Why Knowledge Graphs Stall Before the Clinic: Four Failures and a Way Forward

The short version

Knowledge graphs work. They helped point researchers to a COVID-19 treatment, they connect billions of patient records across dozens of countries, and they sit underneath most of modern bioinformatics. Yet they seldom reach the place where life-sciences organizations feel the most pain: the day-to-day job of running, checking, and defending a clinical trial.

We think the reason is structural. Four failures, built into how the technology grew up, keep knowledge graphs from carrying the one thing regulated work needs most: a record of who said what, when, and on what basis. This paper names those four failures, shows how time makes each of them worse, and explains why a standard built for smart cities is a surprisingly good fit for clinical trials.

Knowledge graphs have earned their place

A knowledge graph stores facts as connected things rather than rows in a table: this patient, this drug, this lab result, and the relationships between them. The idea goes back to 2001, when Tim Berners-Lee, James Hendler, and Ora Lassila described a web of data that software could reason over, in "The Semantic Web" in Scientific American. It went mainstream on 16 May 2012, when Google launched its Knowledge Graph under the banner "things, not strings," with more than 500 million objects and 3.5 billion facts at launch.

Healthcare and life sciences have real wins to show for it, and they are worth celebrating:

• A COVID-19 treatment. Scientists at BenevolentAI used an AI-enhanced knowledge graph, built from machine-read literature and structured databases, to identify baricitinib, an approved rheumatoid arthritis drug, as a COVID-19 candidate (Richardson et al., Vaccines, 2022). In the COV-BARRIER trial of 1,525 hospitalized patients, baricitinib was associated with a 38.2% relative reduction in 28-day mortality (Marconi et al., Lancet Respiratory Medicine, 2021), and the FDA approved it for certain hospitalized adults with COVID-19 on 10 May 2022 (FDA).
• Population-scale research. The OHDSI community's shared vocabularies support more than 331 data sources holding 2.1 billion patient records across 34 countries (JAMIA, 2024).
• The backbone of biology. The Gene Ontology provides 7.5 million human and model-organism annotations (Gene Ontology Consortium, 2023), and open graphs such as Hetionet link 47,031 nodes through 2,250,197 relationships drawn from 29 public resources (Himmelstein et al., eLife, 2017).
• Clinical study design at Novo Nordisk. Novo Nordisk built StudyBuilder, a clinical metadata repository on a graph database that links study designs to the data standards they use, and released it as open source as OpenStudyBuilder in 2022 (OpenStudyBuilder; GitHub). The project describes it as one source of truth for study design specifications, with versioning and a full audit trail, feeding everything from protocol authoring and CRF design to submission data. According to the project, it went live in the business in Q4 2023, with all Phase 2/3 studies using it for protocol authoring, and by October 2025 its schedule of activities was live for all interventional studies, Phase 1 to 4, at Novo Nordisk (OpenStudyBuilder milestones).
• Shared meaning at Roche. Roche runs an in-house semantic infrastructure built around its terminology system (called RTS in earlier publications and RTiS in a 2026 talk), which gives concepts such as substances, indications, and trial phases one persistent identifier wherever they are used. In a 2022 talk, a Roche product manager said more than 100 production applications had been integrated on the terminology side (Knowledge Graph Forum 2022). Roche authors report that because their data catalogs share those identifiers, concepts are identical across them, and integration into their dataset portal "comes without any additional integration and transformation effort" (Pistoia Alliance FAIR Toolkit). Roche has also published how its application ontologies are built in that system and served through JSON-LD APIs (Schwagereit et al., SEMANTiCS 2022), and in 2026 presented its approach to combining central and federated governance of those ontologies (Graphwise AI Summit 2026).

Where an organization can fund the build, the knowledge graph pays off. The question is why that "where" is so small.

And yet they rarely reach the clinic

A systematic review of healthcare knowledge graphs found "only a limited utility in applying the constructed KGs to real-life applications," with many graphs never properly evaluated or made available for inspection (Abu-Salih et al., Journal of Big Data, 2023).

