Validity of the polar h7 heart rate sensor for heart rate variability analysis during exercise in different age, body composition and fitness level groups

sensors Article Validity of the Polar H7 Heart Rate Sensor for Heart Rate Variability Analysis during Exercise in Different Age, Body Composition and Fitness Level Groups Adrián Hernández-Vicente 1,2,3,* , David Hernando 4,5 , Jorge Marín-Puyalto 1,2,3, Germán Vicente-Rodríguez 1,2,3,6,7 , Nuria Garatachea 1,2,3,6,7, Esther Pueyo 4,5,† and Raquel Bailón4,5,†   Citation: Hernández-Vicente, A.; Hernando, D.; Marín-Puyalto, J.; Vicente-Rodríguez, G.; Garatachea, N.; Pueyo, E.; Bailón, R. Validity of the Polar H7 Heart Rate Sensor for Heart Rate Variability Analysis during Exercise in Different Age, Body Composition and Fitness Level Groups. Sensors 2021,21, 902. https://doi.org/10.3390/s21030902 Academic Editor: Marco Altini Received: 22 December 2020 Accepted: 26 January 2021 Published: 29 January 2021 Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations. Copyright: © 2021 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/). 1Growth, Exercise, NUtrition and Development (GENUD) Research Group, University of Zaragoza, 50009 Zaragoza, Spain; [email protected] (J.M.-P.); [email protected] (G.V.-R.); [email protected] (N.G.) 2Department of Physiatry and Nursing, Faculty of Health and Sport Science (FCSD), University of Zaragoza, 22002 Huesca, Spain 3Red Española de Investigación en Ejercicio Físico y Salud en Poblaciones Especiales (EXERNET), Spain 4BSICOS, Aragón Institute for Engineering Research (I3A), IIS Aragón, University of Zaragoza, 50015 Zaragoza, Spain; [email protected] (D.H.); [email protected] (E.P.); [email protected] (R.B.) 5CIBER de Bioingeniería, Biomateriales y Nanomedicina (CIBER-BBN), 50009 Zaragoza, Spain 6Centro de Investigación Biomédica en Red de Fisiopatología de la Obesidad y Nutrición (CIBER-Obn), 28029 Madrid, Spain 7Instituto Agroalimentario de Aragón-IA2CITA-Universidad de Zaragoza), 50013 Zaragoza, Spain *Correspondence: [email protected] † These authors contributed equally to the work. Abstract: This work aims to validate the Polar H7 heart rate (HR) sensor for heart rate variability (HRV) analysis at rest and during various exercise intensities in a cohort of male volunteers with different age, body composition and fitness level. Cluster analysis was carried out to evaluate how these phenotypic characteristics influenced HR and HRV measurements. For this purpose, sixtyseven volunteers performed a test consisting of the following consecutive segments: sitting rest, three submaximal exercise intensities in cycle-ergometer and sitting recovery. The agreement between HRV indices derived from Polar H7 and a simultaneous electrocardiogram (ECG) was assessed using concordance correlation coefficient (CCC). The percentage of subjects not reaching excellent agreement (CCC > 0.90) was higher for high-frequency power (P HF ) than for low-frequency power (P LF ) of HRV and increased with exercise intensity. A cluster of unfit and not young volunteers with high trunk fat percentage showed the highest error in HRV indices. This study indicates that Polar H7 and ECG were interchangeable at rest. During exercise, HR and P LF showed excellent agreement between devices. However, during the highest exercise intensity, CCC for P HF was lower than 0.90 in as many as 60% of the volunteers. During recovery, HR but not HRV measurements were accurate. As a conclusion, phenotypic differences between subjects can represent one of the causes for disagreement between HR sensors and ECG devices, which should be considered specifically when using Polar H7 and, generally, in the validation of any HR sensor for HRV analysis. Keywords: electrocardiography; wearable devices; HRV analysis; cluster analysis; exercise test 1. Introduction Heart rate (HR) variability (HRV) is the oscillation in the intervals between consecutive heartbeats (RR intervals) [ 1 ]. In the last decades, the use of HRV has been popularized, since it allows assessing cardiac autonomic modulation using simple and non-invasive techniques. In general, lower HRV has been associated with poorer prognosis in different clinical conditions, while higher HRV, especially regarding high-frequency oscillations, has been associated with better health. In particular, reduced HRV has been reported in several cardiovascular diseases and has been used for risk stratification, confirming its value as a Sensors 2021,21, 902. https://doi.org/10.3390/s21030902 https://www.mdpi.com/journal/sensors Sensors 2021,21, 902 2 of 14 predictor of total and cardiac mortality [ 2 – 4 ]. Also, low HRV has been described in a wide range of non-cardiovascular diseases, including psychiatric disorders such as depression, anxiety or schizophrenia [5]. Environmental and behavioral factors