What Is a Knowledge Graph — and Why It Matters

Note 1: This post is part 1 of a three-part series on healthcare, knowledge graphs, and lessons for other industries

Note 2: All images by author

Summary

in the first half of the 19th century, and you feel an almost paralyzing ache in your abdomen. You now have a choice. You learn to live with that pain for the rest of your life (which may only be weeks or months away depending on what’s causing that ache) or you venture to the doctor, a nightmarish experience potentially involving tortuous treatments like bloodletting, laxatives, induced vomiting, or downing vials of mercury (Hager 52). 

There is no knowledge about how diseases spread, so going into a crowded hospital could mean exposure to smallpox and cholera (Kirsch and Ogas 80). If you are unlucky enough to need surgery (or have a physician prescribe an unneeded one—again, there is almost no knowledge of disease pathways), there will be no anesthesia. Finding the best surgeon likely means finding the fastest one, who can work as rapidly as possible to minimize the time orderlies have to restrain you while you’re shrieking and writhing on the table. If you survive the surgery, you still have a significant chance of dying of an infection since there’s no knowledge of germ theory and so no aseptic techniques (Kirsch and Ogas 45). And if you’re a pregnant woman, you can expect the maternity ward to be even more fucked up. Nearly 15 percent of babies born in the UK in the mid-19th century died at birth. 

Compare that with the medical care provided in any developed country today, and let’s just say, we’ve come a long way. The infant mortality rate in developed countries is now less than 6 per 1,000 live births, or 0.6 percent. The average life expectancy in developed countries is usually higher than 80 compared to about 40 in the mid-19th century. We have drugs or other treatments for almost all of the most common diseases, and humanity is curing more every day. The future looks even more promising, especially with the increasing capabilities of AI and the funding behind them. The Chan Zuckerberg Initiative (CZI), for example, aims to help scientists cure, prevent, or manage all diseases by the end of the 21st century.   

How has healthcare made this progress? And why does healthcare continue to attract disproportionate investment in AI today? It’s not simply better data; it’s better structure around knowledge. Long before computers, medicine began developing shared understandings of diseases and causal relationships, controlled vocabularies to catalog real-world entities, and data standards to ensure observations were empirical and replicable. Taken together, these frameworks form what we might now recognize as a knowledge graph. 

At a high level, knowledge graphs solve a recurring set of problems that become unavoidable as domains scale:

• Search and retrieval across fragmented systems, formats, and terminologies
• Discovery and design in complex, interconnected systems
• Reuse and repurposing of existing knowledge and assets
• Decision support under uncertainty, with explainable reasoning
• Recommendation and personalization grounded in domain semantics
• Governance, traceability, and regulatory compliance

Mature domain knowledge graphs in healthcare are the reason drugs can be designed to target specific diseases, why your doctor knows about the negative side effects of a drug in Japan even if it goes by a different name there, and why physicians can aggregate and learn from observations from millions of clinical encounters and experiments, often in real-time.

In this three-part series, I hope to provide some context and insights around how knowledge graphs (and their precedents) have worked in healthcare, how healthcare became the industry leader in knowledge graphs, and share some potential lessons for other industries grappling with similar challenges. 

What is a knowledge graph?

A knowledge graph is a layered knowledge system in which ontologies define meaning, controlled vocabularies catalog entities, and observational data provides evidence—allowing knowledge to accumulate, evolve, and be reasoned over as understanding improves.

An ontology defines classes and the relationships between classes; it is the theory underpinning the knowledge graph. In medicine, classes are things like pathogens, diseases, and drugs. The ontology defines the constraints and causal assumptions for how these things relate. For example, pathogens are organisms and can cause diseases. Drugs are chemical substances that can target pathogens and, potentially, inhibit diseases. The ontology deals with classes rather than instances–it doesn’t tell you which pathogens cause which diseases or which drugs inhibit which pathogens. 

The instances are defined as controlled vocabularies. Controlled vocabularies are catalogs of instances of the classes defined in the ontology. For example, there are thousands of known pathogens that can cause diseases in humans: everything from viruses to bacteria to parasites. There are also thousands of drugs and thousands of diseases. These instances of classes are cataloged and maintained by experts and are regularly updated as we learn more about them. Some controlled vocabularies in healthcare are referred to as ‘omics’ because they are about things that end with the suffix “omics” such as genomics, proteomics, and metabolomics.

Note: I am using the broad term “controlled vocabularies” here as an umbrella term that includes taxonomies, glossaries, dictionaries, reference data, and thesauri. There are differences between these, but for the purposes of this high-level article, we are just going to use the term controlled vocabulary for all of them.

