Founder & CTO of UHBSE — building auditable models of longitudinal human biology. Healthcare IT, clinical data and probabilistic simulation. https://nitter.cf/t.co/i7pZnutCLl
Joined January 2022
- Tweets56
- Following5
- Followers3
- Likes2
Konstantin Endreev retweeted
We don’t think human biology should be represented by one giant model that always returns an answer.
A useful biological simulation system needs separate layers for:
*state
*trajectory
*time-to-event
*applicability
*uncertainty
and eventually intervention response.
Different questions require different evidence.
Sometimes the correct output should simply be: not enough support to simulate this
PK-PD and QSP enter as bounded mechanistic modules. The LLM sits above the core: it helps formulate questions and explain results, but it does not invent the numbers.
Only then do we add baselines, trajectory and time-to-event models, applicability checks and uncertainty. Intervention response comes after the data and evaluation design are ready.
Then comes Current State: what we actually know about a person at a chosen moment, how recent the evidence is and what is missing.
UHBSE cannot be built as one giant AI model. Different questions require different tools, and every layer has to be tested independently.
I do not need a model that answers every question.
I need a model that recognizes when a case falls outside its evidence and can stop with a clear reason.
In medicine, “I don’t know” can be the most accurate answer a system gives.
So our order is simple: semantics, provenance and temporal correctness first. Baselines next. More complex ML only after that.
If we silently merge those records, the model will learn our mapping mistakes instead of biology. The more advanced the model is, the more convincing that mistake may look.
The same lab-test name in two clinical systems does not guarantee the same measurement. Units, specimen, method, timing and result structure may all be different.
When we started, I thought the hardest part of UHBSE would be the mathematical model. It turned out that first we had to prove what every input fact actually means.
One prediction is often too confident an answer for biology.
I am interested in a different structure: one real state, several data-supported trajectories, a visible point of divergence and honest uncertainty for every branch.
That is the foundation of UHBSE.
@uhbse_engine
@SergeyNarykov
We are building a research-grade trajectory engine designed to reconstruct a current state, test model applicability and compare possible branches with uncertainty. Clinical use would require separate validation stages.
Let me be clear from the beginning: @uhbse_engine is not a medical chatbot, an electronic health record or a finished “digital twin of the human body.”
UHBSE is currently in the research and validation phase. I want to share not only what already works, but also what we still have no right to claim.
It is based on real longitudinal clinical data, temporal correctness, testable models and explicit uncertainty. Not one polished prediction, but multiple branches with visible limits of confidence.
That question became UHBSE: the Universal Human Biology Simulation Engine. We are building a research engine that reconstructs a current state and compares several possible trajectories.
What could happen to this person’s biological state under a different intervention — before we test it in real life?
