The patient you cannot see
A beginner's journey from one differential equation to the mathematics of hidden biological systems, uncertainty, and personalized medicine.
I give you five numbers: a heart rate, a blood measurement, a biomarker concentration, an imaging reading, and the same blood measurement six hours later. Then I ask: what is happening inside this person's body?
You can't simply look inside and read off the number of diseased cells, the strength of an immune response, how much drug reached a tissue, or the hidden settings of a feedback loop. Medicine gives you a handful of measurements. The biology that produced them stays hidden. Mathematics asks: what hidden system could have produced these measurements?
The animation above is exactly that situation. Coral curves are a real simulated disease and immune response. You only get the white dots. Press Peek inside to cheat, once. For the rest of this issue, you won't get to.
Before the math, your instinct
Lock in an answer for each. They come back at the end.
Start with something ridiculously simple
A sealed box takes an input and returns an output. You can't see inside. I tell you: input 1 gives output 2. You might guess . Congratulations, you've built a model.
But also passes through that point. So does every curve in the family on the left below. Add more observations and see what happens to the family.
More observations don't shrink the family to one. Any number of points leaves infinitely many curves through them. What collapses the family is an assumption, like "it's a straight line". Data plus assumptions make a model; data alone doesn't.
The body is not a calculator
Biology changes. Cells divide, molecules diffuse, hormones rise and fall, drugs enter the body and leave. So instead of , biologists write how things change:
The rate at which the system changes depends on its current state. is the state, is time, and are the parameters that make this system this system. That's a differential equation. Start with the smallest one there is: .
The faint arrows are the rule itself: at every height, the slope the curve must have there. The curve has no choice but to follow them.
You just discovered exponential growth and decay. The solution is . Positive : an infection doubling. Negative : a drug being cleared from the blood, halving at a fixed rate.
Where does x₀e^{rt} come from?
Separate the variables: . Integrate both sides: . Exponentiate: . Doubling time is ; half-life is .
Assumptions hiding here: the rate is constant, nothing else interacts with x, and x is large enough to treat as continuous. Every one of these breaks somewhere in real biology.
The moment biology becomes interesting
Real biology rarely has one variable. Take two: a disease population and an immune population . The disease grows; the immune system kills it; the disease fights back:
The term means the two interact. The system is coupled: change one and you change the other. Now stop looking at curves over time. Put on one axis and on the other. Every point is a possible state of the patient, and the equations draw an arrow at each one: the direction the patient will move next. That's a phase portrait.
Click anywhere to start a patient there.
The dots where nothing changes are equilibria. Two are stable, like valleys: the disease controlled (top left) or the disease escaped (bottom right). The third, in between, is a saddle: balanced like a marble on a ridge. The dashed line through it is the boundary between the two futures. Two patients can start a hair apart, on opposite sides of it, and end up in completely different places. You can understand every term locally and still be surprised by the whole system.
Valleys, ridges and tipping points
Strip the idea to one dimension. A stable state is a valley: nudge the system and it rolls back. An unstable state is a hilltop. The shape of the landscape is set by the biology. Turn up the feedback strength and watch the landscape reshape itself. Then drag the marble.
That jump is the intuition behind a tipping point. Below a certain feedback strength there's one stable state; above it, two appear. Dynamical-systems theory calls that a bifurcation, and medicine is full of them: a fever that breaks, an infection that suddenly takes hold, a heart rhythm that flips.
Now make the problem unfair
So far we cheated. We knew the equations, the variables and the parameters. Real medicine gives you only the measurements , while the true state stays hidden:
The first equation is the hidden system. The second is what you actually see, plus noise . You know . You don't know , or all of . You have to reason backward. A forward problem goes model → predictions. An inverse problem goes observations → possible models. It's like hearing clink… clink… clink from inside a box and trying to work out the shape of the object.
Wait — why can two models fit the same data?
What does "fits as well as it possibly can" even mean? If the model predicts and you measured , add up the squared misses and find the parameters that make that total smallest:
That's optimization, and it turns fitting into geography: every parameter choice is a spot on a map, and the error is the height. Do it by hand first.
What is an inverse problem, really?
