Predicting the outcomes of ongoing process instances remains a central challenge in the field of Predictive Process Monitoring. While deep learning techniques have recently demonstrated strong performance in tackling this task, the emergence of Large Language Models (LLMs) has opened new opportunities for innovation. This research explores the expressive power of LLMs and the richness of textual representations to enhance predictive capabilities. Our method transforms structured event log data into coherent, narrative-style text descriptions. These “semantic stories” capture the context and evolution of each process instance, constructed using domain-specific templates that reflect real-world healthcare workflows. Leveraging a pre-trained LLM, we fine-tune the model using historical data from an Emergency Department in the Turin area, enabling it to learn patterns associated with different process outcomes. Thus, we reframe the prediction task as a natural language prediction task, making it ideal for exploiting the expressive power of LLMs. Finally, we evaluate a test set of process instances. Results show that the method is applicable to complex and variable clinical workflows and holds potential for supporting decision-making in emergency care settings.