Bringing Structured Clinical Evidence and Reasoning to Thumos Care
Doctors do not need more medical text. They need fast, trustworthy answers grounded in clinical guidelines, evidence strength, decision pathways, and current research.
That is why we built a new clinical evidence and reasoning layer inside Thumos Care.
Over the past phase of development, we processed more than 90 clinical guidelines across specialties and extracted not just paragraphs, but structured clinical knowledge:
- recommendations with detailed metadata, including evidence strength
- clinical algorithms
- grouped supporting passages
- links into a biomedical knowledge graph containing entities such as diseases, drugs, pathways, phenotypes, biological processes, cellular components, and molecular functions
We also generated clean SVG diagrams for each extracted clinical algorithm so that decision pathways can be visualized, not just described.
When a doctor submits a query in Thumos Care, the system retrieves relevant guideline recommendations, algorithms, and supporting passages using vector similarity over this structured knowledge layer. It can then search PubMed and other sources for additional research insights. A state-of-the-art reasoning model synthesizes the final response, with the ability to query the guideline layer multiple times through tools during answer generation.
In other words, this is not just chat over medical PDFs.
Why we built this
Clinical guidelines are indispensable, but they are not easy to use under real-world conditions.
They are long. They are distributed across specialties. Important recommendations are often embedded in dense text. Decision pathways are buried in figures and tables. And even when a guideline is clear, clinicians may still want to understand how it relates to newer literature, specific phenotypes, mechanisms, contraindications, or adjacent specialties.
Most medical AI systems flatten all of this into unstructured retrieval. That approach can be useful, but it throws away too much of what makes guidelines clinically valuable.
Guidelines are not just documents. They contain structure:
- recommendation strength
- quality of evidence
- sequencing logic
- escalation pathways
- decision branches
- population-specific nuance
We wanted to preserve that structure and make it directly usable.
What we built
We developed a pipeline that extracts and organizes multiple layers of clinical knowledge from guideline documents across specialties.
1. Structured recommendations
Rather than treating a guideline as a block of text, we extract individual recommendations and preserve detailed metadata, including evidence strength and other supporting attributes.
This makes it possible to retrieve not just "something relevant from a document," but the specific recommendation most relevant to the clinician's question.
2. Clinical algorithms
We also extract clinical algorithms and convert them into clean SVG diagrams.
That matters because many important parts of clinical reasoning are not best represented as prose. They are represented as flows: branching decisions, escalation criteria, diagnostic pathways, therapy sequencing, and treatment adjustments.
By making these algorithms directly retrievable and viewable, Thumos Care can surface not just answers, but decision logic.
3. Grouped supporting passages
In addition to structured recommendations and algorithms, we preserve grouped paragraphs from the source guidelines so that supporting context remains available.
This helps ground answers in the surrounding discussion rather than forcing the system to rely on isolated snippets.
4. A biomedical knowledge graph
All of this is connected to a broader biomedical knowledge graph that already contains entities such as:
- Disease
- Drug
- Pathway
- Phenotype
- BiologicalProcess
- CellularComponent
- MolecularFunction
This creates a richer substrate for retrieval and reasoning across mechanisms, clinical entities, and evidence sources.
How it works at query time
When a doctor submits a search query, Thumos Care retrieves relevant recommendations, algorithms, and supporting passages from this structured evidence layer using vector similarity.
The system can then search PubMed and other sources for additional research insights.
A state-of-the-art reasoning model synthesizes the final response in a constrained multi-step process, with the ability to query the guideline layer multiple times through tools while formulating the answer.
This matters because good clinical answers often require more than a single retrieval pass. A useful system should be able to refine its search, revisit the evidence layer, and pull in the right guideline objects before producing a synthesis.
What makes this different
A lot of medical AI still amounts to some version of this pattern:
- retrieve chunks from documents
- pass them to a model
- generate an answer
That can work, but it is not enough.
We believe clinical AI needs to reason over structured evidence, not just over retrieved text.
That means preserving the distinction between:
- a recommendation and the paragraphs around it
- an algorithm and a descriptive section
- evidence-backed guidance and newer exploratory literature
- biomedical entities and the relationships between them
The goal is not simply to make medical search faster.
The goal is to build a system that is better aligned with how clinical reasoning actually works: guidelines, pathways, evidence quality, biological relationships, and current research all interacting in the same answer process.
Why this matters for clinicians
Real clinical questions are rarely just lookup questions.
A physician may want to know:
- what the guideline recommends
- how strong that recommendation is
- where it sits in a broader decision pathway
- what factors modify that recommendation
- whether newer research adds anything important
Those are not all the same question.
A useful clinical system should be able to support them together, while staying grounded in evidence.
That is the direction we believe clinical AI needs to move in: away from unstructured retrieval alone, and toward structured evidence-aware reasoning.
Now live in production
This new clinical evidence and reasoning layer is now live in production in Thumos Care.
The next step is rigorous benchmarking against existing approaches, including clinician workflows that use established evidence tools, to better understand where this system performs best, where it still needs improvement, and how it can become genuinely useful in day-to-day clinical practice.
More to come soon.