Machine Learning for Live-Birth Prediction After IVF/ICSI From Structured Clinical Data: A Systematic Review, Meta-Analysis, and Critical Appraisal
Organizations: A.R. Shahzad, M.Z. Khan, U. Ajaib, A.U. Rehman and J. Deng are with the School of Computer Science, Wuhan University, Wuhan 430072, Hubei, China. · M. Zhang is with the Reproductive Medical Center, Zhongnan Hospital of Wuhan University, and the Clinical Medicine Research Center of Prenatal Diagnosis and Birth Health in Hubei Province, Wuhan, Hubei, China.
Abstract
Machine-learning models predicting live birth after in vitro fertilisation (IVF) or intracytoplasmic sperm injection (ICSI) from routine clinical data report performance from near chance to near perfect. We systematically reviewed these models under PRISMA 2020, searching OpenAlex, Scopus and Web of Science to 22 October 2026, appraised them with PROBAST+AI and TRIPOD+AI, and pooled their discrimination with random-effects meta-analysis on the logit scale. We define what fixes the meaning of a model's output (decision point, endpoint, unit of analysis, data source) and map the 48 included studies by model family, decision point and validation design. Models evaluated only internally pooled to an area under the receiver operating characteristic curve (AUC) of 0.814, against 0.756 (95% prediction interval 0.546-0.889) for models evaluated on a later period or another centre, and 0.764 for studies not at high risk of bias; neither difference was significant, and heterogeneity was extreme. Models whose decision point was not stated pooled to 0.910, the only significant moderator in meta-regression. Against logistic regression fitted to the same data, the best machine-learning model gained a pooled 0.056 AUC (95% CI 0.017-0.090). Of 48 studies, 47 were at high or unclear risk of bias, only four assessed the calibration slope, and no model has evidence of clinical impact. The certainty of the pooled evidence is low to very low; the gains that survive validation came from staging predictions to decision points and local calibration, not deeper architectures. We provide a TRIPOD+AI-mapped checklist, a benchmark specification and a research agenda.
Figures & tables
| Review | Design | Scope | Outcome | Studies | Appraisal | Synthesis |
|---|---|---|---|---|---|---|
| Leushuis et al. [ 13 ] | Systematic review | Prediction models in reproductive medicine, mainly regression | Pregnancy outcomes | – | Derivation, validation and impact phases | Narrative |
| McLernon and Bhattacharya [ 3 ] | Narrative review | Clinical prediction models in IVF; important covariates | Cumulative live birth | – | Model quality, qualitative | Narrative |
| Tian et al. [ 9 ] | Systematic review, meta-analysis and external validation | Clinical prediction models for IVF outcomes, regression and machine learning | Live birth and other IVF outcomes | 72 studies; 86 live-birth models | Risk-of-bias assessment | Pooled discrimination; external validation in a Chinese cohort |
| Gao et al. [ 2 ] | Narrative review | AI across the IVF pathway by data modality | Several | – | None formal | Narrative |
| Ouyang and Wei [ 11 ] | Review | Multimodal AI for embryo grading and pregnancy prediction | Pregnancy (various definitions) | 17 tabular and 8 multimodal pregnancy studies tabulated | None formal | Tabulated AUCs |
| Yadav et al. [ 12 ] | Systematic review without meta-analysis | AI live-birth prediction, image-based and clinical models | Live birth | 23 studies | QUADAS-2 | Narrative |
| Decision point | Clinical question | Predictor domains admissible (cumulative) | Not admissible (leakage) | Example models |
