View Posters By Category
Monday, July 24, between 18:00 CEST and 19:00 CEST
|
Tuesday, July 25, between 18:00 CEST and 19:00 CEST
|
|
|
|
Session A Poster Set-up and Dismantle
Session A Posters set up:
Monday, July 24, between 08:00 CEST and 08:45 CEST
Session A Posters dismantle:
Monday, July 24, at 19:00 CEST | Session B Poster Set-up and Dismantle
Session B Posters set up:
Tuesday, July 25, between 08:00 CEST and 08:45 CEST
Session B Posters dismantle:
Tuesday, July 25, at 19:00 CEST |
Wednesday, July 26, between 18:00 CEST and 19:00 CEST
|
|
|
|
Session C Poster Set-up and Dismantle
Session C Posters set up:
Wednesday, July 26,between 08:00 CEST and 08:45 CEST
Session C Posters dismantle:
Wednesday, July 26, at 19:00 CEST |
A-222: Identification of ethnicity-derived biases and effects in standard of care clinical cancer sequencing data
Track: Equity-focussed Research Presentations
- Lancelot Seillier, Institute of Pathology, University Hospital RWTH Aachen, Aachen, Germany, Germany
- Nadina Ortiz-Brüchle, Institute of Pathology, University Hospital RWTH Aachen, Aachen, Germany, Germany
- Kjong-Van Lehmann, Joint Research Center Computational Biomedicine, University Hospital RWTH Aachen, Aachen, Germany, Germany
Presentation Overview: Show
Panel-based sequencing is becoming a cornerstone of molecular diagnostics and treatment decisions are based on the outcome of these sequencing results. NGS-derived scores like the tumor mutational burden (TMB) show correlation to higher response rates using immune checkpoint inhibitors or immunomodulating agents. Since most of our current knowledge is derived from patient populations that are rather homogeneous and biased towards few population subgroups, we set out to systematically study the extent to which this may affect results that are included in molecular tumor boards.
We hypothesize that the bias of biased reference populations is particularly amplified in clinical sequencing panels where no germline control is being derived. Thus, we selected samples from three cancer enteties with at least 30 patients per major population subgroup from panels with > 1 MBp coverage and investigated the subgroup-stratified TMB distribution.
Preliminary results showed differences in the TMB range and density in the GENIE cohort which is consistent with previously reported results. We have studied the effect of this bias on specific mutations based on the existing population prevalence for BRCA1 and BRCA2 mutations in the TCGA data. Our results suggest personalized medicine profits from ethnicity-aware data preprocessing for clinical scores.
A-223: Jointly Optimizing for Fairness Outperforms Post-hoc Bias Mitigation Strategies
Track: Equity-focussed Research Presentations
- Brett Beaulieu-Jones, University of Chicago, United States
Presentation Overview: Show
As artificial intelligence (AI) permeates healthcare and biomedical discovery, ensuring algorithmic fairness has become increasingly critical. While various software toolkits for fairness assessment and bias mitigation exist, they typically focus on data sampling strategies and post-hoc mitigation in the univariate setting. This study presents a proof of concept for training a neural network that is jointly optimized for both accuracy and fairness across multiple protected attributes. We utilized the EqualityML toolkit and applied bias mitigation strategies sequentially to protected attributes, and then implemented joint optimization by adding a custom loss function. Results from two tasks (predicting malignancy in tumor cells and acute kidney injury in ICU patients) showed improved average parity and comparable AUC when using joint optimization as opposed to univariate mitigation methods. This work highlights potential benefits of incorporating fairness objectives during model optimization and training rather than post-hoc mitigation.