Ariel Rokem, University of Washington eScience Institute

Follow along at: https://arokem.github.io/2020-02-12-UWIN

Data science education

Development of tools and practices for reproducible research

Building a data science community: open, rigorous and ethical

Data-driven research

Data-driven discovery

Sloan Digital Sky Survey

By Donald Pelletier [CC BY-SA 4.0 ], from Wikimedia Commons

Large Hadron Collider

By Maximilien Brice [CC BY-SA 4.0 ], via Wikimedia Commons

Using machine learning to understand brain function

Glaucoma

A leading cause of irreversible blindness

2020: 76 M people affected

2040 (predicted): 112 M people affected

Several different etiologies all leading to increased intraocular pressure (IOP)

→ Glaucomatous optic neuropathy (GON)

Glaucoma diagnosis

Relies on multiple factors

Requires substantial expertise

Complex, expensive

Early detection of the disease very important

Crucial for successful clinical intervention

Automated glaucoma detection

Objective: build an auomated system for detection of glaucoma

Incorporate information from multiple sources

Provide interpretable results

UK Biobank dataset

863 glaucoma patients

771 healthy controls

55 participants who progress to glaucoma

Medical record models

Baseline: age, gender, ethnicity

Model 1: + cardiovascular, pulmonary variables (e.g., BP, FVC, PEF)

Model 2: + ocular data (e.g., IOP)

What features explain the diagnosis?

SHAP values (Lundberg and Lee, 2017)

See excellent explainer here

Retinal imaging

Retinal imaging

Artificial neural networks

A family of machine learning algorithms

Biologically inspired

Minsky and Papert (1969)

A cascade of linear/non-linear operations

Convolutional networks

Inspired by the visual system

Capitalize on spatial correlations in images

Bosking et al. (1997)

Dumolin & Visin (2018)

Image models

Color fundus photos

OCT

CFP + OCT

CFP + OCT + medical records

SHAP values of individual networks

Saliency maps (Integrated gradients; Sundararajan 2017)

SHAP values of contributions to overall ensemble

Summary of findings

Accurate automated glaucoma detection (AUC: 0.97)

Comparison with clinicians

Validation with PtG

Novel association of pulmonary variables with glaucoma

See also Chua et al. (2019)

Effects of glaucoma in photoreceptor layer of retina

See also Choi et al. (2011)

Limitations

"Found data"

Relatively clean sample (no ocular co-morbidities)

Limited comparison with clinician performance

Correlational

Outlook

Large datasets

+ Machine learning techniques

+ Interpretational methods

= Scientific insight

+ Potential clinical application

Thanks!

Parmita Mehta
(UW CSE)

Aaron Lee
(UW Ophthalmology)