Seven publications reveal the thread in Sofya research: structure clinical language, preserve context, and evaluate models where care actually happens.
editorial · 7 min
Sofya research did not begin with a generative model. It began with a more fundamental question: how can a system represent what happens in care without losing clinical meaning? Seven publications from 2022 to 2025 trace a path from terminologies and ontologies to foundation models and LLM evaluation.
Structure before generation
In 2022, a terminology server presented at CBIS proposed recognizing and coding diseases, procedures, and symptoms from clinical notes. That same year, the ONTOBRAS paper combined SNOMED CT, OWL-Time, and semantic rules to represent fall risk in hospitalized patients. The common principle was already clear: clinical intelligence needs shared language, time, and explicit relationships.
research trajectory
Generation came after clinical structure.
The work advances through four dependent layers: represent concepts, define a clinical problem, evaluate in real language, and only then expand model capability.
2022
Clinical semantics
Terminologies, time, and relationships make meaning computable.
2023
Defined problem
Anaphylaxis becomes a bounded clinical field for investigation.
2024
Applied evaluation
Ontologies and LLMs are challenged against criteria and clinical text.
AAAI · foundation
A model with structure
SoftTiger brings this foundation to one of AI’s leading scientific societies.
A synthesis of the seven publications below. The stages show conceptual dependency, not a standalone product timeline.
A clinical problem as a proving ground
Anaphylaxis became a demanding proving ground. An abstract presented at the ASBAI congress in 2023 explored AI and clinical language processing to support diagnosis. In 2024, a SNOMED CT sub-ontology structured criteria for flagging suspected cases. A second study compared four LLMs for identifying true cases and received an honorable mention at CBIS.
Evaluate before generalizing
The work published in Asia Pacific Allergy expanded the evaluation to 969 Portuguese clinical texts. More important than any isolated result is the discipline represented by the study: define a clinical problem, build a reference, measure sensitivity and specificity, and expose the model to real language. In healthcare, evaluation is not a step after the product; it is part of the architecture.
From structured data to a foundation model
At the 2024 AAAI Spring Symposium, SoftTiger carried that trajectory into a clinical foundation model capable of structuring notes as IPS, clinical impression, and medical encounter data using international standards. Founded in 1979, the Association for the Advancement of Artificial Intelligence describes itself as the premier scientific society dedicated to advancing the understanding of AI and is one of the field’s global reference institutions. Presenting the work in its symposium series placed Sofya research inside an international scientific conversation. The leap does not erase the earlier work. It depends on it: terminologies, interoperability, and provenance make model output usable inside a clinical system.
The thread that remains
These publications are not a linear product roadmap. They form something more durable: a set of principles. Context must be structured; concepts must travel across systems; models must be evaluated against defined clinical problems; and every suggestion must retain a connection to its origin. That thread connects Sofya research to the second clinical perspective that follows every case.
