Opening: Doctor's Decision Logic
In outpatient clinics, we often encounter patients who come for consultation holding embryo grading reports from other centers: "Doctor, this AI says my blastocyst is grade A, but another AI gave it grade B. Who should I believe?" As a reproductive physician, my answer is usually: The AI embryo screening system is currently an auxiliary tool, not the sole criterion for determining an embryo's fate. Its value lies in reducing inter-observer variability in grading, but its limitations also need to be understood.
Direct Definition and Core Value of AI Embryo Screening System
An AI embryo screening system essentially applies deep learning algorithms to image or video data of embryo development, training models to identify morphological and dynamic features associated with high implantation potential. Systems currently used in Chinese reproductive centers are primarily based on two types of data:
- Static Images: Scoring micrographs of embryos at specific time points (e.g., day 3, day 5).
- Time-lapse Imaging: Continuously monitoring embryo development to capture dynamic parameters such as cleavage timings, fragmentation patterns, and blastocoel expansion rate.
Its core value lies in standardization: eliminating subjective differences in grading among embryologists. In internal validations across multiple centers, the concordance rate between AI scoring and senior embryologists can reach over 85%, especially in differentiating "medium quality" embryos (morphology grades 2-3), where AI shows greater stability than manual assessment.
How Good is China's AI Embryo Screening System: Real-World Performance Assessment
From a clinical efficacy perspective, the AI embryo screening systems currently available in the Chinese market (including domestically developed and imported versions) demonstrate quantifiable performance in the following areas:
Doctor's Perspective: What AI Embryo Screening "Can" and "Cannot" Do
In clinical decision-making, reproductive physicians' attitude towards AI systems can be summarized as: "Use its strengths, but know its weaknesses."
- What AI Can Do: When there are a large number of embryos (≥5 blastocysts), it helps prioritize the 2-3 with the best morphology, reducing fatigue errors from repeated manual observation. This is especially efficient for cases with many generally good-quality embryos, such as polycystic ovary syndrome or young oocyte donors.
- What AI Cannot Do: Determine true chromosomal integrity, mitochondrial function, or epigenetic status of the embryo. All current AI systems are essentially "image recognition" and cannot detect chromosomal copy number variations or single gene disorders. For advanced maternal age (≥38 years), recurrent implantation failure, or recurrent pregnancy loss, AI cannot replace preimplantation genetic testing for aneuploidy (PGT-A).
Differences Between Chinese AI Embryo Systems and European/American Products
Currently, the major global AI embryo assessment platforms include:
- Europe (e.g., IVY, ERICA): Training data is predominantly from Caucasian populations. Embryo morphological characteristics differ from East Asian populations (e.g., blastocyst expansion patterns, inner cell mass compactness).
- USA (e.g., Life Whisperer): Uses multi-center data, resulting in better model generalizability. However, the FDA has only approved it as an "adjunctive scoring tool," not as an independent decision-making basis.
- China: Local systems (e.g., models trained on data from centers like Peking University Third Hospital, CITIC Xiangya) show slightly higher scoring concordance for East Asian embryos compared to imported systems. However, Chinese systems commonly suffer from single-source training data – if a model is trained only on data from one center, its accuracy may drop by 5-10% when transferred to another laboratory with different incubators, culture media, or embryologist handling practices.
Selection Advice: If the AI model used by your center is matched to your specific laboratory conditions (incubator brand, culture media batch), its reference value is higher. Otherwise, the AI score should only be considered a "second opinion."
The Most Easily Overlooked Detail: AI Screening Depends on "Raw Data Quality"
Many patients and some physicians overlook a key prerequisite: The output accuracy of an AI model is entirely dependent on the quality of the input images.
- The microscope's objective magnification, illumination uniformity, and focus stability all affect AI's ability to identify fragments, vacuoles, and cell boundaries.
- The capture interval of the Time-lapse system (e.g., every 5 minutes vs. every 10 minutes) changes the resolution of dynamic parameters.
- AI models cannot accurately assess post-thaw survival from a "single snapshot" of morphological changes in blastocysts before freezing and after thawing.
Therefore, the performance of an AI system is relatively stable within the same reproductive center. However, results may differ between different centers, even when using the same AI software. When evaluating an AI system, it is necessary to review the internal validation data from that center for its own laboratory conditions, rather than only looking at the theoretical performance provided by the manufacturer.
The Easiest Pitfall: Overinterpretation of "AI Grades"
Common misconceptions include:
- Misconception 1: "An embryo with an AI grade A will definitely lead to pregnancy." In reality, even the highest-graded embryo by AI has a single-transfer success rate of about 55-65% in young women, meaning there is still a 35-45% chance of implantation failure or biochemical pregnancy.
