P-655 A systematic review to determine the best-performing prediction models for natural conception in couples without major barriers to conception and conception following IUI and IVF
Abstract Study question Which prediction models are best in terms of methodological quality, performance, and low bias risk for estimating pregnancy outcomes following natural conception, IUI, and IVF? Summary answer The following 5 models had good performance and low risk of bias: Hunault, McLernon, and Van Eekelen (natural conception), Steures (IUI), and McLernon 2022 (IVF). What is known already Prediction models constitute an integral part of prognosis-based infertility care, enabling stratified management strategies. Despite the number of models for natural conception, IUI, and IVF, appropriate validation and assessment of clinical utility remains underexplored. Systematic assessments employing standardised frameworks are necessary to ensure these models provide reliable estimates and do not misguide clinical decision-making. ESHRE guidelines emphasise the adoption of prognosis-driven approaches, particularly for unexplained infertility, yet evidence supporting the use of specific validated models in routine clinical practice is limited. This review addresses the gap by systematically evaluating prediction models for quality, performance, and applicability. Study design, size, duration Three systematic reviews assessed prediction models for natural conception, intrauterine insemination (IUI), and in vitro fertilization (IVF). Systematic searches were conducted in OVID MEDLINE, EMBASE, and PsycINFO for studies published from 1978 to June 2024, focusing on the development and validation of models predicting conception, pregnancy, or live birth. A total of 93 studies were included: 19 on natural conception, 20 on IUI (seven incorporating machine learning), and 54 on IVF (11 employing machine learning). Participants/materials, setting, methods Models for natural conception addressed populations with unexplained infertility or no significant barriers to conception, while IUI and IVF models focused on treatment-specific populations. This review updated the evaluation of IVF models from 2019, complementing previous reviews. Both statistical and machine learning models were assessed for calibration and discrimination (C-index). Data were extracted using the CHARMS checklist, and the risk of bias was evaluated using the PROBAST tool, ensuring adherence to rigorous quality standards. Main results and the role of chance Natural Conception: The Hunault, McLernon, and Van Eekelen models showed good calibration and modest discrimination (C = 0.59–0.64). The Hunault model provides a single-point probability estimate useful for initial counselling. The McLernon (2019) and Van Eekelen models dynamically update probabilities, with McLernon estimating conception within six months (revised monthly) and Van Eekelen predicting one-year probabilities (revised every six months). All three models were rated as low risk of bias by PROBAST. IUI: The Steures model exhibited a discrimination of 0.59 (95% CI 0.57–0.61). Calibration showed less than 0.5% error between observed and predicted outcomes. As per PROBAST, it remains the only validated IUI model assessed as low risk of bias. In contrast, seven machine learning models were reviewed and were at high risk of bias. IVF: The McLernon (2022) model demonstrated discrimination of 0.71 (without AMH) and 0.73 (with AMH), with good calibration. The pretreatment model estimated cumulative live birth probabilities over three IVF cycles, while the posttreatment model updated predictions for couples starting their second cycle. Of the 11 machine learning models reviewed, all were at high risk of bias. Recommendations: Hunault, McLernon, and Van Eekelen are suggested for natural conception, Steures for IUI, and McLernon 2022 for IVF. Limitations, reasons for caution The reviews highlight a lack of sufficient external validation and standardised evaluation frameworks for machine learning models, limiting their meta-analysis. While these models are recommended, they should be validated locally and recalibrated, if necessary, before being used in clinical practice to ensure they provide accurate and reliable predictions. Wider implications of the findings Validated models such as Hunault, McLernon, Van Eekelen (natural conception), Steures (IUI), and McLernon (IVF) can optimise prognosis-based care but require further validation. Machine learning models require validation and possible recalibration before adoption. Greater emphasis on validation studies rather than developing additional models is essential to strengthen their clinical utility. Trial registration number No
Read more