An older person living at home, and the family looking after her. The visit writes what the next call asks. Each call adds one point to a pattern only her own history sets the baseline for, and a shift in that pattern is treated as a finding, not a diagnosis. What she raises reaches only the people she authorised.
This walkthrough shows a representative case, on screens that are the product's design. The agent and the record behind them are built, and the record it actually wrote is on the record.
The appointment is already on her phone. Nothing is recording, and the screen says so in the language she reads.
录音前会先出声告知
Nothing records until she taps. The app says so out loud, so the room knows too.
录音前会先出声告知
Her screen shows that it is running and how to stop it. It does not show the words. She is listening, not reading.
录音前会先出声告知
Four things to watch for are written from what the doctor said. This is the handover.
The agent calls in Hokkien and asks those four by name, rather than asking how she is.
The visit's key points are on her screen in large print, an overview rather than a transcript. What was raised is shown alongside who it was routed to, and who it was not.
The visit is where the record is written. The calls are where it is checked, one morning at a time, against the way she sounded the morning before.
The cadence is hers: daily, every other day, or three times a week. Anything slower is refused, because the call is the conversation and the speech sample at once.
7 July, 22 July, 28 July. 41 of 44 calls left a usable sample.
Markers that depend on recognising her words switch off automatically when that confidence can't be trusted. Acoustic and behavioural markers don't need the words to be right, so they keep working.
The first 30 samples set the band, never a population average. Her last 14 days are 2.1x as spread as that period, on 11 samples against 30.
Time spent pausing has become less consistent from day to day over her last 14 days: 2.1x the spread of her own baseline period, on 11 samples against 30. The medication list was checked first and nothing on it had moved.
If she does not answer, that day is left blank rather than estimated. Filling it in would smooth the day-to-day spread, and the spread is the thing being measured.
7 July, 22 July, 28 July. 41 of 44 calls left a usable sample.
Markers that depend on recognising her words switch off automatically when that confidence can't be trusted. Acoustic and behavioural markers don't need the words to be right, so they keep working.
The first 30 samples set the band, never a population average. Her last 14 days are 2.1x as spread as that period, on 11 samples against 30.
Time spent pausing has become less consistent from day to day over her last 14 days: 2.1x the spread of her own baseline period, on 11 samples against 30. The medication list was checked first and nothing on it had moved.
Acoustic and behavioural markers come from pausing, articulation, timing and response. Anything counted from her actual words switches off automatically whenever recognition confidence can't be trusted.
7 July, 22 July, 28 July. 41 of 44 calls left a usable sample.
Markers that depend on recognising her words switch off automatically when that confidence can't be trusted. Acoustic and behavioural markers don't need the words to be right, so they keep working.
The first 30 samples set the band, never a population average. Her last 14 days are 2.1x as spread as that period, on 11 samples against 30.
Time spent pausing has become less consistent from day to day over her last 14 days: 2.1x the spread of her own baseline period, on 11 samples against 30. The medication list was checked first and nothing on it had moved.
A missed call is not counted toward it, and it is never a population average. Rising inconsistency against her own baseline shows earlier than a sinking average does.
7 July, 22 July, 28 July. 41 of 44 calls left a usable sample.
Markers that depend on recognising her words switch off automatically when that confidence can't be trusted. Acoustic and behavioural markers don't need the words to be right, so they keep working.
The first 30 samples set the band, never a population average. Her last 14 days are 2.1x as spread as that period, on 11 samples against 30.
Time spent pausing has become less consistent from day to day over her last 14 days: 2.1x the spread of her own baseline period, on 11 samples against 30. The medication list was checked first and nothing on it had moved.
Her pausing has become less consistent from day to day. The medication list was checked first and nothing on it had moved, so the spread is hers, and it is raised.
7 July, 22 July, 28 July. 41 of 44 calls left a usable sample.
Markers that depend on recognising her words switch off automatically when that confidence can't be trusted. Acoustic and behavioural markers don't need the words to be right, so they keep working.
The first 30 samples set the band, never a population average. Her last 14 days are 2.1x as spread as that period, on 11 samples against 30.
Time spent pausing has become less consistent from day to day over her last 14 days: 2.1x the spread of her own baseline period, on 11 samples against 30. The medication list was checked first and nothing on it had moved.
Her son gets this drift digest but never her verbatim words. Her helper gets her verbatim words but never a drift digest. Only her daughter, shown here, holds both.
7 July, 22 July, 28 July. 41 of 44 calls left a usable sample.
Markers that depend on recognising her words switch off automatically when that confidence can't be trusted. Acoustic and behavioural markers don't need the words to be right, so they keep working.
The first 30 samples set the band, never a population average. Her last 14 days are 2.1x as spread as that period, on 11 samples against 30.
Time spent pausing has become less consistent from day to day over her last 14 days: 2.1x the spread of her own baseline period, on 11 samples against 30. The medication list was checked first and nothing on it had moved.
Time spent pausing has become less consistent from day to day over the last 14 days, against her own baseline. This is a change in how she speaks. It is not a conclusion about her, and it is worth raising at her next appointment.
The call and the record work in English, Mandarin, Malay, Tamil and Cantonese end to end. Where speech recognition is not reliable enough to trust, the product reads how she speaks rather than what she said, so the daily call still works in a language no recogniser handles, and nothing counted off an unreliable transcript is ever charted.