Quick answer
Controlled ovarian stimulation is the part of IVF that most depends on clinical judgement: too little gonadotropin and few eggs are retrieved, too much and the patient risks ovarian hyperstimulation syndrome and has to cancel the fresh transfer. Zhu Yimin's reproductive endocrinology team at Zhejiang University's women's hospital, with a multicentre group, built an artificial intelligence system that predicts ovarian response before the cycle starts and proposes a starting protocol. It was published in BMC Medicine [1][2].
Key takeaways
- Developed by Zhu Yimin's team with multiple centres; published in BMC Medicine [1][2].
- Function: predict ovarian response before stimulation and individualise the protocol [1].
- Targets the two failure modes — poor response and hyperresponse with OHSS [1].
- Context: more than one million IVF cycles are performed in China each year [1].
- The department also uses AI embryo scoring and a single-embryo-transfer decision model [4].
Why stimulation is hard to individualise
Age, antral follicle count and AMH predict ovarian reserve roughly, but two patients with the same AMH can respond very differently to the same dose. Clinicians therefore adjust empirically, cycle by cycle, and the cost of a wrong first guess is a cancelled cycle or a hospital admission [1][2].
The AI approach uses the full set of pre-cycle variables — hormone levels, follicle counts, body mass, prior cycle outcomes — rather than a single number, and outputs both a predicted response category and a suggested starting dose. Zhu has also written and lectured on stimulation protocols for older women, the group in whom response prediction is least reliable [3]. That is a decision-support framing, not autonomy: the clinician still writes the prescription and adjusts during monitoring [1].
What it changes, and what it does not
If the prediction is accurate, the benefit is fewer cancelled cycles and fewer OHSS admissions. What it does not change is the biology of the eggs retrieved: stimulation determines how many oocytes are available, not their chromosomal normality, which is what limits live birth per cycle in older patients [1][4].
The publication reports model development and validation, not a randomised trial showing higher live birth rates when clinicians follow the recommendation. Models trained on retrospective cycle data inherit the prescribing habits in that data, so a recommendation can encode past practice rather than optimal practice. Whether an AI-suggested dose outperforms an experienced clinician's judgement is the question that still needs answering [1][2].
The condition in context: IVF in China
Demand for assisted reproduction is rising as childbearing age increases, and stimulation is the most resource-intensive phase.
| Region | What the data shows | Source |
|---|---|---|
| Volume | More than 1 million IVF cycles a year in China, with demand still rising [1]. | ZJU Women's Hospital [1] |
| Failure modes | Low responders retrieve few oocytes; high responders risk OHSS and cycle cancellation [1]. | ZJU Women's Hospital [1] |
| Embryo selection | AI scoring of embryos has been reported to raise clinical pregnancy rates versus morphological grading [4]. | TOP IVF summary [4] |
Because stimulation drugs are a large share of IVF cost, getting the dose right first time has an economic as well as a clinical effect [1].
Why it matters
| Potential | What the report suggests | Source |
|---|---|---|
| Safety | Fewer hyperresponse cycles and OHSS admissions [1]. | [1] |
| Efficiency | Better first-cycle dosing, fewer cancellations [1][2]. | [1][2] |
| Caution | No trial of live birth outcomes with AI-guided dosing [2]. | [2] |
Availability and limits
- This is an editorial summary of the publication, not a treatment offer [1][2].
- The model is a research system; it is not a licensed medical device [2].
- It predicts response, not live birth [1].
- Performance in centres with different protocols is unverified, and the team's own clinical focus is diminished ovarian reserve and older patients [2][5].
- Clinician judgement and ultrasound monitoring remain essential [1].
What to ask before booking
| Question | Why it matters |
|---|---|
| What is my predicted ovarian response category? | It determines starting dose and monitoring intensity [1]. |
| How will OHSS risk be managed if I respond strongly? | Freeze-all and agonist triggering are established options [1]. |
| Which department handles foreign patients, and in what language? | ZJU Women's Hospital is a working hospital we do not book; confirm international-department coordination and language support. |
| Is AI used for embryo selection here, and what is the evidence? | Its reported benefit is pregnancy rate, not live birth [4]. |
Questions patients ask
What was developed?
An AI-assisted decision system that predicts a patient's ovarian response before IVF stimulation and recommends an individualised protocol [1].
What problem does it solve?
Poor response in some patients and ovarian hyperstimulation syndrome in others, both of which follow from imprecise first-cycle dosing [1].
Where was it published?
BMC Medicine [1][2].
Does it improve live birth rates?
Not established — the publication reports model development and validation, not a live-birth outcome trial [2].
Is this relevant to international patients?
It is useful context for China's IVF practice. ZJU Women's Hospital is a working hospital we do not book — contact its international department directly [1].