Categorical Decision Concordance Between an AI Endometrial-Assessment Platform and Expert Clinicians at the 7-Mm Transfer-Readiness Threshold: A Secondary Analysis of a 2d

Objective: Building on previously reported continuous-measurement performance of an AI-based endometrial assessment platform (ENDOSCAN) versus expert clinicians [1,2], this work evaluates the clinically decisive question that those analyses did not address: categorical decision agreement between AI and clinicians at the 7-mm endometrial-thickness cutoff used for embryo transfer (ET) readiness, benchmarked against inter-clinician agreement on the same cases. Design: Multicentre, retrospective, comparative-accuracy study with paired AI and human reference measurements. Materials and Methods: Two parallel datasets previously described by our group [1,2] were re-analysed for categorical decision concordance: (i) >150 sagittal 2D ultrasound video clips of the endometrium curated from IVF cycles in Poland and the United States, and (ii) a paired 3D ultrasound dataset acquired in mid-luteal or late-follicular phase. The underlying 2D and 3D deep-learning models a thickest-segment localiser and a 3D convolutional segmentation network benchmarked against manual VOCAL have been reported previously and are not the subject of this submission. In the present analysis, each 2D measurement was dichotomised at the 7-mm transfer-readiness cutoff and categorical agreement was quantified using Cohen’s κ for two pairings: (a) AI versus the reference clinician, and (b) independent clinician versus clinician on the same cases (inter-rater benchmark). Continuous-measurement metrics (ICC, MAE, Bland-Altman) and processing time are restated from the prior reports as context. The pre-specified equivalence claim was that AI-clinician κ would not be lower than clinician–clinician κ. Results: At the 7-mm transfer-readiness cutoff, AI-versus-clinician categorical agreement was Cohen’s κ=0.82 (substantial), equalling or exceeding clinician-to-clinician κ=0.78 on the same cases i.e., AI classifications are no more divergent from expert judgement than experts are from one another. The pre-specified equivalence criterion (AI-clinician κ not inferior to clinician–clinician κ) was therefore met. For context, the continuous-measurement performance underpinning these classifications, reported previously [1,2], was: 2D thickness MAE 1.1 mm and ICC 0.94 versus clinicians; 3D volume ICC 0.93 versus VOCAL with processing time reduced from several minutes to <2 seconds. Performance was preserved across clinics and ultrasound platforms. Conclusions: Beyond the continuous-measurement accuracy demonstrated in our prior work, AI-assisted endometrial assessment achieves expert-level categorical decision agreement at the 7-mm transfer-readiness cutoff, with AI-clinician concordance no worse than clinician-clinician concordance on the same cases. This is the clinically operative outcome for FET decision-making and the principal new contribution of the present analysis. Standardising endometrial evaluation in this way may reduce variability in transfer-readiness decisions and support evidence-based optimisation of FET protocols. Support: MIM Fertility, Warsaw, Poland. All authors are employees/co-owners of MIM Fertility. Software is currently Research Use Only. No external funding. Prior dissemination: components of the underlying AI performance data were previously presented [1,2]; the continuous-measurement metrics (MAE, ICC, processing time) cited above are restated from those communications for context. The categorical threshold-agreement analysis (Cohen’s κ at the 7-mm cutoff, with clinician–clinician benchmark) presented here has not been previously reported.