New York Caught the Legionnaires’ Outbreak at Two Cases. I Wanted to Know What Came Before Them.

What two New York outbreaks taught me about early warning, cooling towers, and the data we choose to watch

By Paul Samuel Guarino Independent Researcher · Klein’s Pollution Solutions / Project Promised Land

I have lived in New York my whole life, long enough to stop seeing the machinery.

Cooling towers. Steam. Vents. Metal on the roof.

It becomes the building.

Then people get sick, and you look up.

In the summer of 2026, Legionnaires’ disease appeared in Carnegie Hill and Yorkville. It is a serious pneumonia caused by Legionella. You do not catch it from the person next to you. You breathe mist that already contains it.

On July 2, the city opened an investigation on two confirmed cases.

Two.

A Health Department spokesperson later said past protocol was to start after three confirmed cases in a tight area and time frame.

The city did not wait.

It sampled cooling towers and alerted providers.

That was fast by the city’s previous protocol.

It mattered.

But it was not early warning of exposure.

More than 90 percent of the patients ultimately identified in the cluster had symptom onset before July 7, five days after the investigation began.

The city acted quickly once cases made the danger visible.

The exposure had already happened.

The machinery above us

You cannot see bacteria in a cooling tower from the sidewalk.

You cannot smell the mist.

You cannot tell whether the sampling is current.

Somebody has to be watching the water.

The Health Department tested 183 cooling towers at 160 buildings. Fifty-nine towers at 58 buildings grew live Legionella.

A positive culture is not automatically a source. Epidemiology and genetics still have to connect a tower to a person.

The inspections are what changed the picture for me.

Of the positive towers under Health Department jurisdiction, 60 percent of inspections found at least one violation.

More than 110 violations in all.

A tower at 300 E. 83rd Street was not registered. It had gone more than a year without required sampling and without the required maintenance plan.

Eleven violations.

$11,000 in fines.

By August 31, the cluster stood at 94 cases.

Eleven people had died.

Those are New Yorkers.

The system problem ends there, or it is not a system problem.

The response was real.

More than 260 Health Department staff worked it. More than 180 towers were sampled. More than 360 tests were run. Positive towers were ordered cleaned. Lists were published. Meetings were held.

On July 31, the department said the elevated exposure was over.

One day after the first public list, 74 percent of inspected positive towers had records showing compliance with the new monthly testing rule.

By July 24, that figure was 97 percent.

Names on a page moved the watch.

The city was not asleep.

It also was not watching the danger on a clock that beat the illness.

Nearly all of the outbreak machinery starts when a person has already become a case.

That is the part I cannot leave alone.

I wanted to know if the detector could see it coming

Klein’s Pollution Solutions is a deterministic instrument I have been building under Project Promised Land.

No neural net.

It does not learn what a disease looks like.

It asks whether a measurement stream has left its own normal.

I have run it on records I could replay from the start: Delhi PM2.5, groundwater, mercury in fish, infectious-disease series.

On 27,820 hours of Delhi PM2.5, my reported result was 29 severe episodes detected out of 29, with a median lead of 35 hours.

Could the same locked instrument see the Upper East Side outbreak before the city announced it?

I wanted the answer to be yes.

It wasn’t.

The first strong, outbreak-associated signal landed in CDC epidemiological week 26, the same week the city announced its investigation.

At the resolution of this dataset, Klein detected the event when the city did.

That is a useful detection result. It is not an early-warning result.

I tried to make that result fail

Maybe it had not been given enough normal history.

The public series held 249 weeks of citywide legionellosis surveillance, from 2022 into 2026.

The baseline comes from the beginning of whatever series you hand the locked detector, and the statistical layers will not arm until they have enough history.

Same detector.

Same thresholds.

Same outbreak.

Thirteen lengths of history.

Whenever the baseline reflected ordinary conditions, the Upper East Side event kept resolving around week 26.

Shorten the history too far and the alert moved later.

