2026-09-21 · 8 min read
How to Prepare Laboratory Data for a Controlled AI Pilot
A practical governance checklist for selecting, cleaning and protecting laboratory data before an AI-assisted workflow is tested.
Tie the data to a defined pilot
Start with a narrow intended use and identify the minimum data needed to test it. Collecting extra fields without a defined purpose increases privacy, security and interpretation risks.
Diagnosis, treatment, clinical interpretation and result authorisation remain outside a general operational pilot unless governed and performed by appropriately qualified professionals.
Identify ownership and permission
Record who owns the source, who may access it and what approvals apply. Confirm whether data may be used in the selected environment before transferring anything.
Use approved security, privacy and retention controls. Never paste sensitive records into an unapproved AI service.
Check structure and meaning
Review identifiers, units, dates, reference fields, missing values and duplicated records. Create a data dictionary so a field has the same meaning to the technical and laboratory teams.
Keep source records unchanged and make cleaning steps reproducible.
Build a representative evaluation set
Include routine examples, incomplete records, boundary conditions and known problem cases relevant to the intended use. Separate development material from the controlled evaluation set.
Agree the evaluation criteria before inspecting results to reduce the temptation to choose only favourable examples.
Preserve human review and traceability
Every material output should be traceable to the input and reviewed by a named, qualified person. Record corrections, escalations and rejected outputs.
A pilot supports evidence gathering; it does not by itself establish clinical suitability, regulatory compliance or fitness for routine use.
Frequently asked questions
Should all available laboratory data be used? No. Use the minimum appropriate data for the defined purpose.
Can anonymisation remove every risk? No. Apply the organisation's approved privacy assessment and controls.
Who approves the pilot? Appropriate laboratory, quality, privacy, security and professional owners according to risk and applicable requirements.