Abstract
he short-term scaling exponent alpha1 of detrended fluctuation analysis (DFA-a1) of heart rate variability (HRV) has been shown to be a sensitive marker for assessing global organismic demands. The wide dynamic range within the exercise intensity spectrum and the relationship to es- tablished physiologic threshold boundaries potentially allow in-field use and also open opportunities to provide real-time feedback. The present study expands the idea of using everyday workout data from the AI Endurance app to obtain the relationship between cycling power and DFA-a1. Collected data were imported between September 2021 and August 2023 with an initial pool of 3123 workouts across 21 male users. The aim of this analysis was to further apply a new method of implementing workout group data considering representative values of DFA-a1 segmentation compared to single workout data and including all data points to enhance the validity of the internal-to-external load re- lationship. The present data demonstrate a universal relationship between cycling power and DFA-a1 from everyday workout data that potentially allows accessible and regular tracking of intensity zone demarcation information. The analysis highlights the superior efficacy of the representative-based approach of included data in most cases. Validation
and including all data points to enhance the validity of the internal-to-external load re- lationship. The present data demonstrate a universal relationship between cycling power and DFA-a1 from everyday workout data that potentially allows accessible and regular tracking of intensity zone demarcation information. The analysis highlights the superior efficacy of the representative-based approach of included data in most cases. Validation data of the performance level and the up-to-date relationship are still pending. Keywords:interbeat intervals; RR-intervals; HRV; DFA; monitoring 1. Introduction Knowledge of an athlete’s individual exercise intensity zones is crucial for tailoring effective endurance training plans. To determine physiologic thresholds for exercise inten- sity distribution, various methods have been employed based on different physiological measures, such as gas exchange data, blood lactate concentration, and heart rate (HR) variability (HRV) kinetics [1–7]. The first physiological boundary transition has been de- scribed as the aerobic threshold (AeT) and the second boundary has been referred to as the anaerobic threshold (AnT) to provide the demarcation in a classic three-zone model (e.g., zone 1 as “moderate” intensity:≤AeT, zone 2 as “heavy” intensity: >AeT to≤AnT, zone 3 as “severe” intensity: >AnT). Regularly updating these transitions is essential to avoid training with the wrong intensity targets. The disadvantages of exceeding intensity targets may include glycogen depletion [3], prolonged cardiac parasympathetic recov- ery [8,9], and gastrointestinal barrier disruption [10] along with the potential of overall central and local muscular fatigue [11,12]. However, information about these physiological transitions typically requires laboratory testing and/or specific exercise protocols. In that regard, it could be shown that the short-term scaling exponent alpha1 of de- trended fluctuation analysis (DFA-a1) [13–16] as a non-linear metric of HRV may be a Sensors2024,24, 4468.
Sensors2024,24, 4468 2 of 12 sensitive marker for assessing global organismic demands during acute endurance exercise with a wide dynamic range, encompassing the moderate, heavy, and severe exercise inten- sity domains [17–24]. DFA-a1 quantifies the fractal scale as a qualitative pattern of ANS regulation and represents correlation properties of HR time series in cardiac beat-to-beat intervals with the maintenance of basic stability of the control systems between order (persistence) and disorder (change) in the context of homeodynamics during resting condi- tions with values around 1.0 [25–30]. During exercise, DFA-a1 depicts strongly correlated patterns (values well above 1.0, periodic behavior) at low-intensity exercise, transitions to fractal patterns (value at around and below 1.0) at moderate exercise intensities, and drops to uncorrelated and anticorrelated patterns at the highest intensities in the severe domain (values around and below 0.5, loss of fractal dynamic toward random and anti- correlated behavior). Importantly, this method does not require a maximal exercise test, and data can be derived from a chest belt with a sensor device (e.g., H10 of Polar Electro GmbH (Büttelborn, Germany), Movesense Medical of Movesense Ltd. (Vantaa, Finland)), which, therefore, allows in-field use and also opens opportunities to provide real-time feedback on exercise intensity. Given these properties, a proposal was made based on the signal-theoretical background to utilize this metric as a biomarker for exercise intensity distribution, including discrete numerical values (AeT: DFA-a1 of 0.75, AnT: DFA-a1 of 0.5) that correspond to established physiologic threshold boundaries [24]. The present report tries to expand this idea by using everyday workout data to obtain up-to-date intensity zone demarcation information with the relationship between cycling power and DFA-a1. The lower the DFA-a1, the higher the organismic demands during exercise, and as a measure of global internal load, DFA-a1 is expected to correlate strongly with the external load metric power at comparable internal and external conditions for a given acute and chronic physiological status of the athlete. The aim of this analysis was to further apply a new method of implementing workout group data considering representative values of DFA-a1 segmentation compared to single workout data and including all
