The fertility sector is experiencing an artificial intelligence (AI) boom, aiming to transform processes ranging from time-lapse embryo imaging to complex decision-making in patient management. The possibilities are undeniably exciting, but many question whether these technologies are driving positive change for those on a fertility journey.
AI’S PENETRATION INTO FERTILITY CARE
‘AI for sperm analysis isn’t simply one tool, it is a diverse toolkit. Success will depend on integrating robust biophysical knowledge into transparent, interpretable models that clinicians and reproductive scientists can confidently employ.’
One of the biggest technological success stories in fertility is time-lapse imaging. Time-lapse incubators capture images of developing embryos approximately every 10 minutes, avoiding the need for removal for examination under a microscope. This approach has many advantages, not least that embryos remain undisturbed in a closed environment, while a rich data source is created for analysis.
It seems logical, then, that AI models could be trained on these data to select the best embryo for transfer, thereby improving treatment success. However, repeated studies have shown that the addition of AI to closed-culture systems does not improve the chances of leaving with a healthy baby.1 As a case study, AI embryo selection illustrates the complexities of fertility care – there may be significant value in the underlying technology, but the addition of ‘AI’ is often used more for patient marketing than for truly improving outcomes.
A promising use of AI in fertility care is in the detection of sperm in samples with very low counts.2 Here, AI is being employed as a tool to search tens of thousands of images to find single sperm for intracytoplasmic sperm injection (ICSI). This task can take embryologists hours to perform, and so AI frees resources and supplements and improves workflows, rather than aiming to replace expert insight.
UNDERPINNING MODELS AND DATA ARE CRUCIAL FOR SUCCESS
The true potential of AI in fertility lies not necessarily in its analytical power, but in the opportunity for interpretation of models. Approaches that are purely data-driven risk spurious correlations, particularly in large datasets, which may not translate to an individual’s chance of success. Data quality – in terms of accuracy and quantity – is key, but it is equally as important that measurements with a meaningful relationship to outcomes are included.
A good example of this is in sperm analysis. Here, ‘head-tracking’ parameters (often used to quantify sperm motility) have historically been uncorrelated with clinical outcomes, and yet these same measurements have been used in large-scale AI models to predict treatment success. The problem in this situation is clear: large datasets can inadvertently foster data ‘hallucinations’: patterns confidently identified by algorithms, despite lacking biological validity.
As an alternative, ‘white-box’ approaches aim to include known biophysical, mechanical or pharmacological frameworks in predictive models. These ‘mechanistically informed’ algorithms offer results that are transparent, interpretable and rooted firmly in biology, significantly reducing the risk of overfitting or false positives, and making them uniquely suitable for driving genuine advances in fertility care.
[Below is an ilustration of a model-based learning approach. Videos of swimming sperm (a) can be integrated into an AI model to extract the flagellar waveform (b). This, in turn, can be coupled with biophysical simulations to extract meaning (c–d). In this example, insight is drawn from a mathematical model that calculates the power (related to metabolic output) required by the sperm as it swims.]
WHY IT MATTERS: SPERM AT THE HEART OF FERTILITY
It is only in the last few years that we have begun to really understand the impact men, and indeed sperm, have on the development of a healthy child. A retrospective study of 2425 couples with unexplained infertility undergoing in vitro fertilisation/ICSI treatment in Australia revealed the significant impact that male age has on success.3 At the sperm level, mechanistic analysis of the HABSelect clinical trial (2722 couples in 16 centres across the UK), which compared the impact of changing the individual sperm selected in ICSI, highlighted that a significant proportion of age-related miscarriage after 8 weeks is due to sperm.4 These signs point to factors that AI should be able to help uncover, provided we can employ the right models, with the right data, in the right way.
HOW CAN AI HELP? WHITE-BOX MECHANISTIC MODELLING OF SPERM
AI for sperm analysis isn’t simply one tool, it is a diverse toolkit. Success will depend on integrating robust biophysical knowledge into transparent, interpretable models that clinicians and reproductive scientists can confidently employ. At the Centre for Reproductive Science in Birmingham we are developing white-box AI approaches that combine fluid dynamics, energy metabolism and sperm mechanics to bridge the gap between complex biological phenomena and clinical implementation.
Our driving principle is that we know that sperm affect success, but existing measurements do not have the discriminatory power to draw prognostic insight. Instead, we turn our attention to the sperm flagellum (tail).5 It is the flagellum that responds to environmental signals, reacts to the physical environment, and is sensitive to subtle changes in metabolism and morphology. By integrating new measurements of sperm with clinical factors incorporating both partners, we hope to provide actional insights by grounding predictions in validated underlying mechanisms.
We believe that the key to fully unlocking AI’s potential for integrating understanding of sperm in treatment lies in the interdisciplinary collaborative environment we foster. Here, we will bring together reproductive scientists, mathematicians and data-focused researchers, reproductive biologists, clinicians and patients themselves, to ensure that AI’s transformative potential genuinely enhances fertility care, rather than offering just another technological novelty.
MEURIG GALLAGHER
Associate Professor of Interdisciplinary Healthcare Science, Centre for Human Reproductive Science, Metabolism and Systems Science, College of Medicine and Health, University of Birmingham
REFERENCES
1. Bhide P et al. 2024 Lancet https://doi.org/10.1016/S0140-6736(24)00816-X.
2. Goss DM et al. 2024 RBMO https://doi.org/10.1016/j.rbmo.2024.103910.
3. Horta F et al. 2019 Human Reproduction https://doi.org/10.1093/humrep/dez223.
4. West R et al. 2022 Human Reproduction https://doi.org/10.1093/humrep/deac058.
5. Gallagher MT et al. 2019 Human Reproduction https://doi.org/10.1093/humrep/dez056.
Illustration of a model-based learning approach. Videos of swimming sperm (screenshot seen in a) can be integrated into an AI model to extract the flagellar waveform (b). This, in turn, can be coupled with biophysical simulations to extract meaning (c–d). In this example, insight is drawn from a mathematical model that calculates the power (related to metabolic output) required by the sperm as it swims.
