Author(s): CHEN, Y., TSAI, S.H., LIN, Y.P., KO, K.J., DONG FANG, C.D., Institution: SOOCHOW UNIVERSITY, Country: TAIWAN, Abstract-ID: 718

INTRODUCTION:
Pacing in 24-h ultramarathons typically shows a progressive decline in speed and increasing variability as fatigue accumulates, but athletes may adopt distinct pacing “phenotypes” that are not fully captured by group averages (1,2). Large-scale endurance datasets have highlighted meaningful pacing profiles (e.g., fast starters vs slow finishers) and their performance implications, supporting data-driven pacing classification approaches (3). This study aimed to identify hourly pacing phenotypes using feature-based clustering across two race years and to examine how pacing types distribute across performance tiers.
METHODS:
Lap-derived timing data from the Soochow 24-h ultramarathon (2024–2025) were aggregated into hourly mean speed and hourly normalized speed (hourly speed divided by each athlete’s mean speed across the 24 h). From 62 athletes, 13 were excluded due to non-completion or prolonged stoppage (>30 min), leaving 49 athletes for analysis. Athletes were stratified into three performance tiers: Gold (n=14), Silver (n=16), and General (n=19), harmonized across years. Pacing features were engineered from the 24-point hourly normalized-speed series, capturing pacing level (e.g., early/mid/late means), decline shape (overall and segment slopes), and stability (e.g., SD across 24 h, range, maximum adjacent-hour change, proportion of hours above baseline). Features were z-standardized and reduced using principal component analysis (PCA). K-means clustering was performed on PCA scores for k=2–6. Model selection used internal validity indices (silhouette, Calinski–Harabasz, Davies–Bouldin) and cluster interpretability.
RESULTS:
Across k=2–6, a three-cluster solution showed the best overall balance of internal validity and interpretability (silhouette≈0.41; Davies–Bouldin≈0.97). The derived pacing phenotypes were: (i) Stable pacing (dominant cluster), (ii) Fast start / steep decline, and (iii) Highly variable / stop-go (small cluster). Tier-stratified distributions showed that Gold athletes were predominantly classified as Stable pacing (13/14), whereas the General tier showed a higher proportion of Fast start / steep decline and Highly variable / stop-go patterns (13/19 and 2/19, respectively). Silver athletes were mainly Stable pacing (13/16) with smaller representation of other phenotypes.
CONCLUSION:
Feature-based clustering of hourly normalized speed identified three interpretable 24-h pacing phenotypes and revealed clear differences in phenotype distribution across performance tiers. The predominance of Stable pacing among higher-tier athletes supports the applied use of pacing phenotypes for athlete profiling and monitoring. Future work should link pacing phenotypes to performance outcomes (e.g., distance, late-race speed retention) and to contextual factors such as direction changes and rest patterns.
[1] Takayama 2016 / [2] Bossi 2017 / [3] Smyth 2018