Clinical trust cannot be imported

In late June, Idiap put out word of its work towards artificial-intelligence tools that are reliable and locally relevant for Valais patients. The phrase is worth pausing over, because it does not say what one first believes one reads in it. This time it is not about the debate already conducted in these pages over the existence of a sovereign Swiss layer facing the great foreign laboratories. That debate is largely settled. The one that interests me here plays out on a finer scale, and it is not resolved in the same way.

A general medical tool, however well trained, learns from masses of records, publications and protocols that come from everywhere and from nowhere in particular. It knows a great many things that are true on average. What it does not know, by construction, is how a particular department of the Hôpital du Valais refers a patient from a valley practice to Sion, which examinations are carried out on site and which require a journey, the words that a patient of sixty puts to a pain and that a model trained elsewhere may misread. There is nothing spectacular in this list, and that is precisely why it never surfaces in public debate, even though it determines whether a tool truly helps or whether it adds a layer of translation between the patient and those who care for him.

The Hôpital du Valais employs more than six thousand people. That figure gives the measure of what would have to be replaced were we content merely to plug a general-purpose service into a system of this size without ever adapting it to its own pathways. That is not what Idiap proposes. The institute — established in Martigny since its founding, thirty-five years ago now, and shaped from the outset by speech recognition and language processing, which is to say by the precise manner in which a person formulates what is wrong — does not seek, to my mind, to compete with the great models on their own ground, raw power. It occupies the ground they leave empty: local relevance, clinical validation, the trust that a caregiver can place in a suggestion because he knows where it comes from.

Most hospitals, Valais or otherwise, have neither the means nor the calling to train a medical model from scratch, and the choice between the actor and the shop window is not played out on that caricatured alternative. It is played out between the two: are the adaptation, the validation, the fine tuning of a tool for a given department carried out here, by people who answer for their choices before the patients concerned, or does the hospital accept a closed product, delivered from elsewhere, with no oversight of what it corrects, generalises or ignores?

That is what practical sovereignty in medicine means. Not the abstract capacity of a country to say that it owns its models, but the concrete capacity of a department to say where each recommendation comes from and why it applies to this particular patient rather than to an average statistical one. A valley doctor can explain his decision in a single sentence because he has known his patient for fifteen years. An imported tool, unadapted, will never be able to explain its own in the same way, whatever the quality of the model that produces it — and a patient senses this, even if he could not put it in those terms.

One does not care for a population with a tool that knows it on average. One cares for it with a tool that knows, at least a little, where it comes from.