To see why, start with a question that comes up in clinical research all the time:

What did we know about Participant 204 when Protocol v3.0 was implemented at Site 123?

Picture a safety toxicologist asking it. Answering it means checking re-consent, delegation and training, specimen collection, shipping and cold chain, lab work, and the data captured in the trial systems. Every one of those systems kept careful records. None of them holds the answer. So a person walks the chain by hand: a report from one system, an email to a data manager, a call to the monitor, and by Thursday a call to the site.

We call this digital archaeology: skilled people digging through exports and spreadsheets to rebuild a chain of evidence that no system kept whole. It happens because organizations have structural amnesia. They cannot remember what they never wrote down, and the thing they almost never write down is the link between facts, with who made it and when.

This is not an interoperability problem. The industry has spent two decades getting data from one system to another. The problem is that the context travels badly. Many systems collect their own audit trails with great care, and then the moment data is exported, the record of where it came from is left behind. Knowledge graphs should be the cure for this. Four failures explain why they usually are not.

The four failures

1. The reification crisis

In clinical work, the facts about a fact matter as much as the fact itself. "This participant has Type 2 diabetes" is not useful on its own. You need to know who recorded it, when, on what evidence, and how sure they were.

The classic knowledge graph standard, RDF, stores knowledge as simple three-part statements and has no native place to hang that extra information on a single statement. The standard workaround, called reification, turns one fact into a cluster of extra statements. An empirical study found that it takes four statements to annotate a single fact this way, which "will magnify the size of the KG by four times" while making the queries more complicated (Govindapillai, Soon & Haw, F1000Research, 2021).

So the very information regulators care about most, provenance and confidence, is the information the classic approach makes most expensive to keep. Under budget pressure, it gets left out.

2. The cold-start problem

A new knowledge graph platform starts empty. In a domain like clinical research, the useful questions only become answerable after a great deal of meaning has been modeled: studies, sites, participants, visits, specimens, deviations, and the standards and terminologies that describe them. Until that work is done, the graph cannot answer much, and that work is slow, specialized, and expensive.

This is why the successes cluster in organizations that can afford to wait. Everyone else either never starts, or hires a vendor to fill the empty database for them, which leads straight to the third failure.

3. Proprietary ontology lock-in

Data can be exported. Meaning usually cannot. An ontology, the model that says what things are and how they relate, captures an organization's hard-won understanding of its own work. When that model lives inside a vendor's platform, in that vendor's format, the organization is renting its own institutional knowledge. Leave the platform and the tables come with you. The meaning stays behind.

Lock-in is a cousin of the cold-start problem: the fastest way past an empty database is to borrow someone else's model, and the price of borrowing is that you never own it. Practitioners in the field have made the same point from the owner's side, arguing that a well-built ontology is core intellectual property and that open standards are how you keep it yours (Seale, "Your Ontology, Your IP," 2024).

4. The two-worlds divide

The field split into two camps. RDF and the semantic web stack offer rigor: global identifiers, shared vocabularies, formal reasoning. Property graphs offer practicality: easy to attach details to a relationship, easy to traverse, friendly to developers. Each camp tends to be dogmatic, and organizations end up forced to choose rigor or usability, or to pay for both.

It does not have to be either-or. Our view is simple: linked data should be the shared surface that systems talk through, and teams should be free to store and compute however they like underneath. Developers should not have to become semantic web specialists to read and write meaningful, traceable data.

The clock running through all four

Each failure gets worse with time. Knowledge changes. Identifiers get replaced, definitions get revised, models drift. A graph that cannot record when something was true cannot tell you when its knowledge has gone stale, and stale knowledge does not announce itself.

The clinical record has examples. At Brigham and Women's Hospital, an alert that reminded clinicians to monitor thyroid function in patients on amiodarone silently stopped working when an internal identifier for amiodarone was changed in another system (Wright et al., JAMIA, 2016). When researchers externally validated a widely deployed sepsis prediction model in 27,697 patients, it reached an AUC of 0.63 and failed to identify 67% of patients with sepsis (Wong et al., JAMA Internal Medicine, 2021).