influence HRV [ 6 ]. In psychology, HRV is commonly used due to its modulation by mood states, emotions or cognitive capacity [ 5 ]. Likewise, HRV is a useful measurement in the sports field since it is sensitive to changes in fitness, fatigue and performance [ 7 ]. Healthy habits like exercise, balanced diet, mindfulness or psychological interventions have been shown to increase HRV measures indexing vagal function [ 5 , 6 ]. Additionally, non-modifiable factors, such as age or sex, also influence the autonomic regulation of the heart, with HRV having been reported to progressively decline with age and to present enhanced high-frequency oscillations and attenuated lowfrequency oscillations in females as compared to males [ 8 ]. Because of these high interand intra-individual variations in HRV, it is key to use wearable devices validated for HRV analysis that can allow for precise interpretation of cardiac responses to different autonomic states. A recent systematic review and meta-analysis showed that HRV measurements derived from portable devices are generally accurate when compared to lab-based electrocardiogram (ECG) [ 9 ]. Given the low cost of HR monitors it is not surprising that they are widely used by practitioners and researchers. Particularly, Polar Electro Oy (Kempele, Finland) is one of the most well-established brands in HR monitoring, with Polar H7/H10 HR sensors having been validated both at rest and during exercise [ 10 – 12 ]. Nevertheless, previous Polar validation studies have been carried out in small groups of young, lean, healthy and physically fit volunteers [ 10 – 12 ]. However, the device is commonly used by individuals with various phenotypic characteristics, regardless of how these may affect the accuracy of the measurements [ 13 , 14 ]. In the meta-analysis described in [ 9 ], the absolute error of portable devices was found to vary with the evaluated HRV metric, tilt/recovery position and the percentage of women in the study sample. The characteristics of the subjects and their influence on the measurements provided by portable devices have not been analyzed yet. The ECG records the electrical activity of the heart using electrodes placed on the surface of the body. Therefore, differences between measurements from HR sensors and ECG could vary depending on the characteristics of the population under study. To start with, the age-related myocardial fibrosis present in the cardiac tissue, the amount of subcutaneous fat or the electrode placement are expected to affect the voltage tracings [ 15 ]. Additionally, voltage will be influenced by ventricular size or mass, as observed when comparing trained athletes with non-athletes [ 16 ]. Accordingly, some groups of subjects such as men, athletes or black/African have been reported to have higher QRS voltage and, consequently, RR intervals become easier to be detected [ 16 ]. On the other hand, obesity, older age and sedentary lifestyle may cause lower voltage, which may result in lower accuracy of portable devices. In particular, under such circumstances, some heart beats can be misdetected and, while this may not considerably affect mean HR, it may notably hamper HRV assessment. For these reasons and in light of previous studies, we hypothesized that when the quality of the Polar H7 ECG is compromised by high noise during intense exercise and/or by specific phenotypic characteristics of the subjects, HRV measures can be distorted, particularly those related to high-frequency power [17]. The purpose of the present study was to evaluate the validity of HRV analysis derived from RR intervals recorded by Polar H7 HR sensor at rest and during exercise and recovery in different phenotype groups based on age, body composition and fitness level. 2. Materials and Methods 2.1. Subjects A total of sixty-seven males agreed to participate in the study. The sample consisted of three groups of volunteers: 22 young adults (20–30 years old), 22 middle-aged adults (40–50 years old) and 23 older adults (60–70 years old). Only subjects within the prede- Sensors 2021,21, 902 3 of 14 fined age ranges were included in the study. Subjects were excluded from the study if they were going through an acute disease, were suffering from heart diseases (e.g., heart failure or atrial fibrillation), were on cardiac medication or presented any clinical condition contraindicating physical exercise. However, subjects who were overweight, sedentary or suffering from chronic diseases such as hypertension, diabetes or hypercholesterolemia were included in the study, because of their high prevalence in the society. Table 1shows the descriptive characteristics of the three age groups. The study was approved by the ethical committee for clinical research of Aragón (ID of the approval: PI17/0409), and was conducted by adhering to the Declaration of