One very short run missed the outbreak because the outbreak itself had entered the baseline and started defining normal.

One configuration produced a week-23 alert.

I first threw it out because it came before the first known Upper East Side specimen.

That was the wrong rule.

If “before the first diagnosis” disqualifies a signal, you have defined early warning out of the test.

So I went back.

The week-23 mark is still weak.

Thin baseline.

A citywide count that was not unusually high.

The far edge of a plausible emergence window.

I cannot responsibly call it an Upper East Side warning.

The reason is not the calendar.

The reason is that the evidence tying it to this particular outbreak is weak.

Get attribution wrong, and you can publish an early warning that never happened.

I wasn’t going to do that.

Then Harlem changed the question

The Upper East Side was not New York’s only recent Legionnaires’ cluster.

In 2025, Central Harlem had one.

On July 25, the Health Department’s own spatial surveillance system flagged eight positive urine-antigen reports inside an area of about one kilometer.

Five were confirmed within hours.

That was the day the investigation began.

But Klein, running against the citywide series, had already been marking elevated legionellosis activity.

Waves appeared at the end of May and the beginning of June.

A spike appeared June 14.

Then two marks became harder to wave away:

July 5.

July 12.

Three weeks and two weeks before recognition.

My first instinct was to discard them because diagnosed cluster cases appeared later.

That repeated the same mistake.

A diagnosis is not the beginning of an infection.

Exposure has already happened.

Incubation has happened.

Symptoms have developed.

Someone has gone to a doctor.

Testing has happened.

Reporting has happened.

The Health Department’s published case definition for the Harlem investigation reached back to June 30 for people who lived, worked, or stayed in the affected zone.

July 12 falls inside that window.

July 5 sits near its edge.

They are not equally strong marks.

But they are not disposable either.

The pre-recognition rise was real

There was one obvious problem we had to eliminate.

Maybe 2025 only looked elevated because the Harlem outbreak itself was already being counted.

So we separated the weeks before recognition from the weeks afterward.

The result matters.

The seven elevated citywide weeks from June 7 through July 19 all ended before the city recognized the cluster on July 25.

The recovered Harlem cluster curve contained no diagnosed cluster members in those seven weekly totals.

In other words, the pre-recognition elevation was not simply Harlem’s known cases making their own background look abnormal.

The citywide record was already elevated before officials recognized the Harlem cluster.

Klein marked that movement, including July 12, which falls inside the Health Department’s published exposure-eligibility window.

July 5 sits near its edge.

That makes these marks worth testing.

It does not prove that Klein warned of Harlem.

A citywide series cannot identify which neighborhood produced a case, and the connection to this particular cluster remains unresolved.

That distinction is the whole point.

What the towers actually showed

The published Harlem investigation identified two cooling towers on the same block whose environmental isolates were highly related to clinical isolates from eight patients.

One was an unregistered construction tower on West 137th Street.

It had major monitoring lapses and missed required Legionella sampling.

The other was Harlem Hospital’s tower.

That system met routine monitoring requirements.

Investigators could not determine which tower infected anyone.

That counterevidence belongs here.

Harlem Hospital’s tower had negative cultures in March and June.

It was fully disinfected on July 2.

At the same time, the Health Department’s later review found three intermittent periods in June and July when disinfectant fell below target for two consecutive readings before returning to target after routine biocide dosing.

That does not prove the July 5 or July 12 Klein marks came from Harlem.

A citywide count cannot tell me which neighborhood produced a case.

An unrelated Legionnaires case enters the same number.

A Harlem infection diagnosed somewhere else also enters the same number.

What can be said is narrower.

Before the city recognized Harlem, the citywide record was already moving.

Klein marked that movement.

During the same general period, one tower later genetically linked to isolates from eight patients had short documented disinfectant dips, while a second tower on the same block was not being properly monitored.

I cannot write:

“Klein warned three weeks before the Harlem outbreak.”

I do not know that.

But I also cannot honestly dismiss the pre-recognition marks as meaningless simply because they came before diagnosis.

