Abstract
Purpose :
To evaluate the effectiveness of machine learning (ML) and artificial intelligence (AI) applied to ophthalmological and retinal patients. The data-types required for the latter align closely with core ML/AI algorithmic requirements. Using the Synthetic Analyst AI platform, an array of data sources were evaluated and applied to future patient treatment outcomes.
Methods :
This is an ongoing prospective Retina Metric analysis of an AI, deep learning Synthetic Analytic platform. The platform automatically processes and learns from complex data sources with modern ML/AI algorithms. Deep learning integrates, analyzes patients historical and current examinations with their associated diagnostic retinal testing. This includes imagery and quasi-imagery data (slit lamp, OCT, micro-perimetry(MP1), retina photography, FA/ICG angiography), metric data (visual acuity, IOP, MP1, ERG, VEP), and patient metadata (chief complaint, present illness history, medical history, family/social histories). Foundational, sate-of-the-art convolutional neural networks (CNNs), were selected, tailored, and trained to identify features in the imagery data sources that were predictive of qualitative and quantitative relationships to other data.
Results :
16 terabytes of imagery and quasi imagery data, spanning over a twelve-year period were used to form a relational data set. With AI/ML analysis, the platform appears ready to suggest and predict future morbidity in patients with WET AMD, Clinically Significant Macular Edema and POAG. Applications of the CNN techniques to real-time image analytics are provided. The nature of the extracted features--high-dimensional, non-linear relationships between images and predicted classifications--are characterized, and a roadmap for further applications of ML and AI to similar examination and diagnostic data interrelationships is also described.
Conclusions :
The Retina Metrics-Synthetic Analyst AI platform currently appears comparable and suggests with time and a growing data set, capable of surpassing human interpretation capabilities. The ability to evaluate large volumes of historical diagnostic and patient data may further expose potential non-intuitive data patterns and correlations, potentially changing and improving individual patient outcomes.
This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.