Machine Learning-Based Prediction of Track Sprint Cycling Performance in Elite-Level Male Cyclists: A Multinational Cross-Sectional Analysis

Abstract

Background

Traditional linear approaches may not adequately capture the non-linear relationships and interactions among laboratory-derived sprint test metrics. This study aimed to predict flying sprint performance of elite-level male track cyclists using multiple linear regression and random forest models based on anaerobic ergometer metrics.

Methods

A total of 333 elite male track cyclists completed a 30-s all-out cycle-ergometer sprint test and indoor-velodrome flying 100-m and 200-m sprints. Eight predictors derived from the ergometer test result (body mass, peak power output, 30-s mean power, mean cadence, relative peak power, 5-s maximal mean power normalized to body mass, maximal 5-s power decline, and percentage power drop) were used to develop multiple linear regression and tuned random forest models. Model performance was evaluated on a held-out test set, with random forest hyperparameters optimized via nested cross-validation on the training data. The 100–200-m split time (flying 200-m minus flying 100-m time) was examined in unadjusted and peak power-adjusted models to assess the independent association of power decline metrics with second-half sprint performance.

Results

In mutually adjusted multiple linear regression models, 30-s average power was the only independent predictor of flying 100-m and 200-m times (p < 0.001). Test-set predictive performance was modest for linear regression (R² = 0.326 for 100 m; R² = 0.329 for 200 m) and marginally higher for the tuned random forest models (R² = 0.369 for 100 m; R² = 0.393 for 200 m). Random forest importance analyses consistently ranked relative peak power and 30-s average power as the most influential predictors across both outcomes. Percentage power drop was not independently associated with the 100–200-m split in either unadjusted or peak power-adjusted models.

Conclusions

Sprint performance could be predicted with modest accuracy from routinely collected ergometer metrics, with sustained power and relative peak power emerging as the primary contributors. Power decline metrics showed no independent association with second-half sprint performance after accounting for peak power, suggesting that fatigue resistance, as measured by a 30-s all-out cycle-ergometer sprint test, may not directly translate to competitive second-half sprint outcomes.

Supplementary Information

The online version contains supplementary material available at 10.1186/s40798-026-01058-1.

Key Points

• After mutual adjustment, 30-s average power was the sole independent predictor of flying 100-m and 200-m sprint times of elite male track cyclists.

• The tuned random forest model demonstrated marginally superior test-set predictive performance relative to linear regression, with variable importance analyses consistently identifying relative peak power and 30-s average power as the most influential predictors.

• Percentage power drop was not independently associated with the second-half of the sprint performance after adjustment for peak power, suggesting that ergometer-derived fatigue indices may not directly translate to competitive sprint outcomes.

Supplementary Information

The online version contains supplementary material available at 10.1186/s40798-026-01058-1.

Background

Sprint cycling performance is a complex, multifaceted phenomenon influenced by peak power generation, fatigue resistance, biomechanics, tactical execution, and psychological factors [1–5]. Elite track sprinters must possess exceptional maximal power and the capability to sustain high power outputs through competitive efforts, as peak power in isolation is insufficient for competitive success [2, 3].

Laboratory-based cycle-ergometer sprint testing provides an objective and reproducible means of assessing sprint cycling capabilities, yielding reliable metrics such as peak power output, relative peak power (W·kg⁻¹), and fatigue-related power decline indices, most commonly operationalized as the percentage decline in power output across the sprint effort [7–10]. However, significant discrepancies frequently emerge between laboratory-measured performance parameters and actual competitive outcomes [12–14]. Laboratory testing environments cannot fully replicate the multifaceted conditions of competitive cycling [11], including aerodynamic resistance, tactical decision-making requirements, and dynamic bicycle handling demands [3, 11, 15]. Consequently, athletes demonstrating similar laboratory-measured peak power outputs may achieve substantially different competitive results owing to variations in fatigue resistance profiles and strategic execution capabilities [3].

