Can AI predict miscarriage after IVF?

Artificial intelligence (AI) could help clinicians estimate the risk of miscarriage after IVF, using five measurements routinely collected during treatment.

Researchers in China developed a computer model that uses information about a patient's age, body mass index (BMI), hormone and blood markers to estimate the likelihood of early pregnancy loss.

The authors of the study, published in BioData Mining, said: 'The model is intended to support risk stratification and counselling,' but added that it 'requires external validation' before clinical use.

The team analysed medical records from 1201 couples who had undergone their first cycle of IVF or ICSI at Shanghai First Maternity and Infant Hospital. All had achieved a pregnancy confirmed by ultrasound. Of these pregnancies, 222 (18.5 percent) ended in early loss, defined as miscarriage within the first 12 weeks.

The researchers tested eight different machine learning methods to see how accurately they could predict the outcome of a pregnancy. The best-performing model, called XGBoost, used five factors: the woman's age, BMI, levels of follicle-stimulating hormone (FSH), thyroid-stimulating hormone (TSH), and platelet count. FSH, a hormone involved in the development of eggs, was the strongest predictor, followed by BMI and TSH.

When tested on a separate group of patients from the same hospital, the model showed a high ability to distinguish pregnancies that resulted in a live birth from those that ended in early loss. The researchers also developed an online calculator that allows clinicians to enter the five measurements and obtain an individual risk estimate.

However, the tool is not designed to predict whether a transferred embryo will implant in the first place: it can only be used after an ultrasound has confirmed a pregnancy. Nor can it explain why a miscarriage might happen, as pregnancy loss following fertility treatment can have several underlying causes, bringing significant emotional challenges for patients (see BioNews 1179).

The study also has limitations. It was based on records from a single hospital and did not include factors such as embryo quality, genetic information or immune status. The model therefore needs to be tested in larger studies involving different fertility clinics before it can be considered reliable for wider use.

If further research confirms its accuracy, the approach could eventually help fertility teams identify patients who may benefit from additional support during early pregnancy (see BioNews 1290).

---

The use of AI in fertility and genomics will be explained and explored at this year's PET Annual Conference – AI and Automation in Fertility and Genomics: Pipeline? Or Pipe Dream? – taking place in central London on Wednesday 9 December 2026.

Find out more and register here.