Privacy, safety and limits

Privacy

Local only. The server, the database and every screen run inside the building, on its own network.
No recordings. Sound is analysed in the room; only a score (“distress 0.8”) leaves it. Nothing is recorded.
Pseudonyms. The prototype uses no real patient data, ever.

Never fails silently. Every sensor reports that it is alive every 5 seconds; a sensor that goes quiet, or whose connection drops, raises Sensor offline within seconds. A screen that loses the server says Not live.
Everything is on the record. Approvals, alerts, acknowledgements, pages and rule changes are journaled, with when they happened. The server rebuilds its state from that journal after a restart.
A bell is never ignored. A pressed call bell always raises a critical alert, even on an empty bed.

It is a supplementary alert: it adds to rounds and standard monitoring, never replaces them.
Thresholds are demonstration values until the clinical team signs each one off.
Security is not done: the control room PIN is a demonstration gate; accounts, roles, encrypted links and per-device credentials come before any real use.
Sound recognition today is loudness, not a trained classifier: a clap scores like a call for help. Recognising choking or crying needs real recordings, collected with consent.
Approvals: room microphones need privacy approval, and the regulatory status must be confirmed.