Traditional statistical analyses, such as linear regression models, often inadequately capture the complex non-linear relationships and threshold effects inherent in sprint cycling performance metrics [11, 16, 17]. Machine learning (ML) methodologies, particularly random forest (RF) regression, offer a promising alternative by effectively modeling these non-linear interactions and complex interdependencies without requiring explicit a priori assumptions about relationship structures [16–22]. ML methods have also demonstrated capacity to identify critical performance thresholds and quantify variable importance in multidimensional athletic performance data [16, 17, 23]. Although previous sprint cycling studies have linked laboratory power metrics to sprint outcomes, large multinational datasets combining standardized 30-s ergometer metrics with velodrome flying sprint outcomes remain limited. The present work extends prior modeling efforts by comparing an interpretable linear framework with a tuned RF approach under a transparent held-out test evaluation and by quantifying predictor importance using permutation and impurity-based measures.

This study aimed to predict flying 100-m and 200-m sprint performance in elite-level male track cyclists by comparing multiple linear regression (MLR) and tuned RF models based on eight laboratory-derived ergometer predictors and to determine whether power decline metrics were independently associated with second-half sprint performance after adjustment for overall sprint power capacity.

Methods

Participants

A total of 333 elite-level male track cyclists from South Korea (n = 128), the United Kingdom (n = 101), Australia (n = 55), and Japan (n = 49), were included in this study (mean ± SD: age, 26.4 ± 4.3 years; height, 178.2 ± 5.8 cm; body mass, 86.5 ± 8.9 kg; competitive cycling experience, 7.4 ± 3.1 years). Eligible participants were active competing track cyclists holding a current UCI or national federation license and engaged in structured sprint training programs, consistent with Tier 3 (Highly Trained/National Level) classification as defined by McKay et al. [24]. Participants typically engaged in approximately 16–18 h of total structured weekly training, encompassing on-track sprint sessions, resistance training, technical drills, and recovery-based activities. The inclusion criteria required the participants to be free from musculoskeletal injuries or cardiovascular conditions during the preceding 6 months. The exclusion criteria included current use of medications known to influence physical performance, recent musculoskeletal injuries or cardiovascular conditions within the preceding 6 months, failure to meet Tier 3 eligibility criteria, and incomplete paired data resulting from a laboratory-to-field assessment interval exceeding 7 days.

Participants were recruited through established high-performance track cycling sprint programs and national federation pathways at each site, and those attending scheduled testing sessions who met the eligibility criteria were included. A participant flow diagram with site-by-site counts is provided in Supplementary Figure S1. Descriptive statistics for all laboratory-derived predictors and field-based sprint outcomes are provided in Supplementary Table S1. As part of routine program procedures, all athletes provided consent in accordance with the procedures of their respective high-performance monitoring programs, and only de-identified records were used for secondary analysis. The secondary analysis protocol was approved by the Institutional Review Board of Chung-Ang University (Approval No. 1041078-20250131-HR-025), and this study was conducted in accordance with the Declaration of Helsinki.

Study Design

This study constitutes a secondary analysis of routinely collected performance testing records integrated into scheduled high-performance monitoring sessions conducted across participating sites. The analytic sample represents a convenience sample of all eligible athletes with complete laboratory and velodrome outcomes during the data collection period (N = 333); the site-specific testing timeline is provided in Supplementary Figure S2. To minimize short-term training fluctuations and environmental variability, each athlete's laboratory and velodrome assessments were scheduled within ≤ 7 days of one another, and all records were de-identified prior to pooling and analysis. Testing was conducted by trained site personnel as per a standardized protocol. At the South Korea site, routine cycle-ergometer sprint testing, and on-site data management were administered by one of the authors with staff support, whereas testing at international sites was performed by trained local coaches and test personnel under the same protocol, with the research team responsible for protocol standardization, quality control, and centralized data curation.

Each cyclist completed a 30-s maximal sprint test on a Wattbike ergometer (Wattbike Pro; Wattbike Ltd., Nottingham, UK), from which eight predictors were derived: body mass (kg), peak power output (W), mean power output over 30-s (W), mean cadence (rpm), relative peak power output (W·kg⁻¹), 5-s maximal mean power normalized to body mass (W·kg⁻¹), maximal decline in 5-s power output (W), and percentage power drop (%). Field sprint performance was assessed using a flying 200-m sprint with a 100-m intermediate split on a 250-m indoor velodrome track, recorded using an electronic timer.

