Infertility diagnosis and treatment planning are typically guided by factors such as patient age, ovarian reserve, and sperm condition. However, due to the complexity of these parameters, establishing an effective treatment plan and accurately predicting treatment success—both before and during infertility treatment—is extremely challenging. AI technology can assist healthcare providers in offering more accurate counseling to patients. Additionally, AI can facilitate patient access to fertility-related information and appropriate medical care. Efforts to integrate AI into patient-provider interactions are already underway, and the potential of the large language model-based ChatGPT (Google) service has recently been acknowledged [6]. ChatGPT has demonstrated the capacity to generate responses to infertility-related queries that are comparable in length, accuracy, and sentiment to responses from the US Centers for Disease Control. However, 6.12% of its factual statements were found to be inaccurate, and only 0.68% of responses included citations. Despite its limitations, such as unreliable citation practices and a risk of misinformation, ChatGPT shows promise in addressing clinical questions.

Several AI algorithm models have been developed to diagnose infertility and predict pregnancy success rates using basic patient demographics and data from previous ART cycles. Liao et al. [7] developed a dynamic diagnosis and grading system for infertility by retrospectively analyzing data from 60,648 couples who underwent IVF-ET. Seven key indicators—age, body mass index (BMI), follicle-stimulating hormone (FSH) level, antral follicle count (AFC), anti-Müllerian hormone (AMH) level, number of oocytes, and endometrial thickness—were identified. The weights of these indicators were determined using entropy-based feature discretization and a random forest algorithm. The system categorized infertility into five grades, achieving a predictive stability of 95.94% through cross-validation. The highest grade, A, was associated with a 54% pregnancy success rate, while the lowest, E, had a success rate below 1%. The primary advantages of this scoring system include comprehensive diagnostic capability using various indicators and flexibility to update in real time as new samples are introduced. Additionally, McLernon et al. [8] developed a personalized cumulative live birth prediction model using logistic regression at two critical time points: before starting the first complete IVF cycle and after an unsuccessful initial IVF cycle. Their population-based cohort study included data from 88,614 patients, obtained from the Society for Assisted Reproductive Technology Clinical Outcomes Reporting System (SART CORS) and representing a nationwide United States cohort. The pre-treatment model indicated that cumulative live birth rates declined after age 30 and with higher BMI. AMH levels were significantly correlated with live birth probability up to 2 ng/mL. Key predictors identified included age, BMI, previous full-term birth, male factor infertility, polycystic ovary syndrome (PCOS), diminished ovarian reserve, uterine factor infertility, and unexplained infertility. Each additional IVF cycle reduced live birth odds by 42%, while higher AMH levels increased live birth odds. The post-treatment model indicated that women entering a second cycle had higher rates of diminished ovarian reserve (39%) and PCOS (38%) compared to their first cycle. A higher number of eggs retrieved during the initial cycle was associated with greater live birth odds. Other predictors in the post-treatment model closely aligned with the pre-treatment model. This model is publicly accessible online, allowing patients to input values such as their age, BMI, and the number of previously retrieved eggs to estimate their live birth rate (https://www.sart.org/). The cited studies are detailed in Table 1.

Intrauterine insemination (IUI) has a lower success rate than IVF-ET but is less invasive and more convenient for patients, making it a relatively accessible option for subfertile couples [9]. However, AI applications in ART have predominantly focused on IVF-ET, with limited research addressing its application in IUI. Kozar et al. [9] investigated how AI could improve IUI success rates by selecting appropriate patients for the procedure. They analyzed 1,029 IUI procedures using various stimulation methods. The random forest model exhibited the best performance, achieving an area under the curve (AUC) of 66%, accuracy of 71%, sensitivity of 0.4, and specificity of 0.8. Semen parameters were identified as the most significant predictors. Although this model’s performance was modest compared to AI studies in IVF, the inherently lower success rates of IUI, which have not significantly improved over recent decades, complicate direct comparisons with IVF outcomes.

A large reproductive study introduced a novel method combining complex network-based feature engineering and a stacked ensemble (CNFE-SE) to predict IUI outcomes [10]. This dataset comprised 11,255 IUI procedures and included 20 key predictive features. The model achieved an AUC of 84%, sensitivity of 0.8, specificity of 0.9, and accuracy of 0.9. The most significant predictors were semen parameters (sperm concentration and motility) and female BMI. The superior performance of the CNFE-SE model suggests the potential for its use in a DSS to assist physicians in selecting optimal IUI treatment plans. The cited studies are detailed in Table 2.

