Embryo assessment & selection in IVF

In IVF, embryologists choose which embryo to transfer largely by visual grading, which varies between people and clinics. Time-lapse incubators already record every embryo continuously; deep-learning models turn that video into an implantation-potential score (and, in some products, a ploidy-likelihood estimate) that ranks embryos objectively. The AI supports, but never replaces, the embryologist's decision. Benefit: more consistent grading and faster review, with outcome gains still under study.

Problem & value

• Today: embryologists grade day-3/day-5 embryos by morphology (blastocyst expansion, inner cell mass, trophectoderm) and, where time-lapse is available, by morphokinetic timings. Grading is subjective, with documented inter- and intra-observer variability. Invasive PGT-A (trophectoderm biopsy) is costly, not allowed everywhere and of debated benefit for all patients.
• AI value: standardised ranking of embryos, less annotation time per embryo, shorter time to transfer of the most viable embryo, a non-invasive screening complement, and training support for junior embryologists.
• Evidence: retrospective studies show good discrimination for implantation; randomised trials of AI ranking have mostly shown non-inferiority/time savings rather than clearly higher live-birth rates, so claims should be read against the published trials [1][2]. Do not promise improved birth rates.

Variants

• Implantation / live-birth potential scoring from time-lapse video (core use case).
• Ploidy prediction (euploid vs. aneuploid likelihood) as non-invasive pre-screen; research-grade or limited claims, not a PGT-A replacement.
• Static-image blastocyst grading for clinics without time-lapse incubators (smartphone/microscope).
• Oocyte and sperm assessment (maturity, morphology, sperm selection for ICSI), plus lab quality control and workflow/witnessing support.
• Research: embryo-selection decision support combined with patient data (see HLTH-004).

How it works

1. time-lapse incubator (multi-focal video, per embryo/well)
2. quality control & embryo/well detection (segmentation)
3. feature extraction (3D-CNN / video transformer / frame-level CNN + temporal model)
4. score heads: implantation/live-birth likelihood, ploidy likelihood, morphokinetic events
5. ranking within patient cohort + explanation (saliency / event timeline)
6. embryologist review, decision and transfer order (human-in-the-loop)
7. outcome follow-up (pregnancy, live birth) fed back for monitoring

• Models: typically convolutional or video-transformer networks trained on embryo videos labelled with outcome (implantation with fetal heartbeat) or ploidy. Fully deep-learning systems score from raw video without manual annotation; older systems use rule-based morphokinetic algorithms.
• Labels are the hard part: transferred embryos with known outcome are a biased subset (selection bias); untransferred embryos have no label. Models should be validated per clinic and patient group.
• Human-in-the-loop: the embryologist and clinician decide; the score is one input next to morphology, patient age and clinical history. Every override is logged.

Data preparation

The data are time-lapse videos (multi-focal planes, one frame every 5 to 20 minutes over 5 to 7 days), embryologist annotations and grades, and outcome labels such as implantation, live birth or PGT-A results. They live in the incubator vendor's software and local storage, the clinic's lab information system or Excel sheets, and the patient record, often in separate systems. Typical gaps: outcomes recorded late or not linked to embryo IDs, inconsistent grading between embryologists, differing incubator and imaging settings, large data volumes, and unclear consent for secondary use.

Preparation steps: link embryo IDs to outcomes across systems (PREP-005, PREP-020), restructure spreadsheets into tidy tables (PREP-011), define a grading scheme (PREP-019), build an expert-scored evaluation set (PREP-018), clear consent and legal basis (PREP-025), and pseudonymise (PREP-022). For a bought product the effort is mostly local validation and integration; for own models, data work is 60 to 70 percent because outcome labels are scarce and slow to arrive. See GUIDE-003.

Implementation paths

Which path for whom

• Enterprise (fertility clinics, IVF networks): Path A, because CE-marked scoring tied to the time-lapse incubator ecosystem exists and clinics lack the labelled outcome data and regulatory capacity to build their own.
• Developer: Path C, because open vision models (video classifiers, segmentation) on a single GPU allow research on annotated time-lapse data with full control over sensitive embryo and patient data.

Path A — Buy (recommended)

• Vitrolife (SE): iDAScore (deep-learning scoring in EmbryoScope+/EmbryoViewer; CE-marked).
• Fairtility (IL): CHLOE (embryo assessment, CE-marked, links to several incubators).
• Others: AIVF EMA, Alife Health (Embryo Assist, US), Presagen Life Whisperer (AU), Future Fertility (oocyte assessment, CA), Embryonics/Gaia; Merck/Vitrolife/CooperSurgical/Genea time-lapse incubators. Verify current CE (MDR) certificates and the intended-use statement per product.
• Choose A: always for clinical use. Only a legal manufacturer with a conformity assessment can put a score into a clinical decision.

Path B — Build on cloud APIs

Rarely appropriate for clinical use: general VLM APIs (Claude, Gemini, GPT, Mistral) are not medical devices and have no outcome-validated embryo model. Reasonable only for non-clinical tasks: summarising lab records, drafting patient information or QC reports, with EU-region hosting and a data-processing agreement; no embryo images tied to identifiable patients without a clear legal basis.

Path C — Self-host (open source)

• Research and validation: PyTorch/MONAI with open video backbones (VideoMAE, TimeSformer, 3D ResNet), open embryo datasets such as the Gomez et al. human embryo time-lapse dataset [3].
• Serving: on-prem Triton/ONNX Runtime; HW2 (one 16-32 GB GPU) is enough for inference and fine-tuning small video models; HW1 suffices for static-image grading.
• Status: self-built models used clinically make the clinic a manufacturer (Art. 5(5) MDR in-house exemption applies only under strict conditions).

