Purchase this article with an account.
Soshian Sarrafpour, Bing Chiu, Hardik Parikh, Maria De Los Angeles Ramos Cadena, Hiroshi Ishikawa, Gadi Wollstein, Joel S Schuman, Joshua A Young; Utilizing a J48 Decision Tree to identify Patients at risk for Angle Closure Glaucoma.. Invest. Ophthalmol. Vis. Sci. 2019;60(9):1984. doi: https://doi.org/.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Machine learning has been utilized in various fields from chess and self-driving cars to grading cataracts and screening for diabetic retinopathy. Here, we utilize a J48 decision tree classifier to analyze data from Humphrey Visual Fields (HVF), optical coherence tomography (OCT), and epidemiologic data for markers to identify patients who develop angle closure glaucoma from normal controls.
A retrospective chart review using data from 114 eyes with angle closure glaucoma and 242 eyes from normal participants from the Pittsburgh Imaging Technology Trial performed at Pittsburgh Medical Center from 2007-2016. The WEKA machine learning application from the University of Waikato was used to generate a J48 tree. OCT retinal nerve fiber layer (RNFL) scans, OCT macula scans, and ganglion cell analysis was obtained using commercial spectral-domain OCT (Cirrus HD-OCT, Carl Zeis Meditec, Dublin, CA). HVF was done using standard automated perimetry. Eyes with OCT imaging with signal strength less than 7 or pathology except refractive or visually insignificant cataracts were excluded.
A pruned J48 classifier tree was formed using 65 attributes from 356 eyes using 10-fold cross validation and minimum leaf size of 12. A tree with 16 leaves was generated that correctly classified 238 instances out of 356 instances (accuracy of 66.9%). Of note, in the setting of thicker outer inferior macula thickness on OCT macula, the tree was able to identify 13/17 angle closure glaucoma patients (76.5%) based off central macula thickness. Similarly, in settings of thicker outer inferior macula thickness, and thinner inferotemporal RNFL thickness, the tree was able to predict 10/12 (83.3%) patients as angle closure glaucoma. Finally, in the setting of thinner outer inferior macula thickness and thicker inferotemporal RNFL thickness, the tree was particularly powered at identifying patients as normal, correctly identifying 96/98 patients as normal (98%).
Though limitations including imbalanced class distribution and overfitting of data exist, our tree was able to identify patients with angle closure with 66% accuracy and was able to identify specific features associated with developing angle closure glaucoma versus remaining normal. The decision tree generated by machine learning techniques may lend insight into particular risk factors and possible screening methods for angle closure glaucoma.
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