Gender-sensitive word embeddings for healthcare
Lead author Shunit Agmon, MS, PhD(c), discusses this month's JAMIA Journal Club selection:
Agmon S, Gillis P, Horvitz E, Radinsky K. Gender-sensitive word embeddings for healthcare [published online ahead of print, 2021 Dec 16]. J Am Med Inform Assoc. 2021;ocab279. doi:10.1093/jamia/ocab279
Watch the Recording
Author
Shunit Agmon, MS, PhD (c)
Shunit Agmon is a PhD candidate in the Computer Science Department at the Technion – Israel Institute of Technology, under the supervision of Dr. Kira Radinsky and Prof. Benny Kimelfeld. Her research focuses on mitigating the effects of social biases on machine learning algorithms.
During her doctorate studies she also worked as a research intern at Amazon. Prior to that, she had completed her BSc (summa cum laude) and her MSc at the Technion as well. Shunit is a two-time recipient of the student research prize for cross-discipline collaboration in Data Science, funded by the Israeli planning and budgeting committee.
Manager and Moderator
Harry Reyes Nieva, PhD (c), MAS, MA
Statement of Purpose
In recent years, natural language processing (NLP) models are increasingly being used in clinical prediction tasks. These models are usually trained on a corpus of text describing clinical trials. For decades, clinical trials have excluded woman participants. NLP models, and specifically word embeddings, have been shown to capture and aggravate biases from the texts they are trained on. Therefore, models trained on the corpus of clinical trials will be inherently biased, and lead to poorer performance on women than on men.
Existing efforts to mitigate gender bias in word embeddings focus on removing the gender information from the embeddings. In the clinical domain, this approach is unhelpful: gender is an important feature that should be used in predictive models. We propose a gender-sensitive approach to mitigate the bias in word embeddings for healthcare, where the contribution of each clinical trial to the training process of the embeddings is determined by the number of female participants in that trial. We compare the performance of the gender-sensitive embeddings to a neutral baseline and show that the performance increases for several clinical prediction tasks: comorbidity classification, ICU readmission prediction, and hospital length of stay.
Target Audience
The target audience for this activity is professionals and students interested in health informatics.
Learning Objectives
After participating in this webinar the listener should be better able to:
- Identify the effects of biases in the data on natural language processing model results.
- Apply a data augmentation method to train gender-sensitive word embeddings
Format
- 35-minute presentation by article author(s) considering salient features of the published study and its potential impact on practice
- 25-minute discussion of questions submitted by listeners via the webinar tools and moderated by JAMIA Student Editorial Board members.
Accreditation Statement
The American Medical Informatics Association is accredited by the Accreditation Council for Continuing Medical Education to provide continuing medical education for physicians.
Credit Designation Statement
The American Medical Informatics Association designates this live activity for a maximum of 1 AMA PRA Category 1 Credit(s)™. Physicians should claim only the credit commensurate with the extent of their participation in the activity.
The live webinar only offers CME credit. The recording on our website will be openly available for learners but will not offer CME credit.
Commercial Support
No commercial support was received for this activity.
Disclosures for this Activity
The following planners and staff who are in a position to control the content of this activity disclose that they have no relevant financial relationships with commercial interests/ineligible entities:
- Presenter: Shunit Agmon
- JAMIA Journal Club Planners: Harry Reyes Nieva, Kirk E. Roberts
- AMIA Staff: Susanne Arnold, Pesha Rubinstein
Instructions for Claiming CME Credit
Use the link in the webinar’s chat area to access the claim-credit survey; in a day or two you will receive an email with your CME certificate.
If you require a certificate of participation, contact Pesha@amia.org.