Personalized Clinical Note Generation from Doctor-Patient Conversations
Nathan Brake, Thomas Schaaf
TL;DR
The paper tackles personalized clinical note generation from doctor-patient conversations by introducing physician embeddings that capture individual note styles. It proposes a three-phase framework: train with physician-specific embeddings, adapt to new physicians by selecting the best matching existing embedding, and test using ROUGE-2 and Factuality metrics. Results show substantial improvements over a baseline lacking embeddings, with notable gains when embeddings are applied to the decoder, and effective adaptation for unseen physicians demonstrated by improvements in HPI, PE, and A&P sections. This approach enables rapid enrollment of new physicians without retraining, potentially scaling personalized clinical documentation in real-world settings.
Abstract
In this work, we present a novel technique to improve the quality of draft clinical notes for physicians. This technique is concentrated on the ability to model implicit physician conversation styles and note preferences. We also introduce a novel technique for the enrollment of new physicians when a limited number of clinical notes paired with conversations are available for that physician, without the need to re-train a model to support them. We show that our technique outperforms the baseline model by improving the ROUGE-2 score of the History of Present Illness section by 13.8%, the Physical Examination section by 88.6%, and the Assessment & Plan section by 50.8%.