AI can improve ovarian stimulation outcomes by assisting clinicians with optimized dose adjustments, precise trigger timing, and better patient monitoring. DSS tools, computer algorithms developed for patient management based on electronic medical records, are increasingly used in clinical settings [11]. Recently, a DSS for IVF management demonstrated high accuracy and alignment with evidence-based expert team decisions in four key areas: stopping or continuing stimulation (0.92), triggering or canceling the cycle (0.96), determining the number of days until follow-up (0.87), and adjusting the dose (0.82) [12]. The algorithm exhibited the lowest accuracy in dosage decisions, attributed to the infrequent and physician-preference-driven nature of dosage adjustments during ovarian stimulation. The trained model tended to be more conservative in maintaining dosages compared to expert decisions; notably, however, when experts opted to increase the dosage, the algorithm never suggested a reduction.

Additionally, AI-based studies have aimed to reduce the frequency of monitoring during COS. Robertson et al. [13] sought to identify the optimal follicular tracking day to predict trigger timing and the risk of over-response. By retrospectively analyzing 9,294 ultrasound scans from 2,322 IVF cycles using random forest regression and classification methods, they identified stimulation day 5 as optimal for predicting trigger timing (mean squared error, 2.16±0.12) and risk of over-response (AUC, 0.91±0.01). Furthermore, Letterie et al. [14] developed an AI-based workflow algorithm intended to limit COS monitoring to a single visit while optimizing and distributing oocyte retrieval schedules. This study defined three key objectives: first, to identify the optimal single day for the monitoring visit based on the pre-treatment clinical profile; second, to provide a range of three potential trigger days to maximize scheduling flexibility without compromising outcomes; and third, to predict the number of mature oocytes. The dataset comprised 1,591 autologous cycles, each from a unique patient. The algorithm achieved a mean prediction error of 1.4 days for predicting the optimal monitoring day, with an accuracy of 0.8. The variance in retrieved oocyte count was 0–3 across the recommended trigger days. Additionally, the sensitivity for predicting mature oocyte count based on pre-treatment profiles was approximately 0.8. Future DSS platforms could thus design individualized workflows using pre-IVF clinical data. These workflows could be adaptively modified based on data collected during COS, enabling outcome prediction and personalized treatment. The cited studies are detailed in Table 2.

The use of COS protocols varies widely depending on clinician, ART center, and patient preferences, leading to a lack of standardized guidelines. AI-driven analysis of diverse and scattered datasets can help select appropriate, personalized stimulation protocols for individual patients [15]. Correa et al. [16] developed an ML model to recommend the initial dose of FSH to achieve 12 mature oocytes across patient types. Using baseline information including age, BMI, AMH, AFC, and prior successful pregnancies from over 3,000 patients across five centers, they analyzed the linear relationship between the initial FSH dose and the number of eggs retrieved. The model significantly outperformed clinician-prescribed doses in both development and validation datasets. However, one limitation of these simple linear models is their failure to capture nonlinear dose-response relationships [15]. Another AI-driven analysis compared COS protocols, analyzing approximately 20,000 cycles for poor responders. The results showed comparable outcomes between the antagonist (n=15,000) and flare protocols (n=5,000) [17]. Given varied outcomes across protocols, the conclusion of Wald et al. [18] that protocol choice appears to only minimally impact outcomes remains relevant. However, future AI studies may contribute to the development of methods that directly predict patient-specific outcomes for each protocol, enabling personalized recommendations for injection types, dosages, and egg retrieval protocols [15].

DSS can also assist in determining optimal trigger timing. A comparative study was performed to determine the optimal trigger timing to maximize the yield of fertilized oocytes (2PNs) and blastocysts [19]. Utilizing causal inference with a T-learner and bagged decision trees, the algorithm, despite having physician-model agreement rates of only 53% and 62%, yielded an average of 1.4 more 2PNs and 0.6 more usable blastocysts compared to physician decisions. Another AI model was designed to predict mature oocyte outcomes for triggering either today or tomorrow, employing follicle counts and estradiol levels in a linear regression and estradiol forecasting model [20]. This study, encompassing three ART centers and 30,278 IVF cycles, demonstrated that early or late triggers resulted in fewer mature oocytes (up to 2.7 fewer), 2PNs (up to 2.0 fewer), and usable blastocysts (up to 1.0 fewer) compared to optimal timing. These AI platforms, though promising, are not yet commercially available; furthermore, the research supporting them has limitations due to its retrospective nature and primary focus on decision timing. Nevertheless, the findings suggest that DSS could effectively support daily COS management decisions.

