When a person visits the doctor and finds out they have a certain ailment, they are often referred to a specialist who can analyze their condition further. This specialist works with the report taken by the initial doctor in order to rule out certain illnesses or recommend tests to take. Currently, reports are unorganized and it is likely that crucial information contained in the reports are dismissed by specialists.
My project is an application that can summarize clinical information and tune relevance based on the profession of the user. For example, if the specialist is a cardiologist, the summarized information presented will contain information that most relates to what functions the specialist is observing. In order to summarize this information, I used query based abstractive summarization using word embeddings trained on a database with 6M+ clinical reports with the help of the Rubin Lab at Stanford. The use of word embeddings ensures that it captures the semantic meaning of all the words present by turning them into word vectors and then rearranges them into a format that captures the essential meaning of the sentences. I tuned the relevance of summaries by picking out the words with the closest semantic meaning to the predetermined word sets and used omni-supervised learning in order to achieve the highest possible model accuracy. By consolidating critical information regarding patient history and assessment, the application effectively minimizes the risk of overlooking critical information that can potentially lead to the death of a patient.
What inspired you (or your team)?
As a side project, I used abstractive summarization in order to summarize certain news articles and started looking for real life applications of text summarization. When I started interning for the Rubin Lab over the summer, I stumbled upon this project and was immediately intrigued by it. Through extensive research, I found out that the summarization of clinical information would save doctors valuable time and minimize errors in decision making and reports. I then decided to create the solution with the help of the Rubin Lab.