Purchase this article with an account.
Michael Beyeler, Geoffrey M Boynton, Ione Fine, Ariel Rokem; Interpretable machine-learning predictions of perceptual sensitivity for retinal prostheses. Invest. Ophthalmol. Vis. Sci. 2020;61(7):2202.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To provide appropriate levels of stimulation, retinal prostheses must be calibrated to an individual's perceptual thresholds, despite thresholds varying drastically across subjects, across electrodes within a subject, and over time. Although previous work has identified electrode-retina distance and impedance as key factors affecting thresholds, an accurate predictive model is still lacking. The aim of this study was thus to develop a model that could 1) predict thresholds on individual electrodes as a function of stimulus, electrode, and clinical parameters (‘predictors’), and 2) reveal which of these predictors were most important.
We analyzed a dataset of 4,362 perceptual thresholds measured on 627 electrodes in 13 Argus II patients collected at 7 different implant centers from 2007-2018. Gradient boosting (a powerful ensemble model based on decision trees) was used to predict thresholds of individual electrodes as a function of stimulus (pulse width/polarity), electrode (impedance, electrode-retina distance), and clinical (subject age, years blind, time since surgery, implant site) predictors. Additional predictors were engineered from information routinely collected during system fitting (mean impedance of neighboring electrodes, fraction of unresponsive electrodes, subject’s false positive rate). The model was first fit to all available data to allow for a comparison with previous work, and then trained using a nested leave-one-patient-out cross-validation procedure to test its ability to generalize to novel patient data.
Using stimulus, electrode, and clinical predictors, the model achieved a coefficient of determination of R2=0.47 when fit to all data, which is consistent with previous studies, but dropped to R2=0.05 in the cross-validation procedure. Adding the engineered predictors described above improved the model performance to R2=0.73 in the full dataset and R2=0.61 in cross-validation. Years blind, time since implantation, and mean impedance of neighboring electrodes were the most important predictors, alongside electrode-retina distance and impedance.
Gradient boosting was able to predict perceptual thresholds across a large population of retinal prosthesis measurements. An accurate predictive model of perceptual sensitivity that relies on routinely obtained clinical data has the potential to transform clinical practice in predicting visual outcomes.
This is a 2020 ARVO Annual Meeting abstract.
This PDF is available to Subscribers Only