Topic modeling and sentiment analysis of Greek clinician-patient conversations in hematologic malignancies

Conversations in clinical settings, especially those involving hematologic cancers or palliative care are not only informational but also emotionally charged. Understanding how these conversations are structured could help develop AI systems that support not only information exchange but also personalized emotional care.

This study explored the themes discussed and emotional tones in real conversations between Greek speaking clinicians and patients. Our aim was to contribute toward building digital tools for low-resource languages such as Greek that are not only linguistically appropriate but also emotionally aware.

We analyzed over 52,000 anonymized utterances from real Greek clinical conversations using BERTopic for topic modeling and a domain-specific sentiment classifier fine-tuned on Greek medical data. Topics were labeled using both a large language model (Gemma-3) and extractive keyword methods (KeyBERT).

The analysis revealed 35 thematic clusters, covering areas like symptoms, diagnosis, emotional states, treatment choices, and end-of-life planning. Sentiment analysis revealed that patients expressed more negative sentiment in utterances related to pain, uncertainty or personal loss. Clinicians often responded with a more neutral or even empathetic positive tone, especially when offering support or medical advice.

Our findings show that Natural Language Processing (NLP) can reveal both the content and emotional tone of conversations in medical settings. This approach could support the development of digital health tools such as conversational agents that are more emotionally aware in Greek and other low-resource languages, especially in emotionally charged domains like palliative care.