June 2023
Volume 64, Issue 8
Open Access
ARVO Annual Meeting Abstract  |   June 2023
Transfer learning for reducing data inequality in health disparities studies.
Author Affiliations & Notes
  • TingFang Lee
    Department of Ophthalmology, NYU Langone Health, New York, New York, United States
    Division of Biostatistics, Department of Population Health, NYU Langone Health, New York, New York, United States
  • Gadi Wollstein
    Department of Ophthalmology, NYU Langone Health, New York, New York, United States
    Department of Biomedical Engineering, NYU Tandon School of Engineering, Brooklyn, New York, United States
  • Chisom T. Madu
    Department of Ophthalmology, NYU Langone Health, New York, New York, United States
  • Andrew Wronka
    Department of Ophthalmology, NYU Langone Health, New York, New York, United States
    Medical Center Information Technology, NYU Langone Health, New York, New York, United States
  • Lei Zheng
    Department of Ophthalmology, NYU Langone Health, New York, New York, United States
  • Joel S Schuman
    Department of Ophthalmology, NYU Langone Health, New York, New York, United States
    Department of Biomedical Engineering, NYU Tandon School of Engineering, Brooklyn, New York, United States
  • Jiyuan Hu
    Division of Biostatistics, Department of Population Health, NYU Langone Health, New York, New York, United States
  • Footnotes
    Commercial Relationships   TingFang Lee None; Gadi Wollstein None; Chisom Madu None; Andrew Wronka None; Lei Zheng None; Joel Schuman Zeiss, Code P (Patent); Jiyuan Hu None
  • Footnotes
    Support  P30EY013079; R01-EY013178
Investigative Ophthalmology & Visual Science June 2023, Vol.64, 975. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      TingFang Lee, Gadi Wollstein, Chisom T. Madu, Andrew Wronka, Lei Zheng, Joel S Schuman, Jiyuan Hu; Transfer learning for reducing data inequality in health disparities studies.. Invest. Ophthalmol. Vis. Sci. 2023;64(8):975.

      Download citation file:


      © ARVO (1962-2015); The Authors (2016-present)

      ×
  • Supplements
Abstract

Purpose : An important area of exploration is race disparity in healthcare. One of the major setbacks in these projects is the imbalance in the input data such that minority groups are often underrepresented. We propose to utilize transfer learning that transfers useful information from source domain (large dataset) to the target domain (small dataset) to improve model prediction performance among data-disadvantaged racial groups. We will illustrate the method by evaluating the prediction of rate of visual field change over time from different races.

Methods : N = 1,149 subjects who received ≥3 visual field tests from 2018 to 2021 were enrolled in the study, including 701 (61%) White, 128 (11%) Asian, and 320 (28%) Black subjects. The rate of change in visual field mean deviation (MD) per year is the primary outcome in the prediction task. Predictors include baseline MD, age, gender, ethnicity (Hispanic/non-Hispanic), glaucoma diagnosis (glaucoma, glaucoma suspect, and non-glaucoma), and area deprivation index (ADI; ranking of neighborhoods by socioeconomic disadvantage) (Table 1). We used the dominant race group, White, as the source domain, and Asians and Blacks as the target domains. A two-step transfer learning algorithm was then exploited to transfer knowledge from source task to boost prediction performance in the target task. Mean absolute errors (MAE) and mean square errors (MSE) were used to evaluate the prediction performance of the proposed transfer learning in comparison to that using the original linear regression on only target domain.

Results : The source domain included all White subjects, and the target domains included 70% of Asian and Black subjects. 30% of Asian and Black subjects were used as testing datasets to assess prediction performance. The MAEs and MSEs of predicting MD rate of change per year in Asians and Blacks through transfer learning were lower than the prediction errors through linear regression (Table 2). Transfer learning improved 24.8% MSE in predicting Asians’ outcome, and 2.2% MSE in Blacks’ outcome comparing with original linear regression.

Conclusions : Transfer learning can improve predictive model performance for data-disadvantaged race groups, and the improvement is more competently when target domain is smaller. This approach can help resolving data imbalance in race disparity research especially for groups who are markedly underrepresented in the source data.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.

 

 

×
×

This PDF is available to Subscribers Only

Sign in or purchase a subscription to access this content. ×

You must be signed into an individual account to use this feature.

×