This live panel discussion is part of the AI Happy Hour series, brought to you by Stanford AIMI and friends. We cover hot topics in AI in medicine as well as live questions & comments from attendees. It's casual, insightful, and open to all!
Outline:
03:45 - Panel Introductions
07:18 - Contrast of NLP work in and out of medicine - domain specificity, jargon heavy, noisy, no definitive grammar, lot of negation, speculative, templates/sections/headers, long range dependency, table extraction and generalization
14:19 - Key health applications of NLP - Precision medicine, patient reported information that is not captured in CPT code, outcomes prediction, ambient accurate notes, risk prediction, phenotype extraction
21:00 - Can standardized reports assist with data extraction? - Possibly interfere with recording a patient’s story vs placeholder for standardized measurements
24:27 - Relating image analysis and text - radiology reports as labels, weak supervision, distinguishing training vs validating data sets
30:38 - Foundational vs specific health NLP models - benefits with starting small, leverage larger models appropriately while building smaller ones
36:15 - Downsides of NLP - significant time and cost, data isolation, de-identification, bias, diverse labeling, appropriate crowdsourcing
47:11 - Equity in NLP training sets - biases in patients and providers
48:19 - New areas to explore - English vs other languages, precision medicine, patient goals, shared decision making, time trajectories, radiology ordering summary, generate text from images, improved EHR search, creating patient cohorts, lay language questions
56:52 - Quantify provider reactions/sentiment analysis - interpersonal relationship, identify biases
Panelists:
Tina Hernandez-Boussard - Stanford
Curtis Langlotz - Stanford
Meliha Yetisgen - University of Washington
Rogier van der Sluijs - Stanford
Beliz Gunel - Stanford
Stanford AIMI: [ Ссылка ]
Twitter: [ Ссылка ]
#StanfordAIMI #AI
