medical-ai · calibration · VEX
AXISMED
AI-assisted clinical calibration and decision support system, under development within the AxisDynamics research ecosystem.
- Status
- active
- Authors
- axisdynamics
- Category
- research
- Documentation
- Not available
- Repository
- Not available
Summary
AXISMED is our applied research line in the medical domain: an AI-assisted clinical decision support system whose goal is not to answer more, but to answer with calibration. In a context where a safe but wrong answer can be more dangerous than an abstention, the design focuses on the system knowing when it does not know and declaring it.
The project builds on AxisDynamics’ two research lines: probabilistic calibration and applied metacognition. The question that orders it is simple to state and hard to solve: how does a clinical assistant behave when the available evidence is not enough to support a recommendation?
Approach
Literature-grounded retrieval. Answers are built on a hybrid retrieval layer —dense search in a vector store combined with lexical retrieval (BM25)— over biomedical literature corpora. The goal is for every claim to be traceable to its source instead of depending on the model’s parametric knowledge.
Calibration and uncertainty honesty. On that basis, declared confidence is measured against actual performance (Brier Score, ECE, MCE and associated metrics) and recalibration methods are applied. The system must distinguish between what it knows, what it infers and what it cannot support.
Traceability. Every answer preserves the retrieved evidence, the calibration metrics of the case and the evaluation protocol used, so that the behavior is auditable and reproducible.
Evaluation
Measurement follows the same criterion we apply to AI systems in production: it does not evaluate whether the model answers well, but whether it knows when it does not know. We work with reference medical question sets and with semantic perturbation protocols to detect overconfidence.
Status
Active project, in the research and development phase. It has no clinical validation, regulatory certification or authorization for care use: it is a system under construction within the AxisDynamics research ecosystem. Progress is published as experiments and publications once sufficient traceability is reached.