Gaze responses to false-positive computer-aided detection prompts during colonoscopy: a paired-video and real-time eye-tracking study

Methods

• Event-locked eye tracking assessed attraction, attention occupation, recovery, and time amplification after false-positive prompts across controlled and real-time clinical settings.Retrospective paired-video viewing enabled within-subject comparisons; prospective recordings assessed gaze responses during real-time workflow.Source: Methods · p. 2

Results

• Gaze was attracted by false-positive prompts in 48.6% (68/140) of retrospective observations and 65.2% (533/817) of prospective events.Among attraction events with complete recovery, median attention occupation was 1000 ms and 1100 ms, respectively, versus median prompt durations of 33 ms and 267 ms; time amplification was 17.55-fold and 5.15-fold.Source: Results · pp. 2–3
• In the retrospective paired analysis, visible artifact prompts drew gaze closer to the prompted region than the same-coordinate unassisted reference.Lesion gaze recognition was 98.0% without versus 99.0% with CADe, and first gaze entry into lesion ROIs occurred 147.8 ms earlier with CADe.Source: Results · p. 3

Conclusions

• False-positive prompts frequently captured gaze, with attention persisting beyond prompt visibility; prompt-related attentional burden should be considered in CADe evaluation and design.The conclusion frames this as support for more effective human–AI collaboration.Source: Conclusions · p. 3

Motivation

• CADe’s clinical value also depends on effects on endoscopist behavior and performance, beyond detection outcomes.Improved adenoma detection has not been consistently observed in pragmatic implementation and population-based screening trials.Source: §1 · p. 4

Research gap

• Earlier evaluations quantified false-positive frequency or procedure-level burden but did not establish whether individual prompts capture gaze or hold attention after disappearing.A prior controlled video study examined assessments and decision times after true- and false-positive recommendations.Source: §1 · pp. 4–5

Study aim

• The study shifts analysis from procedure-level outcomes to individual prompts, using event-locked eye tracking to characterize attraction, attention occupation, recovery, and persistence relative to prompt visibility.It combines paired-video experiments with prospective real-time recordings, with lesion-directed gaze as a secondary comparator.Source: §1 · p. 5
• This event-level approach assesses human–AI interaction at the process level alongside conventional detection endpoints and false-positive event counts.Source: §1 · p. 5

2.1 Study design

• The retrospective study paired unassisted and CADe-assisted viewing of the same videos, while the prospective study recorded live CADe-assisted colonoscopies without an unassisted control.The datasets were analyzed and reported separately because their designs, control conditions, and event acquisition differed.Source: §2.1 · p. 5

2.3 Eye-tracking setup and CADe system

• Eye tracking, video frames, and CADe bounding boxes were aligned for event-level analysis, with artifact events defined as prompts outside expert-annotated lesion windows persisting for at least two frames.Gaze was recorded at 30 Hz; single-frame detections were excluded from artifact events.Source: Figure 1 · p. 7, §2.3 · p. 8

2.2 Subjects

• The retrospective paired-video experiment included five endoscopists—three senior and two novice—who completed both unassisted and CADe-assisted viewing sessions; prospective data comprised 42 complete real-time CADe-assisted colonoscopies by nine senior endoscopists.Source: §2.2 · p. 8

2.4 Expert lesion annotation

• Two expert endoscopists independently reviewed videos, resolved disagreements by consensus, and used the resulting lesion intervals as analysis windows.In prospective recordings, cecal intubation time additionally defined the start of withdrawal; lesion-window onset was t=0 for lesion-directed gaze analyses.Source: §2.4 · pp. 8–9

2.5 Retrospective paired-video experiment

• Each participant viewed the same 60 clips, approximately 30 seconds each, in matched unassisted and CADe-assisted versions.The clips were concatenated into a 33-minute, 23-second sequence.Source: §2.5 · p. 9
• Participants completed the unassisted session first and the CADe-assisted session two weeks later, with screen gaze recorded throughout.Source: §2.5 · p. 9
• For lesion comparisons, the visible CADe box defined the assisted ROI, while the same coordinates at the corresponding time defined the unassisted reference ROI.This enabled within-subject comparison of gaze responses to the same lesion region.Source: §2.5 · p. 9
• Artifact candidates were CADe boxes outside expert-annotated lesion windows and had to persist for at least two consecutive detection frames.The final dataset contained 41 lesion events and 28 artifact events, each evaluated in both viewing conditions.Source: §2.5 · p. 9

