What is Ventrecs Lab?+
A laboratory information system (LIS) in the Ventrecs platform: test catalogue, specimen workflows and labels, orders, results and critical-result acknowledgement, instrument interfaces, and a review layer of rules and optional AI.
Does the AI decide results?+
No. Rules run first, with the AI switched off. A facility can run rules only, shadow mode (the model runs and is measured but nobody sees it) or supervised mode (suggestions go to named reviewers). A suggestion is marked unverified, cannot release or change a result, and needs a reason from the reviewer to modify, reject or defer it.
What do the rules check?+
Critical values, implausible values, change since the last result, specimen age and suggested follow-up tests. A follow-up test is only a recommendation until the clinician authorises it. Rule sets are versioned and approved by the laboratory director before they are active.
When is a result not interpreted?+
When the unit is ambiguous, no reliable reference interval exists, quality control is invalid or missing, the specimen is mismatched or rejected, the value is unverified, context is missing or the test is unsupported. The case states the reason. The laboratory still verifies and releases the value as usual.
How do analysers connect?+
Analysers send HL7 v2 or ASTM E1394 messages. Each message is matched to the instrument, the order and the test, with test codes mapped to your tests and a failed or missing quality-control run blocking interpretation. Doubtful messages are quarantined with the reason, a duplicate is applied once, an older measurement after a newer one is ignored, and a signed-off result is never overwritten.
How is the intelligence layer governed?+
Rules, models, prompts, references and unit conversions move from draft to submitted to approved by the director to active, and each decision names its reviewer. A chained audit log can be verified from the screen. A hosted model needs a signed data-processing agreement, a no-training term, a security assessment and validation evidence before activation, sees only de-identified facts, and nothing is trained automatically.