Introduction
In recent years, artificial intelligence (AI) has become a powerful tool in various aspects of our daily lives. One of the most particular areas revolutionized by AI is medicine (1-4), in which AI has the capability of addressing complex challenges both in diagnosis and therapy. Reproductive medicine, as an evolving branch of medicine concerned with the clinical management of infertility, is no exception to this integration (5). Infertility affects approximately 9 to 15% of couples across the world and is associated with profound physical, emotional, and financial complications (6). While traditional diagnostic and treatment approaches often depend on subjective assessments and rigid protocols that fail to reflect unique variations among patients, AI holds immense potential to transform this matter. By utilizing large-scale data and recognizing intricate patterns via algorithms, AI presents an opportunity to advance infertility care (7). Such progress has the potential to advance diagnostic precision, support individualized care, and yield more effective treatment outcomes (7).
Many subtypes of AI can aid in this integration, consisting of machine learning (ML), deep learning (DL), natural language processing (NLP), and computer vision. These technologies support areas such as image analysis, treatment prediction and planning, embryo selection in IVF, and the discovery of fertility biomarkers (5). Predicting semen parameters—such as sperm concentration and motility—based on environmental factors and lifestyle is another significant task that can be accomplished using AI-based decision support tools (8). Similarly, deep learning models have demonstrated high accuracy in selecting embryos for transfer, showing potential to support embryo prioritization in IVF (9). 
Despite the aforementioned advancements, significant challenges remain. Ethical concerns, data privacy issues, validation requirements, and the essential need for a structured framework are among the notable instances. Additionally, since clinicians depend on transparent decision-making pathways, the nature of many AI models and their interpretability issues continue to delay their acceptance and integration into clinical practice (10). The current literature on this topic is fragmented and lacks a comprehensive review of AI’s effectiveness, implementation challenges, and future directions in infertility care. This narrative review was conducted to evaluate the current applications of AI-powered tools and algorithms in the diagnosis and management of infertility, and highlight its potential benefits, limitations, and future directions. By synthesizing existing evidence, our main purpose is to establish a foundation for further development and clinical application of AI-driven solutions in reproductive medicine.

Methods
In this narrative review, studies that focus on the use of AI and AI‑powered tools to aid the diagnosis and management of infertility were examined. The study’s defined keywords were searched using online sources such as PubMed/MEDLINE, Web of Science (WoS), Scopus, Embase, Cochrane Library, Google Scholar, and IEEE Xplore. To identify all relevant publications, comprehensive search queries were created that combined the pertinent keywords using Boolean operators. The search was completed on October 25th, 2025. Two keywords and their synonyms were used, applying one search strategy as detailed below in table 1. The results were compiled in an EndNote library, with duplicates removed. Following this, screening, selection, and data extraction steps were conducted.
The following search query was applied in the PubMed/MEDLINE database: ((Infertility [Title/ Abstract] OR "Assisted reproductive technology" [Title/Abstract] OR IVF [Title/Abstract] OR "Embryo selection" [Title/Abstract] OR "Sperm analysis" [Title/Abstract] OR "Ovarian reserve" [Title/Abstract] OR Subfertility [Title/Abstract] OR Sterility [Title/Abstract]) AND ("Artificial intelligence" [Title/Abstract] OR "Machine learning" [Title/Abstract] OR "Deep learning" [Title/ Abstract] OR "Neural network" [Title/Abstract] OR "Computer vision" [Title/Abstract] OR "Predictive modeling" [Title/Abstract] OR Algorithm [Title/Abstract] OR "Natural language processing" [Title/Abstract] OR "Computer reasoning" [Title/Abstract] OR "Convolutional neural network" [Title/Abstract])).
This phase was divided into two distinct parts. Initially, two researchers reviewed the titles and abstracts of the articles to determine their relevance to the study’s concept. In the second part, the researchers carefully reviewed the full texts of the screened articles. Publications that fulfilled the eligibility criteria were then selected for data extraction. 

Table 1. Overview of keywords 

Original studies, written in English, published between 2013 and 2025, with relevance to the subject of review, were included. Review articles, case series, case reports, duplicate publications, conference abstracts, opinion pieces, editorial letters, preprints, abstracts lacking accessible full texts, articles not written in English, and articles not pertinent to the study’s aim were excluded. After removing duplicates using referencing software, the screening process of titles and abstracts and then the full text of the articles was carried out using a semi-automated workflow by two independent authors.