2.6 Prospective real-time CADe recordings

• For each procedure, endoscopy video, gaze coordinates, and per-frame CADe boxes were synchronized to a common timeline.Cecal intubation marked withdrawal start, and expert-annotated lesion appearance and disappearance times defined lesion windows.Source: §2.6 · p. 9
• Boxes overlapping lesion windows were assigned to the corresponding lesion ROI; boxes outside all lesion windows were candidate false-positive artifact events.Source: §2.6 · pp. 9–10
• The dataset contained 60 lesion windows and 817 artifact events from 42 complete real-time CADe-assisted colonoscopies.Source: §2.6 · p. 10
• Lesion-directed gaze and false-positive prompt-related gaze responses were analyzed separately as descriptive measures of real-time human–AI interaction.Source: §2.6 · p. 10

2.7 Gaze metrics

• Lesion recognition required at least 100 ms of cumulative gaze within the ROI during the expert-annotated lesion window; first-entry time ran from window onset to the first gaze sample inside the ROI.Retrospective assisted and unassisted comparisons used the visible box and the same-coordinate reference ROI.Source: §2.7 · p. 10
• Artifact gaze-hit proportions were calculated in 0.5-second bins using a fixed ROI formed by the union of boxes seen during the first 0.5 seconds after onset.Gaze attraction was assessed during that same initial 0.5-second interval.Source: §2.7 · p. 10
• An event was classified as attraction-triggered when the 0–0.5 s post-onset window contained at least 300 ms of valid gaze and a contiguous direction-gated approach lasting at least 100 ms.Distance was zero inside the ROI and otherwise the shortest pixel distance to the ROI.Source: §2.7 · pp. 10–11
• Immediate attraction magnitude was baseline median gaze-to-ROI distance minus the minimum direction-gated distance after prompt onset, with the minimum’s time recorded as tpeak.Baseline covered −0.5 to 0 seconds; the post-onset minimum was found in the first 0.5 seconds.Source: §2.7 · p. 11
• Attention occupation ran from the onset of a qualifying approach to recovery, defined as stable non-attraction for at least 200 ms after tpeak.When gaze entered the ROI vicinity, recovery also required moving outside it; events without recovery within five seconds were right-censored.Source: §2.7 · p. 11
• Time amplification was occupation duration divided by artifact on-screen duration and was calculated for recovered attraction events.The same event-locked framework was applied separately to retrospective and prospective datasets.Source: §2.7 · p. 11

2.8 Statistical analysis

• All analyses were descriptive, with no formal hypothesis tests, because the study aimed to characterize gaze mechanisms rather than test a prespecified hypothesis.The retrospective experiment included only five endoscopists, with observations nested within subjects and lesion or artifact events.Source: §2.8 · pp. 11–12
• The study reported point estimates, sample sizes, and interval estimates rather than P values, and retrospective comparisons used paired observations where evaluable in both conditions.First ROI-entry times were averaged within each participant before summarizing across five paired participants.Source: §2.8 · p. 12
• Attraction-trigger confidence intervals used the Wilson score method, while event-locked gaze-hit intervals used mean ± 1.96 standard errors.The gaze-hit intervals assumed independent observations and did not adjust for clustering within endoscopists or procedures.Source: §2.8 · p. 12
• Retrospective first-entry times were summarized as participant-level means, while prospective first-entry times were summarized as medians among recognition-positive lesion windows.Source: §2.8 · p. 12
• Right-skewed duration and ratio measures were summarized by medians, and occupation duration and time amplification included only attraction-triggered events with observed recovery.Events without recovery within five seconds were excluded from those summaries and counted separately.Source: §2.8 · p. 12
• Senior-versus-novice results were exploratory, and retrospective and prospective datasets were analyzed separately rather than pooled.Source: §2.8 · p. 12

3.1 Study datasets and event characteristics

• The retrospective paired dataset included 41 lesion and 28 artifact events evaluated by five endoscopists; the prospective dataset included 60 lesions and 817 artifacts across 42 procedures by nine senior endoscopists.False-positive event rates were 0.84 events/min retrospectively and 2.77 events/min prospectively.Source: §3.1 · p. 13

