June 2022
Volume 63, Issue 7
Open Access
ARVO Annual Meeting Abstract  |   June 2022
Federated learning for collaborative clinical diagnosis and disease epidemiology in retinopathy of prematurity
Author Affiliations & Notes
  • Adam H Hanif
    Oregon Health & Science University Casey Eye Institute, Portland, Oregon, United States
  • Charles Lu
    Athinoula A Martinos Center for Biomedical Imaging, Charlestown, Massachusetts, United States
  • Ken Chang
    Athinoula A Martinos Center for Biomedical Imaging, Charlestown, Massachusetts, United States
  • Praveer Singh
    Athinoula A Martinos Center for Biomedical Imaging, Charlestown, Massachusetts, United States
  • Aaron S Coyner
    Oregon Health & Science University Casey Eye Institute, Portland, Oregon, United States
  • James Brown
    Computer Science, University of Lincoln, Lincoln, Lincolnshire, United Kingdom
  • Susan Ostmo
    Oregon Health & Science University Casey Eye Institute, Portland, Oregon, United States
  • R.V. Paul Chan
    Ophthalmology, University of Illinois at Chicago, Chicago, Illinois, United States
  • Daniel Rubin
    Biomedical Informatics Research, Stanford University School of Medicine, Stanford, California, United States
  • Michael F Chiang
    National Eye Institute, Bethesda, Maryland, United States
  • Jayashree Kalpathy-Cramer
    Athinoula A Martinos Center for Biomedical Imaging, Charlestown, Massachusetts, United States
  • J. Peter Campbell
    Oregon Health & Science University Casey Eye Institute, Portland, Oregon, United States
  • Footnotes
    Commercial Relationships   Adam Hanif None; Charles Lu None; Ken Chang None; Praveer Singh None; Aaron Coyner None; James Brown Boston AI Lab, Code S (non-remunerative); Susan Ostmo None; R.V. Paul Chan Phoenix Technology Group, Alcon, Code C (Consultant/Contractor), Genentech, Code F (Financial Support), Siloam Vision, Code O (Owner), Boston AI Lab, Code S (non-remunerative); Daniel Rubin None; Michael Chiang Novartis, Code C (Consultant/Contractor), Genentech, Code F (Financial Support), InTeleretina, Code I (Personal Financial Interest); Jayashree Kalpathy-Cramer Genentech, Code F (Financial Support), Boston AI Lab, Code S (non-remunerative); J. Peter Campbell Boston AI Lab, Code C (Consultant/Contractor), Genentech, Code F (Financial Support), Siloam Vision, Code O (Owner)
  • Footnotes
    Support  This work was supported by grants R01 EY19474, R01 EY031331, R21 EY031883, and P30 EY10572 from the National Institutes of Health (Bethesda, MD), and by unrestricted departmental funding and a Career Development Award (JPC) from Research to Prevent Blindness (New York, NY).
Investigative Ophthalmology & Visual Science June 2022, Vol.63, 2329. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Adam H Hanif, Charles Lu, Ken Chang, Praveer Singh, Aaron S Coyner, James Brown, Susan Ostmo, R.V. Paul Chan, Daniel Rubin, Michael F Chiang, Jayashree Kalpathy-Cramer, J. Peter Campbell; Federated learning for collaborative clinical diagnosis and disease epidemiology in retinopathy of prematurity. Invest. Ophthalmol. Vis. Sci. 2022;63(7):2329.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : To utilize federated learning (FL), a method of collaboratively training deep learning (DL) models without sharing patient data, to differentiate inter-institutional diagnostic patterns and disease epidemiology in retinopathy of prematurity (ROP).

Methods : 5,245 retinal images were obtained from exams of patients in the neonatal intensive care units of 7 institutions. Images were labeled according to both the bedside clinical grading (CL) of plus disease (plus, pre-plus, no plus), and a reference standard diagnosis (RSD) determined by a consensus of three masked graders and the clinical diagnosis. Birthweight (BW), gestational age (GA), and clinical grades for all eye exams were recorded. DL models were trained on clinical labels for plus disease classification using either a centralized multi-institutional dataset, or an FL approach. Area under the receiver operating characteristic curve (ROC) was used as a measure of model performance. A DL-derived vascular severity score (VSS) was calculated for each eye exam. An “institutional VSS” was calculated by averaging the VSS of the most severe eye exam for each baby at each site. Demographics, clinical diagnosis of plus disease, and VSS between institutions were compared with simple linear regression, McNemar-Bowker test and one-way ANOVA.

Results : The performance of FL and central models trained on clinical labels was found to be equivalent (ROC = 0.93±0.06 vs 0.95±0.03, p=0.0175). The proportion of patients diagnosed as no plus and pre-plus by CL and RSD methods varied significantly (p<0.00l). Vascular severity and VSS corresponding to “no plus” diagnoses varied significantly across institutions (p<0.001). We found an inverse relationship between institutional VSS and mean GA (Figure 1, p=0.049, adjusted R2=0.49).

Conclusions : FL enabled the development of an accurate, DL-derived ROP VSS by drawing from multiple institutions’ labels without inter-institutional sharing of data. We identified differences in the clinical diagnosis of plus disease, and levels of objective ROP severity between institutions. FL has promise for objectively assessing differences in clinician diagnostic paradigms and disease severity across institutions.

This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.

 

Relationship between institutional VSS and mean population (A) BW (p=0.049) and (B) GA (p=0.10)

Relationship between institutional VSS and mean population (A) BW (p=0.049) and (B) GA (p=0.10)

×
×

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.

×