That is why the most important question in regulated work is rarely "what is true?" It is "what was true then, and what did we know at the time?" Answering it takes two clocks: one for when something happened in the world, and one for when we recorded or learned it. Engineers tend to avoid that second clock. People end up paying for its absence, usually on a weekend before a deadline.

What a smart city taught us about clinical trials

Our founder, Bo Lora, came to this problem from an unexpected direction: smart cities. His interest grew out of working with smart city technology, in particular the open-source FIWARE community and the standard it builds on, which exist to help the many independent systems in a city share what they know.

The resemblance to a clinical trial is striking. Think of a trial as a city: many buildings, many citizens, many vendors, many rules, many secrets to keep, and no two systems wired the same way. Sites, labs, couriers, sponsors, monitors, and participants are all independent actors, each with their own systems, each holding a piece of the picture. What makes a city "smart" is not one giant database. It is a shared way for all those independent parts to publish what they know, in context, so anyone with the right to ask can get an answer.

Smart cities use an open standard for exactly this, called NGSI-LD, published by the European Telecommunications Standards Institute (ETSI) (ETSI GS CIM 009). ETSI describes its goal as becoming the "glue" between applications and databases for smart cities, smart agriculture, smart manufacturing, and more (ETSI GS CIM 029). It is built on JSON-LD, a W3C standard, so it speaks the language of linked data while still looking like ordinary JSON to a developer.

In plain terms, here is what that brings to a clinical trial, mapped to the four failures:

• Details live on the link itself. In NGSI-LD, properties and relationships can carry their own properties, such as who asserted something, when it was observed, and how confident they were (ETSI GS CIM 009). That speaks directly to the reification crisis.
• Facts carry time. Observation and system timestamps are part of the standard, so "as of" becomes something you can ask, not something you reconstruct.
• The question travels, the record stays. The standard defines centralized, distributed, and federated ways of deploying, so a question can go to the systems that hold the answers instead of copying everything into one place (ETSI GS CIM 009).
• Open and portable. Because it is an open standard on top of JSON-LD, meaning can travel between implementations rather than staying locked in one, and it bridges the two worlds: linked-data rigor underneath, developer-friendly JSON on top.

There is a second reason the fit matters now: the two worlds are converging. Clinical trials increasingly collect data from devices and people outside the clinic. The FDA issued final guidance in December 2023 on using digital health technologies, such as wearables and sensors, to collect data remotely from trial participants (FDA, 2023), and in September 2024 on running trials with decentralized elements, such as telehealth and in-home visits (FDA, 2024). A trial with wearables, home visits, and a dozen vendors looks more like a smart city every year. It makes sense to borrow the tools that were built for one.

What it does not solve on its own

We want to be honest about the limits. In our view, NGSI-LD gives you one clock well: when something was observed. What regulated work also needs is the second clock, a clean way to ask not only what was true on a date, but what we knew on that date. That is work still to be done. We have also not found documented production use of NGSI-LD in regulated clinical settings, so this is an argument from architecture, not a track record. And no standard, however good, removes the need for people who know the domain to say what things mean.

Most of all, clinical teams should never have to learn NGSI-LD. The standards they already use stay where they are. A good shared layer links standards without replacing them, and stays out of sight.

It is about provenance, not plumbing

The standards the industry relies on were asked to say what things are, and they do that well. Now we are asking them to help run clinical trials, which is a different craft. Running a trial means keeping track of how thousands of facts connect, who vouched for each one, and when.

Born FAIR: a birth certificate and a passport

Here is the way we like to picture it. When data is born FAIR, meaning findable, accessible, interoperable, and reusable from the moment it is created (Wilkinson et al., Scientific Data, 2016), it gets a birth certificate: where it came from, who recorded it, and when. You cannot get a passport without a birth certificate, so the data gets a passport too. Then, every time it crosses a border, from one system to the next, from site to sponsor, from a CRO to a regulator, the passport gets stamped.

