Limited tasks and predefined criteria.
Science and validation
We test more than accuracy. We test usefulness and safety.
We assess data quality, algorithm performance, the clinical process and whether people understand the result correctly as separate questions.
Testing the process and safety before claiming an effect.
With methods, limitations and conflicts of interest.
Claims must not get ahead of evidence.
Validation framework
Four independent levels
Strength at one level cannot conceal weakness at another.
Does the system work correctly?
Versions, access, resilience and audit.
Has the data been extracted correctly?
Completeness, units, timing and reproducibility.
Is the result useful to a doctor?
Relevance, safety and supporting evidence.
Is the result understandable?
Limitations must not create false confidence.
Approval stages
From code to a clinical use case
Technical testing
The feature exists and passes tests.
That is not enough for clinical use.
Expert review
The intended use and hazardous operating modes are defined.
Limits are set in advance.
Evaluation study
The sample, reference standard and criteria are defined before the study begins.
No adjusting them after seeing the result.
Clinical pilot
We test operation within a real workflow under specialist supervision.
Access is limited to pilot participants.
Launch
Users, monitoring and a rollback method are defined.
Quality is monitored after launch as well.
What we measure
A pilot tests the process
Metrics are agreed before it starts.
Data
History completeness
Transformation errors
Proportion requiring manual review
Process
Preparation time
Review time
Completed follow-up
Safety
Rate of unsafe suggestions
Quality of handover to a doctor
Rule violations
Understanding
Doctor understanding
Patient understanding
Fit with the workflow
Transparency
Evidence will be published
Protocols, versions, results and limitations — after review. No public effectiveness data is available yet.
Hypothesis generation methodologyA public description of the principles is already available.
→Clinical validation protocolsTo be published after the task and study design are approved.
In preparationPilot resultsPublication only with methods, limitations and participants' permission.
No data yetModel and system cardsPlanned for active AI features and significant versions.
In preparationBIOS welcomes independent validation.
One use case. One owner. Measurable criteria.