|---|---|---|---|---|
| Before treatment | Start IVF? Expected chance of a baby; counselling and funding | Female age, BMI, ovarian reserve (AMH, AFC, basal FSH), infertility duration, type and cause, reproductive history; male age and semen analysis | Stimulation response; oocyte, fertilisation and embryo data | McLernon pre-treatment [ 8 ] ; Qiu et al. [ 52 ] ; Zhang et al. [ 45 ] ; Shao et al. [ 46 ] ; Cai et al. [ 37 ] ; Lantzi et al. [ 22 ] ; Xing et al. [ 26 ] ; stage 1 of Guo et al. [ 51 ] and Liu et al. [ 35 ] |
| After ovarian stimulation | Trigger; proceed to retrieval; plan fresh transfer or freeze-all | + protocol, gonadotropin dose and duration, trigger-day hormones, endometrial thickness | Oocyte yield, fertilisation and embryo data | Peng et al. [ 53 ] ; stage 2 of Guo et al. [ 51 ] and Liu et al. [ 35 ] |
| Before embryo transfer | Fresh transfer or freeze-all; number and choice of embryos | + oocytes retrieved, MII, 2PN, embryo development and morphology, number and stage of embryos transferred | Pregnancy test and any post-transfer event | Stage 3 of Guo et al. [ 51 ] ; Banerjee et al. [ 10 ] ; Benchaib et al. [ 21 ] ; Zhu et al. [ 54 ] ; Wu et al. [ 55 ] ; single blastocyst transfer [ 38 ] ; frozen transfer [ 56 , 36 , 24 , 25 ] ; McLernon post-treatment [ 8 ] |
| After an unsuccessful complete cycle | Start another stimulation? | + full history of the previous complete cycle | Outcomes of the next cycle | Banerjee second cycle [ 10 ] ; Shao after the first cycle [ 46 ] ; US post-treatment model [ 3 ] |
| After a positive pregnancy test (prognosis only) | No treatment decision; counselling about an early pregnancy | + serum -hCG after transfer | – | Mourad et al. [ 48 ] ; Lin et al. [ 47 ] |
| Endpoint | Numerator and threshold | Who is in the data (denominator) | Studies | What the predicted probability means |
|---|---|---|---|---|
| Clinical pregnancy (surrogate) | Pregnancy confirmed clinically; excludes later loss | Transfers | iDAScore trial [ 65 ] ; Sergeev and Diakova [ 66 ] | Chance of reaching clinical pregnancy; overstates the chance of live birth |
| Live birth per fresh transfer | 1 live infant after 24 weeks [ 51 , 67 ] or 28 weeks and alive at 1 month [ 60 ] ; multiple births count once [ 51 ] | Cycles that reached fresh transfer; cancelled cycles excluded | Guo et al. [ 51 ] ; Zhu et al. [ 54 ] ; Wu et al. [ 55 ] ; Liu et al. [ 60 ] ; Benchaib (per blastocyst) [ 21 ] | Chance that this transfer ends in live birth, conditional on reaching transfer |
| Live birth per frozen transfer | 1 live infant from the frozen transfer, often 28 weeks [ 59 ] | Frozen transfers, often one per patient | Zhao et al. [ 56 ] ; Luo and Wang [ 59 ] ; Liu et al. [ 36 ] ; Lin (euploid) [ 47 ] ; Abdala (euploid) [ 61 ] ; Ortiz (mosaic) [ 62 ] ; Liang et al. [ 24 ] ; Bai (PCOS) [ 25 ] ; Liu (single blastocysts, 73% frozen) [ 38 ] | Chance that this frozen transfer ends in live birth |
| Singleton live birth (harm-aware) | One live infant; multiple births counted as a separate outcome | Transfers or cycles | Luke et al. [ 7 ] (live birth and multiple birth modelled jointly) | Chance of a singleton baby; makes the harm of multiple pregnancy visible for transfer decisions |
| Live birth per treatment cycle (registry label) | “1 if there were one or more live births as a result of this cycle” | HFEA cycles with completed embryo transfer, fresh and subsequent frozen–thawed | Zhang et al. [ 45 ] ; Lantzi et al. [ 22 ] ; Priya and Kumble [ 23 ] | Chance per registered cycle; women not identifiable across cycles |