Helsinki. After a clear explanation of the potential risks of the study, all volunteers provided written informed consent. Table 1. Descriptive characteristics of the three age groups. Outcome Young Adults (n= 22) Middle-Aged Adults (n= 22) Older Adults (n= 23) Age (years) 25.46 ±2.85 43.17 ±3.32 63.97 ±2.79 Height (m) 1.75 ±0.06 1.77 ±0.06 1.71 ±0.05 Weight (kg) 72.01 ±11.92 78.19 ±10.30 76.31 ±7.76 BMI (m/kg2)23.43 ±2.95 25.02 ±2.83 26.21 ±2.84 Body fat (%) 15.25 ±5.59 19.69 ±5.65 23.38 ±5.17 Trunk fat (%) 16.31 ±6.32 21.29 ±6.38 25.69 ±6.42 PWC80% (W/kg) 2.00 ±0.64 2.01 ±0.58 1.73 ±0.65 Values are expressed as mean ± standard deviation (SD). BMI = Body mass index; PWC 80% = Physical Work Capacity at 80% of maximum HR (208 −0.7 ∗age in years) in watts/kg bodyweight. 2.2. Procedure All subjects completed one test session. Prior to the test, they were asked to adhere to the following instructions [ 18 ]: (1) avoid exercise or strenuous physical activity the day before the test; (2) drink plenty of fluids over the 24-h period preceding the test; (3) get an adequate amount of sleep (6–8 h) the night before the test; (4) avoid substances such as tobacco, alcohol or stimulants (caffeine, theine, taurine, etc.) in the 8 h before the test; (5) avoid food intake for 3 h prior to performing the test; and (6) wear comfortable, loosefitting clothing. Subjects’ skin was prepared by using a razor to remove any hair from the electrode sites, cleaning the skin with alcohol and drying it with a gauze. A 12-lead high-resolution Holter ECG was acquired, with the 10 electrodes placed as indicated by the manufacturer (H12+, Mortara Instrument, Milwaukee, WI, USA), ensuring that they did not interfere with the HR sensor strap (Polar H7, Polar Electro Oy). The test was conducted in an environmentally controlled room (22–23 ◦ C), between 16:00–20:00, and was divided into 3 consecutive segments: resting (S REST ), cycling (S CY ) and recovery (S REC ). During S REST , volunteers were monitored while seated at rest for 5 min, without any movement or talking. A period of 2–3 min was established to change from the chair to the cycle-ergometer, namely from S REST to S CY , during which the subject rode the electrically braked cycle-ergometer (Ergoselect 200 K, Ergoline; Bitz, Germany) at 50 W workload and chose a cadence which was maintained during the entire test according to the workload and cadence displayed in the cycle-ergometer screen. S CY was a submaximal cycle-ergometer test divided into three stages lasting 5 min each. In order to avoid a maximal exercise test, the maximum heart rate (HRmax) was estimated for each subject by using the formula defined by Tanaka et al. HRmax = 208 − 0.7 ∗ age (years) [ 19 ]. Workload was adjusted during each stage to 60, 70 and 80% of HRmax, with these stages denoted as S CY60 , S CY70 and S CY80 , respectively. Finally, during S REC , volunteers remained seated again for 5 min without any movement or talking. Figure 1shows an example of the temporal evolution of RR intervals from a subject throughout the entire test. Sensors 2021,21, 902 4 of 14 Sensors 2021, 21, x FOR PEER REVIEW 4 of 14 Workload was adjusted during each stage to 60, 70 and 80% of HRmax, with these stages denoted as SCY60, SCY70 and SCY80, respectively. Finally, during SREC, volunteers remained seated again for 5 min without any movement or talking. Figure 1 shows an example of the temporal evolution of RR intervals from a subject throughout the entire test. Figure 1. Example of the RR intervals for one subject throughout the entire test. Dotted lines separate the different test segments: resting (SREST), cycling (SCY) and recovery (SREC). SCY was divided into three stages corresponding to 60, 70 and 80% of HRmax, denoted as SCY60, SCY70 and SCY80, respectively. 2.3. Data Recording Subjects self-reported their birth date, current diseases and medication. The anthropometric characteristics of the subjects were assessed. Stature was measured to the nearest 0.001 m using a portable stadiometer (SECA 225, Hamburg, Germany), with subjects standing with their scapula, buttocks and heels resting against a wall, the feet with the heels touching, forming a 45° angle and the head in the Frankfort’s plane. A portable body composition analyzer (TANITA BC-418MA; Tanita Corp., Tokyo, Japan) was used to measure the body mass to the nearest 0.1 kg, with underwear and after urination. TANITA BC-418MA was also used to estimate the percentage of body fat and trunk fat (r = 0.87, p < 0.001 vs. dual-energy X-ray absorptiometry) [20]. Body mass index (BMI) was calculated dividing weight in kilograms by height in squared meters. Beat-to-beat RR intervals with 1-ms resolution were obtained using a Polar V800 HR monitor simultaneously with a Polar H7 chest Soft Strap (Polar Electro Oy, henceforth referred to as