Experimental Procedures

Laboratory Measurement

Laboratory sprint testing was conducted at the Graduate School of Sports Medicine, CHA University (Seongnam-si, Republic of Korea), Manchester Velodrome (Manchester, United Kingdom), and Joe Ciavola Velodrome (Melbourne, Australia). The participants underwent anthropometric assessments to measure muscle mass, body fat mass, and total body water, using a body composition analyzer (InBody 270, Biospace, Seoul, Republic of Korea). All participants completed a standardized cycle-ergometer sprint testing protocol under consistent conditions. Prior to testing, all ergometers were calibrated using the manufacturer-recommended zero-offset procedure to ensure comparable power measurements across sites. Upon arrival, each cyclist performed a 15-min warm-up comprising low-intensity pedaling (60–70 rpm) on the cycle-ergometer and dynamic stretching exercises targeting the major leg muscles (quadriceps, hamstrings, glutes, and calves) to minimize injury risks. After the warm-up, the participants rested passively in a seated position for 5 min to ensure physiological stabilization. The cycle ergometer was individually adjusted according to each participant’s anthropometric characteristics, including saddle height, handlebar position, leg length, arm reach, and typical cycling posture. The air resistance was uniformly set to level 10 for all tests. The participants then completed a 30-s maximal sprint test on a Wattbike ergometer (Wattbike Pro; Wattbike Ltd., Nottingham, UK) from a stationary standing start position; they were instructed to accelerate explosively from the outset and maintain maximum effort throughout, and strong verbal encouragement was provided. Peak power (W), 30-s mean power (W), mean cadence (rpm), relative peak power (W·kg⁻¹), and percentage power drop (%) were continuously recorded; 5-s maximal mean power normalized to body mass (W·kg⁻¹) and the maximal decline in 5-s power output (W) were additionally derived from the device-recorded power trace for subsequent analyses. Percentage power drop, defined as the relative decline from peak to minimum power output across the 30-s effort, was treated as a fatigue-related power decline index throughout the analysis. The participants adopted their preferred sprinting posture to closely replicate actual race conditions. The detailed setup for cycle-ergometer sprint testing is illustrated in Fig. 1A.

Fig. 1.

Experimental setup for laboratory and field cycling performance measurements. A Laboratory assessment consisted of a maximal 30-s sprint test performed on a calibrated Wattbike Pro ergometer (Wattbike Pro; Wattbike Ltd., Nottingham, UK) under controlled environmental conditions (21 °C, 45% relative humidity). The air resistance setting was standardized at level 10 for all participants, who initiated the sprint from a stationary standing start position. Participants were permitted to sprint in a standing position throughout the effort to enhance ecological validity, because track sprint cyclists typically generate maximal power during standing accelerations; this posture was standardized across all participants. B Field assessment comprised a flying 200-m sprint test conducted on a standardized 250-m indoor velodrome. Prior to entering the timed section, participants completed a 750–800-m self-paced acceleration phase to attain maximal velocity, thus eliminating the influence of acceleration capabilities on the measured sprint performance. Sprint times for the flying 100-m and 200-m intervals were recorded using a high-precision electronic timing system

Field Measurement (Track Cycling Performance)

Field-based sprint performance data were collected at three standardized indoor 250-m velodromes: the Manchester Velodrome (Manchester, United Kingdom), Joe Ciavola Velodrome (Melbourne, Australia), and Yeongju Cycling Training Center (Gyeongbuk, Republic of Korea). All tests were conducted in a controlled, trial-like environment rather than during official competition to minimize external variability. All tests were conducted under the supervision of experienced cycling coaches and official timing personnel, with the research team overseeing the consistency and accuracy of the data collection. Prior to entering the timed 200-m section, riders independently performed a self-paced acceleration over approximately 750–800 m with no motor pacing or external assistance permitted. A standardized starting protocol was enforced under the supervision of experienced coaches and timing officials to ensure procedural consistency across all participants and sites. The participants performed flying 200-m sprint trials, involving an acceleration phase of approximately 750–800 m, to achieve maximal velocity before entering the official timing zone. Timing commenced automatically as participants passed the designated 200-m start line. From this point, the initial 100-m interval (from the start line to the 100-m mark) and the total 200-m interval times were precisely measured using a high-precision electronic timing system (MYLAPS ProChip, MYLAPS Sports Timing, Nijmegen, The Netherlands). All sprint trials followed standardized instructions and protocols, with the coaches ensuring measurement accuracy and participant safety throughout each test. Each participant performed a single maximal trial for laboratory and field tests. The detailed setup for field-based sprint testing is illustrated in Fig. 1B.

