Purchase this article with an account.
Dana Schlegel, Benjamin Katz, Vittorio Bichucher, Richmond Starbuck, Wei Xu, Jacob Durrah, Eman Al-Sharif, Andrew DeOrio, Naheed W Khan, Kanishka Thiran Jayasundera; RetDegenDx: A retinal dystrophy genetic diagnosis prediction tool. Invest. Ophthalmol. Vis. Sci. 2016;57(12):3161. doi: https://doi.org/.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Develop a machine learning program that predicts the most likely genetic cause of a patient’s retinal dystrophy using input information about patient demographics, electroretinogram (ERG) response, visual field, pattern of inheritance, and fundus autofluorescence (FAF) features. This program may help to inform appropriate genetic test panels to order and can be used as a tool with which to interpret genetic testing results, along with existing resources such as PolyPhen, SIFT, and reported causative mutations in the literature.
Data on patient age, sex, ERG response, visual field, family history, FAF imaging, and genetic diagnosis was collected on 152 patients seen at the Kellogg Eye Center. After filtering out mutated genes that affected fewer than 5 patients, 102 patients were usable for machine training purposes. Machine learning algorithms were developed to predict the genes most likely to be mutated and causing the observed clinical features for a given patient. Multiple algorithms were applied, and the support vector machine (SVM) with radial basis function (RBF) kernel was shown to perform best. Machine learning used 80/20 training/testing splits of the data. The classification performance was compared with that of a baseline classifier.
A prototype has been created that has greater than 60% accuracy for predicting the causative mutated gene in a particular patient’s sample. This prototype is currently being shown to other retinal dystrophy clinics to obtain more patient data from these institutions. With the continued collaboration of outside institutions, the functionality of the algorithm will be improved.
The generation of this machine learning program provides physicians with a prediction of a causative mutated gene when evaluating a new patient in their clinic. This may help ophthalmologists refine their clinical diagnosis and identify the most relevant genetic testing to order. Additionally, genetic test results can often be difficult to interpret. Programs such as PolyPhen and SIFT help to determine whether or not a variant detected in a gene is truly pathogenic. The RetDegenDx program is one further tool that can be used to help predict the gene in which mutations are most likely to cause the clinical features of the patient. When used alongside existing programs, it can help to improve the interpretation of genetic test results.
This is an abstract that was submitted for the 2016 ARVO Annual Meeting, held in Seattle, Wash., May 1-5, 2016.
This PDF is available to Subscribers Only