| Cumulative live birth per complete cycle | 1 live birth from the fresh and all frozen transfers of one stimulation; a conservative variant counts untransferred embryos as no live birth [ 34 ] | Women or retrieval cycles | Qiu et al. [ 52 ] ; Fu et al. [ 34 ] ; Chen et al. [ 68 ] ; Liu et al. [ 35 ] ; Fatmawati et al. [ 69 ] ; Chen et al. [ 70 ] ; Fraire-Zamora et al. [ 71 ] ; Brás de Guimarães et al. [ 72 ] ; Zhang et al. [ 73 ] ; Sun et al. [ 74 ] ; Alson et al. [ 58 ] ; Banerjee et al. [ 10 ] ; Cai (within 2 years) [ 37 ] ; Xing et al. [ 26 ] | A woman’s chance from one stimulation, including frozen transfers; time horizon not standardised [ 63 ] |
| Study | Data and period | Population and endpoint | n; live births; female age | Decision point | Selected model (comparators) | AUC (95% CI) | Evaluation data |
|---|---|---|---|---|---|---|---|
| Banerjee et al. [ 10 ] | USA, 1 centre; 2003–2008 | First and second fresh autologous cycles; LB including linked frozen transfers | 1,676 development cycles; 634 + 230 validation cycles; 29% (first cycles, development); n.a. | Before transfer; After a complete cycle | GBM boosted trees (age-only model 0.68) | Ext. 0.80 first cycle; 0.68 second cycle | temporal validation 2007–2008 |
| Qiu et al. [ 52 ] | China, 1 centre | First complete cycle; cumulative LB | 7,188 women; 2,797 (38.9%); mean 32.7 | Before treatment | XGBoost (LR, RF, SVM) | Int. 0.73 | random 30% test |
| Brás de Guimarães et al. [ 72 ] | Portugal, 1 centre; 2012–2019 | IVF/ICSI complete cycle; LB 24 weeks | 1,193 cycles; 387 (32.4%); mean 34.3 | Before transfer | ANN (decision tree 0.749) | Int. 0.752 (0.725–0.775) | 15% hold-out of rebalanced data |
| Goyal et al. [ 78 ] | UK, HFEA register | IVF cycles; LB per cycle | 141,160 after undersampling of 495,630; 50% by design; age bands only | Not stated | RF (classical, deep and ensemble learners) | Int. 0.846 | random split of rebalanced data |
| Amini et al. [ 49 ] | Iran, 1 centre; 2011–2014 | IVF/ICSI cycles with a clinical pregnancy; successful delivery | 6,071 cycles; 4,930 (81.2%); mean 30.9 (LB), 32.3 (no LB) | After pregnancy test | RF (LDA 0.57; LR 0.55; NB 0.53; SVM, XGBoost) | Int. 0.60 (0.55–0.64) | 30% test, mean of 500 random splits |
| Liang et al. [ 24 ] | China, 1 centre; 2012–2015 | Frozen–thawed embryo transfers; LB 28 weeks per transfer | 2,189 transfer cycles; 815 (37.2%); Not reported | Before transfer | Random forest, 100 trees, with genetic-algorithm search over modifiable factors | Int. 0.83 | 10-fold CV (0.83 0.02) after predictor selection on all cycles |
| Study | Data and period | Population and endpoint | n; live births; female age | Decision point | Selected model (comparators) | AUC (95% CI) | Evaluation data |
|---|---|---|---|---|---|---|---|
| Liu et al. [ 36 ] | China, 2 centres; 2016–2022 | Single vitrified–warmed blastocyst transfer, age 35; LB | 1,168 transfers; 352 (30.1%); 35 | Before transfer | Stacking (19 learners) | Int. 0.791 (0.752–0.831) | random 25% test |
| Liu et al. [ 38 ] | China, 1 centre; 2013–2019 | Single blastocyst transfers, fresh (day 5) and frozen; LB per transfer after 27 weeks | 8,383 transfers; 4,027 (48.0%); median 30.0 | Before transfer | XGBoost with cycle features and day-3 morphology (LASSO 0.661, GLM 0.662; without day-3 morphology 0.819) | Int. 0.830 (0.822–0.839) | whole cohort, apparent (Table 3); random-split results only in Table S3 |
| Mourad et al. [ 48 ] | Canada, 2 centres; 2012–2016 | Positive pregnancy test after first single transfer; LB 24 weeks | 1,995 patients; 1,318 (66.1%); mean 33.3 | After pregnancy test | LR selected (classification tree, RF, NN all 0.70) | Int. 0.78 (0.74–0.82) | random 30% validation |
