Subjective Variability

Every part of the assessment that carries weight is a judgement of degree rather than of fact. How tightly is the inner cell mass compacted. How uniform are the trophectoderm cells, and how many of them are there. How thin is the zona, and is it thinning evenly. Each of those is real, visible structure — and each is being estimated against a remembered standard rather than measured against a recorded one.

So the reading inherits everything about the reader. Where they trained. Whose grading they learned alongside. How many blastocysts they have personally scored. Which plane of focus they settled on, and at what minute of the assessment window they looked. None of those inputs is written down anywhere, which is precisely why the resulting variance is unchecked: there is no artefact to audit afterwards, only a letter and a number in a log.

The same forces operate between clinics, only larger. Two labs in the same city can hold genuinely different house standards for the same letter, and neither has any mechanism for discovering the gap. A grade is not portable. It means what the lab that issued it meant by it.

Because the grade governs the ranking, variance in the reading propagates straight into the cycle. Transfer order changes. Which embryos are cryopreserved changes. A patient counselled on the strength of a grade is being counselled on an estimate whose confidence interval was never stated to anyone, including the embryologist who produced it.

It also blocks the thing that would fix it. IVF fails more than half the time, and improving that requires learning from outcomes — but learning requires that the input be stable. If the recorded grade shifts with the observer, then every retrospective analysis built on those grades is measuring the lab’s readers as much as it is measuring the embryos. Variance at the bench becomes noise in the dataset, and the noise is indistinguishable from biology.

The consequence compounds nationally. A grading record that cannot be compared across labs cannot be pooled across labs, which is one reason domestic clinical evidence is so much harder to assemble than the number of cycles run would suggest.

Not the embryologist’s authority — the evidence underneath it. A grade should be reproducible: the same specimen should return the same structural read on Tuesday as on Friday, in one lab as in another, and the read should be inspectable rather than asserted.

That is what Forlivf is built to supply. Semantic segmentation maps the boundaries of each structure on the image itself, so the measurement is a visible object, not a remembered impression. The embryologist still decides. They now decide against something that can be pulled up, checked by eye, disagreed with, and — for the first time — compared with what the same model said about the last four hundred blastocysts the lab ran.