Statistics, Mathematics and Physics Transform Biomedicine

Big data in biomedicine is driving a return of mathematics to the center of medicine, and with it the emergence of computational medicine as a discipline in its own right.

The interfaces are concrete. Causal inference methods, among them instrumental variables, target trial emulation, and Mendelian randomization, allow observational data to support claims that once required trials. High-dimensional statistics addresses the regime where variables outnumber patients, which is now the normal condition of multi-omic medicine. Bayesian methods provide the natural formalism for updating belief about one patient as evidence accrues. Survival and competing-risks analysis handles the multimorbid patient for whom a single endpoint is a fiction.

One concept stands above that list, because it governs how much of the rest can be believed. The replication crisis is usually discussed as a problem of p-values, underpowered studies, or publication bias.

Yoav Benjamini, whose false discovery rate, developed with Yosef Hochberg, is among the most-used statistical procedures of the past half-century, has argued that the deeper culprit is selection: inference drawn on findings chosen because they looked interesting, from among the thousands examined.¹⁷ Selecting and then estimating from the same data inflates the estimate, whether or not any individual test was correct. The problem is attended to when candidates run to thousands and ignored when they run to dozens.

N-of-1 care multiplies the candidate inferences per patient. A multi-omic profile, a continuous physiological stream, and a digital twin running counterfactuals generate thousands of comparisons, from which the physician, or the model, will select the striking one. Every such encounter is a selective-inference problem whether or not anyone names it. The competence implied is not computing a false discovery rate at the bedside, but the reflex to ask how many things were looked at before this one was noticed, a habit belonging alongside sensitivity and specificity.

The corollary at scale is more serious than the bedside habit. If every N-of-1 encounter is a selective-inference problem, then a system conducting millions of them will manufacture false positives industrially, and each will reach a patient dressed as a personal finding rather than as the most striking of ten thousand comparisons. Individual vigilance cannot be the remedy. The correction has to be built into the machinery that surfaces the finding, so that the number of comparisons from which a result was drawn travels with the result, and the models proposing candidates are designed to account for their own selection. That requirement belongs with the curation argument above and with the digital twins of Chapter V, since a twin running counterfactuals overnight is, above all, a selection engine.

Two examples show how statistical innovation reaches practice instead of remaining methodological. The first is the N-of-1 trial: a randomized, blinded, multiple-crossover experiment conducted within a single patient, who serves as their own control. Individually it answers the only question the patient actually has, which is whether this works for me, and, aggregated across a series of such trials through Bayesian hierarchical models, it yields population-level estimates too. It suits exactly the conditions conventional trials handle worst: chronic, fluctuating, symptom-defined disease, and rare phenotypes for which no cohort can be assembled. Its obstacles are administrative rather than mathematical, which is the translational-gap pattern of Chapter II in statistical dress.

The second is the mathematics beneath physical AI. A robot acting on a patient must estimate a state it cannot directly observe, from noisy sensors, and commit to an action before the uncertainty resolves, which makes it a filtering and stochastic-control problem as much as an engineering one. Whether such a system knows when to stop and ask is a question of calibrated uncertainty, and conformal prediction can attach honest coverage guarantees to its judgment that a situation lies within its competence. This is the technical content of the escalation threshold that Chapter IX treats as an ethical matter. The threshold has to be a number before it can be a policy, and deciding what that number should be is a clinical judgment expressed statistically.

For physicians this implies new responsibilities, not new hobbies: interpreting statistical outputs and their uncertainties, communicating them to patients, critically evaluating the assumptions beneath the data, and partnering routinely with data scientists instead of working alone. The physician becomes an evaluator of data quality and of the limits of what a given body of data can support. Few will become statisticians and none need to. The working relationship with statisticians, data scientists, and domain experts does have to be close enough and frequent enough to settle the question that actually arises at the bedside: whether an output is trustworthy enough to act on for this patient.

Technological and technoscientific expertise has been viewed as something to be layered onto a core of specialty expertise. It should instead be viewed as part of the core competencies of each of the specialties, embedded.

This is a subset of a larger issue surrounding multidisciplinarity. Additional disciplines should be treated as integral elements of base training, not as modular add-ons bolted to a fixed core. Intelligence becomes an ecosystem, with hard limits. Pattern recognition is not understanding. Complexity sets a permanent ceiling on prediction. And data-aware medicine becomes a core clinical skill.

With the tools in view, we can now describe what they are for: the personalization of medicine down to the individual.