As part of my Academic Clinical Fellowship, I explored whether natural language processing (NLP) could help summarise GP consultations — automatically, accurately, and meaningfully.
The motivation was simple: GPs are under immense pressure, and documentation takes time away from what matters — the patient in the room. What if we could teach a machine to extract key information from free-text consultation records and generate concise summaries or action points?
Using the “One in a Million” consultation dataset, I worked with colleagues in the Department of Engineering to develop models that could begin to understand the structure and content of primary care conversations. We predated the large language model boom, so this project was both technically challenging and conceptually ambitious at the time.
The results? Promising — but not perfect. The model could identify certain clinical elements quite well (e.g., presenting complaint, advice given), but nuance and context remain hard to encode. Medical dialogue is full of subtlety, uncertainty, and interpersonal dynamics — all things that NLP still struggles with.
Still, it was a valuable proof of concept. It showed how collaboration between clinicians and data scientists can seed tools that ease documentation burden and support better record-keeping. And with the rise of tools like HeidiAI and ScribePro, it’s clear this area will only grow.