Resultant equations for training load monitoring during a standard microcycle in sub-elite youth football: a principal components approach

Introduction

Training load (TL) monitoring has become a research hot topic in youth football (Impellizzeri et al., 2022; Staunton et al., 2021). This is largely due to the growing access to electronic performance and tracking systems (EPTS) that provides valid TL measures (de Dios-Álvarez et al., 2021; Oliva-Lozano & Muyor, 2022). In recent years, the weekly TL variation has been extensively analyzed in elite and sub-elite football contexts (Teixeira et al., 2022a). Training monitoring has been extensively performed using objective and subjective methods to monitor internal training load (ITL) and external training load (ETL) (Impellizzeri et al., 2022). Global positioning system (GPS) devices have become a customary, low-cost and optimal navigation satellite system to extract valid and reliable ETL outcomes (e.g., distances, sprints, accelerations (ACC), change of directions or body impacts) (Beato et al., 2018; Buchheit et al., 2021). Otherwise, the ITL has been usually monitored by heart rate (HR) and perceived exertion using non-invasive wearable sensor systems, rating perceived exertion (RPE) and total quality recovery (TQR) scales (Haddad et al., 2017; Brink et al., 2010). The research has shown a significant correlation between ETL and ITL in young athletes, however it is still difficult to interpret fitness-recovery status (Impellizzeri et al., 2022). Combining ETL and ITL has been reported as a valid strategy to analyse dose-response dissonances, however the major influencing factor remain to be defined (Bourdon et al., 2017; Teixeira et al., 2021a).

Additionally, the emergent tracking tools appears to have created confusion in dose-response considerations given the data analysis requirement to extract relevant information from large amounts of data (Griffin et al., 2021; Scantlebury et al., 2020). This kind of tracking device can provide big datasets express as a thousand data per second expressed by a large number of variables depending on the time-motion technology used (Rojas-Valverde et al., 2020; Ruan et al., 2022). Otherwise, another challenge has been to standardize the physical and psychophysiological data in meaningful information (Impellizzeri et al., 2022; Staunton et al., 2021; Vanrenterghem et al., 2017). As well, capturing the training frequency, intensity, time/duration, type, volume, and progression (FITT-VP) variables is another critical challenge created by tracking systems (Staunton et al., 2021; Scantlebury et al., 2020). Thus, it is more critical than ever to turning datasets into relevant information for athlete-monitoring cycle (Teixeira et al., 2021a; Weaving et al., 2019). Afterwards, the data-reduction techniques has been applied to explain the dimensionality of the TL variables in different football codes such as futsal (Rico-González et al., 2022a), Australian football (Sheehan et al., 2020), rugby (Scantlebury et al., 2020; Weaving et al., 2020) and Gaelic football (Gamble et al., 2019).

Principal component analysis (PCA) is one of the most used data-reduction techniques to extract redundant information from TL data in football (Rico-González et al., 2022b; Rojas-Valverde et al., 2020). Using a PCA approach, a significant percentage of the total variance in a dataset can be extracted (Warmenhoven et al., 2019). Thus, PCA analysis allows to reduce the complexity in a large group of correlated variables by determining the principal components (O’Donoghue, 2008; Rojas-Valverde et al., 2020). Recently, a systematic review conducted in football reported a 77.1% of explained variance in 12.8 extracted variables out of 51.4 variables distributed over 6.4 principal components (Rojas-Valverde et al., 2020). However, the studies with PCA approaches has focused mainly on TL monitoring in professional and elite youth football (Casamichana et al., 2019; Scantlebury et al., 2020; Sheehan et al., 2020). Until now, PCA approaches were only applied in elite football contexts to simplify the TL having regard to different game formats (Casamichana et al., 2019; Zurutuza et al., 2020), contextual factors (Gonçalves et al., 2019; Oliva-Lozano et al., 2021), competition level (Ricotti et al., 2013), positional role (Moura et al., 2015), tactical behaviour (Ric et al., 2016; Rico-González et al., 2022b) and motor skills (Los Arcos, Mendiguchia & Javier, 2017). Recently, some studies have described the application of TL monitoring strategies during a weekly microcycle in sub-elite youth football, expressing by a low seasonal variation and a high weekly variation (Teixeira et al., 2021b, 2022b). Therefore, it is important to establish the major influencing factor for an accurate training monitoring and manipulation during a standard microcyle. Also, an optical TL monitoring can enhance a proper long-term athlete development, injury prevention and training design (Pino-Ortega et al., 2021; Rico-González et al., 2022c; Rojas-Valverde et al., 2020). More specifically, this can help research, practitioners and coaches to prescribe adequate training intensity over a standard microcycle in youth football (Rico-González et al., 2022a). Therefore it is critical to standardize and reduce the dimensionality of the weekly training data in young football players from sub-elite contexts (Teixeira et al., 2022c; Trecroci et al., 2018). Thus, the aim of this study was to reduce the dimensionality of the internal and external load measures, by a PCA approach, in order to describe and explain the resultant equations for TL monitoring during a standard microcycle in a sub-elite youth football players. Additionally, it is intended to identify the most representative measure for each principal component.

