Reproductive Health and AI: Embryo Selection, Sperm Analysis, and IVF Success Prediction

In an IVF laboratory the embryologist looks at day-5 blastocysts and chooses the one with the highest chance of implanting. For decades that choice has been based on morphology — the embryo's appearance — and is inherently subjective: two embryologists can score the same embryo differently. AI, using time-lapse imaging data, has become one of the most active research areas in reproductive medicine by claiming to standardize and perhaps improve this decision. In this article I address the current level of evidence for AI in embryo selection, sperm analysis, and IVF success prediction — including one important caveat.

Embryo Selection: AI's Most Ambitious Promise

Time-lapse incubators image the embryo continuously over hours. This rich data is ideal terrain for deep learning. The notable models in this area include:

STORK and BELA

STORK is a deep neural network built on Google's Inception architecture, trained on time-lapse images from 10,148 embryos. It predicts blastocyst quality with AUC > 0.98, generalizes well to images from clinics outside the U.S., and can outperform individual embryologists. BELA (Blastocyst Evaluation Learning Algorithm), developed by the Weill Cornell team, combines time-lapse data and maternal age to evaluate both embryo quality and chromosomal health.

The critical evidence: the iDAScore randomized trial (2024)

This brings us to the most important point of the article. In 2024, Nature Medicine published a multicenter, randomized, double-blind, noninferiority trial conducted at 14 IVF clinics in Australia and Europe. Women under 42 with at least two early-stage blastocysts on day 5 were randomized to either standard morphological evaluation or the deep-learning algorithm iDAScore.

In 1,066 patients (533 per arm), the clinical pregnancy rate was 46.5% in the iDAScore arm (248/533) and 48.2% in the morphology arm (257/533) (risk difference −1.7%; 95% CI −7.7 to 4.3; p = 0.62). The conclusion: the trial did not demonstrate noninferiority of deep learning to standard morphology in terms of clinical pregnancy rate. iDAScore did, however, significantly reduce evaluation time. (Source: PubMed; Nature Medicine, 2024.)

AI in Sperm Analysis and Selection

On the male-factor side, AI is transforming both semen analysis and selection of individual sperm for ICSI (intracytoplasmic sperm injection). Computer-vision algorithms can move beyond sample-level semen evaluation to analyze individual sperm in real time, assigning categorical scores by progressive motility parameters.

• Sperm selection for ICSI: Deep learning can evaluate low-resolution sperm images and select the most suitable viable sperm for ICSI with high accuracy; one study reported an F1-score of 0.951 with this approach.
• Surgical sperm retrieval in azoospermia: AI image analysis can detect spermatozoa in azoospermic samples faster and with better sensitivity than embryologists — one of the first known clinical applications of machine learning in surgical sperm retrieval.

The core gain on the sperm side is sustaining attention and speed beyond what the human eye can manage for hours and reducing selection variability. Even so, most of these tools are still maturing through studies confirming whether they produce the same clinical outcomes as embryologists.

IVF Success Prediction

Beyond embryo and sperm selection, AI is also trying to predict the success probability of an entire IVF cycle: combining the patient's age, ovarian reserve, prior cycles, and laboratory data to estimate live-birth probability. The clinical value here lies in helping the couple form realistic expectations and personalizing the treatment plan (how many cycles, which protocol). The same caveat applies, however: prediction models that are not externally validated reflect the reality of a single center.

Table: Three Main Applications of AI in IVF

Clinical Perspective: Excitement and Caution Together

Reproductive medicine may be one of medicine's most AI-suitable areas: abundant imaging data, repetitive decisions, inter-observer variability. STORK's generalization power and the speed of sperm-selection algorithms are real gains. But, as the iDAScore study reminds us, laboratory performance and patient outcome are not the same thing. For an embryologist or a reproductive specialist, the honest position today is this: it is reasonable to use AI as a strong assistant that speeds up and standardizes the work; presenting it as a guarantee of "more babies" goes beyond the evidence.