The Night the Fall Detector Went Quiet: Tracing a Missed Alarm to a Frame Cadence Assumption
A fall detector that misses a fall rarely fails loudly. It fails by saying nothing, and in most ambient assisted living architectures, saying nothing looks exactly like all clear. The case here is a room-mounted vision pipeline that reliably caught daytime falls, then missed a resident's fall in the small hours while every service reported healthy. Nothing crashed. The fault was an assumption built into the classifier long before that night: the pose stream arrives at a steady rate, so if no fall signature appears, no fall happened.
The trigger and the assumption are different findings
The trigger was ordinary. The camera switched to infrared night mode, auto exposure lengthened the integration time, and the effective frame rate dropped well below its daytime value. During the descent, motion blur smeared the keypoints. Pose confidence fell under the tracker's gate, the tracker released the track, and the classifier received an empty window instead of a fall. The descent probably lasts a few hundred milliseconds (an estimate), and at that cadence it covered too few frames for the temporal model to recognise anything even if the track had held. The trigger explains this one night. The assumption explains why every similar night would end the same way. Retuning exposure or swapping the camera removes the trigger but keeps the assumption, which will fail again at the next firmware update, smudged lens or unlit bedroom.
Where silence entered the design
The pipeline was built as a chain of capture, pose estimation, tracking, classification and alerting, and every stage was designed to emit events, not states. When the tracker lost the resident it published nothing, and that output is identical to what an empty room produces, so no downstream stage could tell the two apart. Health monitoring watched process liveness and frame throughput. Both stayed green, because frames were still arriving and every process was alive, just blurred and slow. Liveness is not observational coverage, and that distinction was never written down as a requirement. Medical device software lifecycle and risk management standards such as IEC 62304 and ISO 14971 expect hazard analysis to cover degraded inputs. The hazard table, however, listed camera failure as a binary event, so a camera that was alive but degraded fell between its rows.
A rule you can audit
If a stage cannot vouch for its input, measured here as an effective pose rate below 12 frames per second sustained across the classifier's window, it must publish an explicit degraded state that the alerting layer treats as a reportable condition with its own escalation path, never as an absence of events.
Keeping coverage when one part fails
Redundancy only helps if the redundant parts do not share the same assumption. A second camera with the same infrared switching just doubles the blind spot. A mmWave radar or a floor vibration sensor relies on different physics and does not care about lighting. A wearable accelerometer ignores the room completely but fails silently whenever it is left on the nightstand. Fusion should therefore weigh coverage as well as detections: each source reports what it sees and how much of the room and time window it can actually vouch for. When vision degrades, the system can fall back on slower inactivity logic, such as prolonged stillness in the zone where the resident was last tracked. That logic survives frame loss but costs minutes of detection latency. The cost that remains is human. Degraded state notifications are not falls, and if there are too many of them, staff learn to ignore them, which recreates the silence one layer up. The next decision is who owns the escalation threshold for degraded states, because that number sets the trade-off between alarm fatigue and unobserved time, and it is a clinical choice disguised as an engineering parameter.
Technical FAQ: How do I write this up without implicating the person who approved the camera firmware update?
Record the update as the trigger and show that the firmware behaved within its specification. Then show that nobody had any signal the classifier depended on frame cadence, because that dependency was never stated as a requirement or monitored. The corrective action then targets the design: make the cadence assumption an explicit requirement that is checked at runtime, and add degraded but alive inputs as a separate row in the hazard analysis.