Methods

Quasi-experimental approach

Current research has a prospective, observational and cross-sectional design, by applying an individual TL strategy via GPS technology, HR monitoring system, RPE and TQR scales. Resultant equations for TL monitoring in sub-elite youth football was obtained by a PCA approach. The weekly TL was continuously monitored during 2019–2020 in-season, representing a total of 6-week period from 18 training sessions and 324 observation cases (Teixeira et al., 2021b, 2022d). A minimum of 150 observation cases (i.e., 5 to 10 cases per variable) was assured to perform PCA analysis (Jolliffe & Cadima, 2016). Figure 1 summarizes the procedures for quasi-experimental approach.

Results

Data-reduction procedure, eigenvalue and component number

Figure 2 presents the eigenvalue ranged between 1.44% and 5.21%. Overall, five PCA accounted for 68.6% of the total explained variance. The five extracted PCA explained 27.4%, 16.2%, 10.0%, 8.3% and 6.7% of the variance in TL dataset, respectively. Thus, the first PC explained 27.4% of the TL by TD, HMLD, DSL, ACC and DEC. The second PCA explained 16.2% of the TL thought HSRr and SPR. The thirty PCA explained 10.0% of the TL via HR_max, AvHR, %HR and TRIMP. The fourth PCA explained 8.3% of the baseline outset (i.e., sRPE, TQR, maturation offset and chronological age). The fifth PCA explained 6.7% of the accumulated TL (i.e., AvS and MRS). Constantly, PHV, relative age, experience level and BMI were excluded from the PCA (r < 0.4).

Table 4 also shows the data-reduction procedure resulting from rotated component matrix for accumulated TL variables with factor component loadings (eigenvectors). Four variables were excluded from the PCA due to the communalities below 0.4 (i.e., PHV, relative age, experience level and BMI). Also, KMO’s criteria reported a sampling adequacy of sampled data, reporting a considerable proportion of the variance as result of the underlying factors (KMO = 0.73). Furthermore, significant Barlett Sphericity test was significant (p < 0.001).

Table 4:

Principal component analysis: data reduction procedure using varimax for rotated component matrix with factor loadings (eigenvectors) >0.4.

Variables	PC1	PC2	PC3	PC4	PC5	Uniqueness
TD (m)	0.698					0.365
AvS (m·min−1)					0.680	0.321
MRS (m·s−1)					0.790	0.259
HSRr (m)		0.928				0.041
HMLD (m)	0.788	0.501				0.123
SPR (n)		0.895				0.088
SPR (m)		0.940				0.066
DSL (au)	0.705					0.465
ACC (m·s−2)	0.844					0.233
DEC (m·s−2)	0.877					0.184
HRmax (bpm)			0.763			0.366
HRAv (bpm)			0.967			0.055
%HRmax			0.953			0.081
TRIMP (au)			0.692			0.501
sRPE (au)				−0.516		0.629
TQR (au)				−0.553		0.676
OFFSET (y)				0.669		0.343
Age (y)				0.836		0.261

DOI: 10.7717/peerj.15806/table-4

Note:

Abbreviations: ACC, acceleration; AvHR, average heart rate; AvS, average speed; DEC, deceleration; HMLD, high metabolic load distance; HR_max, maximal heart rate; MRS, maximum running speed; SPR, average sprint distance; SPR_N, number of sprints; sRPE, session ratings of perceived exertion; TD, total distance; TQR, total quality recovery; TRIMP, training impulse.