PolarH7). Concomitantly, a 12-lead ECG was recorded at a sampling rate of 1000 Hz using a high-resolution Holter device (H12+, Mortara Instrument, henceforth referred to as ECG and used here as a reference). VO2max can be estimated from submaximal exercise tests, a safe and feasible method showing good validity against maximal tests (correlation coefficients: 0.69 to 0.98) [21]. Rather than commonly used tests with stages of short or variable duration, an Figure 1. Example of the RR intervals for one subject throughout the entire test. Dotted lines separate the different test segments: resting (S REST ), cycling (S CY ) and recovery (S REC ). S CY was divided into three stages corresponding to 60, 70 and 80% of HRmax, denoted as SCY60, SCY70 and SCY80, respectively. 2.3. Data Recording Subjects self-reported their birth date, current diseases and medication. The anthropometric characteristics of the subjects were assessed. Stature was measured to the nearest 0.001 m using a portable stadiometer (SECA 225, Hamburg, Germany), with subjects standing with their scapula, buttocks and heels resting against a wall, the feet with the heels touching, forming a 45 ◦ angle and the head in the Frankfort’s plane. A portable body composition analyzer (TANITA BC-418MA; Tanita Corp., Tokyo, Japan) was used to measure the body mass to the nearest 0.1 kg, with underwear and after urination. TANITA BC-418MA was also used to estimate the percentage of body fat and trunk fat ( r = 0.87 , p< 0.001 vs. dual-energy X-ray absorptiometry) [ 20 ]. Body mass index (BMI) was calculated dividing weight in kilograms by height in squared meters. Beat-to-beat RR intervals with 1-ms resolution were obtained using a Polar V800 HR monitor simultaneously with a Polar H7 chest Soft Strap (Polar Electro Oy, henceforth referred to as PolarH7). Concomitantly, a 12-lead ECG was recorded at a sampling rate of 1000 Hz using a high-resolution Holter device (H12+, Mortara Instrument, henceforth referred to as ECG and used here as a reference). VO 2 max can be estimated from submaximal exercise tests, a safe and feasible method showing good validity against maximal tests (correlation coefficients: 0.69 to 0.98) [ 21 ]. Rather than commonly used tests with stages of short or variable duration, an ad-hoc test with 5-min stages was defined to allow reliable estimation of the low-frequency power of HRV. This enabled assessment of HRV response to increased sympathetic activity with each cycling stage [ 22 ]. Cardiorespiratory fitness was assessed using the approach of “Physical Work Capacity” (PWC). PWC in watts was measured during S CY80 of the submaximal cycle-ergometer test and was subsequently divided by the subject’s body weight (PWC 80% in W/kg). Alternatively to the use of fixed HR thresholds, this method incorporates the Sensors 2021,21, 902 5 of 14 age-dependent decline of HRmax [ 23 , 24 ] and has been previously used as an objective assessment of cardiorespiratory fitness [25,26]. 2.4. Data Analysis and Processing Raw RR interval time series, RR P (i), recorded by PolarH7 were downloaded from the “Polar Flow” web platform. RR interval time series from the ECG, RRE(i), were extracted using a multi-lead approach by a wavelet-based detector [ 27 ] with optimized parameters for noisy environments as described in [ 28 ]. Each beat detection was manually verified by an operator with a dedicated interface. The delay between the RR interval series RR P (i) and RR E (i) was estimated as the time lag maximizing their cross-correlation over the first 3 min of the test when the subject is relaxed. Then, both series were synchronized by compensating for this delay. Since the two RR interval series can have different lengths, due to, e.g., wrong or missed beat detections in the Polar data, an algorithm was developed to match the RR intervals from both series, thus allowing characterization of the agreement between the paired series RR P (ip) and RRE(ip), where ip refers to the indices of beats that are matched in the two series. 2.5. Heart Rate Variability HRV indices were obtained by algorithms specifically developed and previously published by our research group using MATLAB version R2017a (MATLAB, MathWorks Inc., Natick, MA, USA) [17,27–30]. 