The Concentrator Alarms at 3 A.M. Saturday. Its Oxygen Purity Started Drifting Nine Days Earlier

"We already track our equipment." You track hour meters on a spreadsheet

Every agency owner says the same sentence when this comes up, and it is usually true in the narrow sense. Somebody in the DME closet keeps a list: unit serial number, patient, delivery date, last service. The concentrators get cleaned and checked when they come back. The hospital beds get looked at when a bed rail complaint comes in. This is calendar-based maintenance, and it is exactly the thing factories spent the last two years walking away from.

You were probably pitched something in 2024 that promised to fix this, and it was a tracking app that asked your delivery driver to type more things into a phone in somebody's driveway. It did not survive contact with a Friday route. What is different now is that nobody has to type anything: the machine itself is the reporter, and the checking is cheap enough to run on all of it, all the time.

The 3 a.m. failure, and who pays for it

A hospice patient on five liters of continuous oxygen. Saturday, 3:10 a.m. The concentrator alarms — low purity, or the compressor stops, or the unit simply gets loud and hot in a way that terrifies a family. The daughter switches to the backup E-cylinder, if she remembers there is one and if she remembers how, and calls the on-call line.

Now count that one event. The on-call RN spends 40 to 90 minutes on it, most of it coaching a frightened family through a cylinder regulator over the phone. Somebody triggers an after-hours delivery, billed at an emergency rate if you contract your DME and paid in overtime if you own the fleet. There is a real chance the family calls 911 anyway, and for a hospice patient an emergency department visit related to the terminal diagnosis is your cost and can end in a revoked election. And you may never get that family's confidence back, which surfaces months later on a CAHPS Hospice survey.

Unplanned equipment failure in this business is not a maintenance problem. It is a clinical event, a staffing event and a reputation event, all at once, and it always happens at the worst hour because that is when nobody is watching.

What a concentrator tells you nine days before it quits

Predictive maintenance means you stop replacing equipment on a calendar and start replacing it on a signal — in this trade, the signal is usually oxygen purity drifting downward over days on a unit whose sieve beds and compressor have a lot of run hours on them.

Home oxygen concentrators — the EverFlo, the Perfecto2, the Companion 5, the 525 series and their cousins — all carry an oxygen sensor and an hour meter, and most models built in the last decade log alarm history. Purity does not usually fall off a cliff. It sags: 95%, then 94, then 92 across a week or two as the sieve beds lose their ability to hold nitrogen, faster in a humid house or a dusty one or a home with a smoker in it. Compressors get hotter and pull harder before they fail. Filters clog on a schedule set by the carpet, not by the calendar.

Manufacturing figured this out first and at scale — predictive maintenance is now the single most-adopted AI use in manufacturing at roughly 64%, ahead of quality control and demand forecasting — and the reason it took off is that fusing a stream of readings with visual inspection beats fixed service intervals badly. The same fusion works here, and here the second half of it is free: an aide or a nurse is already standing in that living room two or three times a week.

flowchart TD
  A["Concentrator purity and hour readings"] --> E["Risk score per unit, updated nightly"]
  B["Alarm history and compressor run pattern"] --> E
  C["Aide photographs filter and display on routine visit"] --> E
  D["Home conditions: pets, smoker, humidity, altitude"] --> E
  E --> F["Weekly swap list for the DME technician"]
  F --> G["Unit changed on a daytime route with the family present"]
  E --> H["High-risk unit: backup cylinder check moved up"]

The same Saturday, run the other way

Nine days before that 3 a.m. call, the unit's purity reading has slipped two points and its hour meter has crossed the range where this model's sieve beds start to go. The nightly check puts it near the top of the swap list. On Wednesday the technician is already in that ZIP delivering a bed; he swaps the concentrator at 2 p.m. with the daughter standing there, shows her the new unit, checks that the backup cylinder is full and that she knows the regulator, and takes the old unit back to be rebuilt on a bench during business hours instead of condemned at 3 a.m.