| Peng et al. [ 53 ] | China, 1 centre; 2015–2022 | IVF/ICSI cycles; LB per cycle | 11,486 couples; 3,097 (27.0%); 33.2 (SD 5.2) | After stimulation | LR selected (RF 0.674; XGBoost, LightGBM) | Int. 0.671 (0.630–0.713) | 10-fold CV (Table 3; the abstract swaps the LR and RF values) |
| Abdala et al. [ 61 ] | UAE, 1 centre; 2017–2023 | Single euploid frozen transfer after PGT-A; LB | 1,979 transfers (1,459 couples); 1,002 (50.6%); 51% under 35 | Before transfer | LR selected (RF 0.606, XGBoost 0.601, SVM 0.581) | Int. 0.626 | 2,000 repeated patient-grouped 60:40 splits |
| Bai et al. [ 50 ] | China, 1 centre; 2016–2022 | Fresh IVF-ET; LB among women with a clinical pregnancy | 2,031 clinical pregnancies (2,625 women); 1,711 (84.2%); 20–40 (range) | After pregnancy test | LightGBM (LR 0.662; XGBoost 0.718; RF 0.711; 5 others) | Int. 0.746 (0.688–0.804) | 20% test of clinical pregnancies |
| Study | Data and period | Population and endpoint | n; live births; female age | Decision point | Selected model (comparators) | AUC (95% CI) | Evaluation data |
|---|---|---|---|---|---|---|---|
| Wu et al. [ 55 ] | China, 1 centre; 2016–2023 | Fresh cleavage-stage transfer; LB | 11,728 records; 3,971 (33.9%); mean 31.5 (LB), 33.2 (no LB) | Before transfer | RF (XGBoost, GBM, AdaBoost, LightGBM, ANN) | Int. 0.807 | random 20% test |
| Xing et al. [ 26 ] | China, 1 centre; 2017–2022 | IVF; cumulative LB per stimulation, fresh and frozen transfers | 1,792 cycles; 1,036 (57.8%); mean 32.3 | Before treatment | XGBoost (RF 0.776, SVM 0.762, LR 0.728, NN 0.727) | Int. 0.775 | stratified random 30% test (538 cycles); events from the 57.8% rate |
| Yao et al. [ 64 ] | USA, 6 centres | First IVF cycle; LB | 4,635 patients; 56.4–58.5% (test sets); 40 (eligibility) | Before treatment | Centre-specific ML (SART registry model) | ROC AUC not different from registry model; PR-AUC 0.75 vs 0.69 | Centre-specific out-of-time test sets |
| Zhu et al. [ 54 ] | China, 1 centre; 2019–2023 | PCOS, fresh transfer; LB | 1,062 cycles; 466 (43.9%); median 31 | Before transfer | XGBoost (DT, KNN, LightGBM, NB, RF, SVM) | Int. 0.822 (0.777–0.867) | random 30% test |
| Zhu et al. [ 85 ] | China, 1 centre; 2018–2023 | Endometriosis (laparoscopic diagnosis), fresh transfer; LB 28 weeks | 1,836 patients; 739 (40.3%); median 33 | Before transfer | XGBoost (LR 0.805; DT, KNN, LightGBM, NB, RF, SVM) | Int. 0.852 (0.821–0.884) | random 30% test |
| Alson et al. [ 58 ] | Sweden, 1 centre; 2019–2022 | First IVF/ICSI cycle, prospective; cumulative LB from the index cycle ( 22 weeks) | 1,037 women; 424 (40.9%); mean 32.0 | Before treatment | XGBoost (bagged trees post hoc) | Int. 0.66 | stratified 20% test |
| Study | Data and period | Population and endpoint | n; live births; female age | Decision point | Selected model (comparators) | AUC (95% CI) | Evaluation data |
|---|---|---|---|---|---|---|---|
| Lin et al. [ 47 ] | China, 1 centre; 2019–2025 | Euploid frozen transfer, recurrent pregnancy loss; LB 28 weeks | 127 patients; 79 (62.2%); n.a. | After pregnancy test | RF (LR 0.892; seven others) | Int. 0.917 (0.778–1.000) | random 30% test |
| Liu et al. [ 86 ] | China, 1 centre; 2020–2024 | Endometrioma after sclerotherapy; cumulative LB 24 weeks | 194 patients; 97 (50.0%); median 33 | Not stated | XGBoost (DT, RF, SVM) | Int. 0.830 (0.719–0.941) | random 30% test (59 patients) |
| Liu et al. [ 35 ] | China, 1 centre; 2019–2024 | Diminished ovarian reserve; cumulative LB 28 weeks | 1,234 patients; 282 (22.9%); median 38 | Before treatment; After stimulation; Before transfer | CatBoost (DT, LightGBM, GBM, XGBoost, AdaBoost) | Int. 0.759 (0.704–0.810) stage 1; 0.755 (0.699–0.808) stage 2; 0.793 (0.739–0.840) stage 3 | random 30% test |