DFA-a1 is expected to correlate strongly with the external load metric power at comparable internal and external conditions for a given acute and chronic physiological status of the athlete. The aim of this analysis was to further apply a new method of implementing workout group data considering representative values of DFA-a1 segmentation compared to single workout data and including all data points to enhance the validity of the internal-to-external load relationship. 2. Materials and Methods 2.1. Participants The analysis described here does not rely on laboratory data but instead daily work- out data of cycling athletes collected through the AI Endurance platform app between September 2021 and August 2023. An initial pool of 3123 workouts across 21 users was included. All 21 users are male with an average age of 52.5 years (SD±8.5, range: 34–66). For reference, the average number of workouts for each user is 148.7 (SD±129.1, range: 28–380), spanning an average of 312.3 days (SD±198.7, range: 78–689) and yielding an average of 0.47 workouts per day for each user (SD±0.25, range: 0.13–0.91). Written in- formed consent about data availability and processing was obtained from each participant. All participants have agreed to share their data in AI Endurance for research purposes. All participant’s data were anonymized. Due to the applied standards, the local ethics board of MSH Medical School Hamburg waived the requirement for ethical approval for the current study according to the guidelines for ensuring good scientific practice of MSH Medical School Hamburg. Ethical approval for the used approach of heart rate variability analysis was obtained through the local ethics committee of MSH Medical School Hamburg (refer- ence No. MSH-2022/172). This study was carried out in accordance with the principles set forth in the most recent revisions to the Declaration of Helsinki. 2.2. Study Design The data collection was entirely passive as the participants did not follow a particular set of study instructions during their workouts. Focusing on cycling power and utilizing a diverse range of workout data (including ERG mode, a smart trainer mode that auto- matically controls the resistance such that the athlete has to exercise
the Declaration of Helsinki. 2.2. Study Design The data collection was entirely passive as the participants did not follow a particular set of study instructions during their workouts. Focusing on cycling power and utilizing a diverse range of workout data (including ERG mode, a smart trainer mode that auto- matically controls the resistance such that the athlete has to exercise at a pre-determined power) collected across varying intensities and conditions, a new method to elucidate the
Sensors2024,24, 4468 3 of 12 correlation between power and DFA-a1 was devised. The new method is used to define representatives of each DFA-a1 to power a data cluster via a simple averaging procedure that alleviates the effects of cardiac lag that otherwise creates noise in the DFA-a1 to power relationship. This method has been applied to determine the correlation between using data from single workouts first and then data on workout groups. The latter has been obtained by merging data of all workouts executed within a window of 10 days. Only groups containing at least 4 workouts have been selected. 2.3. Data Analysis Timestamped cycling power and RR interval data were imported into the AI En- durance platform via Garmin or Suunto recording devices. The power and RR interval data were recorded with unknown devices. Data were only used if the RR interval data were of sufficient quality (see section “data quality”). In AI Endurance, RR interval data were controlled and corrected for artifacts (long, short, missed, extra, and ectopic beats) [31–33] and afterward detrended [34] with the “smoothness priors” of Lambda = 500. DFA-a1 was calculated with “short term fluctuations” if 4≤n≤16 beats. For the DFA-a1 calculation, the window width was set to 120 s, with a recalculation grid interval performed every 5 s. The same window width of 120 s was applied to calculate the cycling power (in Watts) as the mean over this window, every 5 s. To analyze the correlation between power and the internal load marker DFA-a1, certain prerequisites must be met for the data under scrutiny. •Data Quality: Artifacts—e.g., missed beats by HR monitors—can significantly affect the DFA-a1 calculation. Hence, the focus of the data analysis was on including work- outs with a maximum of 5% artifacts. From the initial data pool, this criterion narrows down the dataset to 2096 workouts (still 21 users), forming 411 10-day workout groups. •Fatigue Consideration: To mitigate the impact of fatigue, only minutes 5 to 20 of each workout were included in the data analysis. The first 5 min of data were excluded as they are more prone