They are candidate precursors.

That is what the record supports.

Two outbreaks. Two different outcomes.

This is where the comparison becomes useful.

The Upper East Side produced a clean same-week result.

The first strong outbreak-associated Klein signal landed in the same epidemiological week the city recognized the event.

Harlem was different.

The citywide record was already elevated before recognition.

Klein marked that movement.

July 12 falls inside the Health Department’s published exposure window.

July 5 sits near its edge.

The connection to the Harlem source remains unresolved.

The contrast is already clear in the available record: the Upper East Side produced same-week detection; Harlem produced candidate pre-recognition signals.

Those are different outcomes.

And they point to a testable question about what the measurement stream captured before recognition.

What the previous outbreak changed

Harlem also let us test another possibility.

Maybe seeing one outbreak could somehow teach the detector the next one.

No.

There is nothing in Klein that learns that way.

What changes is the reference state.

Start the Upper East Side stream after Harlem has gone quiet, and the next event separates cleanly.

Leave the earlier outbreak inside the statistics defining normal, and variability rises.

In some runs, the week-26 signal drops from a spike to a wave.

In others, it slips later.

Build the baseline inside the earlier outbreak, and a later event can disappear into it.

The lesson is not:

“Show it previous outbreaks so it learns the disease.”

The lesson is simpler.

An outbreak is not normal because it happened before.

This time, the measurement was the aftermath

Then I saw what I had been staring through.

In Delhi, I had hourly particles in the air.

In East Palestine, chemicals in water.

On the English–Wabigoon, mercury in fish.

Those measurements sit relatively close to the physical thing being studied.

For Legionnaires’, I had the people who got sick.

Legionella grows.

Water carries it.

Mist leaves the tower.

Someone breathes it.

Infection.

Incubation.

Symptoms.

A doctor.

A test.

A report.

Then a number for the whole city.

Then my detector receives it.

The data are real.

They are downstream.

This time, the measurement was the aftermath.

Harlem adds one complication.

Even a downstream count can start moving before officials have a spatial cluster they are willing to name.

The question is whether that movement can be separated from everything else happening in a city of eight million people.

Why some events warn, and others don’t

That sent us back through the earlier replays.

The runs that produced meaningful lead time tended to have more frequent observations, more local measurements, or measurements closer to the underlying process.

Delhi PM2.5: a 35-hour median lead.

Yap dengue: an alert while most of the epidemic was still ahead.

Geneva influenza: a disruption on the rise.

SARS in Hong Kong: an early wave before the later international alert.

Then there are the Legionnaires results.

Upper East Side:

same-week detection from weekly citywide surveillance.

Harlem:

candidate pre-recognition movement, with outbreak-specific attribution unresolved.

That is not a law.

It is a testable idea.

Early warning depends on whether a precursor was ever recorded.

How often?

How local?

How close to the process?

How much information dies when seven days become one number?

How much disappears when three ZIP codes become an entire city?

What does the baseline call normal?

The detector cannot extract a precursor the measurement system never kept.

I needed another way to test that idea

There was an obvious danger.

I could have invented “precursor observability” because I wanted the Legionnaires results to mean something.

So we looked for another New York disease measured at two different places in the same chain.

COVID gave us one.

Wastewater sits upstream of a diagnosed clinical case.

The city also publishes diagnosed-case data.

Same city.

Same waves.

Same locked detector.

Two different observables.

The result was not clean.

Good.

Clean results make me nervous.

In the Delta comparison, wastewater alerted about 14 days before the case stream.

BA.5 separated by considerably more in the weekly analysis.

BA.2 went the other direction.

Omicron initially looked like a contradiction after the wastewater samples were averaged into weeks.

Then we looked at the native wastewater cadence.

The rise was already underway while diagnosed cases were still comparatively flat.

At that resolution, the first wastewater magnitude alert led the downstream case alert by roughly 13 to 18 days.

But there was a cost.