Beginner: you see the effect and try to infer the cause. Picture: observed data → hidden system. Math: , solve for and .
Deep dive: Hadamard called a problem well-posed if a solution exists, is unique, and depends continuously on the data. Inverse problems routinely fail all three. The toolkit for that: regularization (Level 11), Bayesian inverse problems and posterior distributions (Level 07), and the distinction between structural identifiability (could perfect data determine θ?) and practical identifiability (can your real, noisy, finite data?) (Level 06).
Two explanations, one measurement
Suppose two processes both push the output up: . You measure and . Can you determine ? You only ever learn . and fit identically. A parameter is identifiable if your observations contain enough information to determine it, within the model's assumptions. Try.
This isn't a failure of optimization or of artificial intelligence. It's an information problem. The fix is a different measurement, one that sees on its own, not a cleverer algorithm.
Wait — why can't we just collect more data?
Bayes enters the hospital
Real measurements are noisy: . There's no exact curve, only evidence. So the question changes from "what is the answer?" to "which explanations are plausible, and how uncertain are we?" That's what Bayes' theorem is for:
Here's a real use. A patient got a drug. You want its elimination rate (how fast their body clears it), because that sets their dose. Concentration falls as , but each blood test is noisy. Start with a wide belief about , then draw blood.
Each test narrows the belief. Turn the noise up and it narrows more slowly, and sometimes a noisy test even shifts it the wrong way for a while. The output isn't a single magic number; it's a distribution of possibilities, and the haze around the concentration curve is that uncertainty, carried forward into a prediction.
Wait — why do we need probability?
The geometry of measurement
Shine a light on a 3D object and you get a 2D shadow. The shadow contains information, but not everything, and different objects can cast the same shadow. Measurements work the same way: a full state is mapped to observations with . So a new question appears: which shadow should you look at?
Below is a hidden system: four compartments that exchange a substance along the lines, like four connected organs. You have three sensors. Each one records its compartment six times. Place them, then run the experiment and see how well you can reconstruct all four hidden starting values.
Sensors left: 3. Click a compartment to add a sensor; you can stack more than one on the same spot.
Did you try opposite corners? It feels balanced, and it's a trap. The network is symmetric, so and affect both corners identically, and their difference casts no shadow at all. The math that measures this is called Fisher information; the field is optimal experimental design.
How the reconstruction is computed
The compartments follow , so . A sensor on node at time reads row of dotted with , plus noise. Stack every reading into . With a Gaussian prior, the posterior covariance is . The bars show the square roots of its diagonal. Directions with are invisible: the data never touch them, and the uncertainty there stays at the prior.
Now introduce the monster: nonlinearity
In , growth looks exponential when is small and stalls as nears . One system, different regimes. That's nonlinearity, and it brings multiple equilibria, oscillations, tipping points, and hysteresis: memory.
Below, a system feels an outside stress . The lavender curve shows every equilibrium for every stress level: a bifurcation diagram. Drag the yellow line slowly from one side to the other and back. Watch the coral dot, the actual system.
Up past the low branch simply ends, and the system has to leap to the high one. Coming back, it doesn't drop at the same place; it holds on until . Same stress, two possible states, and which one you're in depends on history. Removing the cause doesn't always undo the effect.
Wait — why can a tiny parameter change create a huge outcome?
One disease, two mathematical worlds
Until now mathematics watched. Control theory asks: what if we change the system? In , is now a dose we choose. Measure, estimate, intervene, measure again. Here are two patients with the same diagnosis and the same first scan: 40% tumor burden. Same equations:
Only their parameters differ: tumor growth rate and drug sensitivity . Choose a schedule. All three give the same total amount of drug.
Patient A
Patient B
That's the mathematical intuition behind personalized medicine: the equations may share a structure, but different parameters make different trajectories, and different best interventions. Estimate from one person's data and you have a patient-specific model: a hypothesis about how this person's system behaves.
A model is not a crystal ball
Here's Patient B's predicted burden at day 84 under each schedule. Clean bars invite you to say "the computer knows which treatment works". But we never know and exactly. Switch on the uncertainty.
