July 2019
Volume 60, Issue 9
Open Access
ARVO Annual Meeting Abstract  |   July 2019
Using machine learning to detect preclinical retina structural changes in diabetics at varying HA1c levels
Author Affiliations & Notes
  • Christopher Anderson Clark
    School of Optometry, Indiana University, Bloomington, Indiana, United States
  • Ann E Elsner
    School of Optometry, Indiana University, Bloomington, Indiana, United States
  • Footnotes
    Commercial Relationships   Christopher Clark, None; Ann Elsner, None
  • Footnotes
    Support  none
Investigative Ophthalmology & Visual Science July 2019, Vol.60, 1541. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Christopher Anderson Clark, Ann E Elsner; Using machine learning to detect preclinical retina structural changes in diabetics at varying HA1c levels. Invest. Ophthalmol. Vis. Sci. 2019;60(9):1541.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : Diabetic retinopathy often times shows blatant changes such as exudates and edema on SD OCT. However there may be subtle architectural changes of the retina that are subclinical but associated with elevated blood sugar. The purpose of this study is to use machine learning to detect retinal features in diabetics based upon HA1C.

Methods : 8,760 SD OCT images (Optovue) from screening centers were used to train and test the machine learning model. All images came from diagnosed diabetics (59.8% female) with a wide range of HA1c. The entire data set was analyzed from 5 to 10 HA1c in 1% increased (6 levels in total). 2,920 images were reserved for testing the model and the remaining 5,840 were used to train the model. 18 models were trained and tested using 500 features at each HA1c level. A fine KNN model was the best at differentiating SD OCT images by HA1c at the time of the screening visit.

Results : The fine KNN model could detect unique structural changes in the SD OCT image regardless of which HA1c cut level was used, even at low HA1c levels as 6. At a HA1c level of 6, the ROC area under the curve was 0.74 using the testing image set. 1,099 correctly predicted above 6 HA1C, 1,049 predicted below. There were 771 images that were categorized in the wrong leve. The ROC area under the curve increased with increasing HA1c levels.

Conclusions : The fine KNN model was not surprising as the best model. KNN (k Nearest Neighbor) takes advantage of multiple groups of features (similar or dissimalar) with a weight for the distance between the features in the image. As such, for things like retinopathy, it works well for seeing fields of exudates and nearby edema. The ability of the model to predict at all levels including relatively low HA1c ones suggests that there are other retinal features occuring at subclinical levels. Future work is needed to exam which features in the model were detected.

This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.

×
×

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.

×