Research prototype — Not for clinical use
This repository implements a PyTorch model to predict IVF outcomes for Fresh vs Frozen Embryo Transfer (ET) using synthetic data derived from a published RCT (n=838; 419 Fresh, 419 Frozen).
The model combines a rule-based risk engine with a neural network trained on Age, BMI, and Infertility Duration to recommend transfer type.
Wei, D., Zhang, Y., Zhao, J., et al. (2025). Fresh versus frozen embryo transfer in women with low prognosis: a multicentre, randomised controlled trial. The Lancet, 405(10335), 289–299. https://doi.org/10.1016/S0140-6736(24)02703-2
| Outcome | Fresh | Frozen | RR (95% CI) | p-value |
|---|---|---|---|---|
| Live birth | 40% | 32% | 1.25 (1.06–1.48) | 0.009 |
| Clinical pregnancy | 47% | 39% | 1.19 (1.02–1.38) | 0.02 |
| Ongoing pregnancy | 43% | 36% | 1.19 (1.01–1.40) | 0.04 |
| Pregnancy loss† | 10% | 17% | 0.60 (0.35–1.03) | 0.06 |
| Ectopic† | 4% | 1% | 3.52 (0.75–16.63) | 0.11 |
† % of clinical pregnancies.
-
Rule-based engine
- Encodes trial outcome rates as priors.
- Adjusts for Age, BMI, Duration via multiplicative risk factors.
- Outputs composite risk score + recommendation.
-
Synthetic data generation
- Samples features to match trial baselines.
- Labels via rule-based engine (weak supervision).
-
Neural network
- 3-layer MLP, CrossEntropy loss, Adam optimiser.
- Learns to reproduce rule-based recommendations.
-
Visualisation
- Training curves, outcome probabilities, risk factor multipliers.
- Rule-based risk engine
- Synthetic cohort generator
- Neural network classifier
- Visualisations: training curves, outcome probabilities, risk factor multipliers
- Model persistence (save/load)
Below is a sample risk analysis plot generated by plot_risk_analysis():
- Labels from heuristics, not patient-level data.
- Multipliers illustrative, not validated.
- Probabilities uncalibrated.
- Educational use only.
pip install torch numpy matplotlib
python embryo_transfer.py