The Linear Algebra Blind Spot: Why Healthcare AI Is Only as Safe as Its Underlying Math
YouTube Video
Open Interactive Presentation
🎧 Listen to the Audio Overview
Acronyms and Terms
AI — Artificial Intelligence, systems that perform tasks requiring human-like reasoning or pattern recognition
EHR — Electronic Health Record, the primary system used to document and manage patient care
HIE — Health Information Exchange, infrastructure for sharing health information across organizations
HL7 — Health Level Seven, a family of healthcare interoperability standards
FHIR — Fast Healthcare Interoperability Resources, a modern API-based healthcare data exchange standard
SQL — Structured Query Language, the language used to manage and query relational databases
LLM — Large Language Model, a neural network trained on vast amounts of text
Embedding — A mathematical representation that converts information into vectors that machines can process
Vector — An ordered set of numbers representing position, direction, or characteristics within a mathematical space
Matrix — A structured collection of numbers used to transform, combine, and analyze vectors
Most healthcare AI discussions begin at exactly the wrong layer.
People talk about models, chatbots, diagnostic assistants, ambient scribes, and clinical copilots. The conversation usually starts where the software becomes visible. The real story starts much earlier, in a place most hospital executives, physicians, and even many software developers never see.
It starts when a patient’s life is squeezed into mathematics.
That sentence sounds dramatic until you remember what an AI system actually receives. It does not receive a frightened patient sitting in an emergency department at 2 a.m. It does not receive uncertainty. It does not receive context. It receives numbers.
Somewhere deep inside the machine, blood pressure becomes coordinates. Lab values become coordinates. Diagnoses become coordinates. Medication histories become coordinates. Entire physician notes become coordinates. The patient narrative is converted into vectors and matrices, and from that point onward the machine interacts only with mathematics.
The fashionable term today is “embedding.” It sounds modern and mysterious, like something invented in a Silicon Valley conference room stocked with artisanal coffee and whiteboards. In reality, embeddings are simply a way of translating human meaning into geometry.
A patient with diabetes, hypertension, kidney disease, and a history of smoking becomes a point in a very large mathematical space. Another patient becomes another point. Similar patients cluster together. Dissimilar patients drift apart. The machine is not reading stories. It is navigating geometry.
That distinction matters far more than most healthcare organizations realize.
One of the most persistent misunderstandings in healthcare AI is the belief that data quality problems begin in the database. Often they begin much earlier, during representation itself.
Suppose two physicians describe the same patient differently. One writes, “shortness of breath.” Another writes, “dyspnea.” A third writes, “difficulty breathing.” Humans immediately understand these phrases are closely related. Machines do not understand anything. They only process representations.
The challenge is not transporting the information. HL7 can transport it. FHIR can transport it. APIs can transport it. The challenge is preserving meaning.
Transport and meaning are not the same thing.
This may be the most important distinction in modern healthcare informatics.
A message can arrive perfectly.
The meaning can still be wrong.
Healthcare spent decades solving transport problems. Messages move between systems better than ever. Yet organizations continue to experience what they call “data quality” failures.
Many of these are not data quality failures at all.
They are representation failures.
The laboratory result may be accurate. The diagnosis may be accurate. The transmission may be accurate. Yet the mathematical representation generated for downstream AI analysis may distort relationships in subtle ways.
Once the representation is distorted, every subsequent calculation inherits the distortion.
The machine is not malfunctioning.
The geometry is.
This is where linear algebra quietly becomes one of the most important patient safety disciplines of the AI era.
Most people encounter linear algebra in school as a collection of unpleasant symbols. Rows. Columns. Determinants. Eigenvectors. Examinations designed by individuals who appeared personally offended by student happiness.
The practical reality is much more interesting.
Linear algebra is the mathematics of transformation.
A matrix answers a simple question: if I have information in one form, how do I transform it into another form?
That sounds suspiciously like healthcare.
Every day healthcare systems transform information. Clinical observations become diagnoses. Diagnoses become billing codes. Notes become quality metrics. Encounters become population health reports.
AI systems perform the same activity at a vastly larger scale.
An LLM predicting the next word is performing an enormous sequence of matrix operations. A clinical prediction model estimating sepsis risk is performing matrix operations. An imaging model identifying tumors is performing matrix operations.
The software layer changes.
The mathematics underneath remains surprisingly similar.
This creates a dangerous blind spot.
Many healthcare leaders evaluate AI systems through software demonstrations. They inspect dashboards, user interfaces, and performance metrics. They rarely ask how the underlying mathematical representations were constructed.
That omission is comparable to inspecting an aircraft cabin while ignoring the engines.
The machine may look beautiful.
The important failures occur elsewhere.
Geoffrey Hinton has repeatedly warned about AI systems producing highly confident falsehoods. In popular discussions these are called hallucinations. I prefer the older term confabulation because it captures something healthcare professionals understand instinctively.
The machine is not lying.
The machine is generating a plausible narrative from its internal structure.
If that structure is flawed, the output can be persuasive, coherent, and completely wrong.
Linear algebra offers no protection against bad assumptions.
It guarantees consistency, not truth.
Feed flawed vectors into a mathematically elegant system and you obtain mathematically elegant mistakes.
The arithmetic succeeds.
The patient may not.
The uncomfortable reality is that healthcare amplifies these risks because errors scale differently than they do elsewhere.
If a music recommendation engine recommends terrible songs, society survives.
If a healthcare model embeds a systematic representational bias into patient risk scoring, the error can spread across thousands of encounters before anyone notices.
A physician’s mistake may affect a handful of patients.
A model’s mistake may become organizational policy.
That is not a software problem.
It is an architecture problem.
A non-obvious insight here is that organizational structure eventually becomes encoded inside mathematical models. If one health system consistently documents conditions differently than another, those workflow differences eventually appear inside embeddings, feature vectors, and model behavior.
The AI is not merely learning medicine.
It is learning the quirks, biases, shortcuts, incentives, and documentation habits of the organizations that generated the data.
Hospitals often assume they are deploying clinical intelligence.
Sometimes they are deploying institutional habits disguised as mathematics.
Unfortunately, there is no clean solution.
Healthcare data is inherently messy because healthcare itself is messy.
Patients are inconsistent. Clinicians are busy. Terminologies evolve. Diagnoses change over time. Context disappears during handoffs. Human language refuses to behave like a database schema.
No amount of mathematical sophistication eliminates those realities.
The practical architectural response is not blind trust and not blanket skepticism.
It is auditability.
Healthcare organizations should understand how data becomes representations, how representations become predictions, and where semantic meaning may be lost during transformation. Governance should focus not only on model outputs but on embedding strategies, provenance, terminology mapping, and representational assumptions.
The future healthcare architect will need to ask a new set of questions.
What information disappeared during vectorization?
Which clinical distinctions were preserved?
Which distinctions were collapsed?
What does the model know?
More importantly, what can it no longer know because the representation discarded it?
Those questions sound mathematical.
They are actually clinical safety questions.
The irony is that modern AI appears astonishingly sophisticated precisely because the mathematics underneath is astonishingly sophisticated. Yet that sophistication can become a trap. The better the model becomes at generating plausible answers, the easier it becomes to forget that every conclusion ultimately rests on vectors, matrices, and transformations that began as imperfect attempts to describe a complicated human being.
The patient enters the hospital as a person.
The machine receives a geometry problem.
Whether those two realities remain connected may become one of the defining healthcare safety challenges of the next decade.