What clinics and studies showed
A decision support platform at a Moscow clinical center raised early-stage disease detection by 14% (2025). A system that offers the doctor three likely diagnoses out of 265 groups has processed 14.4 million cases in Moscow polyclinics and works in 19 regions. In Kenya, according to a study by a clinic network and the model developer, an AI consultant in the medical record cut diagnostic errors by 16% and treatment errors by 13% across 39,849 visits (2025). A research orchestrator of language models, according to its developer, correctly solved 85.5% of 304 complex NEJM cases, against 20% for 21 practicing doctors, but the authors state plainly that it needs validation and regulatory approval before clinical use (2025).
How decision support systems work
The general principle: the system reads data that already exists, such as an image, an ECG, a medical record or complaints, and highlights what is easy to miss. The decision stays with the doctor: this is how registered systems work. A signal must lead to action: at a Toronto hospital, a senior doctor examined a patient with a high-risk signal at once, and unexpected deaths in the general medicine unit fell by 26%.
Where the limit is
Software that the maker intends for diagnosis, for example to interpret health data or assess risk, counts as a medical device under Russian law (Law 323-FZ, Article 38) and is registered with Roszdravnadzor, Russia’s healthcare regulator; in August 2026 the regulator reported 59 registered domestic medical devices with AI. Everything below this line, such as booking, reminders, documents and schedule summaries, can be rolled out without registration in weeks.
What a clinic can take from this
Ask a diagnostic AI vendor for the registration certificate number, the risk class and validation data on the Russian population. Measure not the accuracy in a report but the share of cases found and the number of false alarms on your own patient flow. Hand service tasks to AI in parallel, without waiting for diagnostic AI.