Every hospital has infection prevention protocols. Surgical checklists. Hand hygiene policies. Sterile technique guidelines. They exist because they work, when they are followed.
The problem is that nobody actually knows how often they are followed.
The arithmetic of observation
Take a 20-bed ICU running three shifts. Each patient may undergo multiple procedures per day that require sterile or clean technique: central line maintenance, dressing changes, medication administration through ports, blood draws.
A conservative estimate puts this at 80 to 120 protocol-dependent events per shift. Across three shifts, that is 240 to 360 events per day. In a single unit.
A typical infection preventionist is responsible for an entire hospital, sometimes multiple facilities. Direct observation audits sample a handful of events per month. Those numbers do not add up.
Published compliance rates from audited observations hover around 80 to 95 percent. Unannounced audits and covert observation studies consistently come in 15 to 30 percentage points lower. The presence of an observer changes behavior. Remove the observer and baseline habits return. This is the Hawthorne effect, and it has been documented in healthcare settings since the 1980s.
The operating room has the same problem
Surgical teams follow the WHO Surgical Safety Checklist, scrub protocols, instrument handling procedures, sterile field maintenance guidelines. The compliance auditing model is identical: someone watches, records, reports.
A 2020 systematic review in BMJ Quality & Safety found full compliance with the WHO checklist in only 39 percent of observed cases. Partial compliance was common. But the denominator (total number of surgeries that should have been observed) was never captured. There were not enough observers.
Surgical site infections affect over 157,000 patients per year in the US. Each one costs 60,000. The root cause is not that surgical teams lack training. It is that nobody measures whether they follow the training.
Why sampling does not work
Infection prevention programs sample because continuous observation by humans is physically impossible. But sampling has specific, well-documented failure modes.
The Hawthorne effect means sampled audits measure best-case behavior, not typical behavior. Auditors know this. Everyone knows this. The data gets reported anyway.
Selection bias means audits happen during convenient hours, usually weekday mornings. Nights, weekends, emergencies get missed. These are also when compliance degrades most.
Then there is statistical adequacy. Thirty observations per month in a unit generating 10,000 protocol events per month means you are inspecting 0.3 percent of activity. In manufacturing, a 0.3 percent inspection rate on a process where failure kills people would be considered negligent.
The case for continuous measurement
Computer vision does not fix training. It does not fix protocol design. What it fixes is the measurement gap.
Cameras in procedure areas feed video to an on-premise processing unit. The AI recognizes protocol steps and documents deviations in real time. Every event gets recorded. Every shift. No sampling, no scheduling, no Hawthorne effect.
This is not a replacement for infection preventionists or surgical quality teams. It is the measurement infrastructure they have never had.
The protocols exist. The question has always been whether anyone is watching.