Search for a command to run...
Abstract Study question Can AI-based models reliably predict IVF outcomes by evaluating embryo quality? Summary answer The majority of AI-based models demonstrated high performance in predicting live birth, clinical pregnancy, and embryo ploidy status. What is known already Ensuring precision in embryo assessment is vital for optimizing IVF outcomes, yet traditional methods relying on human judgment remain subjective and introduce interobserver variability. Emerging AI-driven models integrating clinical, demographic, and laboratory data, beyond morphological and morphokinetics data, offer a more objective and comprehensive approach to evaluating embryo potential. Study design, size, duration A systematic review and meta-analysis building upon previous research, incorporating 21 studies published from 2022 onward. This continuation expands the evidence base on AI-driven prediction models in IVF/ICSI cycles, focusing on preimplantation embryos deemed suitable for transfer. The input data is diverse, incorporating demographics data, clinical and laboratory records, morphokinetics and morphological data. The primary outcomes assessed include the predictive performance for live birth, clinical pregnancy, and embryo ploidy status. Participants/materials, setting, methods A systematic literature search was conducted in the PubMed/Medline, Embase, and Web of Science databases, restricted to articles published in English up to January 2025. To ensure continuity with the previous meta-analysis, only studies published from 2022 onward were considered. The initial search identified 1319 studies, of which 57 were duplicates and 1241 were excluded for not meeting the inclusion criteria. After full-text screening and citation mining, 21 studies were deemed eligible for inclusion. Main results and the role of chance For live birth prediction, pooled sensitivity and specificity were 83% (95% CI: 71–90%) and 87% (95% CI: 59–97%), with PPV and NPV of 88% (95% CI: 48–98%) and 82% (95% CI: 76–87%). Clinical pregnancy prediction showed sensitivity of 80% (95% CI: 68–88%), specificity of 78% (95% CI: 69–85%), PPV of 80% (95% CI: 68–88%), and NPV of 72% (95% CI: 67–77%). For embryo ploidy, sensitivity and specificity were 74% (95% CI: 71–78%) and 73% (95% CI: 68–79%), with PPV and NPV both at 74%. Pooled estimates across all outcomes showed sensitivity of 74% (95% CI: 69–79%), specificity of 77% (95% CI: 72–82%), PPV of 78% (95% CI: 70–84%), and NPV of 74% (95% CI: 69–79%). The Area Under the Curve (AUC) of the Summary Receiver Operating Characteristics (SROC) curve was 0.809, and the partial AUC (pAUC) was 0.648, demonstrating consistent and reliable predictive performance across outcomes. Limitations, reasons for caution The limited number of studies meeting the inclusion criteria and the diverse designs used in developing AI models, which can increase heterogeneity, serve as limitations. Including women of all ages is another constraint, as advanced maternal age is linked to reduced IVF success rates. Wider implications of the findings The performance metrics suggest robust predictive capabilities of AI-based models in assisted reproduction. The latest results keep underscoring the potential of AI to complement traditional embryo evaluation methods, providing objective and reliable predictions for IVF outcomes. More research and increased collaboration among developers are essential for further improvements. Trial registration number No