What a Model Actually Learns From Your Cases

Last week we argued that a simulation trained on aggregated data predicts an average result, and that no surgeon operates at the average. The natural follow-up question is a practical one. If personalization is the answer, what does it actually require, and what is a model learning when a surgeon shares their cases?

This article is the mechanical explanation. No metaphors about teaching the software your style, because that phrasing hides the part that matters.

Three layers, not one model

It helps to separate what stays fixed from what adapts.

Layer one is physics. Soft tissue deforms according to mechanical properties that are not specific to any surgeon. Arbrea’s simulation core comes out of research at ETH Zurich, where the problem was making deformable bodies behave correctly under load. Tissue elasticity, volume displacement, and gravitational response belong here. This layer does not need personalizing, and it should not be personalized. Mechanics are mechanics.

Layer two is the general model. This is the learned mapping from pre-operative anatomy plus surgical intent to a plausible post-operative shape, built across a broad population of cases. It knows what a 350 cc moderate-plus implant tends to do in a patient with a given base width and tissue quality. It knows it on average.

Layer three is the adaptation. This is the layer the Research Partner Program builds. It adjusts the general mapping using paired cases from one practice, so that when the model is uncertain between several plausible results, it resolves toward the one that practice actually produces.

The distinction matters because it sets a realistic expectation. Personalization does not rebuild the simulation from scratch on a few hundred cases, which would be neither possible nor safe. It shifts a well-behaved general model into the region of outcome space where a specific surgeon works.

What counts as useful data

Surgeons tend to assume the answer is volume. More cases, better model. Volume helps, but it is not the binding constraint. Consistency is.

Four things make a case useful.

A paired capture. A pre-operative 3D scan and a post-operative 3D scan of the same patient. Without the pair, there is nothing to learn from. A post-operative photograph alone tells the model what the result looked like but not what the starting anatomy was, and the whole signal lives in the relationship between the two.

A consistent post-operative interval. Tissue at six weeks and tissue at nine months are different tissues. Swelling resolves, the implant settles, the lower pole relaxes. A dataset where the interval varies unpredictably teaches the model that the same surgery produces many different results, which is the opposite of the goal. A practice that captures consistently at a defined interval contributes far more per case than one that captures whenever the patient happens to return.

The decisions attached to the case. Implant manufacturer, reference, volume, and profile. Pocket and plane. Any technique detail the surgeon considers material to the result, such as extent of inferior release or fold management. Without these labels, the model sees an input and an output with the reasoning removed, and cannot separate the effect of the implant from the effect of the technique.

Capture conditions that do not drift. Same device, comparable patient positioning, comparable arm position. This sounds pedantic. It is the difference between a model learning surgery and a model learning the idiosyncrasies of how a particular room was set up on a particular day.

A hundred well-labelled, consistently captured pairs from a single surgeon carry more adaptive signal than several times that number captured loosely. This is the part that surprises most partners.

What the model does not learn

Worth stating explicitly, because it is a common concern.

The model is not learning to identify patients. Adaptation operates on anatomical geometry and surgical parameters, not on identity. Patient data is de-identified before it enters any training process, and consent for research use is obtained through the partner’s own patient consent workflow.

What a partner receives

The output of participation is a version of Arbrea whose simulations are calibrated to that practice. Concretely, that means the model resolves toward the pocket the surgeon actually uses, weights the implant families they actually stock, and reflects the tissue behaviour of the population they actually treat.

Two secondary effects tend to matter as much as the primary one.

The first is that building the dataset forces a practice to standardize its own outcome capture. Surgeons who complete the process usually end up with the first systematic longitudinal record of their own results they have ever had, independent of the model. Several partners have described this as the more immediately useful outcome.

The second is calibration in the other direction. When a surgeon compares their own aggregated results against the general model, the difference is a measurement of their technique rather than an impression of it.

The honest version of the effort involved

This is not a passive program. Contributing means capturing scans on a protocol, labelling cases with technique detail, and following patients to a defined post-operative interval. For a practice that already scans routinely, the incremental work is modest. For one that does not, it is a change in workflow.

The trade is explicit. Surgeons who put in that structured effort get a model shaped around their own work, and they get it first, while the field is still using models shaped around everyone’s.

See the data protocol and what participation involves.