June 2023
Volume 64, Issue 8
Open Access
ARVO Annual Meeting Abstract  |   June 2023
Using explainable artificial intelligence (XAI) to explore factors affecting meibomian gland (MG) dropout
Author Affiliations & Notes
  • Andrea Storås
    Department of Holistic Systems, Simula Metropolitan Center for Digital Engineering AS, Oslo, Norway
    Department of Computer Science, OsloMet - storbyuniversitetet, Oslo, Akershus, Norway
  • Fredrik Fineide
    Department of Medical Biochemistry, Oslo Universitetssykehus, Oslo, Norway
    Department of Computer Science, OsloMet - storbyuniversitetet, Oslo, Akershus, Norway
  • Hugo Hammer
    Department of Computer Science, OsloMet - storbyuniversitetet, Oslo, Akershus, Norway
    Department of Plastic and Reconstructive Surgery, Oslo Universitetssykehus, Oslo, Norway
  • Ayyad Zartasht Khan
    Department of Medical Biochemistry, Oslo Universitetssykehus, Oslo, Norway
  • Morten Schjerven Magno
    Department of Ophthalmology, Sorlandet sykehus HF Arendal, Arendal, Norway
    Department of Plastic and Reconstructive Surgery, Oslo Universitetssykehus, Oslo, Norway
  • Xiangjun Chen
    Department of Medical Biochemistry, Oslo Universitetssykehus, Oslo, Norway
    Department of Ophthalmology, Sorlandet sykehus HF Arendal, Arendal, Norway
  • Pål Halvorsen
    Department of Holistic Systems, Simula Metropolitan Center for Digital Engineering AS, Oslo, Norway
    Department of Computer Science, OsloMet - storbyuniversitetet, Oslo, Akershus, Norway
  • Tor Paaske Utheim
    Department of Plastic and Reconstructive Surgery, Oslo Universitetssykehus, Oslo, Norway
    Department of Medical Biochemistry, Oslo Universitetssykehus, Oslo, Norway
  • Michael Riegler
    Department of Holistic Systems, Simula Metropolitan Center for Digital Engineering AS, Oslo, Norway
  • Footnotes
    Commercial Relationships   Andrea Storås None; Fredrik Fineide None; Hugo Hammer None; Ayyad Khan None; Morten Magno None; Xiangjun Chen None; Pål Halvorsen None; Tor Utheim The Norwegian Dry Eye Clinic, Code O (Owner); Michael Riegler None
  • Footnotes
    Support  None
Investigative Ophthalmology & Visual Science June 2023, Vol.64, OD25. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Andrea Storås, Fredrik Fineide, Hugo Hammer, Ayyad Zartasht Khan, Morten Schjerven Magno, Xiangjun Chen, Pål Halvorsen, Tor Paaske Utheim, Michael Riegler; Using explainable artificial intelligence (XAI) to explore factors affecting meibomian gland (MG) dropout. Invest. Ophthalmol. Vis. Sci. 2023;64(8):OD25.

      Download citation file:


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

      ×
  • Supplements
Abstract

Purpose : The MGs produce lipids that reduce tear film evaporation. Artificial intelligence (AI) models show impressive results in the medical field, but may be hard to interpret and trust. XAI can enable more trustable AI support in the clinic. This study investigated the applicability of XAI to determine which factors are the most important predictors for MG dropout grade.

Methods : Data collected from 573 dry eye patients between September 2021 and December 2022 at the Norwegian Dry Eye Clinic were analyzed using AI. We imputed missing values using KNNImputer from the scikit-learn library. StandardScaler was used to normalize numerical features. Models were developed using the ExtraTreesClassifier algorithm. To predict the grade of MG dropout (0-4), two AI models (left and right eye) were trained. Dropout grades 0 and 1 were combined. The data was split into 75% for training and 25% for testing. Model performances were compared with a baseline model which always predicts the majority class. The most important features for MG dropout grade were investigated using XAI. All experiments were performed in Python 3.9.2.

Results : ExtraTreesClassifier models outperformed baseline models, with balanced accuracies of 36% (right eye) and 35% (left eye). Figure 1 shows the most important features for right eye predictions. The number of expressible MGs is the most important feature, followed by meibum expressibility and tear film break-up time. Figure 2 shows corresponding feature importance for the left eye model. The number of expressible MGs is considered most important followed by meibum expressibility and age. Due to a higher ratio of left-eye patients with the worst dropout grade than right-eye patients, the feature rankings of the two models differ. In both models, measurements in the opposite eye are important. This can be explained by the fact that both eyes are usually affected.

Conclusions : By applying XAI methods to the analysis of dry eye disease, new and interesting insights can be gained. Based on the methods applied in this study, the number of expressible MGs is the most important predictor of MG dropout grade. Still, our AI models cannot replace meibography in the clinic. More data should be collected since this might lead to improved models.

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

 

Feature importances for predicting MG dropout in the right eye.

Feature importances for predicting MG dropout in the right eye.

 

Feature importances for predicting MG dropout in the left eye.

Feature importances for predicting MG dropout in the left eye.

×
×

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.

×