One patient, one night. What the machine reads, what it says, and what it cannot check.

What this page is

The rest of this site is written for people who build hospital software. This page is for everyone else: the clinician who has been sold a demonstration before, the engineer from another industry, the reader of the book who wants to know what the machine actually does. No acronyms. One patient, one night, and the exact limits of what happened.

One thing first, because it matters more than anything below it. The patient is invented. Every case on this site is synthetic and every screen says so. What is not invented is the reasoning, which is generated live, in the cloud, at the moment you press Start, over the numbers on your screen and over public federal drug-safety data fetched at that moment. The page shows you that data and the web address it came from, so you can check it in a second tab rather than take anyone's word.

The problem, in one paragraph

Every alarm in a hospital is a line somebody drew in advance. A value crosses it and a machine makes a noise. That design has a flaw built into its foundations: nothing is said until the patient is already on the wrong side of a number. A patient in intensive care produces a continuous stream of information, heart rate and pressure and oxygen and the exact rate of every drip and every lab result, and between eleven at night and seven in the morning the only thing reading that stream is a set of limits. A trend is not a value. It has no line to cross. So the machine that holds the numbers says nothing, and a tired human being is asked to be the monitor.

For forty years my industry, including me, got very good at collecting that stream and moving it. Nobody built anything that could read it and ask a question nobody had written down in advance, because until recently nothing could.

One patient, one night

Call her Ruth. Seventy one, admitted through the emergency department with a serious infection, now in an intensive care unit with one drip holding her blood pressure up, another running an antibiotic, and a catheter measuring what her kidneys are doing. Her nurse has three other patients. A doctor looks carefully at seven in the evening, a different doctor looks quickly at eleven, and at three in the morning nobody looks, because nothing has alarmed and there are two sicker people down the hall.

Every few minutes, all night, the method reads everything about Ruth at once. Not one number: all of them, including the sentences. Then it asks the same question every time. Is what is happening to this person still consistent with what we believe is happening to her, and is what we are doing about it still the right thing to be doing?

At about one in the morning it notices four things that are individually unremarkable. Her urine output has been falling for three hours, gently, never crossing a line. Her potassium came back at the top of the normal range at ten o'clock, which is normal, so nobody was told. She is on an antibiotic that is hard on the kidneys, at a dose that was right for the kidney function she had when she arrived. And she is on a second drug, for her heart, that keeps potassium in the body. Any competent clinician handed all four together would see it immediately. Nobody is handed all four together at one in the morning. That is the entire problem.

So the machine says something. Once, quietly, in about as many words as this paragraph: her kidneys appear to be closing down, she is on two things that will push her potassium up, the last potassium was already at the top of the range, recommend a repeat potassium now and a review of the antibiotic dose. Then it says what it is unsure about, and it names the public evidence it used. A blood test at half past one comes back higher. The dose is changed at two. Ruth is fine in the morning, and nobody writes anything down, because nothing happened. That is the product: an event that did not occur, on a night nobody remembers.

What the demonstration on this site actually does

Press Start on the home page and a simulated patient's numbers begin to move through a compressed encounter. At each moment the method decides to speak, the encounter pauses and you read three things.

• The advisory. The risk now, not at a limit already crossed. The reasons, each traceable to a number on the screen or a public source. Options, for a person to accept or reject. It never acts on its own in the standard cases; a clinician reads it and decides.
• The evidence it read. For every active drug: what the drug is and what class it belongs to, from the National Library of Medicine's RxNorm; what has gone wrong with it across the population, from the FDA's adverse-event reports, with the date of the data; whether the official label is on file, from DailyMed; and how many registered studies mention it, from ClinicalTrials.gov. Each line carries the web address that produced it.
• What it could not check. Every response states how many of the eight standard medication safety checks it was able to perform and names the rest. On one vendor's certified test record the answer was two of eight. That half is the point. A machine that will tell you what it does not know is the only kind worth listening to.

Each alert is labeled. If it was generated live it says so. If the day's allowance has been used, or a call fails, the page says that out loud and labels the note as one written in advance, not generated now. There is no quiet fallback, because a system that will tell you it is live when it is not is a worse thing than no system at all. I know that because the first version of this demonstration did exactly that to me, and I took it down.

What is new about it, stated carefully

Not the data. The federal reference layer has been public and free for decades; my own company loaded it off CDs in 2001. Not the connectivity, which was solved, badly and expensively, over twenty years of standards work I sat in on. What arrived, from outside medicine, is a machine that can be handed the whole picture, the numbers and the sentences and the timing and the gaps, and asked a question nobody wrote down in advance. Not calculate. Reason.

Closed loops in medicine are not new either. Target-controlled infusion and the artificial pancreas have existed for decades, and each is an excellent controller that manages one variable toward one target: a very good thermostat. The method here is different in the one way that matters. The controller is reasoning over the entire context, not a rule mapping one number to one action. One case on this site shows a supervised version of that loop, clamped to limits a clinician sets, and the technical datasheets say plainly what it is and is not.

And the cost. One reasoning pass over one patient costs on the order of a cent, and less as context is reused. The expensive part of clinical decision support was never the software. It was licensing the knowledge. When the knowledge is public and the reasoning is a commodity, the question stops being what can we afford to check and becomes what should we be checking. My industry has barely started having that conversation.

What it is not

• Not a medical device. A research demonstration, advisory only, not cleared by anybody, not for clinical use. The FDA's own guidance makes clear that a system watching a patient continuously and speaking in time-critical moments would be a device, and I would make the same argument in their position.
• Not fed real patients. Synthetic data only, everywhere, always.
• Not a product. There is nothing to buy here. I am retired. The point of the site is to show that the method is doable and credible, in public, where anyone can test it and tell me where it fails.
• Not finished. The hard part, connecting a recommendation to a named patient at a real bedside safely, lawfully and in time, is not solved by this or by anyone. The documents list everything still wrong with it.

The demonstration is open to anyone and capped each day, because every live run costs real money and it comes out of one pocket. When the allowance is used the page says so. Run itThe technical datasheet (PDF, 2 pages)