Resultant equations and paths from principal components analysis

The weightings (eigenvectors) of the PCA analysis are represented by a path graph in Fig. 3. Overall, the weightings ranged between −0.52 to 0.97. The highest weightings were observed in AvHR (bpm) (PCA 3) and the lowest weightings in sRPE (au) (PCA 4). Considering the highest factor in each principal component, the variables considered were TD (0.698), SPR (0.940), AvHR (0.967), Age (0.836) and MRS (0.790) for PCA 1 to PCA 5.

The resultant equations from extracted principal component are presented in Table 5. On this basis, the resultant equations for TL monitoring during a weekly microcycle can be expressed into five principal components determine the equations for the baseline variables: (1) explosiveness and impacts; (2) HSR; (3) HR measures; (4) baseline characteristics; (5) average running velocity.

Table 5:

Resultant equations from extracted principal component analysis.

PCA	Construct	Variables	Calculation
1	Explosiveness, accelerations and impacts	TD (m), HMLD (m), DSL (au), ACC (>3 m·s−2), DEC (<3 m·s−2)	0.698 × TD (m) + 0.788 × HMLD (m) + 0.705 × DSL (au) + 0.844 × ACC (m·s−2) + 0.877 × DEC (m·s−2)
2	High intensity running	rHSR (19.8–25.1 km · h−1), SPR (n), SPR (m)	0.928 × rHSR (km · h−1) + 0.895 × SPR (n) + 0.940 × SPR (m)
3	Heart rate	HRmax (bpm), AvHR (bpm), %HRmax, TRIMP (au)	0.763 × HRmax (bpm) + 0.967 × AvHR (bpm) + 0.953 × %HRmax + 0.692 × AkubatTRIMP (au)
4	Baseline characteristics	TQR (au), sRPE (au), Offset (y), Age (y)	−0.553 × TQR (au) + −0.516 × sRPE (au) + 0.669 × Offset (y) + 0.836 × Age (y)
5	Average running velocity	AvS (m·min−1), MRS (m·s−1)	0.680 × AvS (m · min−1) + 0.790 × MRS (m·s−1)

DOI: 10.7717/peerj.15806/table-5

Note:

Abbreviations: ACC, acceleration; AvHR, average heart rate; AvS, average speed; DEC, deceleration; HMLD, high metabolic load distance; HR_max, maximal heart rate; MRS, maximum running speed; SPR, average sprint distance; SPR_N, number of sprints; sRPE, session ratings of perceived exertion; TD, total distance; TQR, total quality recovery; TRIMP, training impulse.

Discussion

The aim of this study was to reduce the dimensionality of the internal and external load measures, by a PCA approach, in order to describe and explain the resultant equations for TL monitoring during a standard microcycle in a sub-elite youth football players. Additionally, it is intended to identify the most representative measure for each principal component. After data reduction, five principal components were extracted from TL dataset explaining 68.7% of the total variance. The TL measures with the highest weight in each PCA were DEC, SPR distance, average HR, chronological age and MRS.