Statistical and Machine Learning Analysis

Data are presented as mean ± SD unless otherwise stated, and 95% confidence intervals (CIs) are reported where appropriate. All analyses were conducted in Python (v3.13.3) using numpy, pandas, scikit-learn, and stats models, with statistical significance set at p < 0.05 for inferential tests.

MLR models were fitted on the training set to estimate mutually adjusted associations between predictors and sprint outcomes. Model diagnostics included assessment of linearity, homoscedasticity, and multicollinearity using the variance inflation factor (VIF). In parallel, RF regression was implemented to capture potential non-linear associations and quantify predictor importance. A fixed train–test split was applied (training n = 266; test n = 67), with RF hyperparameters tuned on the training set via randomized search within a nested cross-validation framework; final hyperparameter configurations and cross-validation summaries are provided in Supplementary Tables S2, S3, and S4. Generalization performance for both models was evaluated on the held-out test set using R², RMSE, and MAE, with median absolute error (MedAE) and Pearson r reported as complementary indices. Predictor importance was quantified using impurity-based (mean decrease in impurity) and permutation-based methods to provide complementary perspectives and mitigate the known limitations of impurity-based ranking alone.

As an exploratory analysis, second-half sprint performance was defined as the 100–200-m split time (flying 200-m minus flying 100-m time). The association between percentage power drop and split time was examined using bivariate linear regression and a peak power-adjusted model to account for potential confounding by overall sprint power capacity. Data-quality checks, outlier screening, and distributional diagnostics are reported in Supplementary Table S5.

Results

Multiple Linear Regression Analysis

The results obtained from the mutually adjusted MLR models are presented in Table 1 (Flying 100-m) and Table 2 (Flying 200-m). After mutual adjustment across eight pre-defined predictors, 30-s average power was the only predictor independently associated with sprint time for both outcomes (p < 0.001), with higher average power associated with faster sprint times. Predictive performance on the held-out test set was modest for both outcomes (Flying 100-m: R² = 0.326, RMSE = 0.136 s, MAE = 0.107 s; Flying 200-m: R² = 0.329, RMSE = 0.286 s, MAE = 0.223 s). VIF values ranged from 1.37 to 158.71 across predictors (Tables 1, 2), indicating substantial multicollinearity among power-derived metrics. Standardized regression coefficients for all predictors are illustrated in Fig. 2.

Table 1.

Multiple linear regression coefficients for 100-m flying sprint time

Predictor	β	SE	t	p	95% CI	VIF	Standardized β
Intercept	6.558	0.780	8.41	< 0.001	(5.023, 8.093)	–	–
Body mass (kg)	− 6.68 × 10⁻4	0.0084	− 0.08	0.937	(− 0.017, 0.016)	79.28	− 0.035
Peak power output (W)	2.64 × 10⁻4	4.20 × 10⁻4	0.63	0.530	(− 5.64 × 10⁻4, 0.0011)	158.71	0.388
30-s average power (W)	− 7.69 × 10⁻4	1.49 × 10⁻4	− 5.15	< 0.001	(− 0.0011, − 4.75 × 10⁻4)	4.51	− 0.535
Average cadence (rpm)	3.93 × 10⁻4	0.0013	0.31	0.755	(− 0.0021, 0.0029)	1.37	0.018
Relative peak power output (W·kg⁻1)	− 0.036	0.036	− 0.98	0.328	(− 0.107, 0.036)	119.94	− 0.525
5-s maximal mean power (W·kg⁻1)	− 0.0040	0.014	− 0.28	0.780	(− 0.032, 0.024)	18.82	− 0.059
Max–min 5-s power drop (W)	5.18 × 10⁻5	2.62 × 10⁻4	0.20	0.843	(− 4.64 × 10⁻4, 5.67 × 10⁻4)	55.12	0.072
Percentage power drop (%)	− 0.0013	0.0034	− 0.37	0.713	(− 0.0080, 0.0055)	16.38	− 0.073

β regression coefficient, SE standard error, CI confidence interval, VIF variance inflation factor

VIF and standardized β are not applicable to the intercept

Table 2.