- Misconception 2: "An embryo with an AI grade C should be discarded." When the number of embryos is small (only 1-2 blastocysts), embryos with lower AI grades can still result in a live birth, especially for young women with low risk of aneuploidy.
- Misconception 3: "AI can replace embryologists." Currently, no reproductive center uses AI to completely replace human expertise. AI is more like an "assisted reading system"; the final decision still requires the embryologist to integrate the embryo's developmental process, patient age, and previous cycle history.
Practical Workflow of AI Embryo Screening
- Embryo Culture: After routine oocyte retrieval and fertilization, embryos develop in the incubator to the blastocyst stage on days 5-6.
- Image Acquisition: Brightfield images of blastocysts (sometimes including polarized light or Hoffman modulation contrast) are captured using a microscope or Time-lapse system.
- AI Scoring: The system automatically outputs a score (typically A/B/C/D or 0-10) and identifies the main features influencing the score (e.g., inner cell mass grade, trophectoderm grade, expansion degree).
- Manual Review: The embryologist combines the AI score, dynamic development record, patient age, and previous transfer history to make the final embryo ranking.
- Decision: For embryos planned for transfer, the AI score serves as one reference; for embryos planned for PGT-A, the AI score can be used to prioritize higher-graded embryos for biopsy.
Interpretation of Key Indicators in AI Scoring
Understanding the specific assessment dimensions of an AI system helps evaluate its reliability:
- Inner Cell Mass (ICM) Grade: Grade A (compact, large), Grade B (loose, medium), Grade C (sparse). AI recognition accuracy for ICM is higher than manual assessment because the grayscale levels and boundaries of the ICM are relatively clear in images.
- Trophectoderm (TE) Grade: Grade A (dense, continuous cells), Grade B (sparse, discontinuous). AI assessment error for TE is larger, as trophectoderm cells are morphologically diverse and easily confused with fragments.
- Blastocyst Expansion Degree: Ranges from stage 1 (early blastocyst) to stage 6 (fully hatched). AI recognition accuracy for expansion degree is nearly 100%, but this indicator itself has limited independent predictive value for live birth.
- Fragmentation Percentage: AI is relatively accurate in identifying fragmentation ≥10%, but tends to miss small fragments <5%.
Frequently Asked Questions: When is AI Screening Suitable?
Based on clinical practice in several domestic reproductive centers, the AI embryo screening system offers higher reference value in the following scenarios:
- ≥10 oocytes retrieved, ≥4 blastocysts formed, needing to prioritize the order for transfer.
- Previous repeated failures with morphologically "good" embryos, hoping to reassess remaining embryos with AI.
- Large number of embryos, but patient's financial resources are limited, making PGT-A for all embryos unfeasible; AI can help select 2-3 for testing.
When is it Not Suitable to Rely on AI Screening?
- When there are only 1-2 blastocysts, the risk of "exclusion" based on AI score is too high; a comprehensive assessment by a senior embryologist considering the developmental process is recommended.
- Female age ≥40 years, or known chromosomal structural abnormalities; AI cannot assess chromosomal issues, and PGT-A is mandatory.
- Recurrent implantation failure ≥3 times, or recurrent pregnancy loss ≥2 times; AI does not help identify the underlying cause.
Practitioner Observation: Real Penetration Rate of AI Screening in Chinese Reproductive Centers
As of early 2025, approximately 30-40% of assisted reproductive centers in China have introduced some form of AI embryo assessment system, but less than half have truly integrated it into routine clinical decision-making pathways. The main reasons are threefold:
- Cost: The annual maintenance fee for a Time-lapse system plus AI software is approximately 150,000-250,000 RMB, leading some small to medium-sized centers to perceive a low return on investment.
- Trust: Senior embryologists (with over 10 years of experience) generally find that AI scores align with their experience about 80% of the time, but "that 20% disagreement often involves details I consider important."
- Data Barriers: China lacks a unified multi-center AI training database. Models vary significantly between centers, making it difficult to establish industry standards.
In the next 3-5 years, the value of AI may be more evident in remote embryo assessment and patient communication: using AI-generated intuitive images and scores to help patients understand differences in embryo quality, rather than serving as a direct clinical decision-making tool.
Risk Reminder
The AI embryo screening system is currently still classified as "assisted decision-making" level and has not yet obtained registration approval from the China National Medical Products Administration (NMPA) as an "independent medical device." Any claims of "AI screening guarantees live birth" or "AI grading replaces embryo biopsy" are not supported by current clinical evidence. When choosing a reproductive center, patients should check whether the center provides internal validation data for its AI scores and whether there is clear written informed consent stating that "AI scores are for reference only."
Comments (0)