Path D — Custom training

Worthwhile for manufacturers and academic centres: thousands of embryos with linked outcomes (ideally multi-clinic, multi-incubator) plus PGT-A labels. Requires site-level external validation, calibration and fairness checks (age, ethnicity, protocol), MLOps under ISO 13485 / IEC 62304 and a clinical evaluation.

State of the art (2026)

• Video deep learning (3D CNNs, video transformers, self-supervised pre-training) has replaced hand-crafted morphokinetic models; commercial systems score without manual annotation.
• Trials: randomised studies comparing AI-ranked selection to standard morphological grading generally report comparable clinical outcomes with less evaluation time; no consistent evidence yet of higher live-birth rates per transfer [1][2].
• Ploidy prediction from images remains limited and population-dependent; PGT-A still the reference.
• Limits and failure modes: domain shift between incubators and clinics, outcome-label bias, scores correlating with patient age (confounding), poor calibration, over-reliance by staff, and lack of transparency about training populations.

Shared resources & combinations

• Same stack (image-classification, segmentation, video models): HLTH-002 digital pathology, HLTH-026 lab diagnostics, HLTH-001 radiology; SCI-008 scientific microscopy.
• Combine with: HLTH-004 clinical decision support (treatment planning), HLTH-018 chart summarisation, PERS-021 cycle tracking on the patient side, laboratory witnessing/traceability systems.
• Training pipelines for medical image models (annotation tools, MLOps, validation) are reusable.

Effort, cost & KPIs

• Path A: licence typically bundled with or added to time-lapse incubator/software; 1-3 months to validate locally and train staff. Time-lapse hardware is the main cost driver.
• Path C/D: 12+ months, specialised team, regulatory work costs more than modelling.
• KPIs: AUC/calibration on local data, embryologist-AI agreement, time per case, inter-observer variability, time to pregnancy, cumulative live birth rate per cycle, override rate and outcomes of overrides, drift over time.

People & change

Involve clinic management and the medical lead first, then embryologists as daily users and owners of the grading process, IT for storage and integration, the DPO, and legal/compliance for MDR status and sensitive ethical questions. Patients are affected by the output and must be informed how AI supports embryo ranking and that the embryologist decides.

Works council: depends. If the system records which embryologist accepted or overrode a score, it is suited to monitor performance and triggers co-determination under §87(1) no. 6 BetrVG; inform the works council when processes change with AI (§90(1) no. 3) and consider an expert (§80(3)). Where no works council exists, this does not apply; public clinics involve the Personalrat.

Data protection: reproductive and genetic data are special categories, so a DPIA is likely (Art. 35 GDPR), with AVV for vendors (Art. 28) and DPO involvement. AI literacy: support staff with training on score meaning, limits and automation bias (Art. 4 AI Act; GUIDE-006). Pilot in shadow mode with a small group of embryologists, compare scores with their decisions, and keep the embryologist as final decision maker. Typical concerns are deskilling and ethical doubts; address them with open communication and documented oversight. See GUIDE-008, GUIDE-005, GUIDE-014.

Risks, regulation & compliance

• EU AI Act: high risk. An AI system that is a medical device (or safety component) requiring third-party conformity assessment under MDR is high-risk via Art. 6(1)/Annex I; obligations (risk management, data governance, human oversight, logging) apply, with the staged timeline for Annex I products running to 2 August 2028 [4] (deferred from August 2027 by the Digital Omnibus on AI, Reg. (EU) 2026/1744). Not a prohibited practice, but selecting embryos on non-medical traits is ethically and legally off-limits (see below).
• MDR (EU 2017/745): software that supports selection of embryos for transfer is generally classified at least IIa (Rule 11); check the notified-body certificate and intended use [5].
• Germany: Embryonenschutzgesetz (ESchG) strictly limits embryo handling, including selection and genetic testing (PID only under §3a and an ethics committee); clinics must verify that product features fit the legal framework. Rules differ across Member States.
• GDPR: embryo videos linked to patients, genetic and health data (Art. 9); explicit consent, DPIA, processor contracts, no secondary training use without a legal basis.
• Ethics: no selection for sex or non-medical traits; transparency to patients about AI use and uncertainty; avoid presenting a score as a guarantee.
• Mitigations: local validation, human decision authority, override logging, post-market monitoring.

Getting started

1. Define the goal (consistency, time saving, ranking support) and audit current grading variability.
2. Confirm the legal frame (ESchG or national law, ethics committee) and the product's MDR certificate.
3. Pilot one CE-marked product in shadow mode on retrospective cases from your own clinic.
4. Compare AI ranking with embryologist choice and known outcomes; define override rules.
5. Set up post-market follow-up with outcome tracking and consent/DPIA documentation.

Sources

1. Vitrolife — iDAScore product and clinical evidence — https://www.vitrolifegroup.com/
2. Kragh MF, Karstoft H et al., "Automatic grading of human blastocysts from time-lapse imaging" and iDAScore studies — https://pubmed.ncbi.nlm.nih.gov/?term=iDAScore+embryo
3. Gomez T et al., "A time-lapse embryo dataset for morphokinetic parameter prediction" (Data in Brief 2022) — https://doi.org/10.1016/j.dib.2022.108258
4. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
5. Regulation (EU) 2017/745 (MDR) — https://eur-lex.europa.eu/eli/reg/2017/745/oj
6. Fairtility — CHLOE embryo assessment — https://www.fairtility.com/
7. Embryonenschutzgesetz (ESchG) — https://www.gesetze-im-internet.de/eschg/
8. ESHRE — Good practice recommendations on add-ons / time-lapse in reproductive medicine — https://www.eshre.eu/Guidelines-and-Legal