Clinical AI that keeps data inside the institution
When health institutions evaluate AI, they often ask where their data goes before they ask how accurate the system is. This article explains why that question should come first and what a good answer looks like.
Conversations about clinical AI often begin not with accuracy or features but with the question of where the data actually goes. It is a reasonable thing to ask. Health data is among the most sensitive information an institution holds. With cloud-hosted, API-based models, the honest answer is often that the data is processed somewhere else, by someone else, under rules the institution does not fully control.
Data sovereignty means that an institution, and often a country, keeps authority over where its data is stored, who can access it and which laws govern it. In healthcare, this is increasingly a legal and ethical requirement rather than a preference. A hospital cannot send patient records to a third-party cloud on another continent and ignore the consequences. A ministry building national infrastructure often has to assume the data never leaves the country at all.
MINA was built with this constraint in mind from the start. It runs on-premise, or in an isolated cloud assigned to the institution. Clinical data is processed locally and does not leave the institution or, where required, the country. Every answer it gives is traceable back to a source and carries a confidence signal, so a reviewer can see not only what the system suggests but also why. Privacy is a basic part of the deployment rather than a setting added later.
This approach also has a strategic side. For a buyer evaluating AI across borders, data sovereignty is often the deciding factor. A platform that keeps data inside the institution removes a whole class of risk that cloud-only tools cannot avoid. Deployment follows the requirements of each market's rules: KVKK in Türkiye, GDPR in Europe and their equivalents elsewhere. The approach used to localize the model's language is applied to regulatory requirements in the same way.
Keeping data in place only matters if the system using it is careful. By design, MINA does not make decisions; it supports them, and every critical action stays with a clinician or operator. Safety is also measured. In our internal evaluation on synthetic and anonymized data, the rate of clinically critical hallucination stayed very low, and the system caught the large majority of deliberately planted clinical traps that a less careful model could miss. Task-level evaluation results for MINA v3.7 are on the MINA v3.7 page.
Accuracy is an important criterion when a system is evaluated. But when a health institution decides whether to bring AI into its workflow, it usually looks at data sovereignty, traceability and the system's cautious safety behavior. When these conditions are in place from the start, the later stages of the evaluation become much easier.