| Luo and Wang [ 59 ] | China, 1 centre; 2021–2024 | Frozen–thawed transfer; LB 28 weeks | 3,470 development + 1,467 prospective; 41.2% / 46.0%; median 31 (LB), 33 (no LB) | Not stated | XGBoost (LR 0.818; RF) | Int. 0.843. Ext. 0.809 | random 30% test; prospective temporal cohort 2024 |
| Minano Masip et al. [ 87 ] | Canada, 1 centre; 2012–2019 | Single blastocyst transfers with time-lapse culture; LB | 355 transfers (259 women); 127 (35.7%); mean 35.4 | Before transfer | SVM on clinical data (combined with a CNN on time-lapse images: 0.76) | Int. 0.65 | patient-level 15% test |
| Nowak et al. [ 88 ] | Poland, 1 centre; 2013–2022 | IVF cycles with time-lapse annotation; LB | 375 cycles + 144 synthetic records; n.a.; n.a. | Before transfer | XGBoost (LR, SVM, NN) | Int. 0.927 | 20% test of real plus synthetic records |
| Stage | Question answered | Evidence required | Status in the literature reviewed |
|---|---|---|---|
| 1. Development | Does the model discriminate and calibrate in its own data? | Internal validation with optimism correction; calibration; TRIPOD+AI reporting [ 14 ] | Abundant; mostly single-centre, most at high risk of bias (Sections 4 and 5) |
| 2. External validation | Does performance hold in other centres and later periods? | Temporal and geographic validation with 100–200 events; calibration [ 98 ] | Few: McLernon models [ 111 , 9 , 37 ] ; Banerjee et al. [ 10 ] ; Guo et al. [ 51 ] ; Fu et al. [ 34 ] ; Luo and Wang [ 59 ] ; Shao et al. [ 46 ] ; Sun et al. [ 74 ] ; Zhang et al. [ 73 ] ; centre-specific models [ 64 ] |
| 3. Local updating | Is the model calibrated for this clinic now? | Recalibration or refitting on recent local data; monitoring for drift | Recalibration improved IVFpredict [ 77 ] ; Platt scaling [ 51 , 35 ] ; McLernon model refitted locally [ 46 ] ; centre-specific modelling [ 64 ] |
| 4. Clinical utility | Would using the model improve decisions? | Decision-curve analysis at pre-specified thresholds [ 97 ] | Reported in 12 of the 48 included studies; thresholds not pre-specified |
| 5. Impact | Does using the model change decisions and outcomes? | Prospective comparative study, ideally randomised | None identified; nearest are the iDAScore trial [ 65 ] and the IDoser protocol [ 2 ] |
| 6. Implementation and monitoring | Does the model stay safe and calibrated in routine use? | Post-market monitoring, change control and regulatory conformity [ 104 , 107 , 109 ] | Not reported; EU high-risk obligations for AI in medical devices apply from 2 August 2028 [ 105 ] |
| Direction | Promise | Evidence for live birth so far | Enabling methods | Main risk | A convincing study would show |
|---|---|---|---|---|---|
| Multimodal fusion (images + clinical data) | Adds embryo and endometrial information that records lack | Fusion beat images alone (0.77 vs 0.70) [ 112 ] ; fusion beat tabular data alone by 0.033 [ 113 ] | Decoupled or late fusion; image foundation models [ 114 ] | Comparison against images alone; equipment-specific inputs | Gain over a tuned tabular model, externally validated, with calibration |
| Tabular foundation models | Strong small-sample performance without tuning | Strong on public benchmarks [ 92 ] ; one live-birth cohort, in a preprint [ 18 ] | In-context learning on synthetic priors | Prior may not match clinical data | Head-to-head comparison across cohorts, with calibration |