of 5% artifacts. From the initial data pool, this criterion narrows down the dataset to 2096 workouts (still 21 users), forming 411 10-day workout groups. •Fatigue Consideration: To mitigate the impact of fatigue, only minutes 5 to 20 of each workout were included in the data analysis. The first 5 min of data were excluded as they are more prone to HRV artifacts due to the HR monitor potentially not having built up sufficient moisture from the athlete’s sweat and the potentially suboptimal position of the chest strap that gets adjusted during the first few minutes of exercising. •Data Consistency: Inconsistent data, especially during periods without pedaling (where power equals 0), can skew the results. Indeed, while stopping spinning for a few seconds occasionally is not a problem, it becomes one when it is performed for longer times and/or very frequently. In such a situation, a lot of data points with the same power value of 0 and different values of DFA-a1 would be included, significantly spoiling the correlation. To maintain consistency, workouts with prolonged or frequent periods of no pedaling (exclusion of datasets pedaling < 90% of the dataset) were excluded, and data points with a power of 0 were discarded. • Data Range and Intensity: For a meaningful correlation between power and DFA-a1, a sufficient range of the internal load marker is necessary. Data inclusion focuses on datasets where at least 50% of the data points were within the dynamic range (DFA- a1 < 1.0), which indicates a significant effort level. Data with mostly DFA-a1 > 1.0, typically corresponding to very easy rides, breaks, or resting conditions in healthy individuals, were not processed [22]. The last criterion is the most restrictive, reducing the number of eligible single workout datasets to 554 across 17 users and workout groups to 73 across 11 users. For reference, the average number of workouts for each user is now 32.6 (SD±55.3, range: 1–225), while the list of workout groups for each user is 1, 36, 5, 14, 1, 2, 6, 3, 1, 3, and 1, yielding an average number of groups per user
datasets to 554 across 17 users and workout groups to 73 across 11 users. For reference, the average number of workouts for each user is now 32.6 (SD±55.3, range: 1–225), while the list of workout groups for each user is 1, 36, 5, 14, 1, 2, 6, 3, 1, 3, and 1, yielding an average number of groups per user of 6.6 (SD±10.5). The reduction in the number of workout groups is also influenced by the decision to exclude groups with fewer than four workouts to minimize the influence of individual workouts on the overall group analysis. Even after meticulous data selection, directly correlating power and DFA-a1 using all data points available typically yields weak correlations (for a definition of correlation coefficients, see Section), as illustrated by the blue histograms in Figure. This is because while data points are a list of (P, DFA-a1) pairs collected at the same time, there
Sensors2024,24, 4468 4 of 12 is a physiological lag in the cardiac response to an increase/decrease in power output as the external load measure. (Determining the optimal lag between power time series and DFA-a1 series through cross-correlation presented methodological challenges, leading to ambiguous criteria to pinpoint the “best correlation” and its corresponding lag value. Consequently, specific lag determination methods are not detailed in this study.) This physiological lag means that when an athlete suddenly increases power, there are data points with high power but still high DFA-a1 values, as the latter takes time to catch up. Similarly, during a decrease in power (especially after intense effort), there are points with low power but still low DFA-a1. These inconsistencies lead to poor correlations. An example would be an athlete attempting a short power burst (severe domain) after riding at low exercise intensity (moderate domain) and going back to riding at low intensity (moderate domain) after this effort. The raw (P, DFA-a1) data of this example exhibits data points outside the expected negative correlation.Sensors 2024, 24, x FOR PEER REVIEW 4 of 12 typically corresponding to very easy rides, breaks, or resting conditions in healthy individuals, were not processed [22]. The last criterion is the most restrictive, reducing the number of eligible single workout datasets to 554 across 17 users and workout groups to 73 across 11 users. For reference, the average number of workouts for each user is now 32.6 (SD ± 55.3, range: 1– 225), while the list of workout groups for each user is 1, 36, 5, 14, 1, 2, 6, 3, 1, 3, and 1, yielding an average number of groups per user of 6.6 (SD ± 10.5). The reduction in the number of workout groups is also influenced by the decision to exclude groups with fewer than four workouts to minimize the influence of individual workouts on the overall group analysis. Even after meticulous data selection, directly correlating power and DFA-a1 using all data points available typically yields weak correlations (for a definition of correlation coefficients, see Section 2.4), as illustrated by the blue histograms in Figure