The way we learn more about the world is through observation, and in healthcare those observations are treated as evidence. Clinical trials and laboratory experiments produce observational data that justify, refine, or refute claims about how entities in our controlled vocabularies relate to each other. How do we know that the pathogen Treponema pallidum causes the disease syphilis? Because scientists did an experiment and measured the outcome and produced evidence. How do we know that Salvarsan targets and destroys Treponema pallidum and cures syphilis? Because scientists ran clinical studies and measured the effects of treating syphilis patients with Salvarsan.

Connecting entities like this creates a graph. Entities in a graph are sometimes called nodes, and the connections are called edges. Graphs can contain millions of nodes and edges, and with this structure, patterns start to emerge. For example, you can identify the most important or impactful nodes in a graph, distinguish clusters of nodes that are deeply connected, or find the shortest paths between different entities. These techniques (often referred to as graph analytics) are widely used in medicine as part of what is known as network medicine to identify disease mechanisms and potential therapeutic targets (Barabási, Gulbahce, Loscalzo, 2011). This is all possible with a graph, but since we have an ontology, we have more than just a graph. We have a knowledge graph.

Connections in a knowledge graph represent explicit assertions about the world: facts. The knowledge graph isn’t just saying, “Salvarsan is connected to Treponema pallidum.” It is saying “Salvarsan inhibits Treponema pallidum.” It also states that “Treponema pallidum causes syphilis.” These two facts, combined with the logic encoded in the ontology, enable the knowledge graph to infer a new relationship or fact—namely, that Salvarsan may treat or cure syphilis. This is known as reasoning or the ability to derive “logical consequences from a set of facts or axioms.” Knowledge graphs excel at this because they make both the facts and the rules for combining them explicit. 

Medicine has been using this knowledge management structure for decades. Scientists do experiments and learn new things. The findings of these experiments lead to updates in the controlled vocabularies and/or relationships between entities in the controlled vocabularies. Gene X is related to protein Y, which is involved in the biological process Z. As the number of entities and relationships grow, so does our knowledge. Sometimes, but much less frequently, the ontology changes. A substantial change in an ontology is not just an incremental increase in knowledge, but often a change in the way we understand the world.

Healthcare is the leader in knowledge graphs because it excels in all three of these layers. It has spent decades refining causal models for how the natural world works; meticulously cataloging millions of diseases, drugs, proteins, and everything else relevant for medicine; and conducting empirical, replicable experiments with standardized data outputs. These foundations were reinforced by strong regulatory pressure that mandated standardization and comparability of evidence, widespread pre-competitive collaboration and public funding, and early adoption of open, vendor-neutral semantic standards. Combined, these factors created the conditions in which knowledge graphs could thrive as core infrastructure rather than experimental technology.

What problems do knowledge graphs solve?

Once you have entities mapped together, validated with real-world evidence, and grounded in causal pathways, you have a knowledge graph, and you can do all kinds of cool stuff. I will go through some of the most prominent use cases of knowledge graphs in healthcare today and how they may apply to other domains. 

Search and retrieval

Probably the most common use case for knowledge graphs is search. Modern healthcare requires the ability to retrieve relevant, connected context across heterogeneous and multimodal data. Suppose you work at a large pharmaceutical company and you want to know everything about a given drug. You might want to repurpose this drug, assess its safety risk, or compare it with a competitor. Or, maybe the FDA asked you for information about it. You’d have to search in relational databases for experimental data, content management systems for clinical trial reports, and multiple third-party databases for established public or industry knowledge. Not only is the data scattered across disconnected systems and in different formats (relational, text, slides, audio), the drug may also go by different names. The company may have outsourced clinical trials to a UK company who called it by its generic name, for example. 

As generative AI has become more widely adopted, retrieval has emerged as a critical capability in every industry. Large Language Models (LLMs) were trained on a lot of data, but not your data, so the ability to retrieve relevant internal context is crucial when working with these models. We now call this context engineering: “the art and science of filling the context window with just the right information at each step of an agent’s trajectory,” as described by Lance Martin of LangChain. 

Healthcare is uniquely well positioned to take advantage of this new era of AI because of its longstanding investment in knowledge graphs. Tasks like filing regulatory reports are a lot easier if you are able to retrieve the relevant internal context, evidence, and facts. There are companies, like Weave, who are using knowledge graphs to do exactly this. They use the power of the graph to retrieve the relevant information and an LLM to summarize and answer the regulatory questions, enabling automated report generation.
Large financial organizations like Morgan Stanley, Bloomberg, HSBC, and JPMorgan Chase are also using knowledge graphs to unify data silos to build research assistants and advanced search capabilities for their employees and clients.

