Real-world implementation of AI-assisted starting gonadotropin dosing in routine ovarian stimulation: an exposure-defined, propensity score-matched observational cohort study

Introduction Starting gonadotropin dose selection in controlled ovarian stimulation is conventionally guided by clinician experience. This pilot, single-centre, exposure-defined, matched observational cohort study assessed the feasibility of deploying an artificial intelligence-assisted gonadotropin dosing tool (POvaStim) in routine practice, comparing clinicians using the tool versus conventional experience-based judgement. Because AI use was completely confounded with clinician group and clinician experience was not prospectively quantified, the independent contribution of the decision-support tool cannot be isolated. Methods Consecutive women undergoing IVF/ICSI with a GnRH-antagonist protocol were treated between August 3 and December 31, 2022; only the starting dose was AI-assisted. One-to-one propensity score matching on body weight, ovarian reserve score and gonadotropin preparation yielded 119 matched pairs. Primary outcomes were cumulative clinical pregnancy and live birth per oocyte retrieval (data-lock 31 December 2024). Results Cumulative clinical pregnancy occurred in 67 of 119 (56.3%) in the AI-assisted arm versus 76 of 119 (63.9%) in the control arm (risk difference −7.6%, 95% CI −20.0 to +4.8; P = 0.234), and cumulative live birth in 59 of 119 (49.6%) versus 60 of 119 (50.4%) (risk difference −0.8%, 95% CI −13.5 to +11.9; P = 0.897), counting the 16 women without an ascertainable outcome by the data-lock date as non-events. A complete-case analysis excluding these 16 women gave materially unchanged results. These wide confidence intervals do not exclude clinically important differences; the study was not an equivalence or non-inferiority trial. Starting-dose selection was substantially less variable in the AI-assisted arm (interquartile range 43% narrower; variance 2.43-fold lower, 95% CI 1.69–3.50). Gonadotropin exposure was nonetheless higher in the AI-assisted arm (+27.4 IU/day). POvaStim recommended similar doses in both arms: the AI-assisted dose was rounded upward from it in 94.1% of cycles, while control-arm clinicians, unaware of the recommendation, more often prescribed below it. Elective freeze-all, chiefly for hyperstimulation prevention, was more frequent in the AI-assisted arm (46.2% versus 34.5%; risk difference +11.8%; P = 0.064), although none developed severe hyperstimulation requiring admission; the study was not powered for safety endpoints. Discussion AI-assisted starting-dose selection was feasible, with less variable dosing, but given the confound between AI assistance and clinician experience, these differences cannot be attributed to the tool itself.