Multiple linear regression coefficients for 200-m flying sprint time

Predictor	β	SE	t	p	95% CI	VIF	Standardized β
Intercept	12.836	1.607	7.99	< 0.001	(9.672, 16.000)	–	–
Body mass (kg)	0.0016	0.017	0.09	0.925	(− 0.033, 0.036)	79.28	0.041
Peak power output (W)	3.78 × 10⁻4	8.67 × 10⁻4	0.44	0.663	(− 0.0013, 0.0021)	158.71	0.265
30-s average power (W)	0.0017	3.08 × 10⁻4	− 5.43	< 0.001	(− 0.0023, − 0.0011)	4.51	− 0.556
Average cadence (rpm)	0.0033	0.0026	1.28	0.203	(− 0.0018, 0.0084)	1.37	0.072
Relative peak power output (W·kg⁻1)	− 0.062	0.075	− 0.83	0.407	(− 0.210, 0.086)	119.94	− 0.438
5-s maximal mean power (W·kg⁻1)	− 0.0080	0.030	− 0.27	0.788	(− 0.067, 0.051)	18.82	− 0.056
Max–min 5-s power drop (W)	6.24 × 10⁻5	5.40 × 10⁻4	0.12	0.908	(− 0.0010, 0.0011)	55.12	0.041
Percentage power drop (%)	− 8.76 × 10⁻4	0.0070	− 0.12	0.901	(− 0.015, 0.013)	16.38	− 0.024

β, regression coefficient, SE standard error, CI confidence interval, VIF variance inflation factor

VIF and standardized β are not applicable to the intercept

Fig. 2.

Standardized regression coefficients (β) from mutually adjusted multiple linear regression models predicting flying 100-m (left panel) and 200-m (right panel) sprint times. All eight laboratory-derived predictors are displayed. Negative coefficients indicate that higher values of a given predictor were associated with shorter (faster) sprint times, whereas positive coefficients indicate association with longer (slower) times. 30-s average power and relative peak power (W·kg⁻¹) yielded the largest negative coefficients for both outcomes, consistent with their role as primary determinants of sprint performance. Percentage power drop exhibited a negative coefficient, and peak power output a positive coefficient, reflecting the influence of multicollinearity among correlated predictors in the mutually adjusted model

Random Forest Regression Analysis

The tuned RF models demonstrated marginally superior generalization performance relative to MLR for both outcomes (Flying 100-m: R² = 0.369, RMSE = 0.132 s, MAE = 0.102 s; Flying 200-m: R² = 0.393, RMSE = 0.272 s, MAE = 0.214 s; Table 3). Robustness during hyperparameter selection was assessed via nested cross-validation, with outer-fold performance reported as a conservative estimate of generalization across the resamples (Flying 100-m: outer R² = 0.295 ± 0.200; Flying 200-m: outer R² = 0.301 ± 0.181); detailed results are provided in Supplementary Table S4. Final tuned hyperparameter configurations are reported in Supplementary Tables S2 and S3, and a baseline RF model using default hyperparameters is documented in Supplementary Table S6.

Table 3.

Model performance comparison between multiple linear regression and tuned random forest models

Outcome	Model	R2	RMSE (s)	MAE (s)	MedAE (s)	Pearson r
100-m flying	MLR	0.326	0.136	0.107	0.089	0.572
100-m flying	RF (tuned)	0.369	0.132	0.102	0.096	0.609
200-m flying	MLR	0.329	0.286	0.223	0.197	0.574
200-m flying	RF (tuned)	0.393	0.272	0.214	0.176	0.630

R² coefficient of determination, RMSE root mean square error, MAE mean absolute error, MedAE median absolute error

r, Pearson correlation coefficient (reported as complementary indices); test-set metrics were computed on the held-out test split

Predictor importance was assessed using impurity-based importance (mean decrease in impurity, MDI) and permutation importance to provide complementary perspectives. Across both outcomes, relative peak power output (W·kg⁻¹) and 30-s average power (W) were consistently the two most influential predictors, followed by 5-s maximal mean power normalized to body mass (W·kg⁻¹) and peak power output (W) (Table 4; Fig. 3). Importance estimates from both methods were broadly concordant, supporting the robustness of the predictor ranking.

