Abstract
Background
The six-minute walk test (6MWT) provides a practical measure of submaximal functional exercise capacity in adolescents and reflects the integrated response of cardiovascular, respiratory, and musculoskeletal systems during walking. Traditional age-based linear models inadequately capture nonlinear, maturation-dependent interactions among anthropometric, behavioral, and physiological factors. This study aimed to (i) identify stage-specific predictive contributors to 6MWT performance using supervised machine learning, (ii) improve interpretability through explainable artificial intelligence, and (iii) explore data-driven functional and cardiovascular profiles across peak height velocity (PHV) stages.
Methods
In this cross-sectional school-based study, 3,166 Tunisian adolescents (11–18 years) were categorized as pre-PHV (
n
= 959), around-PHV (
n
= 953), or post-PHV (
n
= 1,254). Submaximal functional exercise capacity was assessed using the 6MWT distance. Random forest regression models were developed globally and stratified by maturation stage. Predictive contributors were interpreted using permutation importance and Shapley Additive Explanations (SHAP). Exploratory k-means clustering based only on 6MWT and cardiovascular response variables was used to identify data-driven profiles within each PHV stage.
Results
The global model explained 40% of the variance in 6MWT distance (R
2
= 0.399; RMSE = 121 m). Predictive performance varied markedly by maturation stage, with low explanatory power pre-PHV (R
2
= 0.039), intermediate performance around-PHV (R
2
= 0.178), and higher predictability post-PHV (R
2
= 0.347). Physical activity was the highest-ranking predictor globally and across all PHV stages, followed by chronological age, particularly around-PHV. SHAP analyses showed progressively larger mean absolute contributions of physical activity across PHV stages, reaching approximately 45 m in post-PHV participants. Although the gap statistic favored a one-cluster solution, exploratory global clustering with k = 2 described a lower-response profile [6MWT distance = 490.4 ± 132.1 m; exercise-induced heart rate response (ΔHR) = 68.4 ± 12.6 bpm] and a higher-response profile (6MWT distance = 599.0 ± 157.1 m; ΔHR = 91.3 ± 13.3 bpm), both of which were represented across all PHV stages.
Conclusions
In this cross-sectional cohort, 6MWT performance showed maturation-stage-specific prediction patterns, but model performance remained modest, particularly before PHV. Physical activity had the strongest model contribution after PHV and may represent an important modifiable correlate, although causal inference cannot be made from these data. The exploratory functional and cardiovascular profiles should be interpreted as data-driven clusters rather than stable biological states. These findings support maturation-sensitive assessment of adolescent 6MWT performance and highlight the need for longitudinal and interventional studies.