Discovery and design

By understanding the way different entities interact, both in theory and in the lab, scientists working in drug discovery can design drugs for purpose. Rather than testing different compounds blindly, hoping they find something useful, drug hunters can now work backwards from a desired outcome (such as lowering blood pressure) to identify candidate compounds, while accounting for patient differences (genetics, age, sex), interconnected systems, and potential adverse effects, all while complying with regulatory constraints. Many of the world’s largest pharmaceutical companies, including AbbVie, AstraZeneca, GSK, Pfizer, Merck, Novartis, Novo Nordisk, Roche, and Sanofi use knowledge graphs for drug discovery. There are also companies who focus exclusively on curating healthcare knowledge graphs for drug discovery like BioRelate and BenevolentAI.    

This same type of problem appears in many other industries. Banks often need to create financial products (e.g.,  structured notes) that achieve a desired outcome (e.g., higher yield with limited downside) while accounting for interconnected systems, mitigating adverse effects, and complying with regulatory constraints. Likewise, public policy practitioners often need to create interventions that achieve a desired outcome (e.g., reducing poverty) while accounting for various local contexts (e.g., geography, culture, climate), interconnected systems, and potential adverse effects.

Reuse and repurposing

Rather than designing an entirely new drug to achieve an outcome, it is sometimes easier to repurpose an existing drug. When Dr. David Fajgenbaum was diagnosed with a rare immune disorder while still in medical school, he was told he had weeks to live and a priest was called in to read him his last rites. While there was not enough time to design a new drug, there was time to repurpose something off the shelf. That’s exactly what he did. He found a drug originally meant to prevent organ transplant rejection and used it on himself. His disease has been in remission for 11 years, he finished medical school, and started the nonprofit Every Cure to “ensure that patients don’t suffer while potential treatments hide in plain sight.” Every Cure uses, among other techniques, knowledge graphs.  

Drug repurposing is about taking an existing product, understanding its underlying structure, and safely applying it in a new context. Public policy follows the same pattern. Practitioners identify interventions that worked in one context, understand why they worked, and reapply them elsewhere. Likewise, many companies are sitting on a gold mine of data, collected for some purpose long forgotten. But by understanding the meaning and context of the data, it can be repackaged and reused for different purposes.

Decision support

Healthcare professionals often rely on decision support systems to assist in making decisions that include many interconnected factors and incomplete data (Yang, et al., Al Khatib et al., Zhang et al.). Every day, physicians need to make decisions about how to treat and diagnose their patients based on limited, evolving information. An individual patient’s electronic health records (EHR) can be sparse and have limited predictive power (Yang, et al.). Knowledge graphs give the physician the ability to connect EHRs with controlled vocabularies (diseases, symptoms, drugs) and observational data from previous studies and, increasingly, patient-generated data from wearables (Al Khatib, et al.). 

This helps the physician make more informed diagnoses and treatment recommendations by grounding decisions in what is known from related cases, populations, and clinical evidence, while still accounting for the specific context of the patient. These are especially valuable because the underlying reasoning can be made explicit and explainable, in contrast to many black box AI solutions. Companies like Evidently are building decision support tools, powered by knowledge graphs and AI, to connect patient data across EHRs and existing clinical insights to help clinical practitioners make better, more informed, and explainable decisions in real time.  
Other industries are also using knowledge graphs to power decision support tools. The MITRE Corporation, the R&D organization, publishes MITRE ATT&CK, a knowledge graph of adversary tactics and techniques for decision support in cybersecurity operations. OpenCorporates, is an open legal-entity knowledge graph that is used by companies like Encompass for decision support regarding due diligence.

Recommendation and personalization

While decision support focuses on diagnostic accuracy, safety, and adherence to clinical guidelines, recommender systems in healthcare focus on personalizing and prioritizing options for patients. These systems often rely on patient-centric knowledge graphs (sometimes called Individualized Knowledge Graphs or Personalized Health Knowledge Graphs) to integrate medical history, EHR data, reference knowledge, and data from wearables. Rather than determining whether a clinical decision is correct, recommender systems surface and rank relevant options such as treatment plans, lifestyle interventions, follow-up actions, or care pathways that are most appropriate for a specific patient at a given moment.

Other industries use recommender systems powered by knowledge graphs and semantic technology even more than healthcare. Almost everything you buy and everything you watch is fed to you via recommendation systems. Online retailers like Amazon use them to suggest stuff you might like to purchase, streaming services like Netflix use them to serve up your next binge-watch, and LinkedIn uses them to recommend jobs to candidates and candidates to recruiters.

