At 3 a.m. a ward is a different hospital: low lights, quiet bays, a handful of staff covering patients who, by day, would have a ward full of eyes on them. When she falls, there is often no witness and no sound anyone connects to danger. She cannot always reach the call bell. So she waits. Clinicians call that wait a long lie. Time on a cold floor does its damage quietly: dehydration, pressure injury, hypothermia, and fear. A camera that records perfectly but raises no timely response has documented failure.

The night is the risk

In a 2026 Basel study, night-shift fall risk was 7.6 times higher with contact mats. A Japanese study found 1.47 nursing hours per patient at night, 1.99 by day. A Welsh study found 54% of hip-fracture falls at night, 22% between 02:00 and 05:59. A daytime demo proves the easy half. The question is what it does at 03:20, in the dark.

We studied the long lie at home. Then we stopped at the hospital door.

Almost everything we know about long lies comes from homes. In the BMJ study of people over 90, 80% could not get up, 30% lay on the floor over an hour, and 82% were alone. In Gurley's NEJM study, people helpless for over 72 hours died at 67%, against 12% for those found within an hour. Yet a 2025 review found only six articles on long lies in 45 years. On a ward, a patient with delirium may be less able to summon help.

Four clocks are running. None of them measures discovery.

Ask a hospital for its fall response time and you get one of four numbers. A medical review within 30 minutes, as NICE requires when injury is suspected. A remote observer reacting in 15.8 seconds. Staff reaching a patient in 29 seconds. A call-light wait of up to 17 minutes 27 seconds. Four clocks, four definitions, never compared. The National Audit reports 77% checked within 30 minutes. But that standard clocks nothing before discovery. The audit's own example begins with a patient "found on the floor by her bedside 15 minutes ago". Those 15 minutes, the most dangerous in the sequence, belong to no clock.

Name the chain, and it becomes a number you can manage.

A fall is not one event with one response time. It is a chain of six stages: event, detection, notification, ownership, arrival, assessment. Work the night example honestly. Detection takes 2 seconds, routing 3, acknowledgement 15, travel 60 to 120 at night. Best case, help arrives in 80 to 140 seconds. Compare discovery at the next round: 30 to 180 minutes. That comparison is the case for intervention. The chain deserves to be a KPI, split by day and night. Most wards know last month's fall count. Far fewer know their median fall-to-arrival time.

Do the arithmetic vendors would rather skip.

Moving from lying to sitting takes 2.6 seconds for a healthy young adult, 4.4 for an older person, 9.8 for a skilled-nursing patient. One leading system advertises a lead time of 31 to 65 seconds. Even at the generous end, if the responder is 60 seconds away, detection cannot prevent it. At night, its principal value is not prevention. It is a shorter lie: minutes of exposure converted into seconds.

Accuracy is the wrong headline. Precision is the story.

Every product leads with accuracy. It is the least informative number on the page. A Frontiers review of 25 studies found AI sensitivity of 97.4% to 98.4%; only 2 used a concurrent comparison group. Falls make up less than 2% of monitoring footage. A 2026 Nature study: at 82.4% recall, precision was 32.2%. Roughly two alerts in three needed no action. Kwong Wah Hospital reached 100% sensitivity with a 24.1% false-alarm rate. False alarms are the adoption risk. Black Book Research, surveying 202 nurses in September 2026, found 49% use workarounds to reduce alerts, and 65% feel individually tracked. Report false alerts per bed, per night, every week.

For developers, the pipeline is the product.

No major vendor publishes inference or alert latency. Each hop adds time and can fail. Mature teams set a budget in seconds for event to owned alert. LookDeep's peer-reviewed work, 1 frame per second with detection F1 of 0.92, is serious engineering. Edge works offline; cloud scales at the cost of a hop.

Surviving deployments use the ward's nurse-call systems, with structured alerts carrying room, event and time on the floor. Rauland, Ascom and Austco integrations all exist in the wider market. Escalation is timed: 90 seconds, then 3 minutes. An unacknowledged alert that simply expires is a design decision to accept a long lie. Most hospitals already own the sensor network: cameras, a video system and a control room. The question is whether it stays a recording network or becomes a response network. Systems that run on cameras already installed in the building, as 360Alert does, need no capital hardware programme. A camera estate that cannot be patched, segmented and audited is attack surface with a lens.

Prove it in the dark, before anyone depends on it.

No fall-detection vendor publishes a shadow-mode protocol. Run it for 30 to 90 days detecting and logging, sending no live alerts. Then compare, room by room: events, false alerts per bed per night, and the night split. Agree go or no-go thresholds in writing before the pilot starts. A vendor who resists them has told you something the brochure did not.

Privacy is architecture, not reassurance. And be honest about the bathroom.

In the UK, monitoring a person in bed needs a Data Protection Impact Assessment under UK GDPR, a lawful basis, consent, short retention, and regard to the Surveillance Camera Code. Most NHS trusts state CCTV covers non-clinical areas only. Moving sensing into the room is a governance step-change. Kwong Wah matters twice over: thermal imaging with no patient-identifiable information still delivered perfect sensitivity. Around half of inpatient falls are elimination-related. The ensuite, at night, alone, is where toileting falls happen, and where a camera can never go. Trust will belong to whoever is candid about the blind spot.

Measure the minutes

Falls will still happen. The wait that follows is not inevitable. It is a number. It can be defined and reported like the fall count itself. The fall takes a second. The failure is everything that does not happen in the 40 minutes after it. Measure those minutes, and hold every system to them. The patient on the floor cannot do it for us.

About the author: 360Alert develops AI fall detection for hospitals and nursing homes, using cameras already installed.

Sources: NHS; NAIF / NICE QS86; Fleming, BMJ 2008; Gurley, NEJM 1996; Basel 2026; Welsh study; Quigley 2019; SmartPeep 2025; Frontiers, Oct 2026; Nature 2026; Kwong Wah 2026; LookDeep 2025; Black Book, Sep 2026.