The aide's part of this is one photograph. On her Tuesday visit she photographs the front of the concentrator — the display and the intake filter — through the visit app. The picture answers questions no sensor can: the filter is grey, the unit has been moved into a closet with no clearance, there is a cat asleep on it. Those are the three most common reasons a healthy machine starts failing, and none of them are in the readings.

The arithmetic on a fleet of 240 units

Assumptions, illustrative: 240 concentrators in homes, unplanned after-hours failures running 14 a month, an after-hours emergency delivery costing $185 whether you pay a DME contractor's rate or your own overtime, on-call nurse time on each event valued at $60, and a planned daytime swap costing $45 because the technician was going to be in that ZIP anyway.

Two honest caveats about that table. First, you will never get unplanned failures to zero; power cuts, dropped units and a grandchild unplugging something are not predictable. Five a month is a target, not a floor. Second, the cheapest line in the whole exercise is the checking itself — running the readings on 240 units every night costs a few dollars a month at current prices, which is why this is worth doing on a 240-unit fleet and not only on a 24,000-unit one.

Where this stops and a respiratory professional starts

The swap list is a maintenance list, not a clinical one. Nothing in this should ever touch a patient's prescribed flow rate, decide whether someone needs a different oxygen delivery method, or replace the respiratory therapist's assessment. If purity drift and worsening shortness of breath show up in the same week, the answer is a clinical visit, not a new machine — and it is a real risk that a well-run maintenance program lets a clinical change hide behind an equipment explanation.

Two more limits. If you contract your DME rather than owning it, you cannot do any of this without your supplier's cooperation, because the readings are theirs. Put it in the next contract conversation: ask for unit-level purity, hours and alarm history for your patients, and ask what their own service interval policy actually is. Second, older units in the field simply do not report anything, and no software fixes a machine with no way to tell you how it is doing. Those get handled the old way, by putting them on the shortest visit interval and replacing them soonest.

And keep the technician's judgement in the loop. A ranked list is a suggestion. The technician who has been swapping these for eleven years knows which serial number range gives trouble and which house needs the unit moved out of a closet before anything else helps.

Monday's version of this needs one number per concentrator: hours on the meter. Sort them. Then pull six months of after-hours equipment calls out of your on-call log and see where those units sat on that sorted list. If the failures cluster in the top third, you have proved the idea to yourself for the price of an afternoon, and you can run the swap list by hand while you decide whether to automate it.

Frequently asked questions

We contract all our DME. Is this our problem or theirs?

The after-hours call is yours, the family's fear is yours, and the emergency department trip is billed to you. So it is your problem regardless of whose truck it is. What you can do is stop treating equipment reliability as invisible in the contract: ask for failure counts by month, ask for their service interval policy in writing, and ask for unit-level data on your patients.

Does this apply to anything besides oxygen?

Hospital beds and alternating-pressure mattresses are the next two, and the failure signals there are different — a bed motor that has been running longer per cycle, a pump that cannot hold its pressure overnight. Ambulatory infusion pumps used for continuous subcutaneous infusion sit in their own category, because the manufacturer's service schedule is not something to freelance around. Start with concentrators; they are the highest-volume and highest-consequence item you have.

How many false alarms should we expect?

Early on, a lot — expect the first month's list to be dominated by units that just need a filter and a clearance fix. That is not a failure of the idea; a filter change on a daytime route is far cheaper than the alternative. Track what each flagged unit actually turned out to need, and the list gets more useful within a quarter.

Is there a survey or accreditation angle?

There is. Hospice conditions of participation expect the equipment you supply to be maintained and safe, and accrediting bodies will ask how you know. A dated swap record with the reason attached is a better answer to that question than a maintenance calendar with checkmarks on it.

A note on the calls this generates

Better maintenance moves work from 3 a.m. to Wednesday afternoon, but it does not remove the phone calls — families still call about equipment, about deliveries, about what the alarm means, and most of those calls are not clinical. CallSphere builds AI voice and chat agents that answer business phone lines and web chat around the clock, capture what the caller needs, and book or route it. It does not read a concentrator's sensors and it does not do the maintenance. It keeps the non-clinical half of those calls off your on-call nurse at 3 a.m.