2.5.1. Temporal Domain The following temporal HRV indices were studied [ 1 ]: mean HR (MHR), standard deviation of normal-to-normal RR intervals (SDNN) and root mean square of successive differences of adjacent normal-to-normal RR intervals (RMSSD). MHR was obtained as the inverse of the mean of the RR intervals. SDNN is considered a measure of the total power of HRV and was calculated from the standard deviation of the NN intervals, i.e., normal RR intervals after correcting for ectopic beats [ 29 ]. RMSSD is a measure of short-term variability and was computed by the root mean square of successive differences between adjacent NN intervals. These indices were obtained from RR P (i) and RR E (i) in each segment of the test. 2.5.2. Frequency Domain The instantaneous HR signal, dHR(n) , was derived from both RR P (i) and RR E (i) and sampled at 4 Hz. The integral pulse frequency modulation (IPFM) model was used while dealing with the presence of ectopic beats [ 29 ]. This signal was high-pass-filtered (0.03 Hz) to remove the very low-frequency components, dMHR(n) , and it was also corrected by it: m(n)= (dHR(n)−dMHR(n))/dMHR(n)[30]. The smoothed pseudo Wigner–Ville distribution (SPWVD) was applied to m(n) to estimate its time-varying spectrum. Time and frequency smoothing windows were chosen as described in [ 17 ]. The instantaneous power in the low-frequency band, P LF (n), was extracted integrating the SPWVD from 0.04 to 0.15 Hz for each time instant. The instantaneous power in the high-frequency band, P HF (n), was computed in a band centered on the respiratory frequency with a bandwidth of 0.25 Hz. Figure 2shows an example of dHR(n) , P LF (n) and P HF (n) obtained from RR E . In some analyses, mean P LF and P HF were calculated from PLF (n) and PHF (n) for each segment of the test. Sensors 2021,21, 902 6 of 14 Sensors 2021, 21, x FOR PEER REVIEW 6 of 14 𝑑(𝑛), PLF (n) and PHF (n) obtained from RRE. In some analyses, mean PLF and PHF were calculated from PLF (n) and PHF (n) for each segment of the test. Figure 2. Example of 𝑑(𝑛), PLF (n) and PHF (n) obtained from RRE for one subject: Resting segment (left) and cycling segment (right). Note that the axes have different scales. 𝑑(𝑛)= instantaneous HR signal; PLF (n) = Instantaneous low-frequency power; PHF (n) = Instantaneous high-frequency power; RRE= RR intervals series from the ECG. 2.6. Statistical Analysis The normality of data was checked with the Kolmogorov-Smirnov test. Since the data distribution violated the assumption of normality of the parametric tests, and such a condition was not achieved by commonly employed transformations, a non-parametric analysis was performed. Descriptive values are presented as mean ± standard deviation (SD) and HRV values are reported as median and interquartile range. Statistical analyses were performed using IBM SPSS (version 25; Chicago, IL, USA). The significance level was set at p ≤ 0.05. Wilcoxon test for paired samples, the non-parametric equivalent of the paired samples t-test, was used to determine differences between the temporal domain HRV data obtained from PolarH7 and from ECG. The magnitude of the differences was calculated by determining the effect size (ES): 𝐸𝑆 = 𝑍/√𝑛 where Z represents the Z-score for the Wilcoxon statistic and n is the total number of observations [31]. Differences were considered small when ES < 0.2, small to medium when ES = 0.2–0.5, medium to large when ES= 0.5–0.8 and large when ES > 0.8 [32]. Lin’s concordance correlation coefficient (CCC) was used to study the agreement between the following PolarH7-derived and ECG-derived signals: RR (ip), PHF (n) and PLF (n). CCC determines how much the observed data deviate from the perfect concordance line at 45° on a square axis scatter plot [33]. CCC was evaluated in each segment (SREST, SCY60, SCY70, SCY80 and SREC). A CCC value greater than 0.90 was considered “excellent” [34] and the percentage of subjects with CCC values below this threshold was reported for each segment. Figure 2. Example of dHR(n) , P LF (n) and P HF (n) obtained from RR E for one subject: Resting segment ( left ) and cycling segment ( right ). Note that the axes have different scales. dHR(n) = instantaneous HR signal; P LF (n) = Instantaneous low-frequency power; P HF (n) = Instantaneous high-frequency power; RRE= RR intervals series from the ECG. 2.6. Statistical Analysis The normality of data was checked with the Kolmogorov-Smirnov test. Since the data distribution violated the assumption of normality of the parametric tests, and such a condition was not achieved by commonly employed transformations, a non-parametric analysis was performed. Descriptive values are presented as mean ± standard deviation (SD) and HRV values are reported as median and interquartile range. Statistical analyses were performed using IBM SPSS (version 25; Chicago, IL, USA). The significance level was set at p≤0.05. Wilcoxon test for paired samples, the non-parametric equivalent of the paired samples t-test, was used to determine differences between the temporal domain HRV data obtained from PolarH7 and from ECG. The magnitude of the differences was calculated by determining the effect size (ES): ES =Z/√n where Zrepresents the Z-score for the Wilcoxon statistic and nis the total number of observations [ 31 ]. Differences were considered small when ES < 0.2, small to medium when ES = 0.2–0.5, medium to large when ES= 0.5–0.8 and large when ES > 0.8 [32]. Lin’s