influenced by the decision to exclude groups with fewer than four workouts to minimize the influence of individual workouts on the overall group analysis. Even after meticulous data selection, directly correlating power and DFA-a1 using all data points available typically yields weak correlations (for a definition of correlation coefficients, see Section 2.4), as illustrated by the blue histograms in Figure 1. This is because while data points are a list of (P, DFA-a1) pairs collected at the same time, there is a physiological lag in the cardiac response to an increase/decrease in power output as the external load measure. (Determining the optimal lag between power time series and DFA-a1 series through cross-correlation presented methodological challenges, leading to ambiguous criteria to pinpoint the “best correlation” and its corresponding lag value. Consequently, specific lag determination methods are not detailed in this study.) This physiological lag means that when an athlete suddenly increases power, there are data points with high power but still high DFA-a1 values, as the latter takes time to catch up. Similarly, during a decrease in power (especially after intense effort), there are points with low power but still low DFA-a1. These inconsistencies lead to poor correlations. An example would be an athlete attempting a short power burst (severe domain) after riding at low exercise intensity (moderate domain) and going back to riding at low intensity (moderate domain) after this effort. The raw (P, DFA-a1) data of this example exhibits data points outside the expected negative correlation. Figure 1. Distribution of correlation coefficients (r) obtained in all single workouts (A) and workout groups (B). Blue shows the correlations with the standard approach obtained using all data points, red shows the correlation obtained with the representative method. Since two correlations of representatives were computed for each workout group, the red histogram in (B) refers to the strongest correlation between these two values. A sound approach to the correlation analysis might be used to concentrate on periods where DFA-a1 values demonstrate stability—defined as intervals of 90–120 s with minimal variance in consecutive DFA-a1 readings—and to analyze corresponding power series within these intervals.
representatives were computed for each workout group, the red histogram in (B) refers to the strongest correlation between these two values. A sound approach to the correlation analysis might be used to concentrate on periods where DFA-a1 values demonstrate stability—defined as intervals of 90–120 s with minimal variance in consecutive DFA-a1 readings—and to analyze corresponding power series within these intervals. While theoretically effective under controlled conditions, the practical application of this methodology is significantly limited by the unstructured Figure 1.Distribution of correlation coefficients (r) obtained in all single workouts (A) and workout groups (B). Blue shows the correlations with the standard approach obtained using all data points, red shows the correlation obtained with the representative method. Since two correlations of repre- sentatives were computed for each workout group, the red histogram in (B) refers to the strongest correlation between these two values. A sound approach to the correlation analysis might be used to concentrate on periods where DFA-a1 values demonstrate stability—defined as intervals of 90–120 s with minimal variance in consecutive DFA-a1 readings—and to analyze corresponding power series within these intervals. While theoretically effective under controlled conditions, the practi- cal application of this methodology is significantly limited by the unstructured nature of the data. Among 2096 analyzed workouts, only 101 exhibited the specified stable intervals, with a mere 17 containing more than 1 stable period. This scarcity of suitable data points renders comprehensive statistical analysis unfeasible. The method proposed in this study offers a simpler solution to the aforementioned challenge, preserving the volume of total workout data and effectively uncovering the correlation between power and DFA-a1. This new method (“representative method”) was developed to smooth out problematic data points for the comparison using all data points (“standard approach”). 1. Data Segmentation: In this study, data points (depicted as blue and light blue data points in Figures) were segmented into intervals based on DFA-a1 levels, delineated by yellow (and purple) dashed lines in these figures. The segmentation methodology varied depending on the dataset. For individual workouts, due to a high sparseness of data points outside the dynamic range (DFA-a1 > 1.0),
Segmentation: In this study, data points (depicted as blue and light blue data points in Figures) were segmented into intervals based on DFA-a1 levels, delineated by yellow (and purple) dashed lines in these figures. The segmentation methodology varied depending on the dataset. For individual workouts, due to a high sparseness of data points outside the dynamic range (DFA-a1 > 1.0), the data were divided into eight equal-length intervals within the range [a1min, a1*], where a1minrepresents the minimum DFA-a1 value and a1* is the lesser of DFA-a1 = 1.2 and the maximum DFA-a1 value. For workout groups, two distinct regions were
Description
This study explores the relationship between cycling power and heart rate variability using everyday workout data.