approach. Sci. Rep. 9(1), 2180 (2019).

    Article  PubMed  PubMed Central  Google Scholar 

  45. Barr, D. J., Levy, R., Scheepers, C. & Tily, H. J. Random effects structure for confirmatory hypothesis testing: Keep it maximal. J. Mem. Lang. 68(3), 255–278 (2013).

    Article  Google Scholar 

  46. Rascon, J., Trujillo, E., Morales-AcuÑa, F. & Gurovich, A. N. Differences between males and females in determining exercise intensity. Int. J. Exerc. Sci. 13(4), 1305 (2020).

    PubMed  PubMed Central  Google Scholar 

  47. Deschenes, M. R., Hillard, M. N., Wilson, J. A., Dubina, M. I. & Eason, M. K. Effects of gender on physiological responses during submaximal exercise and recovery. Med. Sci. Sports Exerc. 38(7), 1304–1310 (2006).

    Article  PubMed  Google Scholar 

  48. Tarnopolsky, M. A. Sex differences in exercise metabolism and the role of 17-beta estradiol. Med. Sci. Sports Exerc. 40(4), 648–654 (2008).

    Article  CAS  PubMed  Google Scholar 

  49. Cano, A. et al. Analysis of sex-based differences in energy substrate utilization during moderate-intensity aerobic exercise. Eur. J. Appl. Physiol. 122(1), 29–70 (2022).

    Article  CAS  PubMed  Google Scholar 

Download references

来源:Nature 健康护理主题 · nature.com