concordance correlation coefficient (CCC) was used to study the agreement between the following PolarH7-derived and ECG-derived signals: RR (ip), P HF (n) and PLF (n) . CCC determines how much the observed data deviate from the perfect concordance line at 45 ◦ on a square axis scatter plot [ 33 ]. CCC was evaluated in each segment (S REST , S CY60 , S CY70 , S CY80 and S REC ). A CCC value greater than 0.90 was considered “excellent” [ 34 ] and the percentage of subjects with CCC values below this threshold was reported for each segment. Cluster analysis was performed to identify groups of subjects with similar characteristics in terms of the following three variables of interest: age, body composition (trunk fat percentage) and fitness level (PWC 80% ). Trunk fat percentage was selected among all body composition variables, since it is the most specific to the electrode placement area. Sensors 2021,21, 902 7 of 14 Following the methodology described in previous studies [ 35 , 36 ], two types of cluster analyses were combined: hierarchical clustering (Ward’s method) and k-means clustering. First, individual and multivariate outliers (according to Mahalanobis distance) were detected to reduce the sensitivity of the Ward’s method to outliers. Second, hierarchical cluster analysis was used, as the number of clusters in the data were unknown beforehand. Examination of dendrograms showed that a four-cluster solution produced good differentiation between groups. Finally, k-means cluster was performed with four possible solutions. Compared to hierarchical methods, k-means cluster analysis is considered less sensitive to outliers and has been found to result in greater within-cluster homogeneity and between-cluster heterogeneity [35]. To assess differences in the percentage of error for each HRV index between the four cluster groups, a Kruskal-Wallis test (non-parametric equivalent of one-way analysis of variance, ANOVA) with Bonferroni correction was performed. The Dunn-Bonferroni post hoc method was used for pairwise comparisons. The relative error in HRV indices was calculated as the absolute error of the PolarH7 with respect to the ECG measurement divided by the reference ECG measurement, e.g., (SDNNECG −SDNNPolarH7)/SDNNECG , which was then multiplied by 100 to obtain the percentage of error (%Error). In the case of the frequency HRV variables, %Error was calculated from the mean value for each segment of the test. To evaluate the magnitude of the differences, ES was calculated as: ES =H/(n2−1/(n+1)) , where Hstands for the Kruskal-Wallis test statistic and nis the total number of observations [31]. 3. Results Table 2shows the descriptive characteristics of the 4 cluster groups, which were described as CLUSTER A (High PWC 80% ), CLUSTER B (Low PWC 80% and low age), CLUSTER C (Low PWC 80% , high age and medium trunk fat percentage) and CLUSTER D (Low PWC80%, high age and high trunk fat percentage). Table 2. Descriptive characteristics of the four cluster groups. Outcome CLUSTER A (n= 19) CLUSTER B (n= 13) CLUSTER C (n= 18) CLUSTER D (n= 17) Main Effect pEffect Size Age (years) 38.99 ±13.31 B,D 24.35 ±2.20 A,C,D 52.61 ±10.20 B57.47 ±12.23 A,B <0.001 * 0.586 Height (m) 1.77 ±0.05 D1.72 ±0.06 1.74 ±0.06 1.72 ±0.07 A0.020 0.149 Weight (kg) 72.65 ±9.13 69.18 ±12.27 D76.32 ±7.13 82.72 ±8.78 B0.007 0.182 BMI (m/kg2) 23.00 ±2.24 D23.23 ±2.77 D25.17 ±1.40 28.04 ±2.79 A,B <0.001 * 0.430 Body fat (%) 14.09 ±4.11 C,D 15.45 ±4.54 D 20.40 ± 2.36 A,D 27.67 ±2.42 A,B,C <0.001 * 0.743 Trunk fat (%) 14.76 ±5.11 C,D 16.64 ±5.03 D 22.19 ± 2.86 A,D 30.69 ±2.12 A,B,C <0.001 * 0.725 PWC80% (W/kg) 2.73 ±0.39 B,C,D 1.61 ±0.37 A1.79 ±0.28 A,D 1.35 ±0.22 A,C <0.001 * 0.704 Values are expressed as mean ± standard deviation (SD). BMI = Body mass index; PWC 80% = Physical Work Capacity at 80% of HRmax (208 − 0.7 ∗ age in years) in watts/kg bodyweight. Clusters were based on: age, body composition (trunk fat percentage) and fitness level (PWC 80% ). * = Significant differences between clusters (p ≤ 0.05, Kruskal-Wallis test). A = Different to CLUSTER A; B = Different to CLUSTER B; C= Different to CLUSTER C; D= Different to CLUSTER D. Table 3shows the values of HRV indices obtained from PolarH7 and ECG. Mean P LF and P HF were calculated from P LF (n) and P HF (n) for each segment (differently from Table 4 , where the instantaneous series were used). Wilcoxon test for paired samples revealed that P HF and temporal domain HRV indices (MHR, SDNN and RMSSD) were lower at all cycling stages (S CY60 , S CY70 and S CY80 ) when measured by PolarH7, with P LF being lower at the highest intensity (S CY80 ) when measured by PolarH7. The magnitude of all these differences was small to medium, i.e., 0.2–0.5 according to the effect sizes. Sensors 2021,21, 902 8 of 14 Table 3. HRV indices obtained from PolarH7 and ECG data (n= 