Cognitive-decline pattern
Someone whose speech gradually shifts the way studies describe early cognitive decline: longer pauses, slower talking, more "um"s, and more retold stories.
Noticing the speech changes linked to early cognitive decline
Simulated: Demo data created for this presentation, not a real call.Research links early cognitive decline to gradual changes in everyday speech: longer and more frequent pauses, slower talking, more "um"s, and the same stories told again. We simulated people whose speech changes this way, and people whose speech doesn't, to see whether Remi notices the pattern and tells the family, without worrying families when nothing has changed.
99.5%
of simulated cognitive-decline speech patterns were noticed within six months, typically 11 weeks in
1 in 6 yrs
an unneeded note for a steady person, 10× fewer than our first rule
0%
of low-mood weeks were mistaken for a change in talking (first rule: 22%)
Who we simulated
Steady
Someone who doesn’t change, apart from ordinary ups and downs and one week with a cold.
Sudden slowdown
Someone whose talking suddenly becomes much slower (about 30 words a minute) and stays that way.
Two low weeks
Someone who feels down for two weeks, talks a little slower, then recovers.
Each person was simulated 200 times.
Cognitive-decline pattern over six months
The cognitive-decline pattern unfolding at a realistic pace (weeks 3–26). In this typical run the first family note comes in week 13, about halfway through the change. While the change continues, reminders come no more than once every 4 weeks.
| Week | Weekly score | Smoothed | State |
|---|---|---|---|
| 1 | +2.1 | +0.0 | learning his usual |
| 2 | +0.7 | +0.0 | learning his usual |
| 3 | -1.0 | +0.0 | learning his usual |
| 4 | -2.4 | -1.3 | steady |
| 5 | -0.5 | -0.8 | steady |
| 6 | +2.7 | +1.3 | watching |
| 7 | -0.1 | +0.6 | steady |
| 8 | -0.1 | -0.1 | steady |
| 9 | +2.9 | +1.4 | watching |
| 10 | +1.5 | +1.6 | watching |
| 11 | +2.1 | +1.8 | watching |
| 12 | +2.1 | +2.0 | watching |
| 13 | +5.2 | +3.6 | family note |
| 14 | +0.5 | +2.1 | watching |
| 15 | +2.5 | +2.4 | watching |
| 16 | +5.9 | +4.1 | watching |
| 17 | +3.3 | +3.7 | family note |
| 18 | +5.1 | +4.5 | watching |
| 19 | +4.5 | +4.6 | watching |
| 20 | +5.8 | +5.1 | watching |
| 21 | +5.8 | +5.4 | watching |
| 22 | +4.1 | +4.8 | watching |
| 23 | +6.0 | +5.5 | watching |
| 24 | +6.3 | +6.1 | family note |
| 25 | +5.8 | +6.0 | watching |
| 26 | +7.2 | +6.6 | watching |
Weekly score Learning his usual Smoothed trend His usual range Note line (1.5)
| Change over six months | Remi's rule: noticed | Typical weeks until note | First rule: noticed / weeks |
|---|---|---|---|
| Full size (as above) | 99.5% | 11 (9–13.5) | 100% / 9 |
| Half as large | 73.5% | 17 (14–20) | 90.5% / 13 |
The first rule speaks up a few weeks sooner, but, as below, it sends about 10× more notes to families with nothing to worry about. Smaller changes take longer to notice, and some aren't noticed within six months.
A sudden slowdown in talking speed
A big drop in pace shouldn't wait weeks for the weekly score. After every call, Remi also checks talking speed on its own: if it is far below his usual on 3 chats in a row, the family gets a check-in right away. Requiring a few chats in a row keeps one tired or noisy call from setting it off.