| Language models | Use of free text and prior knowledge; automated pipelines | Few-shot strength on public tasks [ 93 ] ; pregnancy only [ 115 ] | Serialisation, fine-tuning, information extraction | Loss of numerical structure; leakage hidden in automated pipelines [ 116 ] | Live-birth endpoint, admissible predictors, comparison with logistic regression |
| Federated learning | Multicentre training without moving records | Clinical-pregnancy AUC 0.57–0.72 across four hospitals [ 117 ] | FedAvg [ 118 ] ; FedProx [ 119 ] ; secure aggregation [ 120 ] ; differential privacy [ 121 ] | Between-centre heterogeneity; leakage through updates | Live-birth model compared with central and local training, with per-centre calibration |
| Dynamic prediction | Updates as treatment progresses | Staged 0.678 0.731 [ 51 ] ; discrete-time BART over six cycles [ 46 ] ; adaptive per-transfer models [ 122 ] | Discrete-time and multi-state models [ 123 ] ; landmarking [ 124 ] ; joint models [ 125 ] | Leakage across stages; informative discontinuation | Stage-isolated predictors; cumulative endpoint over all transfers |
| Prediction under intervention | Chance of live birth under each treatment option | None identified | Predictimands [ 82 ] ; meta-learners [ 126 ] ; causal forests [ 127 ] | Confounding by indication | Randomised-trial data or stated causal assumptions |
| Element | Specification |
|---|---|
| Data | The anonymised HFEA register as a public seed; contributed single-centre and multicentre datasets under controlled access, or evaluated in place by federated evaluation |
| Schema | A minimal data dictionary of the four predictor domains (Table 2), with each variable time-stamped to the decision point at which it becomes known, and the cycle, transfer and woman identifiers needed for grouping |
| Tasks | T1 before treatment: cumulative live birth per complete cycle. T2 after stimulation: live birth per fresh transfer. T3 before transfer: live birth per transfer, fresh and frozen. T4 after an unsuccessful cycle: live birth in the next complete cycle |
| Endpoints | Live birth with a stated gestational threshold (International Glossary definition by default); multiple births counted once; singleton live birth at term as a secondary endpoint for T2 and T3 |
| Splits | Patient-grouped; a temporal hold-out of the most recent two years; a geographic hold-out of at least one centre; no tuning on hold-out data |
| Baselines | Penalised logistic regression with splines; the McLernon models for T1 and T4; gradient-boosted trees; a tabular foundation model; equal tuning budgets |
| Stage | Recommendation | Section | TRIPOD+AI item | Key reference |
|---|---|---|---|---|
| Question | 1. State the decision point and intended user; admit only predictors routinely available at that point. | 3.1, 5.1 | 3, 9 | [ 15 , 51 ] |
| Question | 2. Define the endpoint fully: gestational threshold, numerator, denominator; report the composition of the non-live-birth group and, for transfer decisions, singleton live birth. | 3.2 | 8, 20 | [ 57 , 63 ] |
| Data | 3. Use one cycle per woman, or group data splits by patient; report how repeated cycles were handled. | 3.2, 5.2 | 6, 12 | [ 94 ] |
| Data | 4. Report the data-collection period and prefer a temporal test set. | 5.1 | 5, 12 | [ 77 , 99 ] |
| Data | 5. Report missingness per variable; impute with methods fitted on training data only. | 5.1 | 7, 11 | [ 76 ] |
| Sample size | 6. Justify the development sample size with formal criteria; size external validation for precise calibration (at least 100 events, usually more). | 5.2, 8.1 | 10 | [ 95 , 137 , 98 ] |
Clinical Prediction
Clinical prediction uses patient records, signals, and images to estimate diagnoses, risks, outcomes, and treatment responses for care.
See all 233 papers