Artificial Intelligence (AI) Applications in Ophthalmology Robert - PDF document

Artificial Intelligence (AI) Applications in Ophthalmology Robert Chang, MD

IDx -- First US FDA Approval for Autonomous AI DR Detection April 2018 81‐91 88‐92 2017 Pivotal Trial N=900 10 sites • Requirements: • > 22 y/o, Topcon NW400 nonmyd • First Dx of more than mild DR: FDA cutoff • Screening at PCP/Endocrine clinics • Contraindicated if patient has visual problems, known retinal issues or prior retinal treatments, poor scan quality • Cannot find other concomitant eye disease Abramhoff MD, et al. NPJ Digit Med. 2018 Aug 28;1:39.

2019 Pivotal Trial N=942 15 sites Multicamera Bhaskaranand M, et al Diabetes Technol Ther. 2019 Aug 7. [Epub ahead of print] The Past = Define Rules (Features) How Does Deep Dot blot or flame Y/N Learning hemorrhages, exudates, Referrable IRMA, venous beading, NV DR AI Y/N Algorithm Referrable DR Work… convolutional neural network (CNN) The Present = Provide Classified Images

How Higher Order Information ( most important pixels only ) Can Still Help You Make the Correct Classification (Who Is this?) How Does It Work Mathematically? Linear Algebra & Derivative Math Produces Nonlinear Classification (Backprop to Find Most Important Pixels to Retain Classification) https://theclevermachine.wordpress.com/2014/09/11/a‐gentle‐introduction‐to‐artificial‐neural‐networks/

Dec 13, 2016 Landmark Publication JAMA. 2016;316(22):2402‐2410. doi:10.1001/jama.2016.17216 JAMA. 2016;316(22):2402-2410. Training Set 128,175 Images 75% Gradable Images 54 MDs 3-7 Grades Per Image To Reach Ground Truth

Performance on Validation Set for Moderate or Worse DM Retinopathy Colored Circle = MDs Black Curve = Algorithm Is the DR Model Generalizable? June 13, 2019 JAMA Ophthalmol. 2019;137(9):987-993.

How About for Glaucoma? • Can Glaucoma Be Detected by a Single Photo? JAMA Ophthalmol. 2019 Sep 12. doi: 10.1001/jamaophthalmol.2019.3501. [Epub ahead of print] http://www.ccri.com/2016/11/10/explainable‐artificial‐intelligence‐xai/ But Wait How Do You Explain The Algorithm Output?

Example Occlusion Mapping Scientific Reports volume 8 , Article number: 16685 (2018) What Biases or Confounders from Data Collection are Incorporated—Hidden Inside the Data? • Train an AI Algorithm Using Data to Predict Who is Going to Commit a Crime? • Blacks > White in prison, a naive A.I. system will infer that a black person is more likely than a white person to commit a crime • Even if race were ignored, imagine other proxies such as which neighborhood one lived, a single parent home, or lack of high school education became stereotypical predictors • Finding a pattern in the data to train great identification accuracy in the data set does not convey meaning • i.e. Husky vs. Wolf

So What is the Role of AI in Eyecare? Make Eye Screening Less Expensive, More Help Summarize More Data Simultaneously Scalable than Reading Centers Than an MD or OD Has Time or Ability What Needs to Be Solved for Mass Market AI? • Tested Only “In Silico” vs. Real World (Safety) • Reaching Clinical Endpoint Consensus Definitions, Integration with Human Values • Sorting Out Biases in Training Sets, Protection of Data / Privacy • Collecting Enough Data to Reach Actionable Endpoints Longitudinally • AI Explainability and Transparency • AI Teamwork and Liability • Sustainable Payment Method or Business Model