Professional road cycling is an endurance sport marked by considerable technical and physiological complexity. Cycling stages exhibit substantial variation in factors such as total distance, accumulated elevation gain, average gradient, and terrain composition (1–3). These technical characteristics significantly shape the tactical strategies employed by teams and directly influence the physiological distribution and collective performance dynamics within the peloton (4).

Traditionally, cycling stages have been classified into broad categories such as “flat”, “mountainous”, or “time trial”. However, this conventional classification tends to be subjective and often lacks empirical precision, potentially overlooking relevant technical combinations observed in actual racing conditions (5, 6). Recent studies have highlighted that objective technical variables, such as accumulated elevation, relative elevation per kilometer, and surface composition, critically impact muscular fatigue, sustainable power output, and recovery between consecutive efforts in professional cycling (7, 8). These variables define the “external load”, a central concept in performance physiology that determines the intensity and specificity of physical demands during prolonged competitions (9–11).

Recently, researchers have paid attention to how these technical characteristics affect not only individual performance but also the cohesion and collective durability of group performance during stage races (12). This variability may increase the risk of fatigue-related injuries and highlight the need for better strategic adaptation (13, 14).

In parallel, recent advances in data science and machine learning algorithms have revolutionized methodologies for classifying and analyzing sports data. Clustering methods, in particular, enable empirical classification of sports events based on objective data patterns, offering more precise and reproducible typologies than traditional classifications (15, 16). Beyond cycling, unsupervised learning has been increasingly applied across endurance sports. For example, in running to identify technique-based subgroups and their relation to running economy (17), in collegiate cross-country cohorts using hierarchical clustering to profile mechanics and risk factors (18), and in swimming to partition inertial measurement unit (IMU)-derived functional data into skill-related patterns (19), thereby broadening the methodological context relevant to our approach. In cycling, the application of these methodologies remains limited, though initial studies have demonstrated their potential to generate empirically grounded typologies of cycling stages, thus facilitating improved strategic planning and more effective training load management (20).

Nevertheless, despite the practical relevance of the relationship between objective technical stage classifications and variability in collective performance, this connection has received limited empirical exploration in recent sports literature. The coefficient of variation (CV) of finish times emerges as a key indicator for assessing how specific technical features affect peloton performance homogeneity or dispersion (21, 22).

Therefore, the primary objective of this study is to develop an empirical and objective classification of professional cycling stages using unsupervised learning methods, and to evaluate how these technical categories relate to collective performance dispersion, measured through the CV of finish times. The central hypothesis of this research is that stages with higher technical demands (high elevation, mixed surfaces) are significantly associated with greater collective performance dispersion, reflecting increased physiological and tactical fatigue.

This study provides a robust quantitative framework useful for strategic and physiological planning in professional cycling, directly contributing to the optimization of specific training approaches, the prevention of cumulative fatigue-related injuries, and an improved understanding of how external technical loads influence the internal physiological dynamics of professional cyclists. Practically, this typology helps coaches, sports scientists, and teams tailor stage-type–specific training, pacing/fueling, and roster/equipment choices, while anticipating fatigue to minimize performance decrements and injury risk across multi-stage races.

Figure 1 summarizes the analysis pipeline: data (2017–2023); rule-based QC/preprocessing (completeness, duplicate/neutralized removal, plausibility bounds); outlier screening via the IQR rule with conservative handling of missing entries; z-score standardization (StandardScaler) and collinearity checks (VIF); K-means clustering with optimal k by silhouette and stability by bootstrap (1,000 resamples); small-cluster rule (n < 5, descriptive only); PCA visualization; and statistical analyses interpretation and reporting.

This study builds an objective; empirically derived technical classification of professional road-cycling stages using unsupervised learning and examines its association with variability in collective performance. Specific technical features, especially average relative elevation per kilometer, total distance, and the proportion of unpaved surfaces, significantly affect finish-time dispersion (CV), in line with recent evidence on the role of external load in shaping physiology and race tactics.

Technically demanding groups (clusters 2 and 5), defined by high relative elevation and extensive unpaved sections, consistently exhibited higher CV, indicating greater tactical fragmentation and accumulated fatigue. Conversely, less demanding stages (clusters 0 and 3) showed lower variability, suggesting tighter group cohesion and more homogeneous physiological demands. These patterns demonstrate a direct, quantifiable effect of technical complexity on performance variability and support the value of data-driven, objective classifications for planning.

