This paper describes our participation in the MultiClinNER subtask of the MultiClinAI shared task, part of the #SMM4H-HeaRD Workshop at ACL 2026. The task requires identifying DISEASE, SYMPTOM, and PROCEDURE mentions in clinical case reports across seven languages: Czech, Dutch, English, Italian, Romanian, Spanish, and Swedish. We compare two BERT-based sequence labeling methods: (i) sentence-level token classification with a fixed train/validation split, and (ii) paragraph-level chunking with 5-fold cross-validation and checkpoint merging, using language-specific BERT models and multilingual XLM-RoBERTa-large as backbones. Our results show that 5-fold training with checkpoint merging consistently outperforms the fixed split strategy, with further analysis suggesting that the gains are primarily driven by improved training-set coverage rather than by differences in input granularity. Language-specific BERT encoders prove most effective for Spanish and English, while XLM-RoBERTa-large yields the strongest results for the remaining five languages through cross-lingual transfer.