A Randomized Two-Period Crossover Reader Study of Artificial Intelligence-Assisted Anaerobic Threshold Interpretation in Cardiopulmonary Exercise Testing

The 150 de-identified CPET examinations will be divided into two non-overlapping case libraries, X and Y, with 75 examinations in each library. The libraries will be balanced, as far as feasible, by clinical severity and acquisition characteristics. None of the reader-study examinations will be used for model training, hyperparameter selection, or checkpoint selection.

Within each seniority stratum, physicians will be randomized in a 1:1 ratio to Sequence A or Sequence B. In period 1, Sequence A will interpret library X without AI assistance and library Y with AI assistance; Sequence B will receive the opposite conditions. Following a 14 ± 3 day washout, the assistance conditions will be reversed in period 2. Case order will be randomized within each period, and new blinded task identifiers will be assigned in period 2. Thus, every physician will interpret every examination once without AI assistance and once with AI assistance, yielding 4,800 interpretations.

In the assisted condition, the interface will display the frozen model's predicted AT time. In the unassisted condition, no model output will be shown. Readers will be masked to case source, routine-report and reference-standard labels, other readers' annotations, and their own annotations from the previous period. For each examination, the reader will classify AT as confirmed, not reached, or unable to confirm. When AT is confirmed, the reader will record the AT time and the interpretive method used. Interpretation time will also be recorded.

The reference standard will be prespecified from unassisted senior-physician annotations. AT will be considered present when at least 6 of 8 senior physicians confirm it, and the reference AT time will be the median of the confirming senior estimates. AT will be considered not reached when at least 6 of 8 senior physicians make that classification. Discordant cases will undergo adjudication by an additional senior physician who is masked to model output and reader identity. For analyses of senior readers, a leave-one-reader-out reference will be used.

The primary analysis will compare absolute AT localization error between assisted and unassisted interpretations using a crossed mixed-effects model. Fixed effects will include assistance condition, physician seniority, the assistance-by-seniority interaction, period, randomized sequence, and case library. Random intercepts will be included for physician and examination. Estimates will be reported with 95% confidence intervals. No simulated or pilot-result data will be included in the prospective study analysis.