Joann Elmore
University of California-Los Angeles, CA·Medicine & Clinical Research
Artificial intelligence is arriving in clinics faster than evidence about whether it actually helps patients. Research here asks a practical question: when a computer suggestion is placed in front of a physician, does the diagnosis get better, worse, or just different?
- Fields
Medicine & Clinical Research
Public Health & Epidemiology
Computer Science
- Topics
artificial intelligence
computer-aided detection
computer-aided diagnosis
- Grants
- 5 known grants, through Jul 2031
What they work on
- Run a two-year pragmatic randomized trial of more than 154,000 3D screening mammograms assigned to interpretation with or without an FDA-cleared AI tool, measuring cancer detection and recall rates across 16 breast imaging facilities.
- Link randomized mammography exams to state cancer registries to assess one-year sensitivity, specificity, interval cancer rate and tumor molecular subtypes with versus without AI assistance.
- Test whether radiologists vary in how much AI helps them based on experience, interpretive volume, tolerance of ambiguity and trust in AI, and look for automation bias, learning curves and carry-over effects.
- Randomize 250 pathologists to different computer-aided diagnosis cue types (image feature cues versus feature plus diagnosis cues) and timings (immediate versus delayed) to measure effects on interpretive accuracy using well-characterized biopsy cases.
- Use eye-tracking on a remote image viewing platform to reveal the perceptual and cognitive mechanisms by which CAD cues shift pathologists' visual search and decisions.
- Benchmark breast cancer recurrence and survival prediction at 5, 10 and 15 years by testing the PREDICT clinical model and retraining SlideGraph, a neural network that learns from whole slide pathology images.
- Build a Survivor Model by extending QuiltNet, the team's vision-language model, to combine whole slide image features with clinical prognostic factors, and compare it against clinical and image-only benchmarks.
- Develop an Explainer Model that turns survival predictions into pathologist-readable text descriptions of the driving image features, refined with pathologist feedback and validated on the CALGB and PATHWAYS datasets.
- Assemble the first multisite database of electronic health record communication in dementia care across three University of California Epic health systems, quantifying provider and caregiver portal network structures with social network measures.
- Test with social network analysis and machine learning-assisted visual analytics whether specific health-record communication structures predict preventable emergency department visits and unplanned hospitalizations among AD/ADRD patients.
More about the research
The work spans breast cancer screening, cancer pathology, prognosis prediction, and dementia care coordination, and it shares a focus on the human-AI interface, how clinicians actually read, trust and act on algorithmic cues, and on rigorous real-world evaluation including randomized trials and fairness monitoring.
Why it matters
Breast cancer screening and cancer pathology reach millions of patients, and AI tools are being adopted with thin evidence; these trials could show directly whether AI-assisted reading finds more cancers without more false alarms, and how to present computer cues so clinicians are helped rather than misled. Better breast cancer prognosis models with explanations could guide treatment choices for individual women. For people living with dementia, mapping how care teams communicate may reveal fixable gaps behind preventable emergency visits and hospitalizations.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Closest labs
Christoph Lee
University of Wisconsin-Madison, WI·Medicine & Clinical Research
Validate a 2D mammography AI algorithm from the DREAM Digital Mammography Challenge on UCLA's Athena Breast Health Network registry, and refine it with radiologist supervision plus genetic mutation and tumor subtype data.
John Mafi
University of California-Los Angeles, CA·Medicine & Clinical Research
Build governance and infrastructure for rapid, reproducible evaluation of AI clinical decision-support tools for dementia, using UC Health Data Warehouse EHR and imaging data and the ATLAS genotyped biobank.
Monica O'Reilly-Jacob
Columbia University, NY·Public Health & Epidemiology
Identify people living with dementia who receive care from nurse practitioners serving as main or ancillary primary care providers, by correcting Medicare claims for incident-to billing and applying the Primary Care Provider Identification Algorithm (PCP-IA).
The school
University of California-Los Angeles: 84 labs on PI Pages, $1.9B a year on research, 803 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.