67). PolarH7 ECG pES SREST PLF (e−4)9.78 (3.88 to 24.74) 9.76 (3.85 to 24.78) 0.074 0.155 PHF (e−4)5.77 (2.30 to 10.91) 5.74 (2.22 to 10.84) 0.067 0.159 MHR (bpm) 62.55 (53.25 to 71.95) 62.84 (53.36 to 72.00) <0.001 * 0.378 SDNN (ms) 60.27 (40.50 to 75.22) 60.28 (40.33 to 75.26) 0.570 0.049 RMSSD (ms) 39.48 (22.48 to 60.04) 39.26 (22.39 to 60.66) 0.336 0.083 SCY60 PLF (e−4)1.17 (0.62 to 2.01) 1.17 (0.65 to 2.23) 0.112 0.137 PHF (e−4)0.51 (0.26 to 1.29) 0.86 (0.44 to 2.07) <0.001 * 0.419 MHR (bpm) 106.00 (100.45 to 112.91) 106.66 (100.75 to 113.17) <0.001 * 0.433 SDNN (ms) 15.94 (12.06 to 20.77) 16.21 (12.85 to 20.32) 0.010 * 0.222 RMSSD (ms) 6.48 (4.34 to 8.51) 8.38 (5.67 to 10.65) <0.001 * 0.482 SCY70 PLF (e−4)0.45 (0.21 to 0.87) 0.46 (0.22 to 0.88) 0.851 0.016 PHF (e−4)0.30 (0.15 to 0.55) 0.41 (0.24 to 0.75) <0.001 * 0.377 MHR (bpm) 124.24 (115.77 to 130.89) 124.32 (115.76 to 130.92) <0.001 * 0.482 SDNN (ms) 10.22 (8.19 to 12.93) 10.48 (8.40 to 13.10) 0.001 * 0.280 RMSSD (ms) 3.74 (2.96 to 4.91) 4.37 (3.64 to 6.68) <0.001 * 0.443 SCY80 PLF (e−4)0.13 (0.09 to 0.23) 0.18 (0.10 to 0.26) <0.001 * 0.364 PHF (e−4)0.23 (0.14 to 0.36) 0.37 (0.25 to 0.66) <0.001 * 0.362 MHR (bpm) 141.01 (130.62 to 148.81) 141.09 (130.69 to 149.05) <0.001 * 0.451 SDNN (ms) 8.11 (6.20 to 9.92) 8.10 (6.34 to 10.65) <0.001 * 0.378 RMSSD (ms) 2.90 (2.32 to 3.90) 3.75 (3.16 to 5.52) <0.001 * 0.405 SREC PLF (e−4)4.75 (1.83 to 10.25) 4.59 (1.88 to 9.97) 0.881 0.015 PHF (e−4)2.06 (0.65 to 4.71) 2.12 (0.82 to 4.70) 0.308 0.102 MHR (bpm) 99.76 (90.76 to 112.15) 98.39 (90.12 to 111.35) 0.002 * 0.268 SDNN (ms) 33.11 (23.27 to 58.30) 33.67 (23.04 to 57.40) 0.094 0.147 RMSSD (ms) 12.87 (7.65 to 23.78) 12.90 (7.99 to 23.32) 0.603 0.046 Values are expressed as median and interquartile range. Segments are based on the test phases: resting (S REST ), cycling (S CY ) and recovery (S REC ). S CY was divided in three stages at 60, 70 and 80% of HRmax, denoted as S CY60 , S CY70 and S CY80, respectively. P LF = low-frequency power; P HF = high-frequency power; MHR = mean HR; SDNN = SD of the NN intervals; RMSSD = root mean square of successive differences between NN intervals. ES = Effect size. * = Significant differences between devices (p ≤ 0.05, Wilcoxon test for paired samples). Table 4. Agreement between devices in: RR (ip), PLF (n) and PHF (n). CCC mean and percentage of subjects not reaching excellent agreement for each segment. SREST SCY60 SCY70 SCY80 SREC Whole sample (n= 67) RR (ip) 0.9929 (1%) 0.9560 (6%) 0.9467 (13%) 0.9319 (16%) 0.9612 (14%) PLF (n) 0.9885 (1%) 0.9713 (4%) 0.9677 (9%) 0.9106 (19%) 0.8251 (30%) PHF (n) 0.9813 (3%) 0.9494 (13%) 0.8858 (27%) 0.6661 (60%) 0.5262 (75%) CLUSTER A (n= 19) RR (ip) 0.9970 (0%) 0.9844 (0%) 0.9472 (21%) 0.9243 (21%) 0.9778 (6%) PLF (n) 0.9999 (0%) 0.9990 (0%) 0.9911 (5%) 0.9169 (16%) 0.7440 (47%) PHF (n) 0.9982 (0%) 0.9645 (16%) 0.9316 (11%) 0.7970 (37%) 0.4859 (88%) CLUSTER B (n= 13) RR (ip) 0.9828 (8%) 0.8996 (15%) 0.8690 (23%) 0.9258 (23%) 0.9473 (15%) PLF (n) 0.9423 (8%) 0.8544 (23%) 0.9757 (8%) 0.8601 (15%) 0.8838 (15%) PHF (n) 0.9284 (8%) 0.8710 (23%) 0.9112 (23%) 0.7455 (62%) 0.7402 (54%) CLUSTER C (n= 18) RR (ip) 0.9943 (0%) 0.9363 (11%) 0.9665 (11%) 0.9035 (22%) 0.9615 (12%) PLF (n) 0.9990 (0%) 0.9998 (0%) 0.9792 (6%) 0.9164 (28%) 0.7791 (31%) PHF (n) 0.9909 (6%) 0.9670 (6%) 0.8906 (28%) 0.5914 (67%) 0.3843 (88%) CLUSTER D (n= 17) RR (ip) 0.9945 (0%) 0.9884 (0%) 0.9844 (0%) 0.9751 (0%) 0.9541 (24%) PLF (n) 0.9999 (0%) 0.9996 (0%) 0.9233 (18%) 0.9361 (18%) 0.9045 (24%) PHF (n) 0.9929 (0%) 0.9739 (12%) 0.8100 (47%) 0.5381 (76%) 0.5366 (65%) Values are expressed as CCC mean and (percentage of subjects under 0.9 threshold). Segments are based on the test phases: resting (S REST ), cycling (S CY ) and recovery (S REC ). S CY was divided in three stages at 60, 70 and 80% of HRmax, denoted as S CY60 , S CY70 and S CY80 respectively. The characteristics of each cluster were the following: CLUSTER A = high fitness; CLUSTER B = low fitness and low age; CLUSTER C = low fitness, high age and medium trunk fat percentage; CLUSTER D = low fitness, high age and high trunk fat percentage. RR (ip) = paired RR interval series; PLF (n) = instantaneous low-frequency power; PHF (n) = instantaneous high-frequency power. Table 4shows CCC values for RR (ip), P LF (n) and P HF (n) and outlines the percentage of subjects not reaching excellent agreement (CCC > 0.90) for each segment of the test. The number of subjects not reaching excellent agreement was clearly higher for P HF (n) than for Sensors 2021,21, 902 9 of 14 P LF (n) ( χ2 (degrees of freedom); χ2 (1) = 45.52; p< 0.001), it increased with exercise intensity ( χ2 (2) = 38.47; p< 0.001) and was lower during exercise than during S REC ( χ2 (1) = 42.31; p< 0.001 ). When performing the analysis separately for each identified cluster, CLUSTER A