More temporal resolution produced more alarms.

Earlier information came with more noise.

So the conclusion is not:

“Wastewater always predicts cases.”

It doesn’t.

The conclusion is narrower.

The warning horizon is whatever precursor information survives the way we measure a system.

The detector and the stream are one system.

Now I know which file I want

I want the Harlem tower record.

Timestamped disinfectant residuals.

pH.

Temperature.

Biocide dosing.

Legionella sampling dates and concentrations.

Before.

During.

After.

Then put those measurements beside July 5 and July 12.

Do not tune Klein until the lines match.

Freeze it.

Replay the tower record chronologically.

Ask one question:

Did the environmental system begin changing before the citywide disease stream did?

If the deterioration clearly comes first, the candidate-warning interpretation gets stronger.

If it comes afterward, that interpretation gets weaker.

Either result is the result.

For the harder test, make the next one blind.

Give me cooling-tower data from an outbreak I have never seen.

Keep the event labels from me.

Lock the detector before the file arrives.

Replay the measurements in time order.

No future values.

No moving thresholds afterward.

No selecting whichever run gives the best headline.

Then reveal the record and score what came back.

New York had already tightened routine cooling-tower surveillance before the 2026 outbreak.

Beginning May 8, operating towers had to be sampled for Legionella at least every 31 days, with the sampling date reported.

After the Upper East Side outbreak, the city also said it would investigate faster culture methods.

Same question.

Other side of the table.

How do we know sooner?

Maybe Klein finds a precursor.

Maybe it doesn’t.

That is what a test is for.

Two cases can be fast and still be late

New York identified the Upper East Side cluster at two confirmed cases.

By its previous protocol, that was early.

The people who did that work should keep the credit.

Surveillance and environmental warning are not the same job.

An epidemiologist can be doing excellent work when two cases trigger an investigation.

The city’s Harlem alarm was excellent work of that kind: eight reports concentrated spatially, recognized in a day.

An environmental warning system has to ask whether the water changed before either person became a case.

Harlem suggests the human count may sometimes begin moving before the spatial alarm.

The Upper East Side shows that it may not.

The open question is whether the tower can speak sooner than either.

Not every outbreak will have a precursor we can measure.

Some records will be noise.

Some systems will turn too quickly.

Some detectors will miss.

Some early marks will belong to something else entirely.

That is why attribution is the job, not the obstacle.

If an upstream signal exists, I want it read there.

Not only at the end of the chain, where the measurement is a New Yorker.

I still look at the towers differently

Not because every cooling tower is a threat.

It isn’t.

Not because mist above a roof is a reason to panic.

It isn’t.

I look up because I can see the distance now between that machine and the person under it.

Maintenance.

Sampling.

Disinfectant.

Lab time.

A report.

An inspection.

A diagnosis.

An investigation.

And trust, which is what you have when you cannot see the water yourself.

The city can move hard once illness makes danger visible.

The harder job is making it visible before illness has to.

In Delhi, the air did not wait for the report.

At Love Canal, the ground did not forget.

On Long Island, we had the aftermath because we did not have the measurement.

On the Upper East Side, the downstream record gave us the same week the city already had.

In Harlem, the citywide record was moving before recognition.

We do not yet know whether that movement belonged to the outbreak.

That is exactly why the next experiment matters.

The Upper East Side result does not support an early-warning claim. The Harlem result does justify a focused test.

Put the timestamped tower record beside the candidate dates.

Freeze the detector.

Replay the measurements in order.

If the environmental signal leads, the case for precursor observability gets stronger.

If it does not, that weakens the interpretation.

Either way, the next test should be upstream—

before exposure becomes illness,

and illness becomes a citywide count.

New Yorkers should not have to become the instrument that tells us the environment changed.

We should be watching the environment first.

Paul Samuel Guarino Independent Researcher · East Northport, New York Klein’s Pollution Solutions / Project Promised Land U.S. Non-Provisional Patent Application 19/700,318