Clinical notes are essential to documenting patient care from initial evaluations in admission notes to care outcomes in discharge summaries. These documents often contain incomplete or inconsistent information, affecting care continuity and clinical decision-making. Analyzing their structure using Natural Language Processing may reveal patterns to improve documentation quality. This study demonstrates that Transformer-based models can semantically distinguish between admission and discharge notes written in Spanish, focusing on patients with Pulmonary Thromboembolism and Pneumonia. Regardless of whether a general-domain (XLM-RoBERTa) or domain-specific (RoBERTa-biomedical-clinical-es) model is used, contextual embeddings clustered in an unsupervised manner consistently form two distinct groups (optimal performance at k = 2). Notably, both models identified the same two clusters, as confirmed by a perfect Adjusted Rand Index (ARI = 1), highlighting the robustness and semantic consistency of the separation. These findings provide empirical evidence that Transformer-based models can detect differences in clinical discourse, even in low-resource settings, offering a promising tool for analyzing and standardizing medical documentation in Spanish. Limitations include the absence of sensitivity analyses and limited generalizability due to sample size and institutional biases. Nonetheless, it lays the groundwork for interpretable analyses of semantic patterns within the discovered clusters.