Walter isn’t a real patient, and the numbers in this piece are illustrative.
Walter went back into hospital 14 days after he was discharged. In those two weeks, his home monitoring didn’t send his care team a single alert.
He’s 74 and has heart failure with a weak pump. He had been admitted because fluid had built up in his body, which is the usual reason people with heart failure end up in hospital. The team took the fluid off with diuretics, and while they did, his kidney function dipped. So when he went home, they halved his diuretic dose. It was a sensible decision. It also gave the fluid room to come back.
He left with a 30-day monitoring kit: a scale, a blood pressure cuff that also records heart rate, a pulse oximeter, and a short questionnaire on his phone each morning. His program would alert the care team if his weight went up by 3 lb in a day or 5 lb in a week, if his heart rate went over 100, if his systolic blood pressure fell under 90, or if his oxygen saturation dropped below 90%.
This is what his second week at home looked like.
Six days of drift. Not one alert.
Walter, days 7 to 12 after discharge. Dashed line: the program alert threshold.
Weight gain over 7 days
lb · alert if above 5
Resting heart rate
bpm · alert if above 100
Systolic blood pressure
mmHg · alert if below 90
Oxygen saturation
% · alert if below 90
Daily steps
steps · no alert rule
Every signal drifted the wrong way. None crossed its line.
Nothing crossed a line. His weight crept up by about 4 lb over five days, never more than a pound on any single day. His resting heart rate went from 72 to 86. His blood pressure eased down a little, and he was walking about half as much as he had been.
His morning check-ins had more to say, if anyone had been reading them closely. On day 9 he wrote that he felt a bit more tired. On day 10 his ankles were a little puffy, which he marked as mild. On day 11 he mentioned that he’d needed an extra pillow to sleep. Every morning he answered no when the app asked if he was short of breath at rest.
On day 14 he was short of breath at rest, and he was readmitted that afternoon.
What an experienced nurse would have seen
Put the same week in front of an experienced heart failure nurse and I’d expect a phone call to Walter by day 11, maybe sooner. None of the numbers would have worried her on its own. What she would have noticed is that they were all moving in the same direction, and that she knew why.
She’d have seen a steady climb in weight in a patient whose diuretic had just been cut in half. For Walter, that slow gain was the most predictable way for things to go wrong. She’d have tied it to his rising heart rate, his falling activity and his drifting blood pressure, each of which is easy to wave away by itself. And she’d have known that puffy ankles and needing an extra pillow at night are early signs of fluid coming back, even though no rule had been set up for either.
The warning sign was the combination, read against what happened on the day he left hospital.
Why the alarms stayed quiet
A threshold asks one simple question: has this number crossed this line right now? That works well for sudden events. It works badly for Walter. The rule looks at one reading at a time, so a slow climb never sets it off. It checks weight, heart rate and symptoms separately, so it can’t notice when they all point the same way. It has no idea his diuretic was just cut. And it can’t read what he types into a free-text box.
The research fits this picture. Chaudhry and colleagues studied heart failure patients from a home weight-monitoring program of more than 10,000 people. They found that weight started to rise about a month before a heart failure admission and sped up in the final week. Gains of just 2 to 5 lb were linked to nearly three times the odds of being admitted (Circulation, 2007). That’s under the “5 lb in a week” rule patients are commonly given (ColumbiaDoctors).
Why lower thresholds don’t fix it
The obvious response is to lower the lines. The trouble is that daily weight bounces around with meals, clothes and the scale itself. Set the limit low enough to catch Walter and you’ll also flag a lot of patients who are perfectly fine. Nurses learn to tune out the noise, and the one alert that matters gets lost with the rest. Most programs end up trading one problem for the other: they miss patients like Walter, or they bury their nurses in alerts.
I don’t think better tuning gets us out of this. Checking a single number against a fixed line just isn’t how clinicians judge whether someone is getting worse.
What would actually help
To catch Walter on day 11, software would need to do roughly what that nurse does.
It would need to remember why his diuretic was halved and keep that in view as each new reading comes in. It would need a clinician to tell it what to watch for in patients like him, something along the lines of “after a diuretic cut, watch for steady weight gain even under 5 lb, especially if heart rate is rising or they’re having trouble lying flat.” And when the picture starts to add up, it should put a suggestion in front of a person instead of acting on its own:
Walter is up about 4 lb in five days, 12 days after his diuretic was halved. Heart rate up 14. Says he’s sleeping propped up. Worth a call?
The nurse decides what happens next. The software never changes a dose or contacts the patient by itself.
My co-founder Keith Deutsch calls this kind of system a context layer. By that he means shared infrastructure that keeps a living picture of the patient, instead of rebuilding it from scratch every time someone asks a question. He sets out the architecture in his white paper, The Context Window and the Context Layer, and one line from it describes Walter’s week well:
“Individually inconclusive signals may accumulate and interact until their combined significance crosses a contextually determined threshold.”
We haven’t built this yet. We’re working through the specification carefully, the prototype comes next, and clinicians are shaping what it watches for.
Where this leaves us
Walter’s monitoring did what it was set up to do. Every reading was taken, sent and checked. Nobody had asked it to notice a story, and for patients like Walter, the story is often where the trouble shows up first.
Sources
- Chaudhry SI et al. Patterns of weight change preceding hospitalization for heart failure. Circulation. 2007;116:1549-1554. Summary via Relias Media.
- Heart failure: checking your weight, ColumbiaDoctors patient education.
- Keith Deutsch. The Context Window and the Context Layer: An Architectural Distinction for Agentic AI. October 2026.
