Clinical event logs are often characterised by very high dimensions, high heterogeneity, and behavioural complexity, making it difficult to apply process mining techniques without a preliminary reduction phase. In this work, we propose and evaluate a stratified sampling strategy that combines trace length with variant frequency, enriched with a refinement based on Direct-Follows Graph (DFG) coverage. The approach was applied to three real-world clinical datasets (COVID-19 Data for Shared Learning, MIMIC-ED, and Sepsis Cases), comparing complete and sampled logs across an extensive set of quantitative, structural, behavioural, and entropic complexity metrics. The results show that the proposed sampling significantly reduces log size (magnitude and support) while maintaining activity variety, trace granularity, and behavioural diversity. The most marked variations are observed in the temporal and quantitative dimensions, highlighting a trade-off between dimensionality reduction and dynamic fidelity. These findings suggest that targeted sampling strategies can improve the scalability of process mining in clinical settings while preserving essential properties of process behaviour.