Resultant equations for TL monitoring during a standard microcycle in sub-elite youth football was split into: (1) explosiveness, ACC and impacts (27.4%); (2) HSR (16.2%); (3) heart bate-based measures (10.0%); (4) baseline characteristics (8.3%); (5) average running velocity (6.7%). Considering the highest representative factor in each principal component, the variables considered were DEC (PCA 1), SPR distance (PCA 2), average HR (PCA 3), chronological age (PCA 4) and MRS (PCA 5). In football, Pino-Ortega et al. (2021) also determined conditional dimensions such as angular velocity, speed displacements, HMLD, HSR, SPR, TD covered, metabolic power, DSL, jumps, impacts, ACC and DEC. The first PCA complies TD, HMLD, DSL, ACC and DEC, being grouped as explosiveness, ACC and impacts. Although there is a definite correlation between body impacts, ACC, and DEC. Otherwise, the TD may be due to an inverse relationship between training volume and intensity (Castillo et al., 2020). Also, the metabolic power was rather than speed-based zones to express running intensity (Osgnach et al., 2010). Nevertheless, the TD could fall outside this construct at first sight. An interaction effect between TD and DEC had already been documented for sub-elite football players (Teixeira et al., 2021b). The second PCA extracted HSRr and SPR, wherefore the HSR is an excellent variable to give meaning about training intensity (Harper et al., 2020). Zurutuza et al. (2020) combined peak velocity and distance covered at different velocities in the same principal component, confirming our results on high intensity demands. The third PCA complied the HR-based measures (i.e., HR_max, AvHR, %HR_max and TRIMP), confirming the correlation between HR-based measures and ETL outcomes (de Dios-Álvarez et al., 2021; Ellis et al., 2021). The fourth PCA was explained by TQR, sRPE, maturation offset and chronological age. Although the fourth PCA has a lower variance explained it is fundamental to consider the influence of chronological age, biological age and perceived exertion (Teixeira et al., 2022a). In line with this component, the perceived exertion seems to be better explained with trainability, maturation and stage of development (Malina et al., 2019). Also, the TL could be influenced by acute: chronic workload ratio, training monotony and well-being variations (Clemente et al., 2021a, 2021b; Rico-González et al., 2022c). Indeed, the literature reported that greater acute: chronic workload ratio and training monotony levels are normally associated with an increased risk of injury or health issues. These measurements might be utilized to comprehend how the data changes throughout in-season phases (Rico-González et al., 2022a). Effectively, perceived exertion in young football players may be also influenced psychophysiological determinants as self-perception of competence and practice experience (Branquinho et al., 2021; Ferraz et al., 2017, 2018). Leading biological maturation in youth sports has become a research-practice gap still lacking knowledge about sub-elite environments using data reduction approaches (Cumming, 2018; Teixeira et al., 2021b, 2022c). Finally, the fifth PCA explained 6.7% of the accumulated TL thought AvS and MRS. Pacing behavior was also reported as a key point to football performance (Ferraz et al., 2018, 2020).

Research findings was slightly small than previous research in futsal (Rico-González et al., 2022a), Australian football (Sheehan et al., 2020), rugby (Scantlebury et al., 2020; Weaving et al., 2020) and Gaelic football (Gamble et al., 2019). Wherefore, the comparisons with current research would consider the differences between football codes. Scantlebury et al. (2020) reported a cumulative explained variance of 91%, 96% and 91% variance in TL in rugby union, field hockey and soccer. Casamichana et al. (2019) reported an explained variance of the external training intensity between 39% and 44%. Also, the eigenvalue of this study ranged between 1.44% to 5.21% by setting up values of accumulated TL substantially lower compared to other studies (i.e., eigenvalues between 1.0% and 68.0%) (Pino-Ortega et al., 2021; Scantlebury et al., 2020). Albeit, current research represents the first time that this statistical approach has been used in a sub-elite youth football, specifically using training data (Rico-González et al., 2022b; Rojas-Valverde et al., 2020).

Current applied PCA determine the resultant equations from individual-based principal components, expressing by major component weightings (Rico-González et al., 2022a; Sheehan et al., 2020; Teixeira et al., 2021a). Indeed, this is the traditional PCA algorithm that computes the principal components based on the covariance matrix or the singular value decomposition the data. It is widely used methods in team sports for dimensionality reduction, data visualization, and feature extraction (Pino-Ortega et al., 2021; Rico-González et al., 2022b; Rojas-Valverde et al., 2020). Other ratios, scores and equivalent equations have already been proposed to measure the TL, by emphasizing training intensity, volume or locomotion profile (Clemente et al., 2019; Owen et al., 2017; Rago et al., 2019). However, the PCA algorithms are diverse and some have not yet been implemented in football (Rico-González et al., 2022b; Rojas-Valverde et al., 2020). Hence, future perspective can explore other PCA algorithms such as incremental, Kernel, sparse and robust PCA approaches (Rojas-Valverde et al., 2020). Incremental PCA allows for incremental updates to the principal components as new data points are added in large datasets or when new data is continuously acquired, such as in real-time monitoring of football players’ performance or training data (Jokiniemi, Pietilä & Mikkonen, 2021). Kernel, sparse and robust PCA has been mainly applied for nonlinear dimensionality reduction, sparsity constraints and noisy or incomplete data (Teixeira et al., 2022c).