Ultimately, applying AI in COS requires extensive testing and validation across diverse regions and populations [21]. Recent efforts have focused on developing algorithms that predict the number of mature oocytes, incorporating not only basic patient demographics and endocrine hormone profiles but also genetic information [22]. Therefore, international multicenter prospective studies are essential to ensure that AI algorithms are broadly applicable across clinical settings. The cited studies are detailed in Table 3.

AI research is widely employed in imaging studies, and in the field of ART, AI technology can be applied to follicular monitoring using ultrasound. A randomized controlled trial compared oocyte yield using AI-based three-dimensional (3D) automated volume calculation with traditional two-dimensional (2D) manual ultrasound follicle tracking in women undergoing IVF cycles [23]. The study found comparable ART outcomes between the two groups, with the 2D manual ultrasound group requiring significantly more time for scans (p<0.01). These results suggest that automated 3D ultrasound can be effectively used for follicle tracking, offering time-saving benefits. A DL-based segmentation algorithm using follicle volume measured by 3D ultrasound could also assist in assessing oocyte maturity, optimizing trigger timing, and predicting ovarian hyper-response [24]. The study identified a follicle volume of 0.5 cm³ or greater on triggering day as optimal for predicting mature oocytes, while a threshold of 3.0 cm³ for the leading follicle was significantly associated with a higher yield of mature oocytes compared to traditional 2D ultrasound. Furthermore, a multi-layer perceptron model outperformed traditional 2D ultrasound methods in predicting ovarian hyper-response. In conclusion, AI technology using 3D ultrasound may offer a promising alternative strategy for follicular monitoring beyond traditional 2D ultrasound methods, potentially improving ART success rates and broadening the range of available techniques. The cited studies are detailed in Table 3.

The cumulus surrounding the oocyte poses challenges in assessing oocyte maturity, and selection based on morphology reduces the number of usable oocytes, making oocyte selection uncommon in standard IVF procedures. However, in cases where cumulus removal is necessary for selecting mature oocytes prior to intracytoplasmic sperm injection (ICSI), or specifically in fertility preservation scenarios, non-invasive AI-based methods can be valuable for assessing oocyte maturity and predicting reproductive potential [4].

Violet (Future Fertility), an oocyte AI image analysis tool trained with a convolutional neural network (CNN), predicted fertilization with 82% accuracy, blastocyst development with 67% accuracy, and live birth potential with 75% accuracy per oocyte cohort [25]. These findings suggest that AI-based analysis of oocyte morphology could inform clinical outcomes in IVF and support patient counseling, especially in fertility preservation. Similar software, Magenta (Future Fertility), analyzes oocyte images and scores each oocyte on a scale from 1 to 10 to predict blastocyst development quality [26]. Magenta scores revealed a significant difference in blastocyst development rates between the highest- and lowest-scored oocytes (46% vs. 27%, p<0.005). In a recent study, images of mature oocytes combined with clinical data from patients at eight fertility clinics across six countries were analyzed using AI to assess developmental competence in oocytes likely to progress to the blastocyst stage [27]. The study found the highest predictive performance in the 38 to 39 age group (AUC, 0.68), with minimal impact from male factors. Compared to previous studies on oocyte image analysis, this research incorporates clinical data and extends globally, highlighting the potential to standardize oocyte quality assessment. Although currently challenging to implement as an exclusive tool for oocyte selection or exclusion, it holds promise for application as part of a broader selection toolkit when combined with other clinical indicators, such as sperm evaluation and oocyte gene expression profiling, to accurately predict blastocyst formation and quality [4]. The cited studies are detailed in Table 4.

Computer-aided sperm analyzer (CASA)-based semen evaluation aims to reduce inter-observer variation and has demonstrated good correlation with manual assessments for analyzing sperm concentration and motility. However, for morphological evaluation, CASA typically shows the lowest correlation with manual assessment, likely due to semen sample heterogeneity and the subjective nature of morphological interpretation. Therefore, integrating AI techniques has the potential to improve the accuracy and consistency of sperm assessment [4].