3.2 Gaze responses to false-positive CADe prompts

• False-positive prompts attracted gaze in 48.6% (68/140) of retrospective observations and 65.2% (533/817) of prospective events; visible prompts shifted gaze 25.7 px closer to the prompted region.The paired comparison used the same-coordinate unassisted reference.Source: §3.2 · p. 13, Figure 2 · p. 15
• Among attraction events with complete recovery, median attention occupation was 1000 ms retrospectively and 1100 ms prospectively, versus prompt durations of 33 ms and 267 ms.Median time amplification was 17.55-fold and 5.15-fold, respectively; right-censored events were excluded from this recovered-event analysis.Source: §3.2 · pp. 13–14, Figure 2 · p. 15

3.3 Lesion-directed gaze recognition and first-entry time

• Retrospective lesion gaze recognition was 98.0% (201/205) unassisted and 99.0% (203/205) with CADe; prospective recognition was 86.7% (52/60).The prospective median first ROI-entry time among recognition-positive events was 1633.5 ms (n=52).Source: §3.3 · p. 16
• Mean first lesion ROI entry across five retrospective participants occurred 147.8 ms earlier with CADe, at 1310.5 ms versus 1458.3 ms unassisted.The figure presents event-level distributions and paired subject means.Source: §3.3 · p. 16, Figure 3 · p. 17

3.4 Experience-stratified gaze responses

• Senior endoscopists’ mean first lesion ROI entry was 295.6 ms earlier with CADe (1542.4 vs 1246.7 ms), while novice entry was 74.0 ms later (1332.1 vs 1406.1 ms).These experience-stratified findings were exploratory given the small subgroups of three seniors and two novices.Source: §3.4 · p. 18
• Senior endoscopists had higher retrospective artifact attraction than novices (52.4%, 44/84 vs 42.9%, 24/56) and greater median time amplification (23.26-fold vs 14.05-fold).These subgroup findings were exploratory given the small sample of three senior and two novice endoscopists.Source: §3.4 · p. 18

4 Discussion

• Attention occupation persisted for approximately 1000–1100 ms across datasets, substantially longer than the false-positive prompts themselves.The authors treat occupation duration as the primary temporal measure and time amplification as complementary.Source: §4 · pp. 18–19
• 17.55-fold retrospective and 5.15-fold prospective median time amplification should be interpreted alongside prompt durations of 33 ms and 267 ms, respectively.Time amplification is sensitive to visual-stimulus duration.Source: §4 · p. 19
• Lesion gaze recognition was high with and without CADe, while first gaze entry into lesion regions occurred 147.8 ms earlier with CADe.The authors caution that earlier gaze entry is not evidence of earlier lesion detection or improved diagnostic performance.Source: §4 · p. 19
• Exploratory subgroup analyses suggested that senior endoscopists had higher false-positive attraction-trigger rates and greater time amplification than novices.The authors note the small subgroup sample of 3 senior and 2 novice endoscopists.Source: §4 · pp. 19–20
• The study characterizes gaze responses during CADe use but does not establish whether repeated exposure changes unassisted detection performance.Longitudinal studies are proposed to examine possible changes in visual search or lesion detection after repeated exposure.Source: §4 · p. 20
• Event-level attention measures may complement alert-frequency metrics because systems with similar false-positive rates may differ in attention consumed per prompt.The discussion identifies reducing unnecessary capture while preserving rapid orientation toward true lesions as a potential design objective.Source: §4 · p. 20
• The paired-video and prospective datasets provide complementary evidence, with the latter showing gaze attraction and persistence during real-time clinical workflow.The datasets differed in setting and prompt duration and were analyzed separately rather than pooled.Source: §4 · p. 20
• Interpretation is bounded by the small retrospective sample, no matched prospective unassisted comparison, 30-Hz gaze recording, and evaluation of one CADe system.Whether attention occupation translates into measurable changes in lesion detection remains undetermined.Source: §4 · pp. 20–21

• False-positive CADe prompts frequently attracted gaze, and attention persisted beyond prompt visibility in controlled and real-time clinical settings.Source: §5 · p. 21
• Event-level gaze metrics may complement conventional detection endpoints and false-positive counts by capturing how individual prompts shape visual attention.Source: §5 · p. 21
• Incorporating prompt-related attentional effects into future CADe evaluation and design may support more effective human–AI collaboration.Source: §5 · p. 21