Futures research should expand the resultant equations by considering other well-being, technical-tactical and match-related contextual factors. Also, PCA approach must also consider the principal component in TL monitoring when considering training mode (i.e., small-sided and conditioned games), training day (i.e., MD-3, MD-2, and MD-1), age group (i.e., U15, U17, and U19) and maturational bands (i.e., pre-, mid- and post-PHV) (Teixeira et al., 2021a). Additionally, the training data represents only a specific sub-elite football academy and must be considered carefully when applied to another to other teams and contexts. As study limitations, the sample size and number of factors was rather small than previous studies with longer monitoring period (Rojas-Valverde et al., 2020). Also, the total variance was also relatively smaller for this PCA paths than other reports in football codes (Pino-Ortega et al., 2021; Rojas-Valverde et al., 2020). However, it must be ensured that football had the lowest percentage of the variance comparing with other football codes (Rojas-Valverde et al., 2020). Furthermore, choosing a higher threshold for total variance (%) may result in fewer retained principal components and a higher degree of data reduction with a consequent loss, noise or redundant information (Jokiniemi, Pietilä & Mikkonen, 2021; Jolliffe & Cadima, 2016). In general, there is no strict rule for the minimum value for percentage of total variance in PCA, as it depends on the specific application and the trade-off between data reduction and information retention (Rojas-Valverde et al., 2020). Furthermore, a commonly used threshold for retaining a principal component is to choose those components that explain at least 60–80% of the total variance, depending on the specific data analysis requirements (Jokiniemi, Pietilä & Mikkonen, 2021; Jolliffe & Cadima, 2016). Finally, the TL strategies applied in this quasi-experimental approach for only compiles GPS, HR and perceived exertion, however more objective measure of fatigue and recovery should be considered in futures reports, such as HR variability, electromyography signal intensity, biochemical markers and other well-being measures (Clemente et al., 2021a, 2021b). Also, further PCA approaches are needed to consider the principal components when integrating physical, physiological and tactical factors in football under an integrative perspective (Teixeira et al., 2022c).

Practical applications

• Current resultant composite equations can be applied to relative contribution of the ITL and ETL measures for monitoring and management load in sub-elite youth football.

• Data reduction techniques decrease the redundant information and dimensionality of the training data, expressing in the following principal components: explosiveness and impacts, high-speed running, heart bate-based measures, baseline characteristics and average running velocity.

• Considering the highest factor in each principal component, DEC (PCA 1), sprint distance (PCA 2), average HR (PCA 3), chronological age (PCA 4) and maximal speed (PCA 5) are the conditional dimension to be considered in TL monitoring during a standard microcycle in sub-elite youth football players.

• Maturational status should be carefully considered in the TL monitoring together with relative age effect, chronological and baseline characteristics.

• Self-perception and practice experience may affect the variance explained by perceived exertion and pacing behavior.

• Training intensity and volume can be more accurately measured by current resultant composite equations and/or most representative factor for a standard microcycle in sub-elite youth football players.

• Futures research should expand the resultant equations for TL monitoring in sub-elite youth football with well-being, technical-tactical and match-related contextual factors.

Conclusion

Using a PCA approach, five principal components could be applied to extract to describe and explain resultant equations for TL monitoring during an in-season standard microcycle in sub-elite youth football. Current research provides the first composite equations to extract the TL in this specific population expressed as explosiveness and impacts, high-speed running, HR-based measures, baseline characteristics and average running velocity. Considering the highest factor in each principal component, DEC (PCA 1), SPR distance (PCA 2), average HR (PCA 3), chronological age (PCA 4) and maximal SPR (PCA 5) are the conditional dimension to be considered in TL monitoring during a standard microcycle in sub-elite youth football players.

Future research should expand the resultant equations within the microcycle, by considering other well-being measures, technical-tactical factors and match-related contextual factors.

Supplemental Information