Symposia 14 Program Schedule
02/17/2024
09:00 am - 10:30 am
Room: West Side Ballroom - Salon 1
Symposia 14: Advances in cognitive screening and neuropsychological assessment of cognitive decline and dementia in individuals with low education/low literacy levels
Simposium #3
Mitigating Bias in Machine Learning-based Cognition Classifier for Low-Education Participants
Jiaqing Zhang, University of Florida, Gainesville, United States
Category: Assessment/Psychometrics/Methods (Adult)
Keyword 1: academic achievement
Keyword 2: aging (normal)
Keyword 3: computerized neuropsychological testing
Objective:
In this study, we investigated potential machine learning (ML) bias in the widely-used Clock Drawing Test (CDT) against people with fewer years of education and methods to mitigate this bias.
Participants and Methods:
Clock drawings from 840 participants were gathered in a pre-operative hospital setting. The data was processed using a semi-supervised deep learning model and incorporated demographic details to predict cognitive function measures, including Attention (ATT), Memory (MEM), and Mini-Mental State Examination (MMSE).
Results:
The model's performance was evaluated for low-education (LE, ≤8 years, n=36) and high-education (HE, >8 years, n=804) groups separately. Classifier bias against the LE group was observed in MMSE and ATT predictions. To mitigate this bias, we employed the AI-Fairness 360 (AIF360) toolbox, which significantly improved the performance balance between the low and high education levels.
Conclusions:
This study demonstrates the need for fairness in ML classifiers, highlighting how tools like AIF360 can address the issue.
| 