Governance, traceability, and regulatory compliance

Healthcare is a highly regulated industry. Drug companies need to comply with regulations to ensure they are monitoring and assessing any potential adverse effects of their drugs; something called pharmacovigilance. They also store individuals’ health data, which is incredibly private and sensitive, and need to comply with regulations covering this like the California Consumer Privacy Act (CCPA) or the General Data Protection Regulation (GDPR). To do this, they focus on something called data lineage—the systematic tracking of how data is generated, transformed, and used across systems. Knowledge graphs facilitate good data governance by connecting domain knowledge to knowledge about the organization itself, such as business processes, org structure, ownership, roles, and policies. Organizations can then trace how data moves through systems, identify who is responsible for it, understand which teams are allowed to use it and for what purposes, and enforce governance rules (Oliveira, et al.).

Financial services firms, like those in healthcare, rely on knowledge graph approaches to support enterprise data governance. Recent research proposes extending these same foundations to AI governance by linking data, policies, and decisions in a unified semantic layer. In regulated environments, governance is not a secondary concern—it is the mechanism by which trust, accountability, and explainability are enforced at scale.

Conclusion

Knowledge graphs are not a recent invention, nor are they a side effect of modern AI. They are a way of organizing knowledge that allows meaning to be shared, evidence to accumulate, and reasoning to remain explicit as understanding evolves. By separating theory (ontologies), instances (controlled vocabularies), and evidence (observational data), knowledge graphs make it possible to build systems that do more than store facts—they support discovery, explanation, reuse, and trust.

Long before large language models, healthcare invested heavily in defining shared concepts, cataloging the natural world, and standardizing how observations are documented and evaluated. Over time, these practices created dense, interconnected knowledge structures that could be extended, queried, and reasoned over as new discoveries emerged. Modern AI systems are powerful precisely because they are now being layered on top of this foundation, not because they replace it.

In the next part of this series, I’ll look more closely at how healthcare became the global leader in knowledge graph maturity. That story includes regulatory pressure, pre-competitive collaboration, public funding of shared knowledge, and early commitment to open standards. In the final part, I’ll step back from healthcare entirely and explore what other industries (finance, policy, manufacturing, energy, and others) can learn from this trajectory as they attempt to build AI-ready systems of their own.

The central claim is simple: progress at scale depends less on smarter models than on better structure. Healthcare learned this lesson early. Others are now being forced to learn it quickly.

About the author: Steve Hedden is the Head of Product Management at TopQuadrant, where he leads the strategy for EDG, a platform for knowledge graph and metadata management. His work focuses on bridging enterprise data governance and AI through ontologies, taxonomies, and semantic technologies. Steve writes and speaks regularly about knowledge graphs, and the evolving role of semantics in AI systems.

Bibliography

Al Khatib, Hassan S., et al. “Patient-centric knowledge graphs: a survey of current methods, challenges, and applications.” Frontiers in Artificial Intelligence 7 (2024): 1388479.

Barabási AL, Gulbahce N, Loscalzo J. Network medicine: a network-based approach to human disease. Nat Rev Genet. 2011 Jan;12(1):56-68. doi: 10.1038/nrg2918. PMID: 21164525; PMCID: PMC3140052.

Hager, Thomas. Ten Drugs: How Plants, Powders, and Pills Have Shaped the History of Medicine. Harry N. Abrams, 2019.

Isaacson, Walter. The Code Breaker: Jennifer Doudna, Gene Editing, and the Future of the Human Race. Simon & Schuster, 2021.

Kirsch, Donald R., and Ogi Ogas. The Drug Hunters: The Improbable Quest to Discover New Medicines. Arcade, 2017.

Oliveira, Miguel AP, et al. “Semantic Modelling of Organizational Knowledge as a Basis for Enterprise Data Governance 4.0–Application to a Unified Clinical Data Model.” arXiv preprint arXiv:2311.02082 (2023).

Rajabi, E.; Kafaie, S. Knowledge Graphs and Explainable AI in Healthcare. Information 2022, 13, 459. https://doi.org/10.3390/info13100459

Yang, Carl, et al. “A review on knowledge graphs for healthcare: Resources, applications, and promises.” arXiv preprint arXiv:2306.04802 (2023).

Yong Zhang, Ming Sheng, Rui Zhou, Ye Wang, Guangjie Han, Han Zhang, Chunxiao Xing, Jing Dong. “HKGB: An Inclusive, Extensible, Intelligent, Semi-auto-constructed Knowledge Graph Framework for Healthcare with Clinicians’ Expertise Incorporated.” Information Processing & Management (2020). https://doi.org/10.1016/j.ipm.2020.102324.