Results
A total of 4376 records were recognized through a research strategy. After removing 505 duplicates, 3871 titles and abstracts were screened, of which 3561 were removed due to lack of relevance to the topic, and 310 articles remained for full-text review. In the next phase, 292 records were excluded after full-text review, and finally, 18 studies met the eligibility criteria and were included in the final review.
After finalizing the screening process, 18 studies met the eligibility criteria for our review and were scrutinized. These studies were explored, aiming to extract data such as the year they were conducted, the country of origin, their main objectives and study type, the AI subset that was utilized in the study, population, and finally, the key points and main conclusions. The findings are presented in table 2. The included studies were primarily conducted in China, followed by the USA and Australia, with the remainder originating from various countries across Asia, Europe, and the Americas (Figure 1). 
Among the various AI technologies used, machine learning was the most prevalent, appearing in 10 studies. Deep learning was utilized in six studies, and a convolutional neural network was employed in two studies. Artificial neural networks and deep neural networks were used in one study each. Articles were classified into four categories based on clinical application. The majority focused on "Embryo Selection and Implantation Prediction (IVF/ICSI [intracytoplasmic sperm injection])" (8 studies), followed by "Ovarian Stimulation, Reserve, and Fertilization Outcomes" (3 studies). Other applications included "Male Infertility" (2 studies) and "Broader Reproductive Outcomes" (5 studies) (Figure 2). The 18 studies were conducted between 2013 and 2025, with 2025 being the most frequent year, accounting for 6 studies, followed by 2023 and 2024, accounting for 3 studies each. The majority of studies employed observational designs, often retrospective, though a few randomized controlled trials (RCTs) were also reported.

Table 2. Description of the findings reported in eligible studies 

 AI: Artificial intelligence, ML: Machine learning, PCOS: Polycystic ovary syndrome, IVF: In vitro fertilization, ICSI: Intracytoplasmic sperm injection, DL: Deep learning, CNN: Convolutional neural network, KID: Known implantation data, FSH: Follicle stimulating hormone, OAT: oligoasthenoteratozoospermia, ART: Assisted reproductive technology, SBT: Single blastocyst transfer

