Gaze responses to false-positive computer-aided detection prompts during colonoscopy: a paired-video and real-time eye-tracking study
Te Luo, Yan Zhu, Peiyao Fu, Ruijie Yang, Xian Yang, Quanlin Li, Pinghong Zhou, Shuo Wang
cs.AI
Quick verdict
Claims false-positive CADe prompts capture endoscopist gaze for ~1 s via event-locked eye tracking; well-designed but small, descriptive, resolution-limited.
Summary
This study quantifies how individual false-positive CADe bounding-box prompts attract and hold endoscopists' gaze during colonoscopy, using event-locked eye-tracking analyses of two datasets. In a retrospective paired-video experiment, 5 endoscopists (3 senior, 2 novice) viewed 60 clips with and without CADe prompts; prospectively, gaze was recorded during 42 real-time CADe-assisted colonoscopies by 9 senior endoscopists, using an EndoAdd system and a 30 Hz Tobii eye tracker. Prompts outside expert-annotated lesion windows, persisting at least two consecutive frames, defined false-positive events. Gaze was attracted in 48.6% (68/140) of retrospective and 65.2% (533/817) of prospective events. Among attraction events with complete recovery, median attention occupation lasted 1000 ms (retrospective) and 1100 ms (prospective), versus median prompt durations of 33 ms and 267 ms — median time amplifications of 17.55-fold and 5.15-fold. Lesion gaze recognition was high (98.0% unassisted vs 99.0% with CADe), and first lesion-ROI entry occurred 147.8 ms earlier with CADe. All analyses are descriptive with no hypothesis tests; confidence intervals treat clustered observations as independent. Main limitations: only 5 retrospective endoscopists, fixed (unassisted-first) viewing order, a single proprietary CADe system, and 30 Hz sampling whose ~33 ms resolution equals the retrospective median prompt duration, making the 17.55-fold amplification there fragile.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- False-positive CADe prompts attracted endoscopists' gaze in 48.6% (68/140) of retrospective paired observations and 65.2% (533/817) of prospective real-time events.
- Among attraction events with complete recovery, attention occupation lasted a median 1000 ms (retrospective) and 1100 ms (prospective), far exceeding median prompt durations of 33 ms and 267 ms (time amplification 17.55× and 5.15×).
- In paired retrospective viewing, visible artifact prompts drew gaze closer to the prompted region than the same-coordinate unassisted reference (mean paired difference +25.7 px, n=121), establishing prompt-specific capture.
- CADe assistance did not impair lesion gaze recognition (98.0% unassisted vs 99.0% assisted) and accelerated first gaze entry into lesion ROIs by 147.8 ms.
- Prompt-related attentional burden should be treated as a complementary CADe evaluation and design dimension alongside false-positive event counts.
Strengths
Novel prompt-level unit of analysis
Shifts false-positive evaluation from procedure-level event counts to individual prompts, with clearly operationalized event-locked metrics (§2.7): direction-gated attraction ≥100 ms within 0.5 s, attention occupation until ≥200 ms stable recovery, 5-s right-censoring — a genuinely new process-level measure for CADe evaluation.
Complementary dual design with a clean paired comparison
The retrospective within-subject design comparing visible artifact prompts against same-coordinate unassisted reference ROIs (mean paired difference +25.7 px, n=121, §3.2) isolates prompt-specific gaze capture from tissue motion; prospective real-time recordings (42 procedures, 817 events) then test whether the mechanism persists in clinical workflow.
Transparent, test-free statistical reporting
Honest, self-aware limitations section
Traceable data provenance
Weaknesses
Tiny, deeply nested retrospective sample
Headline amplification sits at the measurement floor
Uncontrolled order/familiarity confound
Construct validity of 'attention occupation'
Unexplained prospective recognition drop and no code release
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This paper moves false-positive CADe evaluation from procedure-level event counts to **single-prompt attentional dynamics**. Using event-locked gaze analysis on two complementary datasets — a retrospective paired-video experiment (5 endoscopists × 60 clips, unassisted vs CADe-assisted) and 42 prospective real-time EndoAdd-assisted colonoscopies (9 senior endoscopists, 817 artifact events) — it claims: (i) false-positive prompts attract gaze in 48.6% (68/140) and 65.2% (533/817) of events (§3.2); (ii) attention occupation persists a median 1000–1100 ms versus prompt durations of 33/267 ms, i.e., time amplifications of 17.55× and 5.15× (Table 2); (iii) visible artifact prompts pull gaze closer to the prompted region than same-coordinate unassisted references (+25.7 px, n=121); (iv) lesion gaze recognition is preserved (98.0% → 99.0%) with first ROI entry 147.8 ms earlier under CADe (§3.3). The scope is honestly framed: gaze mechanisms only, explicitly not detection outcomes.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
How long does a false-positive AI prompt hold an endoscopist's attention during colonoscopy?
In this eye-tracking study, roughly one second. Median attention occupation was 1000 ms in a controlled paired-video experiment and 1100 ms during 42 real CADe-assisted colonoscopies, even though the on-screen bounding-box prompts themselves lasted only 33 ms and 267 ms respectively.
Does this study show that CADe false positives hurt polyp detection?
No. Detection outcomes were not measured. The paper shows that false-positive prompts capture gaze in about half (48.6%) to two-thirds (65.2%) of events and hold it for about a second; lesion gaze recognition stayed high with CADe (99.0% vs 98.0% without) and gaze reached lesions 147.8 ms earlier. Any detection-level consequence remains untested.
What eye tracker and CADe system were used, and on how many procedures?
A Tobii Pro Nano eye tracker at 30 Hz and the EndoAdd CADe system (June 2023 version, Xuan Wei Technology). Data comprise a retrospective experiment (60 paired clips viewed by 5 endoscopists with and without prompts) and 42 live CADe-assisted colonoscopies by 9 senior endoscopists.
Is the underlying data or code publicly available?
The eye-tracking recordings come from a dataset previously published in Scientific Data (Zhu et al., 2025), which supports data reuse. The paper itself does not state release of its analysis code or of the derived event tables and lesion annotations.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-3 are only available as captions in the extracted text; visual panels, bin-level trajectories, and boxplots could not be verified.
- The abstract is truncated mid-sentence at the end ('...false'), though the full body text through references appears complete.
- No data/code availability statement is visible in the extracted text; availability is inferred only from the cited Scientific Data source paper.
- Supplementary materials and the underlying Zhu et al. dataset contents (e.g., whether lesion windows and CADe boxes are included) could not be inspected.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.