Your RTLS Knows Where the Cart Is. Not What Is On It.
A 2026 buyer's guide to hospital indoor positioning lists what these systems are typically bought for: asset utilisation, staff safety, patient flow through the emergency department, hand-hygiene compliance, capacity reporting. It is an accurate list, and it is worth reading twice for what is not on it.
Nothing in it involves knowing what is inside a container. That is not an omission on the part of the guide. It is a fair description of what the technology does, and the distinction decides which problems a hospital can already solve with the systems it owns and which ones are still open.
What RFID and RTLS actually are
The two terms get used as though they were interchangeable. They are not, and the difference is the whole point of this article.
RFID stands for radio frequency identification. A tag carrying a small chip and an antenna is attached to a thing. A reader sends out radio energy, the tag answers with its identifier, and a record is written. A passive tag has no battery and only answers when it is close enough to a reader, which in practice means passing through a doorway portal, going into a tunnel reader, or being swept with a handheld wand. The record it produces is an event: this identifier was seen at this reader at this time.
RTLS stands for real time location system. It is a layer above identification. Instead of writing an event when a tag passes a fixed point, it estimates where the tag is on a floor plan, continuously. That requires an active tag, meaning one with a battery that transmits on its own, and a grid of receivers through the building.
Accuracy depends entirely on the radio technology underneath, and the spread is large. A published 2026 hospital buyer's guide from Crowd Connected gives Bluetooth mesh at 2 to 3 metres, Wi-Fi positioning at 5 to 8 metres, which it notes is "often not enough to distinguish between adjacent rooms", and ultra-wideband at under 30 centimetres. Bluetooth direction-finding implementations sit between those figures. Cost runs the same way: the same source puts passive RFID labels and inlays at 0.04 to 0.25 US dollars each, on-metal passive tags at 0.75 to 5.00 dollars, and active tags, which are what RTLS needs, at 15 to 50 dollars and upwards per tag before any infrastructure. So a hospital with a good RTLS can find an infusion pump in the right room. That is genuinely useful and it is why these systems get bought.
What the tag reports, and what it does not
Here is the sentence that matters. A tag reports the position of the tag.
If the tag is on a case cart, the system knows where the cart is. If the tag is on a rigid container, the system knows where the container is. If the tag is on a shelf unit, the system knows where the shelf unit is, which is where it has always been.
None of that is a statement about contents. The link between the identifier and what is physically inside is made once, by a person, at the moment the thing is packed or restocked, and it is recorded on a count sheet or a pick list. Every reading after that inherits whatever that person concluded. The audit trail stays complete and the timestamp stays accurate. If an item was missing when the sheet was signed, the system will report a correctly located, correctly processed, complete container, and it will do so with a name and a time attached.
This is the same structural gap we wrote about at the sterile services assembly bench in Sterile Processing Tray Completeness: Can a Camera Do the Second Check?, where published observational work put visualisation failures at 88.6 per cent of observed sterile processing errors. That article deals with the bench. This one deals with everywhere the container goes afterwards.
Where the gap shows up
Two places, and they are usually owned by different departments.
The case cart, between sterile services and the theatre. The cart is picked against a list, wrapped or closed, tracked, and moved. Location tracking confirms it arrived. Nothing between the pick and the theatre door confirms it arrived complete. The consequence is discovered in the most expensive room in the building, at the point where the alternative to waiting is finding a substitute instrument or running an immediate-use sterilisation cycle.
The ward or theatre supply room. Consumables sit on shelves and in cupboards. Somebody counts them, on a rota, by looking. Between counts, the stock level is an estimate. A ward that runs short discovers it at the point of use, and the response is a phone call, a walk to another ward, or a rush order. A ward that overstocks discovers it when something expires. Neither is visible to a location system, because nothing in the cupboard is tagged and nothing about the cupboard has moved.
Note what these two have in common. In both cases the hospital is not missing a location. It is missing a count.
Why item-level tagging does not close it everywhere
The obvious answer is to tag everything. It is a fair answer, it works in specific places, and it is worth being straight about where it stops.
Instrument-level RFID is real and it survives reprocessing. Autoclavable tags exist and are in use. But the vendors themselves scope it carefully. Xerafy, which sells exactly this, describes tray-level identification as "a scalable foundation" and says item-level identification "can then be added selectively where the instrument, workflow and validation requirements justify it". That is an accurate description of a per-item programme: tag selection, position and attachment tested on representative trays, validated, then rolled out to the sets where the case is strong enough.