Table 4.

Predictor importance in tuned random forest models for 100-m and 200-m flying sprint times

Predictor	100-m permutation importance (%)	100-m MDI (%)	200-m permutation importance (%)	200-m MDI (%)
Relative peak power output (W·kg⁻1)	35.241	27.313	35.171	27.566
30-s average power (W)	32.661	23.299	35.735	26.807
5-s maximal mean power (W·kg⁻1)	17.028	14.577	17.317	14.139
Peak power output (W)	4.941	14.144	4.871	14.199
Body mass (kg)	4.705	7.007	3.265	6.204
Max–min 5-s power drop (W)	2.303	5.210	2.074	4.971
Percentage power drop (%)	1.842	4.202	1.626	4.183
Average cadence (rpm)	1.307	4.247	0.741	1.931

MDI mean decrease in impurity; permutation importance values are normalized to sum to 100% within each outcome

Fig. 3.

Permutation importance shares (%) from tuned random forest models predicting flying 100-m and 200-m sprint times. Each bar represents the relative contribution of a predictor to out-of-sample predictive accuracy, normalized to sum to 100% within each outcome. Relative peak power (W·kg⁻¹) and 30-s average power (W) were the two dominant predictors for both outcomes, with combined contributions of approximately 68% (100-m: 35.2% and 32.7%, respectively) and 71% (200-m: 35.2% and 35.7%, respectively). 5-s maximal mean power normalized to body mass (W·kg⁻¹) ranked third for both outcomes (100-m: 17.0%; 200-m: 17.3%). The remaining five predictors each contributed less than 5% to total permutation importance

Power Drop (%) and Second-Half Sprint Performance

In the exploratory analysis of second-half sprint performance, percentage power drop was not associated with the 200-m minus 100-m split time in the bivariate model (β = − 0.00144 s·%⁻¹, p = 0.227). This association remained non-significant after adjustment for peak power (β = 0.000876 s·%⁻¹, p = 0.426), whereas peak power itself showed a strong independent association with second-half performance (β = − 0.000361 s·W⁻¹, p < 0.001). Full results are reported in Supplementary Table S7.

Discussion

The principal finding of this study is that, in mutually adjusted linear regression models incorporating eight correlated sprint test predictors, 30-s average power was the only independent predictor of flying 100-m and 200-m sprint performance. Within the predictive modeling framework, tuned RF models demonstrated a modest improvement in held-out test-set performance relative to linear regression, and consistently identified relative peak power and 30-s average power as the most influential predictors across both outcomes. The substantial multicollinearity observed among power-derived predictors, reflected by elevated VIFs and attributable to the inclusion of mathematically and physiologically related variables such as absolute and body mass-normalized power indices, warrants cautious interpretation of individual regression coefficients. Under these conditions, 30-s average power is best understood as an integrative marker capturing maximal power capacity and the ability to sustain high power output across a short all-out effort, which likely accounts for its robustness as the sole independent correlate of sprint performance after mutual adjustment.

The machine learning procedure was structured to limit overfitting and assess generalizability under a small sample with highly inter-related predictors, following the principles outlined by Vos et al. [29] for small-sample cycling performance prediction studies. Accordingly, a fixed train–test split was used for final test-set performance, RF hyperparameters were tuned via nested cross-validation, baseline and tuned RF models were compared to verify overfitting, and both impurity-based and permutation importance were reported. The RF results should therefore be interpreted as reflecting a conservative model prioritizing generalizability over training data fit.

