Grant Details
Grant Number: |
1R01CA288824-01A1 Interpret this number |
Primary Investigator: |
Elmore, Joann |
Organization: |
University Of California Los Angeles |
Project Title: |
Optimizing the Human-Computer Interaction in Pathology: Understanding the Impact of Computer-Aided Diagnosis Tools on Pathologists' Interpretive Performance |
Fiscal Year: |
2025 |
Abstract
SUMMARY
Accurate pathologic diagnoses, such as a cancer diagnosis, are the cornerstone of quality patient
care. Accurate diagnoses require pathologists to perform a complex series of perceptual and cognitive tasks
including visual search, pattern recognition, and decision-making. New artificial intelligence (AI) and computer-
aided diagnosis (CAD) technology shows promise for assisting pathologists and could improve diagnostic
accuracy. Rigorously evaluating the ways CAD cues can influence clinicians’ behavior, positively or negatively,
is necessary to ensure that CAD improves outcomes as intended without introducing unanticipated but
dangerous downsides. Our prior work in radiology showed that CAD tools can capture attention, induce
overreliance, and actually lead to worse accuracy when used in practice. We hypothesize CAD may exert
similar effects in pathology if employed ineffectively. By studying the impact of CAD cues on a large sample of
pathologists, the proposed research will inform best practices for deploying CAD tools into clinical practice and
have a profound impact on cancer diagnoses and patient care.
Our project will randomize 250 pathologists to the presentation of different types and timing of CAD cues in 3
phases of interpretation. We will examine the impact on pathologists’ interpretive behavior and accuracy of two
different types of CAD cues (Feature cues vs. Feature + Diagnosis cues; Aim 1), and two different timings of
CAD cue presentation (Immediately upon review of the case vs. Delayed until after completing initial review;
Aim 2). Finally, we will leverage eye-tracking to obtain a more granular understanding of the perceptual and
cognitive mechanisms underlying the effects of CAD on interpretive behavior and accuracy (Aim 3).
Strengths of our application include: 1) our experienced, multidisciplinary team with a history of successful
physician recruitment; 2) access to existing and uniquely well-characterized biopsy cases; 3) use of a machine
learning algorithm that our team developed in prior NIH funded research to create different CAD cues; 4)
established relationships with 12 clinical sites across the country for participant recruitment; 5) an innovative
remote image viewing and tracking platform for data collection; and 6) a scientific advisory board comprised of
industry partners and clinical experts who will provide their insight and expertise to ensure that our results are
immediately actionable for clinical practice.
The proposed work is innovative, pioneering, and timely, addressing fundamental questions that
accompany the introduction of new AI/CAD tools into the clinic. Our findings will evaluate the impact of CAD
and illuminate the cognitive mechanisms that underlie successes and errors that occur during the diagnostic
process as physicians interact with CAD cues. The results will inform solutions to guide the successful
implementation of CAD into clinical practice to enhance its benefits, mitigate harms, and optimize benefit for
patients.
Publications
None