The model depends on its assumptions, its equations, its parameters, the quality of its measurements, the variables it includes and the ones it ignores, and whether it still holds outside the conditions it was built in. A model is an instrument for reasoning, not reality itself.
Mathematics refusing to overreact
Try to estimate 50 parameters for one person from 10 measurements and you've made a monster: more unknowns than information. That's an ill-posed inverse problem. A solution may not exist, may not be unique, or may swing wildly when the data change slightly. More mathematical sophistication doesn't automatically create more information. One defense:
The first term says fit the data. The second says don't choose ridiculous solutions. sets the balance. Below, ten noisy measurements from one patient. Give the model more freedom and see whether it gets better at predicting new patients.
The best model isn't the one that hugs every data point. It balances evidence against complexity: Occam's razor, written as math. And a model can be wrong in two different ways. Wrong parameters: better data fixes it. Wrong structure: the equations themselves are wrong, and no amount of fitting can save it. You can have a tiny fitting error and still a biologically misleading model. That's why validation matters.
Wait — why doesn't a perfect fit mean a correct model?
Black box vs. glass box
A machine-learning model learns measurements → prediction from huge datasets. A mechanistic model encodes biological mechanism → dynamics → observations. They answer different questions, and modern computational medicine increasingly uses them together.
Black box
Glass box
Which would you trust more?
Wait — why can't AI simply learn everything?
Descend through a human being
A body can be modeled at many scales: molecules, cells, tissues, organs, physiology, the whole patient. The scales talk to each other in both directions. That's a multiscale system, and connecting models across its levels, with honest uncertainty at each one, is one of the frontiers of systems biology. Zoom down through the levels and open any node.
This is where the buzzword arrives: a digital twin, a computational model linked to one real person and updated as new data come in. Patient-specific 3D anatomy models already help surgeons plan operations; research is exploring "virtual patients" for testing treatment strategies. But a digital twin is not a perfect copy of a human. It's a model, and models have boundaries. The real question is how much useful information about one person can be encoded in a model, and how reliably that model can support a decision.
If the model says , that's not an 82% chance reality behaves exactly as predicted. It's a probability according to this model, given this evidence, under these assumptions. Time to see that for yourself.
You are the mathematician
A fictional patient has a growing tumor. Treatment starts on day 30. You have 5 measurements, a choice of 3 models, 2 drugs, and one question at the end: how confident are you? There's no single score. You get a dashboard.
The final paradox
Can mathematics tell us what's happening inside a patient? Sometimes. But the more interesting discovery: mathematics doesn't make the invisible visible by magic. It gives you a structured way of reasoning about it. You observe, model, infer, simulate, test, collect better data, revise, and repeat.
Differential equations describe change. Dynamical systems describe behavior. Optimization estimates parameters. Probability describes uncertainty. Bayesian inference updates beliefs. Identifiability says what the data can actually reveal. Control theory asks how interventions alter trajectories. Multiscale modeling connects levels of biology. Machine learning finds patterns. Mechanistic modeling encodes hypotheses about causes. Personalized modeling asks whether all of it can be brought together for one person.
That direction is easy. The reverse is the whole field:
A hundred years from now a doctor might ask a patient's model "what happens if we perturb this system?" and watch thousands of simulated futures. The fundamental question will remain: how do we know the model is describing the person, rather than our assumptions about the person? That's a medical question, a mathematical one, a philosophical one, and finally a question about the limits of knowledge.
You estimated the parameters. You simulated the future.
What, exactly, have you learned?
The power of mathematical medicine isn't that equations eliminate uncertainty. It's that equations give us a language for reasoning about uncertainty, mechanisms, and possible futures.
How did your gut do?
Further reading: Kirk, Thorne & Stumpf, "Model selection in systems and synthetic biology", Curr. Opin. Biotechnol. (2013); Raue et al., "Structural and practical identifiability analysis", Bioinformatics (2009); Stuart, "Inverse problems: a Bayesian perspective", Acta Numerica (2010); Laubenbacher et al., "Building digital twins of the human immune system", npj Digital Medicine (2022). All models on this page are illustrative teaching models, not clinical tools.