Clinical utility beyond detection rates: Interpreting artificial intelligence in FIT‐positive colonoscopy

Dear Editor,

Robles de la Osa et al. deserve commendation for evaluating computer‐aided detection and diagnosis in a clinically relevant faecal immunochemical test (FIT)‐positive population [1]. Nevertheless, several issues warrant cautious interpretation.

The absence of statistical significance should not be equated with equivalence or lack of efficacy. The sample‐size calculation assumed a 40% baseline adenoma detection rate (ADR) and a 15% absolute improvement, whereas the control ADR reached 69.8%. This ceiling effect substantially reduced the study's ability to detect a modest incremental benefit. Moreover, ADR was numerically lower with artificial intelligence (AI) than without it (61.5% vs. 69.8%; p = 0.097). Reporting the between‐group effect with its 95% confidence interval and conducting an appropriately designed non‐inferiority or equivalence analysis would better define the range of plausible effects.

The association between withdrawal time and advanced‐neoplasia detection (OR 1.33 per minute) also requires caution. Because withdrawal time included lesion assessment and resection, detection itself may have prolonged the procedure. Therefore, reverse causation precludes interpreting this association as evidence that longer inspection independently improved detection.

Although CADx produced greater specificity and positive predictive value, its specificity of 66.7% and negative predictive value of 57.6% remain inadequate to support ‘leave‐in‐situ’ practice. Furthermore, every lesion was resected; consequently, reductions in unnecessary polypectomy, histopathology use, complications or costs were not directly demonstrated.

These findings reinforce that AI effectiveness is context‐dependent. Our review of AI in gastrointestinal endoscopy highlighted dataset bias, operator interaction, variable real‐world performance and the need for multicentre validation [2]. Similarly, mapping AI innovations to the IDEAL framework demonstrates that technical performance must progress through comparative assessment, workflow evaluation, patient‐centred outcomes and long‐term surveillance before routine adoption [3].

These concerns extend across clinical disciplines: evidence from dentistry similarly indicates that biased datasets, limited reproducibility, inadequate algorithmic transparency, and residual diagnostic errors necessitate continued human oversight [4].

Thus, this trial should not be interpreted as proving that AI is ineffective. Rather, it demonstrates that incremental benefit depends on clinical context and operator performance, while meticulous technique and accountable human oversight remain indispensable.

AUTHOR CONTRIBUTIONS

Divyesh A. Patel: Conceptualization; methodology; writing – original draft.

FUNDING INFORMATION

No funding was received for this work.

CONFLICT OF INTEREST STATEMENT

The author declares no conflicts of interest.

ETHICS STATEMENT

The author has nothing to report.

DATA AVAILABILITY STATEMENT

Data sharing not applicable to this article as no datasets were generated or analysed during the current study.