Other power metrics, including peak power and relative power, did not retain independent statistical significance after mutual adjustment, which is attributable to the substantial shared variance among mathematically and physiologically related power indices. In the exploratory split-time analysis, percentage power drop was not associated with second-half sprint performance in the bivariate model and remained non-significant after adjustment for peak power, whereas peak power showed a strong independent association. This pattern suggests that power decline metrics derived from a 30-s cycle-ergometer sprint may not directly translate to performance within the 100–200-m segment of a flying sprint once overall sprint power capacity is accounted for, and that any apparent unadjusted association may reflect confounding by overall sprint power capacity. The application of RF regression modestly improved predictive performance relative to linear regression (Flying 100-m: R² = 0.369 vs. 0.326; Flying 200-m: R² = 0.393 vs. 0.329), suggesting that accounting for non-linear interactions may yield incremental gains in generalization performance. Collectively, these results indicate that sustained sprint power capacity and relative power characteristics are primary contributors to short-distance sprint cycling success, although a substantial proportion of performance variance remains unexplained by ergometer-derived metrics alone.

These findings reinforce the critical role of peak anaerobic power and the capacity to maintain power output in determining sprint performance [3, 6]. For the shorter 100-m sprint (typically lasting 5–6 s in elite track cycling), performance relies predominantly on explosive peak power output. Despite the relatively short duration of the flying 100-m, 30-s average power emerged as the most robust predictor in the mutually adjusted models, suggesting that sustained power capacity assessed over 30 s may serve as an integrative marker reflecting multiple sprint-relevant qualities, including maximal power output, early acceleration capability, and resistance to within-effort power decline. This interpretation aligns with the established literature, indicating that power output and sprint-specific endurance are key determinants of sprint performance [1–4, 6, 25]. Of note, in the presence of highly correlated power indices, mutually adjusted models may preferentially retain predictors that best summarize shared sprint capacity rather than implying that peak power is of lesser physiological importance. Our results are consistent with those of Smirniotou et al. [26], who reported that lower-body strength and power measures explained approximately 46% of the variance in 100-m dash performance, comparable to the 0.33–0.39 test-set R² achieved by our models. This convergence of evidence underscores that a substantial proportion of sprint performance can be attributed to power-generating capabilities, whereas a meaningful component remains attributable to technical, biomechanical, and environmental factors not captured by laboratory testing.

The modest test-set R² (0.33–0.39) should be viewed in light of the predictor set being restricted to 30-s Wattbike-derived variables. More comprehensive predictor sets yield higher explanatory power: Van der Zwaard et al. [28] explained normalized sprint performance up to R² = 0.65 using muscle fiber and morphological variables, and Vos et al. [29] predicted 4-km time-trial performance with R²≈0.79–0.88 using physiological, training load, and body composition predictors. Given that these values were derived from different disciplines, predictor sets, or populations, they are not directly comparable with the present elite track sprint predictions, but they indicate that incorporating aerodynamic, technical, biomechanical, and neuromuscular variables may improve future models.

Although body mass was not an independent predictor in the mutually adjusted models, relative peak power (W·kg⁻¹) was consistently ranked among the most influential predictors in the RF importance analyses, supporting the practical relevance of power-to-mass characteristics in sprint cycling performance [3, 9, 27]. In practical terms, two cyclists with identical absolute peak power will not perform equally if one carries additional body mass—the lighter athlete (or one with the higher W·kg⁻¹) can achieve superior acceleration and speed, with all other factors being equal. This implies that improvements in power production should ideally be accompanied by appropriate weight management strategies to maximize competitive performance.

RF regression consistently outperformed the linear model in predicting sprint times, albeit modestly, implying the presence of non-linear relationships and interactions among the predictors [5, 16, 17]. The improvements in held-out test-set R² were modest (approximately 0.04–0.06), indicating that although the RF approach added incremental predictive value, a substantial proportion of performance variance remained unexplained. Nevertheless, even marginal gains in predictive accuracy may carry practical significance at the elite level, where competitive margins are often small. Because individual MLR coefficients can be unstable under conditions of substantial multicollinearity, RF-based importance rankings, particularly permutation importance, provide a complementary perspective that is less dependent on coefficient estimation and therefore more robust under collinearity. In applied sprint cycling settings, such interpretable machine learning outputs may prove useful for prioritizing training targets, including relative peak power and sustained power capacity, even in contexts where gains in predictive accuracy remain modest [5, 16, 17]. Given that impurity-based importance measures are susceptible to bias, impurity-based and permutation-based importance were reported; the two approaches were broadly concordant in prioritizing relative peak power and 30-s average power across both outcomes.