The classification proved stable under bootstrap cross-validation (average silhouette = 0.62 ± 0.03), underscoring reproducibility and validity. Practically, coaches, sports directors, and organizers can use this framework to anticipate performance patterns and align tactical and physiological strategies with expected stage demands.

A notable temporal finding is a significant global reduction in CV from 2017 to 2023, most marked in the technically demanding clusters (2 and 5). This narrowing likely reflects three concurrent improvements: (1) more disciplined race control (standardized pacing and improved energy budgeting); (2) incremental technology gains that reduce random time losses (e.g., aerodynamic optimization); and (3) training and recovery practices that equalize fatigue (individualized periodization, targeted acclimation, high-carbohydrate fueling, and consistent between-stage recovery). Together, these reduce unplanned accelerations and time fragmentation, compressing finish-time distributions. Although causal claims are not warranted, the trend plausibly aligns with continued professionalization of pacing and fueling in the WorldTour era: on-bike power meters enable tighter real-time intensity control and evidence-based pacing (28, 29), while contemporary carbohydrate strategies, multiple-transportable blends delivering ≥60–90 g·h⁻¹ with gut training, stabilize late-race power and mitigate performance drift, consistent with narrower distributions (30, 31).

Among all predictors, relative elevation per kilometer exerted the strongest influence on CV (β = 0.42; p < 0.001), with an even larger effect in the most technical stages (cluster 5: β = 0.62; p < 0.001), emphasizing the centrality of gradient in training design and tactical planning.

The analysis relies on publicly available secondary data, which may introduce coverage and measurement biases. Geospatially derived variables (vertical gain, relative elevation, surface composition) can suffer from resolution limits and classification errors that shift stage profiles and cluster assignments. Event-specific timing protocols (e.g., neutralizations, timing resolution) may also affect CV estimates. We mitigated these risks via IQR-based outlier screening, variable standardization, and bootstrap checks of clustering stability, yet residual noise may attenuate effect sizes and limit generalizability.

A further limitation is the absence of direct physiological measurements (power output, heart rate, perceived exertion), which constrain mechanistic interpretation of internal responses to external technical loads. Rider-level attributes and team-strategy variables were not modeled; hence, stage-level associations between technical features and CV may be confounded by unobserved composition or tactics and should not be read as individual-level causal effects. Future work should incorporate direct physiological markers—threshold metrics (lactate threshold/ventilatory threshold 2 or critical power) to stratify metabolic intensity; heart-rate variability (e.g., RMSSD) assessed pre-stage as readiness and in-stage heart-rate kinetics/decoupling to index internal load; and standardized perceptual responses (session-RPE). Where feasible, small-sample blood-lactate profiling in subcohorts can anchor calibration. Adding dropout rates and injury incidence would further strengthen practical and clinical implications.

Our dataset spans road events from 2017 to 2023 across men's and women's calendars, including one-day and stage races. Even so, the learned typology and CV associations may not transfer unchanged to contexts that deviate from the observed joint distribution—e.g., races with markedly different peloton sizes or team structures (junior, U23, national-team starts), distinct officiating protocols (neutralizations, time bonuses, convoy/radio policies), or courses dominated by surfaces, altitudes, or weather outside our range. Such factors can inflate or dampen CV independently of our predictors. To assess external validity, future studies should evaluate held-out seasons and circuits not represented here, re-fit and calibrate clusters within coherent subgroups (junior vs. U23 vs. senior; time trials vs. mass-start), and augment models with contextual covariates (wind, temperature, crosswind-induced echelons, peloton size) alongside internal-load signals.

This study provides a robust and objective empirical classification of professional cycling stages using advanced unsupervised learning techniques. Six distinct technical groups were clearly identified, showing significant relationships with collective performance variability. Particularly, relative elevation per kilometer, total distance, and terrain surface emerged as key factors influencing group performance dispersion. A significant reduction in the coefficient of variation of performance was observed between 2017 and 2023, especially in more technically demanding stages, reflecting specific advances in training methods, applied technology, and strategic management in professional cycling.

This objective technical classification offers a practical, quantitative tool directly applicable in real-world professional cycling contexts. Coaches, sports scientists, and team directors can leverage this empirical typology to optimize competitive strategies, tailor specific training loads according to stage types, and prevent risks related to accumulated fatigue and injury. Future research integrating direct physiological measurements and additional variables on injury incidence or dropout rates will enable a deeper understanding of the physiological, tactical, and clinical dimensions of professional cycling performance.