Deep Learning for low Vision Enhancement (DELVE)
Funded Grant
Overview
Affiliation
View All
Overview
description
Project Summary or Abstract Low vision rehabilitation (LVR) can improve functional ability and quality of life for people with visual impairment, but referral rates for LVR are inconsistent and can be extraordinarily low. Current guidelines for LVR referral are too broad and there is no system to identify candidates for LVR to gain value from the services. Providers are unable to predict which patients are likely to benefit from LVR, given the variety of LVR interventional options and diverse characteristics of patients with low vision. Thus, there is a critical need to understand the provision of LVR services and develop algorithms that can identify patients who may be candidates for LVR in an automated manner to facilitate appropriate, timely referrals. To achieve that long-term goal, we propose to use recent advances in artificial intelligence and natural language processing that have enabled analysis of data in electronic health records (EHR), including clinical findings embedded in unstructured free-text notes, containing the essence of medical documentation with rich information. This R34 planning grant will establish the Deep Learning for low Vision Enhancement (DELVE) multicenter network of LVR providers at seven clinical practice sites who will coordinate and harmonize their EHR data for future research involving deep learning methods to enhance LVR practices. We will create the DELVE multicenter EHR datasets at each site, tools for data management and analysis, and finalize future clinical study protocols, including the preparation of a manual of operations for data harmonization across sites and deep learning model validation. First, we aim to validate and refine deep learning prediction models at two ophthalmologic academic centers to determine visual prognosis (i.e., predict whether vision will improve to better than 20/40 within 1-year). Next, we plan to develop and evaluate deep learning models to predict factors related to LVR provider recommendations and patients’ acquisition of visual aids to provide magnification for television viewing or near reading. We aim to characterize patients who could benefit from such aids and would be candidates for LVR. Findings from two ophthalmologic academic centers will serve as feasibility data to refine methods for future multisite studies. To improve the referral basis for LVR and patient outcomes, we aim to develop novel systems and multicenter DELVE databases to utilize clinical information and criteria generated from deep learning model data. Future work beyond the current R34 would validate our developed algorithms to determine applicability to other optometric LVR practices. Our long-term initiative is to leverage deep learning techniques to identify best practices and gaps in LVR services to drive new intervention development and significantly enhance patient care for visually-impaired adults.
