On 9 September 2026, Margarita Pirinen took part in a round table of the MCCI Committee for the Development of Preventive, Integrative and Functional Medicine and Nutrition on digital technologies and AI in preventive medicine. Following committee chair Anton Polyakov's presentation, doctors and industry professionals discussed AI applications. The paper below develops the points made by Margarita Pirinen in that discussion.
Medical AI works with fragments; a decision concerns a person
A discharge summary, test, examination report or symptom description records individual observations. A language model can extract facts and prepare an explanation, but this is not yet an analysis of a person's condition.
A measurement's meaning depends on its method, reference range, treatment, symptoms and prior trajectory. Without this context, one cannot establish whether a change is new, persistent, expected after a prescription or inconsistent with other data.
The object of analysis is a state trajectory
This approach distinguishes medical data from a person's condition. Data consists of observations of that condition from different sources at different times. Its meaning emerges through connections with the person's history and other observations.
The system must identify what changed, which measurements changed together, which facts support or contradict one another and what information is insufficient. Time is the primary analytical coordinate. A new observation extends a trajectory rather than constituting a separate episode.
From source to testable hypothesis
A document, measurement, symptom or event is stored with its date and provenance.
Data is made comparable with regard to method, units and original context.
An observation is compared with the person's history and related changes.
A hypothesis is presented with its supporting evidence, gaps and limitations.
Every conclusion must lead back to specific sources and comparison rules. Without this, the result cannot be verified.
Normalisation must preserve medical meaning
Measurements with identical names from different laboratories may differ in method, units and reference intervals. Moving values into a common table without these attributes creates false comparability.
Every fact retains its original document, date, method, units, reference range and version. Transformation is recorded as a separate action: the original value, applied rule and comparison limitations remain visible. A correction creates a new version without deleting the previous one.
A language model is a component, not a source of evidence
An LLM can read documents, extract facts, classify and prepare clear text. It does not establish clinical truth. A model may plausibly invent a missing fact, combine incomparable observations or lose the connection between a conclusion and its source.
The generative layer therefore operates within formal rules. Data provenance, the validity of comparisons and the adequacy of evidence are checked separately. Automated analysis stops when the conditions are not met.
The intended result is a testable hypothesis
An autonomous diagnosis is not considered a system output. AI compares heterogeneous observations, identifies significant changes, recurring combinations, contradictions and gaps.
A hypothesis is generated only when sufficient grounds exist and is passed to the professional with the facts used, sources, temporal relationships, alternative explanations and confidence limits. When data is insufficient, the system states what is missing and does not continue interpreting.
The doctor remains part of the decision process
AI retains a long history and performs comparisons that are difficult to make manually across large datasets. The doctor assesses clinical meaning, confirms or rejects the hypothesis and determines the next step.
The medical assessment is stored in the history. Confirmation, correction or rejection of a hypothesis becomes context for the next analysis.
Regulatory constraints are part of the architecture
The meeting discussed cross-border transfer, de-identification of medical data and AI in research. Every scenario needs a defined processing purpose, storage and computation locations, permitted data, access rights and assessment of re-identification risk.
De-identification does not permit arbitrary reuse. Working with an individual's history and constructing a research cohort are separate activities. A research hypothesis requires a dedicated protocol and validation; an individual analysis result is not a scientific conclusion.
Conclusion
Medical AI must work with context and show the basis for its results.
The intended system connects observations over time, preserves provenance, distinguishes facts from interpretation and stops when data is insufficient. Its output is a testable hypothesis for a doctor's assessment.
How this approach is implemented in BIOS