The nested cross-validation showed relatively large variability in outer-fold R². This is largely attributable to the small size of each outer fold (n ≈ 26–27), a well-recognized source of unstable performance estimates in small-sample prediction, together with the narrow performance range of elite cyclists and absence of field-specific factors (e.g., aerodynamics, gearing, technical skill) from the predictor set. The variability therefore reflects a conservative estimate of generalization performance rather than model failure and cautions against over-interpreting single test-set R² values [29, 30].

This study has some limitations. First, our analysis was limited to anaerobic metrics derived from a standardized 30-s all-out cycle-ergometer sprint test, whereas technical, aerodynamic, and psychological factors that may influence competitive performance were excluded. Second, because only male athletes were included, the generalizability of these findings to female track sprint cyclists remains limited; future studies should prospectively validate and, where appropriate, recalibrate the proposed models in female cohorts. Third, the cross-sectional design of this study restricts causal inference capabilities, highlighting the need for longitudinal investigations. Finally, the predictor set comprised mathematically and physiologically related power indices resulting in substantial multicollinearity; individual regression coefficients should therefore be interpreted cautiously as mutually adjusted associations rather than isolated causal effects.

Practical Implications

The findings carry several practical implications for elite sprint cycling. Coaches should prioritize sustained sprint power capacity, because 30-s average power was the most robust independent predictor across both sprint distances. Relative peak power (W·kg⁻¹) was consistent among the most influential predictors, underscoring the importance of developing explosive power relative to body mass alongside appropriate body composition strategies. Percentage power drop should be interpreted cautiously as a standalone fatigue indicator, because it was not independently associated with second-half sprint performance after adjustment for peak power. Finally, the modeling framework presented here may serve as an objective tool for tracking longitudinal changes in key power metrics and supporting individualized training prioritization.

Conclusions

In this multinational cohort of elite-level male track cyclists, sprint performance was predicted with modest accuracy from routinely collected 30-s ergometer metrics. After mutual adjustment, 30-s average power was the only factor that was independently correlated with flying 100-m and 200-m times, whereas RF models consistently identified relative peak power and 30-s average power as the most influential predictors. Percentage power drop was not independently associated with second-half sprint performance after adjustment for peak power. Future studies should externally validate these findings in female cohorts and incorporate additional technical and biomechanical variables to improve predictive performance.

Abbreviations

CI

Confidence interval

MAE

Mean absolute error

MedAE

Median absolute error

ML

Machine learning

MLR

Multiple linear regression

RF

Random forest

RMSE

Root mean square error

VIF

Variance inflation factor

Author Contributions

TK performed data collection across all participating sites, including direct measurement administration at the South Korea site, and contributed to the statistical and machine learning analyses. TK also prepared the original draft of the manuscript. SBP conceptualized and designed the study, led the statistical and machine learning analyses, contributed to the interpretation of results, critically revised the manuscript, and approved the final version for submission. All authors read and approved the final manuscript.

Funding

This research received no external funding.

Data Availability

The datasets analyzed during the current study are not publicly available due to confidentiality restrictions governing the use of de-identified performance monitoring records obtained through national federation high-performance programs, as stipulated by the terms of the institutional ethical approval (Chung-Ang University IRB Approval No. 1041078-20250131-HR-025). Data may be made available to qualified researchers upon reasonable request to the corresponding author, subject to applicable privacy regulations and institutional data governance requirements. The datasets generated and analyzed in the current study are not publicly available because of confidentiality restrictions regarding participant privacy and ethical approval. However, these data can be made available upon reasonable request by contacting the corresponding author.

Declarations

Ethics Approval and Consent to Participate

This study was conducted in accordance with the Declaration of Helsinki. Ethical approval was granted by the Institutional Review Board of Chung-Ang University (Approval No. 1041078-20250131-HR-025) for the secondary analysis of routinely collected, de-identified athlete-monitoring data. All athletes provided consent in accordance with the procedures of their respective high-performance monitoring programs prior to data collection.

Consent for Publication

Not applicable. This manuscript does not contain data from any individual person.

Competing Interests

Taenam Kim and Seung-Bo Park declare that they have no competing interests.