The point is not the paperwork. It is infrastructure that keeps the birth certificate and the passport with the data at every step down the pipeline, so nobody ever has to go back and reconstruct the journey. That is the opposite of digital archaeology. When the toxicologist asks what we knew about Participant 204 when Protocol v3.0 reached Site 123, the blood draw's birth certificate and every stamp on its passport are already there to read.

A practical test for any shared layer is whether it lets you say yes to six things:

1. Every thing has a name any system can resolve: the visit, the tube, the protocol version, the site, the person who drew the blood.
2. Every link between things is data, not a join someone writes later.
3. Every link and every value carries who asserted it, when, and on what basis.
4. Every fact carries when it was true, and you can ask "as of August 3."
5. The systems stay where they are, and the question travels to them.
6. It is an open standard nobody owns, that any vendor can implement.

When those hold, the toxicologist's question about Participant 204 stops being a week of digital archaeology and becomes a query.

Where to go next

Fixing the four failures is not something any single company should own. It starts with shared meaning: reference models built in the open, anchored in the vocabulary of the people who actually do the work, and portable by design, so no organization has to begin from a vendor's seed or rent back its own knowledge.

That is the idea behind The Ontology Project, an open, community-governed effort that ScientixAI sponsors. Its manifesto lays out the problem, the values, and the commitments: open licensing, versioned releases, rotating governance, and a door out. The first reference graph it commits to is clinical research.

Read the manifesto at top.scientix.ai/manifesto.html. Argue with it, mark it up, and tell us where we are wrong.

If you would like to talk about any of this, write to us at info@scientix.ai.

About the authors

Bo Lora is the founder of ScientixAI. He is a veteran, led UX at Amazon Business and served as a Principal Design Technologist in Healthcare and Life Sciences at AWS. His interest in knowledge graphs for clinical research grew out of his smart city work: a clinical trial, like a city, is many independent systems and people that have to share context they can trust.

Marcie Ogilvy is ScientixAI's Acting Chief Marketing Officer and a digital teammate: an AI agent, not a person. She works alongside Bo and the company's other human and digital teammates. On this paper she cut down a longer research manuscript, checked every claim against a primary source and cut what she couldn't verify, and shaped the argument for a life-sciences audience. Bo set the thesis, supplied the ideas, and reviewed and approved the final text.

We publish this way on purpose. ScientixAI builds infrastructure so that every claim can show where it came from and who stands behind it. A paper about provenance should do the same.

