What the AI sees.
A deep-dive into the 9-layer clinical AI engine that decodes every Medivive lab panel and informs every protocol. Every layer is measurement and synthesis — and a licensed provider reviews and signs the result, never an algorithm alone. Most platforms order a panel and read you the headline numbers; Medivive runs every result through nine layers of analysis before a provider ever sees it, turning a column of values into a picture of how your biology is actually behaving.
The nine layers
- Trend analysis — the engine reads direction, not a single snapshot. Velocity and projections are computed across every prior collection, so a value drifting toward a threshold is flagged long before it crosses one.
- 64 derived ratios — relationships between markers that carry more signal than either marker alone, such as triglyceride-to-HDL, free-testosterone-to-SHBG, and neutrophil-to-lymphocyte. These are where a lot of early dysfunction first becomes visible.
- 146 syndrome patterns — a constellation of several in-range-but-trending values that together describe a recognized clinical pattern is the kind of thing a marker-by-marker read tends to miss.
- Personalized optimal ranges — a lab's reference interval is built from a broad population and answers whether you are sick. The engine computes ranges adjusted for your age, sex and activity context.
- Biological age — PhenoAge and BioAge25, two validated models, estimate how your biology is aging relative to the calendar; running both and reconciling them is more robust than any single algorithm.
- Literature research — live research synthesis pulls relevant peer-reviewed findings against your specific marker constellation, reasoning over sources rather than keyword-matching.
- Drug and supplement interaction analysis — contraindications and additive effects across current and proposed medications are surfaced before anything is recommended.
- 5-model adversarial consensus — five frontier models analyze the same results in deliberately different roles, then reconcile.
- Patient education — every finding is translated into plain language drawn from a structured clinical knowledge base.
Why five models — and why they argue
A single language model, however capable, has a single set of blind spots and a tendency to confidently fill gaps. The fix is not a bigger model — it is structured disagreement. Averaging hides disagreement; adversarial roles surface it. Where models disagree, the case is escalated for closer provider review instead of being smoothed into a confident-sounding middle. A claim all five models independently support is far stronger evidence than one model's assertion, and a claim only one model makes is flagged rather than adopted.
- A reasoning-frontier model carries the primary structured clinical reasoning over the full panel and the derived ratios.
- A second frontier model re-analyzes independently — a different architecture with different blind spots.
- A long-context synthesis model holds the full history, knowledge base and literature to keep the read grounded in the patient's own data.
- A research-reasoning model performs the live literature synthesis against the specific pattern.
- An adversarial critic role actively tries to falsify the emerging read and forces the consensus to defend itself.
What the AI does not do
An engine this capable is only trustworthy if its limits are explicit. These are hard constraints on how the system is allowed to operate, not aspirations.
- It never prescribes alone. The engine surfaces patterns, computes ranges and synthesizes literature. Every protocol is decided by a licensed provider in your state.
- Provider attestation is always required. Nothing reaches you as a recommendation until a licensed clinician has reviewed the output, weighed it against your history and attested to it — and the provider can override, revise, or order more testing.
- Anti-hallucination safeguards run throughout. The 5-model adversarial design exists specifically to catch confident errors; points of model disagreement are escalated for human review instead of being averaged away.
Evidence base
The methods above draw on published, peer-reviewed work — phenotypic biological-age estimation from standard clinical chemistry, triglyceride-to-HDL as an early marker of insulin resistance, neutrophil-to-lymphocyte as a derived inflammatory signal, multi-model and ensemble approaches to reducing single-model error in clinical AI, and personalized reference intervals versus broad population ranges. A consolidated reference list is maintained on the page; specific citations are being finalized for publication. Content is provided for education and is not medical advice.