Epidemiological surveillance since John Snow: why continuous ingestion follows

The pump on Broad Street

Epidemiological surveillance is a specific discipline with a specific definition: the continuous, systematic collection of health events, tied to place and time, used to guide action. Each clause is load-bearing. Continuous, because a single study ends and surveillance does not. Systematic, because it depends on case definitions and reporting duties, not on whoever happens to notice. Tied to place and time, because a rate without a location is not yet evidence of anything. Used to guide action, because surveillance is not a research programme; it exists so that someone can close a well, recall a batch, or issue a warning before the next case.

The founding image is Soho, 1854. Cholera had killed several hundred people in ten days in a few streets around Broad Street. John Snow plotted 578 deaths as stacked bars at their addresses on a street map and watched them thicken around one water pump. The plotting was not the argument. Snow then went looking for exceptions that would break the pattern, and found two that strengthened it instead: workers at the Broad Street Brewery, who drank the brewery's own supply and had almost no deaths among them, and a widow in Hampstead who had water carted weekly from the Broad Street pump because she liked its taste, and who died of cholera far from any other case. The map proposed a cause. The anomalies tested it. That sequence — proposal, then arriving evidence permitted to falsify it — is the discipline in miniature.

Snow's map rested on a standing resource that already existed: the Registrar-General's mortality returns, compiled from 1839 by William Farr, who did not believe Snow's theory and argued against it for years using the same data. That detail matters more than it first appears. The returns were a continuous stream before anyone had the right hypothesis for what to do with them. Surveillance, properly understood, is older than the discovery it is best known for enabling.

What Snow was up against

The prevailing account of cholera in 1854 was miasma theory: disease arose from foul air, from the accumulated stench of crowded, dirty places. It was coherent, widely held by people of standing, and consistent with a great deal of casual observation — poor districts were smelly and sickly at once, and correlation did the rest of the work. The problem Snow faced was not a lack of theory but a surplus of a wrong one, entrenched enough that the Board of Health investigating the outbreak initially favoured it. Miasma theory could absorb almost any pattern of deaths by pointing to bad air somewhere nearby. What it could not survive was a specific, falsifiable spatial claim, tested against named addresses that did or did not fit.

Snow's 1849 and 1855 essays on the communication of cholera made the inversion explicit. Where miasma reasoning fit observations to an established cause, Snow let the arriving pattern of cases constrain which causal account could survive. He removed the pump handle on 8 September 1854, by which point the outbreak was already subsiding, so the intervention proved less than the legend suggests. What it demonstrated conclusively was the method, not the immediate victory. Later work, including his study of the Southwark and Vauxhall water companies serving intermixed households from different intakes, made the water-borne case far more solidly than the pump-handle story alone.

The turn

None of this has anything to do with machine learning, and it should be allowed to stand on its own before it is asked to do double duty. But set the shape of the problem beside the current sequence of large models and something aligns without being forced.

Miasma theory, as an object, has the structure of a Large Language Model: a body of authoritative belief, internally coherent, fixed by everything collected and reasoned before the current question arose, and wrong in ways that no further reasoning inside it could reveal. It could accommodate new smells and new slums indefinitely, the way a frozen corpus can accommodate any question phrased for it, without ever being touched by an event that occurred after its formation.

Snow with his notebook in Soho is the Large World Model position. He gathered direct, richly located evidence of a live scene — one outbreak, one parish, one investigation with a beginning and an end. It was vastly better than reasoning from the prior corpus, because it was sensed rather than inherited. It was also bounded. When the outbreak subsided, the collection stopped. A second outbreak in a different parish would need its own Snow, walking its own addresses.

What came after Snow is the harder achievement, and it is the one the sequence is really about. The Registrar-General's returns, formalised notifiable disease law after 1889, sentinel physician networks, and now molecular typing and wastewater sampling, form a standing arrangement: streams that do not stop, feeding beliefs that are revised as each new reading arrives, with the origin of every reading recorded so that its reliability can be weighed. That is the Large Universe Model position on intake. It is not a cleverer instrument. It is a different relationship to evidence — permanent, provenanced, revisable — and nothing has replaced it since public health institutionalised it.

What the third position closes off

The claim under this page is specific: on the axis of intake, there are exactly three positions, and epidemiological surveillance had found all three by 1900. Reason from a fixed prior corpus. Observe a live scene directly. Or maintain permanent, provenanced streams that revise belief as they arrive. The third is terminal not because progress ends but because its complement is empty. Once intake means every available stream, continuously, with origin attached, there is no further category of evidence left to admit — only more streams, faster arrival, and better-trusted sources.

The history since supports this reading better than it flatters it. Israel detected wild poliovirus in sewage in 2013 with no paralytic cases reported at all — environmental sampling saw circulation the clinical stream structurally could not, since only about one infection in two hundred causes paralysis. The same wastewater method caught vaccine-derived poliovirus in London sewage in 2022, again ahead of any clinical case. Aviation's flight data monitoring reads the same shape outside public health entirely: operators download thousands of parameters per flight and flag exceedances continuously, revising belief about which risks matter week to week rather than after a hull is lost. In each case the advance is scale, latency and trust in the stream. None of them is a fourth kind of evidence.

Three objections, one that narrows the claim

Continuous intake without a prior hypothesis is a machine for manufacturing false clusters. Snow succeeded because he had a theory worth testing.

This is correct, and the argument concedes it entirely rather than around it. Snow's map did not replace theory; it adjudicated between two theories already on the table. Intake does not generate hypotheses. It bounds them, and the set of hypotheses that survive contact with evidence shrinks monotonically as intake widens from corpus to scene to standing stream — provided each reading carries the identity of the stream that produced it, since that is what stops a wastewater false-positive from being weighed like a clinical diagnosis.

Surveillance history is a story of expanding legal authority, not expanding sensory reach. Compulsory notification after 1889 changed practice by changing penalties, not by revealing new facts.

Largely right, and it sharpens rather than damages the claim. Once intake reaches the third position, the remaining work is precisely trust: who must report, who may see the data, which stream is believed over another. Naming trust as the successor problem to intake is part of this argument, not an evasion of it.

Real surveillance systems are sparse — sentinel practices, short notifiable lists, discarded detail — so 'everything, continuously' has never actually existed.

True as description, and this is the objection that narrows the claim most. 'Everything, continuously' names a limit, not an achieved state. No system occupies it in full. But a sentinel network sampling one per cent of cases, which never stops and revises belief with each new reading, is categorically different from a bounded study that ends, even at far higher completeness. The claim is about which category of evidence has been admitted, not how much of it currently arrives.

The misreading to disown

The weak version of this argument says Snow proved data beats theory, and that enough intake eventually makes theory optional. Snow believed in waterborne transmission before he drew the map; the map tested a hypothesis he already held, and he never isolated the organism responsible. Filippo Pacini had already described Vibrio cholerae in 1854 and was ignored for thirty years — a plain demonstration that evidence sitting in the record settles nothing by itself. Wide intake constrains which theories can survive. It does not supply the theory, and it does not manufacture the institutional trust required to act on what survives.

What this establishes, and what it does not

This history establishes that continuous, provenanced intake is a distinct and terminal position on one specific axis — a position public health found by trial across five decades and has spent a further seventeen improving without ever needing a fourth category. It does not establish that surveillance produces understanding by itself, that more streams always mean better decisions, or that the institutions carrying those streams are trusted by the people they watch. Those are separate arguments, decided on separate grounds. The axis of intake has a top rung. What is built on top of it is a different question, and a longer one.