obtained the highest CCC values, with CLUSTER D being the group with less subjects showing optimal agreement between devices in P HF (n). Due to the presence of noise in the RR P (i) series during S REC , the instantaneous power could not be properly extracted in 4 volunteers and the final sample for SREC was N = 63. Table 5shows %Error for each HRV index. Kruskal-Wallis test demonstrated significant differences between clusters in P HF at S REST and during exercise (S CY70 and S CY80 ). With regards to temporal domain HRV indices, SDNN showed significant differences between groups at S REST and during exercise (S CY60 and S CY80 ) and RMSSD showed significant differences at the highest intensities (S CY70 and S CY80 ). Both for P HF and for temporal domain HRV indices, CLUSTER D was the group with the highest %Error. The magnitude of all these differences was small, i.e., <0.2 according to the effect sizes. Table 5. Percentage of error (%) for each HRV index and comparison between clusters. CLUSTER A (n= 19) CLUSTER B (n= 13) CLUSTER C (n= 18) CLUSTER D (n= 17) Main Effect pEffect Size SREST PLF 0.0 (−0.2 to 0.3) 0.2 (−0.3 to 0.6) 0.1 (−0.2 to 0.2) 0.3 (−0.1 to 0.4) 0.534 0.034 PHF −0.5 (−2.3 to 0.2) −0.2 (−0.7 to 0.4) −0.5 (−1.4 to 0.4) 0.5 (−0.3 to 2.0) 0.049 * 0.121 SDNN 0.0 (−0.2 to 0.5) B−0.1 (−0.8 to 0.0) A,D 0.0 (−0.2 to 0.2) 0.1 (0.0 to 0.2) B0.021 * 0.147 RMSSD −0.2 (−1.0 to 0.3) −0.1 (−2.3 to 0.4) −0.1 (−1.3 to 0.4) 0.2 (−0.2 to 2.3) 0.150 0.081 SCY60 PLF 0.8 (−0.7 to 18.7) 0.0 (−0.8 to 1.3) 0.4 (−4.0 to 3.0) 0.8 (−0.5 to 4.4) 0.444 0.041 PHF 25.7 (10.3 to 48.3) 2.1 (−12.5 to 17.7) 21.7 (−2.4 to 44.6) 27.0 (10.9 to 65.6) 0.164 0.077 SDNN 0.1 (−2.5 to 0.9) D−0.2 (−1.2 to 1.4) 1.2 (−0.1 to 4.2) 2.2 (1.1 to 8.5) A0.018 * 0.153 RMSSD 14.4 (4.6 to 32.4) 1.8 (−3.2 to 14.6) 21.7 (4.2 to 29.8) 15.8 (7.9 to 44.1) 0.266 0.060 SCY70 PLF 1.6 (−0.5 to 7.5) −0.2 (−2.8 to 0.9) −0.4 (−1.7 to 2.1) −0.6 (−7.8 to 2.8) 0.125 0.087 PHF 9.2 (−6.2 to 50.8) −0.9 (−42.6 to 25.6) C,D 25.3 (17.5 to 57.2) B51.8 (14.5 to 71.7) B0.007 * 0.186 SDNN 0.6 (−0.5 to 2.8) 0.2 (−1.4 to 1.6) 1.5 (−1.2 to 7.0) 2.3 (0.3 to 8.0) 0.104 0.093 RMSSD 15.4 (−2.0 to 29.5) 0.6 (−27.9 to 19.3) D15.6 (14.2 to 36.5) 27.1 (14.2 to 56.0) B0.010 * 0.172 SCY80 PLF 1.5 (−2.6 to 32.0) 2.8 (−1.0 to 3.6) 2.7 (−1.1 to 12.6) 5.9 (0.3 to 16.1) 0.596 0.029 PHF 31.6 (−27.4 to 87.7) −9.0 (−113.1 to 56.4) D28.0 (8.5 to 45.2) 44.7 (37.2 to 77.7) B0.047 * 0.121 SDNN 4.0 (−0.8 to 13.7) −0.4 (−2.9 to 6.1) 1.1 (−0.3 to 4.1) 5.7 (1.6 to 14.3) 0.050 * 0.119 RMSSD 19.2 (−9.9 to 67.0) 8.4 (−43.0 to 26.1) D19.0 (4.7 to 32.7) 32.9 (22.1 to 58.2) B0.028 * 0.137 SREC PLF −0.4 (−33.8 to 0.1) 0.5 (−0.5 to 6.2) 0.2 (−17.6 to 3.9) 0.2 (−4.9 to 12.7) 0.161 0.105 PHF −1.6 (−8.2 to 5.8) 1.6 (0.1 to 5.9) −1.2 (−8.3 to 10.1) 9.3 (−0.2 to 37.2) 0.053 0.157 SDNN −0.1 (−0.9 to 0.1) 0.1 (−1.5 to 0.3) −0.1 (−1.0 to 0.4) 0.0 (−1.8 to 0.7) 0.933 0.007 RMSSD −0.9 (−2.3 to 2.0) 0.1 (−9.8 to 3.0) 1.5 (−1.4 to 5.5) 2.2 (−2.9 to 11.5) 0.226 0.068 Percentage error (%) values are expressed as median and interquartile range. The characteristics of each cluster were the following: CLUSTER A = high fitness; CLUSTER B = low fitness and low age; CLUSTER C = low fitness, high age and medium trunk fat percentage; CLUSTER D = low fitness, high age and high trunk fat percentage. Segments are based on the test phases: resting (S REST ), cycling (S CY ) and recovery (S REC ). S CY was divided in three stages at 60, 70 and 80% of HRmax, denoted as S CY60 , S CY70 and S CY80 respectively. PLF = low-frequency powe r; P HF = high-frequency power; SDNN= SD of the RR intervals; RMSSD = root mean square of successive differences between NN intervals. * = Significant differences between clusters (p ≤ 0.05, Kruskal-Wallis test). A = Different to CLUSTER A; B= Different to CLUSTER B; C= Different to CLUSTER C; D= Different to CLUSTER D. 4. Discussion In this study, HRV analysis from RR intervals provided by PolarH7 at rest and during various exercise intensities has been validated against the same analysis from a simultaneous ECG recording. Wilcoxon test showed a large number of significant differences between devices in HRV indices during exercise. However, the effect size was small to medium and of little practical relevance in the case of MHR. When observing RR (ip), P LF (n) and P HF (n) signals, the percentage of subjects not reaching excellent agreement between devices (CCC > 0.90) increased with exercise intensity and was higher for P HF (n) than for P LF (n). Cluster analysis revealed that phenotypic characteristics like age, body composition and fitness level influenced HRV measurements as well as the differences between PolarH7 and ECG. In particular, CLUSTER D, composed of subjects with low fitness level, high age and high trunk fat percentage, was the group with the lowest number