Several studies assessing sperm motility (progressive, non-progressive, and immotile spermatozoa) have utilized the VISEM dataset, a multimodal video dataset of human spermatozoa [28]. Analyzing sperm video sequences using CNN methods or linear support vector regression can provide consistent and efficient predictions of sperm motility within 5 minutes [29,30]. The VISEM dataset was extended to VISEM-Tracking, which includes longer video clips with annotated bounding boxes and sperm tracking information, making it more valuable than the original dataset for training complex ML models [31]. Recently, AI has been used to analyze sperm motility patterns in real time to select sperm for ICSI, reducing the time spent by embryologists during selection and providing an objective method that may improve fertilization and blastocyst rates [4]. Open access to the Modified Human Sperm Morphology Analysis (MHSMA) dataset facilitates studies on sperm morphology analysis. CNN-based DL methods have demonstrated high accuracy, achieving detection rates of up to 94% for morphological deformities in the sperm acrosome, head, and vacuole in real time without staining [32,33]. Furthermore, attempts have been made to integrate AI technology into the operator-dependent field of testicular sperm extraction to increase sperm identification efficiency in testicular biopsy samples and predict the success of sperm retrieval [34,35]. However, these studies have demonstrated only moderate performance, necessitating further external validation. Various methods exist for sperm selection in ICSI, but none have yielded significant improvements in pregnancy rates [36]. A non-invasive AI approach could help build an extensive database and establish a global consensus on sperm selection. Mojo-AISA, a commercially available AI-based sperm analysis system, was developed to address the limitations of traditional CASA systems [37]. It provides accurate semen analysis results in half the time required by manual methods, improving the efficiency of embryologists. Despite challenges with extremely low-concentration samples, it has potential to minimize human error and improve the accuracy of semen analysis. The cited studies are detailed in Table 4.

Embryo grading traditionally involves visual morphological assessment. However, this subjective method leads to inconsistent results among embryologists, contributing to variability and potentially lowering IVF success rates. The introduction of time-lapse imaging (TLI) is known to allow for more comprehensive embryo evaluation by integrating morphokinetic data with the traditional morphological grading system [21]. Consequently, numerous studies have explored AI methods combined with TLI to assess embryos, including the prediction of embryo quality, implantation potential, and pregnancy outcomes.

STORK, an automated blastocyst selection algorithm based on deep neural networks (DNNs), was developed to predict blastocyst quality for single embryo transfer [38]. Using extensive TLI data, blastocysts were classified into good or poor quality based on the likelihood of pregnancy. The DNN was trained over 50,000 iterations using Google’s Inception model. STORK achieved an impressive AUC exceeding 98% in predicting blastocyst quality. Considering patient age, the model predicted a 66% pregnancy success rate for women under 37 years old. STORK demonstrated robustness across datasets from other clinics and outperformed individual embryologists. Another retrospective multicenter study developed IVY, a fully automated DL model that utilizes raw time-lapse videos (TLV) without manual annotation [39]. This study analyzed over 10,000 embryos across four countries. IVY demonstrated an AUC of 93% in predicting whether a blastocyst would result in a fetal heartbeat pregnancy. A recent systematic analysis of 20 studies on AI in embryo assessment demonstrated that AI models consistently outperformed embryologists in predicting clinical pregnancy (accuracy, 77.8% vs. 64%) and embryo morphology grade (accuracy, 75.5% vs. 65.4%) [40]. However, the findings were limited by the lack of prospective clinical evaluations and variability in AI models, databases, and study designs. Additionally, TLV can support DL models in recognizing blastocyst ploidy status. Lee et al. [41] investigated TLV metadata combined with known preimplantation genetic testing for aneuploidy (PGT-A) outcomes. The model, trained to classify blastocysts as either aneuploid or other (euploid and mosaic embryos), achieved an AUC of 74%. This result modestly exceeded those of previous studies, suggesting that further research could contribute to non-invasive PGT-A assessment.

Recently, Illingworth et al. [42] conducted the first multicenter, randomized, double-blind, noninferiority trial comparing DL with manual morphology-based embryo selection in IVF. Among 1,066 patients, the DL group exhibited a pregnancy rate of 46.5%, while the morphology group had a rate of 48.2% (risk difference, −1.7%; p=0.62). This result did not demonstrate noninferiority. However, the algorithm showed a significant and nearly tenfold reduction in evaluation time compared to the morphology group (21.3±18.1 seconds vs. 208.3±144.7 seconds, p<0.001), regardless of blastocyst count. The absence of significant improvement in clinical pregnancy rates with AI technologies raises questions about their cost-effectiveness. Thus, future well-designed randomized controlled trials and robust regulatory measures are essential to prevent premature adoption, protect patients from exploitation, and ensure that the true clinical value of AI is responsibly evaluated [43]. The cited studies are detailed in Table 5.