For consumables the arithmetic is simply different. A tag at 4 to 25 cents is a rounding error on a 200 euro item and a material cost on a 60 cent one, and every tag has to be applied by somebody. Tagging a shelf of low-value consumables in order to learn whether the shelf is full is not a proportionate answer to the question.
And even where everything is tagged, a passive tag answers when it is read. A cupboard is not a tunnel reader. Knowing the contents of a cupboard continuously means either a reader in every cupboard or a person with a wand walking the same round they walk now.
What a camera adds, and what it does not
A camera looking at a defined space answers a different question from a tag. It does not identify a specific instrument by serial number, and it should not be sold as though it does. What it does is count and classify what is visible against what is expected, whenever it looks.
That makes it a poor replacement for a tracking system and a reasonable complement to one. The tracking system holds identity, custody and history. The camera holds presence and count. A cart that is tracked and also confirmed complete before it leaves the packing area is a different object from a cart that is only tracked.
Three conditions decide whether this is worth anything in a hospital, and all three are architectural rather than commercial.
It has to run on cameras the estate can actually install. Conventional CCTV, RTSP and USB cameras. A proposal that requires purpose-built imaging hardware for every cupboard does not clear capital approval, and it should not.
The people who know the sets have to be the ones training it. A sterile services technician can label a photograph of an orthopaedic set correctly. A supply chain lead can label a full shelf and a short one. An external integrator cannot, and a system that needs an integrator every time a set composition changes will be abandoned inside a year, because set compositions change constantly. This is the reason MAKRR is built as a no-code platform: the person with the domain knowledge labels the examples and deploys the detector.
Inference has to happen on site. Models are trained in the cloud and deployed to an edge device the hospital controls. The video stays inside the building and only the result leaves, meaning a count, a flag, an exception record. That is a data minimisation argument rather than a promise about compliance, and it is short enough to explain in an operational meeting, which matters as much as the technical position. We set out the architecture in more detail in Edge vs Cloud Computer Vision.
This is general information about system architecture and it is not legal advice. Confirm your own position with your data protection officer.
Keep the detection target on the object
One boundary, stated plainly.
The detection target here is the tray, the cart and the shelf. Not the member of staff standing at it. Counting the contents of a cupboard does not require identifying anybody, and a system built for it should not be able to. The tracking system already knows who was on shift. Adding face recognition to a supply room camera creates a staff monitoring problem in exchange for information the hospital already holds.
MAKRR makes no clinical, diagnostic or patient-outcome claim, and patients are never the detection target.
A first project that proves or disproves it
The way to settle this without taking anybody's word for it is narrow.
Pick one thing. One high-turnover set that has a documented history of coming back marked short, or one supply room that runs out of the same three items. Record or photograph enough examples to cover normal variation, including the short cases. Label them. Train the detector. Then run it alongside the existing check for two or three weeks, changing nothing operationally, and compare what it flags against what people found.
At the end of that you have your own numbers rather than ours. The model itself is typically working within a day. The parallel run takes weeks, and it should.
Contents verification is a narrow question and it is testable. If it is one you are looking at, we are glad to talk it through.
FAQs
-
RFID identifies a tag when it comes within range of a reader and writes an event. RTLS estimates where an active, battery-powered tag is on a floor plan, continuously, using a grid of receivers. RFID answers "was this seen here". RTLS answers "where is this now".
-
It depends on the radio technology. Published 2026 figures put Bluetooth mesh at 2 to 3 metres, Wi-Fi positioning at 5 to 8 metres, and ultra-wideband at under 30 centimetres. Room-level certainty needs better than the room dimension, which is why Wi-Fi positioning often cannot separate adjacent rooms.
-
Not on its own. It reports the location of tagged items. A stock level for untagged consumables on a shelf is not something a location system observes.
-
No, and it would be the wrong move. Identity, custody and reprocessing history are what those systems are for and they do it well. The open question is contents verification, which sits beside them.
-
Not in this architecture. Models are trained in the cloud and deployed to an edge device on site. Video is processed locally and only counts, flags and exception records leave the device.