Hallucinations
Generating medical information, facts or recommendations that are unsupported, inaccurate or not grounded in a reliable source.
AI Medical Model for Clinical Reasoning Simulation
This research investigates the problems that cause hallucinations in some current medical LLMs and presents a medical model designed to reduce and detect such hallucinations before the generated output reaches a specialist physician. The research findings led to a patent application for a safe, comprehensive medical AI model intended to support physicians across all medical specialties.
Engineering Clinical Diagnosis
A visual introduction to the behavioural and engineering disorders that may limit the reliability of current medical AI models.
Clinical AI Disorders
Each term below describes a distinct reliability problem rather than a comparison between models.
Overconfidence
Presenting an answer with excessive certainty despite ambiguity, insufficient information or an inability to verify the conclusion.
Catastrophic Forgetting
Loss or displacement of previously learned medical knowledge when newer information or model updates alter existing representations.
Unsupported Outputs
Producing clinical statements without adequate evidence, source verification or a transparent basis for the answer.
Clinical Risk
The potential for unreliable AI output to mislead interpretation, influence decisions or compromise patient safety.
Current Model vs Proposed Direction
The visual comparison introduces the difference in research direction before the simplified comparison table.
Research Comparison
A concise comparison of the current limitations and the proposed research direction without exposing implementation details.
Current Medical AI
- Language prediction
- Unorganised medical information
- Weak source verification
- Answers despite insufficient information
- Potential clinical risk
Proposed Research Model
- Clinical reasoning simulation
- Organised medical information
- Mechanisms to reduce catastrophic forgetting
- Governance behavioural protocol
- Continuous verification process
- Patient-safety-oriented design
Validation Phase
CPIP–Python Code
A dedicated validation phase for testing the behavioural protocol through Python-based evaluation, while the internal engineering implementation remains undisclosed.
CPIP_PYTHON_CODE / VALIDATION_PHASE