Sources

1. Berners-Lee T, Hendler J, Lassila O. "The Semantic Web." Scientific American, May 2001. https://www.scientificamerican.com/article/the-semantic-web/
2. Singhal A. "Introducing the Knowledge Graph: things, not strings." Google, 16 May 2012. https://blog.google/products/search/introducing-knowledge-graph-things-not/
3. Richardson PJ, Robinson BWS, Smith DP, Stebbing J. "The AI-Assisted Identification and Clinical Efficacy of Baricitinib in the Treatment of COVID-19." Vaccines 10(6):951, 2022. https://pmc.ncbi.nlm.nih.gov/articles/PMC9231077/
4. Marconi VC et al. "Efficacy and safety of baricitinib for the treatment of hospitalised adults with COVID-19 (COV-BARRIER)." Lancet Respiratory Medicine 9(12):1407, 2021. https://pmc.ncbi.nlm.nih.gov/articles/PMC8409066/
5. U.S. FDA. "FDA Approves Treatment for COVID-19 in Certain Hospitalized Adults." Drug Information Update, 10 May 2022. https://content.govdelivery.com/accounts/USFDA/bulletins/31738e6
6. Reich C et al. "OHDSI Standardized Vocabularies: a large-scale centralized reference ontology for international data harmonization." JAMIA 31(3):583-590, 2024. https://pmc.ncbi.nlm.nih.gov/articles/PMC10873827/
7. Gene Ontology Consortium. "The Gene Ontology knowledgebase in 2023." Genetics 224(1):iyad031, 2023. https://pmc.ncbi.nlm.nih.gov/articles/PMC10158837/
8. Himmelstein DS et al. "Systematic integration of biomedical knowledge prioritizes drugs for repurposing." eLife 6:e26726, 2017. https://pmc.ncbi.nlm.nih.gov/articles/PMC5640425/
9. Novo Nordisk A/S. OpenStudyBuilder project site (description and milestones). https://www.openstudybuilder.com/
10. Novo Nordisk. openstudybuilder-solution repository, GitHub. https://github.com/NovoNordisk-OpenSource/openstudybuilder-solution
11. Romacker M (Roche). "Building a Semantic Capability Stack to Support FAIR Knowledge Graphs at Scale." Knowledge Graph Forum 2022, abbreviated presentation hosted by Ontotext. https://www.ontotext.com/blog/building-a-semantic-capability-stack-to-support-fair-knowledge-graphs-at-scale/
12. Romacker M, Balakrishnan R, Steiner O, De Schepper H, Kuah Y. "FAIR data by design: Roche." Pistoia Alliance FAIR Toolkit use case. https://fairtoolkit.pistoiaalliance.org/use-cases/fair-data-by-design/
13. Schwagereit F, Romacker M, Richard F, Trypuz R, Liener T, Roche O. "FAIR Data APIs in the FAIR in Vivo Data Sharing Platform." SEMANTiCS 2022 Posters, Demos and Workshops, CEUR-WS Vol-3235. https://ceur-ws.org/Vol-3235/paper25.pdf
14. Graphwise AI Summit 2026. "Harmonizing Centralized and Federated Governance: How to build FAIR Minimal Viable Ontologies" (Roche session, 7 October 2026; abstract public, recording behind free registration). https://graphwise.ai/summit-videos/
15. Abu-Salih B et al. "Healthcare knowledge graph construction: a systematic review of the state-of-the-art, open issues, and opportunities." Journal of Big Data 10:81, 2023. https://pmc.ncbi.nlm.nih.gov/articles/PMC10225120/
16. Govindapillai S, Soon L-K, Haw S-C. "An empirical study on Resource Description Framework reification for trustworthiness in knowledge graphs." F1000Research 10:881, 2021. https://pmc.ncbi.nlm.nih.gov/articles/PMC8634049/
17. Seale T. "Your Ontology, Your IP." The Knowledge Graph Guys, 6 December 2024. https://www.knowledge-graph-guys.com/blog/your-ontology-your-ip
18. Wright A et al. "Analysis of clinical decision support system malfunctions: a case series and survey." JAMIA 23(6):1068-1076, 2016. https://pmc.ncbi.nlm.nih.gov/articles/PMC5070518/
19. Wong A et al. "External Validation of a Widely Implemented Proprietary Sepsis Prediction Model in Hospitalized Patients." JAMA Internal Medicine 181(8):1065-1070, 2021. https://pmc.ncbi.nlm.nih.gov/articles/PMC8218233/
20. FIWARE Foundation. https://www.fiware.org/
21. ETSI. Context Information Management (CIM); NGSI-LD API. ETSI GS CIM 009 V1.9.1, July 2025. https://www.etsi.org/deliver/etsi_gs/CIM/001_099/009/01.09.01_60/gs_CIM009v010901p.pdf
22. ETSI. Context Information Management (CIM); NGSI-LD Implementation Conformance Statement. ETSI GS CIM 029 V1.1.1. https://www.etsi.org/deliver/etsi_gs/CIM/001_099/029/01.01.01_60/gs_CIM029v010101p.pdf
23. W3C. JSON-LD 1.1. W3C Recommendation, 16 July 2020. https://www.w3.org/TR/json-ld11/
24. U.S. FDA. Digital Health Technologies for Remote Data Acquisition in Clinical Investigations. Final guidance, December 2023. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/digital-health-technologies-remote-data-acquisition-clinical-investigations
25. U.S. FDA. Conducting Clinical Trials With Decentralized Elements. Final guidance, September 2024. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/conducting-clinical-trials-decentralized-elements
26. The Ontology Project. A Manifesto. https://top.scientix.ai/manifesto.html
27. Wilkinson MD et al. "The FAIR Guiding Principles for scientific data management and stewardship." Scientific Data 3:160018, 2016. https://www.nature.com/articles/sdata201618