Discussion
Infertility remains a significant global health concern, affecting millions of couples of reproductive age worldwide, with reported cases and clinical detection steadily increasing due to improved awareness and enhanced diagnostic capabilities (27). ARTs serve as the cornerstone of infertility management; however, traditional approaches are often limited by subjective assessments, inter-observer variability, and rigid protocols that fail to account for individual patient variations (28). The integration of AI into this field offers a paradigm shift by introducing objectivity, precision, and personalization into fertility care. Our review synthesizes evidence from 18 studies to map the current landscape of AI applications in infertility management and critically evaluate their potential, limitations, and future directions.
Embryo selection and implantation prediction (IVF/ ICSI): Our findings demonstrate that AI has shown promising performance compared with conventional morphology-based grading in several retrospective studies. The high accuracy achieved by CNN-based models (up to 90%) and the STORK deep neural network in blastocyst quality assessment underscores AI’s capacity to reduce subjective bias and inter-observer variability (11, 12). This is clinically significant because embryo selection is arguably the most critical determinant of IVF success, and even marginal improvements in accuracy can translate into meaningful increases in live birth rates and reductions in multiple gestation rates. 
However, the recent randomized controlled trial by Illingworth et al. (2024) that failed to demonstrate non-inferiority of AI-based selection compared to standard morphology (25), raises important questions about the generalizability and clinical readiness of these tools. This discrepancy between retrospective and prospective findings is not unique to reproductive medicine; it mirrors challenges seen in other AI applications in healthcare, where models trained on curated, single-center datasets often underperform in real-world, multi-center settings (10). The modest predictive accuracy reported by Boucret et al. (2025) (18) further reinforces that model performance is highly context-dependent and influenced by dataset characteristics, imaging protocols, and laboratory conditions.
These findings collectively suggest that while AI holds immense promise for standardizing embryo assessment, its current evidence base is insufficient to recommend widespread clinical adoption without rigorous prospective validation. The absence of standardized benchmarks for embryo quality remains a critical gap, as it impedes direct comparison of AI models and complicates regulatory approval pathways (28).
Ovarian stimulation, reserve, and fertilization outcomes: The integration of machine learning models to predict ovarian reserve and fertilization outcomes represents a promising avenue for personalizing ART protocols. Our review indicates that AI can accurately predict ovarian reserve using routine clinical and biochemical parameters (13), effectively forecast fertilization failure (16), and guide FSH dosing and trigger timing (26). These capabilities are particularly valuable because ovarian stimulation protocols are currently guided by generalized algorithms that often result in suboptimal responses, either through under-stimulation (leading to poor oocyte yield) or over-stimulation (increasing OHSS risk).
The ability to predict ovarian reserve and fertilization outcomes at the outset of treatment could enable clinicians to stratify patients into risk categories, select optimal stimulation regimens, and avoid unnecessary cycles or cancellation. This aligns with the broader movement toward personalized medicine, where treatment decisions are informed by individual patient characteristics rather than population averages (29). However, as noted by AlSaad et al. (2024) (30), most AI models in this domain are derived from single-center datasets, limiting their generalizability across diverse patient populations and clinical settings.
Emerging AI technologies are moving beyond static predictions toward dynamic, full-cycle stimulation management. By integrating multi-parametric data, these models have the potential to predict individualized responses to agonist/antagonist regimens, refine gonadotropin dosing in real-time, and accurately forecast OHSS risk (30, 31). Compared with conventional clinical stratification tools such as the POSEIDON and Bologna criteria, AI-based systems demonstrate improved precision in differentiating poor and hyper-responders, suggesting that they may eventually replace or augment traditional risk assessment tools.
Male infertility and broader reproductive outcomes: Our review highlights AI’s emerging role in male infertility assessment, particularly through automated prediction of semen parameters (8), varicocele classification (24), and potential applications in sperm morphology and DNA fragmentation analysis. These tools could address a significant unmet need, as semen analysis remains largely subjective and variable across laboratories. By introducing standardization and objectivity, AI-powered sperm assessment could improve diagnostic accuracy and facilitate early intervention.
The application of AI to broader reproductive outcomes, including live birth prediction in PCOS patients, couple-level infertility classification (22, 23), and miscarriage risk assessment and live birth outcomes (19, 20), represents an exciting frontier. These studies collectively suggest that AI can synthesize complex, multi-dimensional data (clinical, biochemical, and demographic) to generate individualized prognostic information that could enhance patient counseling, treatment planning, and shared decision-making. The ability to predict live birth outcomes with machine learning models, for example, could help manage patient expectations and guide decisions about whether to pursue additional IVF cycles.
Our findings are broadly consistent with recent systematic reviews and meta-analyses in this field. Wu et al. (2025) (28) similarly noted that AI demonstrates the potential to enhance accuracy and implantation prediction, yet large-scale prospective validation remains limited. Kakkar et al. (2025) (32) and Olawade et al. (2025) (33) emphasized the gap between retrospective AI model development and clinical implementation, highlighting the need for standardized benchmarks, multi-center trials, and regulatory frameworks. The inconsistency between retrospective and prospective findings observed in our review mirrors concerns raised across the broader AI in healthcare literature, where models that perform exceptionally in controlled settings often fail to replicate in real-world clinical environments (10, 34).
The fusion of AI with this field presents both opportunities and ethical challenges. The main issue that leads to skepticism among both patients and clinicians is the "black box" nature of these algorithms. These technologies lack interpretability, and in a setting where the patient is already vulnerable due to the infertility issue at hand, they create an atmosphere of anxiety and helplessness. Transparent decision-making and treatment planning are a necessity. On the other hand, clinicians can either over-rely on AI outputs or dismiss them altogether, consequently worsening the issue (34).
Secondly, small sample sizes and retrospective datasets limit the generalizability of these approaches and introduce various algorithmic biases. Lack of structured frameworks on this integration leaves questions on data privacy unresolved (34). Critically, these technologies risk reinforcing value judgments, such as "good" versus "bad" embryos and lead to social consequences of AI-driven embryo ranking (34). The ethical concerns in this area extend beyond the "black box" nature. Researchers must also scrutinize issues of interpretability, clinical accountability, and algorithmic bias. Factors such as limited data diversity and the absence of regulatory frameworks should be addressed before moving further. AI-powered technologies are a significant augmentation tool, but without proper oversight, biased value judgments still exist. In light of that, the development of a structured and standard framework to clarify clinical responsibility is a crucial requirement to ensure that AI supports, rather than compromises, ethical fertility care (7). To ensure AI becomes an augmentation tool rather than a harmful tool, future directions must focus on the previously mentioned challenges.
A limitation of the current study was the narrow scope of the studies scrutinized. A total of 18 studies may not fully represent the growing body of literature with this focus. English-language studies were only included which raises the probability of selection bias. Eligible studies were mainly single-centered and were mostly retrospective. This restricts the generalizability of evidence compared with large samples. Moreover, sample sizes of included studies were often modest, which exaggerates the aforementioned bias. Furthermore, most AI-powered tools and algorithms have a "black-box" nature, meaning there is no clarity on how a prediction is made by a model. Additionally, no formal risk-of-bias assessment or quality grading (e.g., GRADE) was performed. Thus, the potential for publication bias should also be acknowledged. Finally, the necessity for a regulated framework to address ethical issues of AI integration remains. 
The fusion of AI with the realm of infertility management holds great promise for revolutionizing this subspecialty. Future studies should focus more on multicenter trials with larger sample sizes and standardized frameworks to validate AI-powered models across diverse populations. To gain clinicians’ trust, a more structured framework for ethics, data privacy, and transparency must be implemented. By addressing these issues, the way towards the utilization of AI-based models from experimental studies into clinical routine will be paved.

Conclusion
This review shows that AI—particularly machine learning and deep learning—has been applied across four domains of infertility management: embryo selection and implantation prediction, ovarian stimulation/reserve and fertilization outcomes, male infertility, and broader reproductive outcomes. Deep learning models such as CNN-based systems and the STORK network achieved high accuracy in blastocyst quality assessment and often outperformed embryologists in retrospective settings; machine learning models predicted ovarian reserve and fertilization failure, and showed promise for live birth and miscarriage prediction.  Current evidence supports AI as a decision-support tool that augments—rather than replaces—clinician and embryologist judgment, with multi-center prospective trials and standardized, transparent frameworks needed before routine clinical adoption. To ensure the safe implementation of artificial intelligence in clinical settings, it is recommended that future studies focus on multicenter prospective trials, transparency, and standardized frameworks.

Acknowledgement
AI tools were used only for language refinement and improving the manuscript’s readability. The authors performed all data collection, interpretation, and referencing without reliance on AI-generated content.
Funding: This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

Conflict of Interest
The authors declare that there is no conflict of interest regarding the publication of this manuscript.