Pace on each chat (words/min) Check-in sent His usual range Much slower than usual
| Day | Words per minute | State |
|---|---|---|
| 2 | 146 | |
| 5 | 145 | |
| 7 | 144 | learning his usual |
| 9 | 144 | learning his usual |
| 11 | 145 | learning his usual |
| 16 | 133 | learning his usual |
| 18 | 153 | learning his usual |
| 19 | 142 | learning his usual |
| 21 | 138 | learning his usual |
| 23 | 124 | learning his usual |
| 26 | 131 | |
| 28 | 142 | |
| 32 | 140 | |
| 33 | 141 | |
| 35 | 134 | |
| 37 | 141 | |
| 39 | 147 | |
| 40 | 144 | |
| 42 | 105 | |
| 44 | 107 | |
| 46 | 104 | check-in sent |
| 47 | 113 | |
| 49 | 130 | |
| 51 | 107 | |
| 53 | 105 | |
| 54 | 111 | |
| 60 | 114 | |
| 61 | 107 | |
| 63 | 110 | |
| 65 | 97 | |
| 67 | 99 | |
| 68 | 107 | |
| 70 | 121 | |
| 72 | 117 | |
| 74 | 119 | |
| 75 | 102 | |
| 77 | 136 | |
| 79 | 119 | |
| 81 | 113 | |
| 82 | 126 | check-in sent |
91%
of 30-words-a-minute slowdowns caught, typically after 3 chats (5 days)
1 in 5 yrs
an unneeded check-in for a steady person
2%
of low-mood fortnights set it off (they get the wellbeing card)
| Size of slowdown | Caught | Typical chats until check-in | Weekly score alone would have noticed |
|---|---|---|---|
| 20 words/min slower | 68% | 4 (7 days) | 12% |
| 30 words/min slower | 91% | 3 (5 days) | 17% |
| 40 words/min slower | 98% | 3 (4 days) | 21% |
The weekly score blends seven markers, so one marker changing on its own barely moves it; that's why the pace check runs separately. Readings are computed from each chat's transcript within about 30 seconds of hanging up.
Three people side by side (8 weeks, compressed)
Weekly score Learning his usual Smoothed trend His usual range Note line (1.5)
Cognitive-decline pattern
| Week | Weekly score | Smoothed | State |
|---|---|---|---|
| 1 | +0.5 | +0.0 | learning his usual |
| 2 | +0.4 | +0.0 | learning his usual |
| 3 | -0.1 | +0.3 | learning his usual |
| 4 | -0.2 | +0.1 | steady |
| 5 | +0.7 | +0.5 | steady |
| 6 | +2.4 | +1.2 | watching |
| 7 | +3.0 | +1.8 | watching |
| 8 | +3.0 | +2.5 | family note |
Family note in week 8
Steady
| Week | Weekly score | Smoothed | State |
|---|---|---|---|
| 1 | -0.6 | +0.0 | learning his usual |
| 2 | +0.9 | +0.0 | learning his usual |
| 3 | -0.2 | +0.0 | learning his usual |
| 4 | -0.2 | +0.1 | learning his usual |
| 5 | -0.2 | -0.1 | steady |
| 6 | +0.0 | -0.1 | steady |
| 7 | +1.2 | +0.7 | steady |
| 8 | -1.0 | -0.4 | steady |
No note: stayed within his usual
Two low weeks
| Week | Weekly score | Smoothed | State |
|---|---|---|---|
| 1 | -1.2 | +0.0 | learning his usual |
| 2 | -0.2 | +0.0 | learning his usual |
| 3 | +0.9 | +0.1 | learning his usual |
| 4 | -0.2 | -0.0 | steady |
| 5 | -0.6 | -0.2 | steady |
| 6 | -0.6 | -0.3 | steady |
| 7 | -0.2 | -0.2 | steady |
| 8 | -1.0 | -0.7 | steady |
No check-up note; wellbeing card from week 4
Compared with simpler rules
| Rule | Decline pattern noticed (8 wk) | Unneeded notes / year | Low mood mistaken for change |
|---|---|---|---|
Alert on any unusual call Any single chat with any marker 2 spreads from his usual | 100% | 74.3 | 100% |
Our first rule 2 or more kinds of markers above 1.5, two weeks in a row | 93.5% | 1.5 | 22% |
Remi’s rule Weekly score ≥ 1.5 in 3 of the last 4 weeks, still rising, and mood not low | 92% | 0.16 | 0% |
Unneeded notes are counted over full simulated years of the steady person. Alerting on single calls would message a healthy family more than once a week, which is why Remi waits for weeks of evidence.
Every setting is a trade-off
| Setting | Caught by week 8 | Median wait (weeks) | Unneeded notes per year | Mood dip read as change |
|---|---|---|---|---|
| default (8-call baseline, skip 2, ≥1.5 in 3 of 4) | 92% | 5 | 0.16 | 0% |
| threshold 1.25 | 95.5% | 5 | 0.31 | 0% |
| threshold 2.0 | 80% | 5 | 0.06 | 0% |
| 2 of 4 weeks | 100% | 4 | 0.795 | 0% |
| 2 of 3 weeks | 100% | 4 | 0.765 | 0% |
| 6-call baseline | 92.5% | 5 | 0.33 | 0% |
| 12-call baseline | 77% | 6 | 0.095 | 0% |
| no trend requirement | 94% | 5 | 0.39 | 0% |
| 6-call baseline, 2 of 3 weeks, threshold 1.75 | 98% | 4 | 0.8 | 1% |
Each dot is a variation of the rule (8-week runs). Up and to the left is better. A stricter line or a longer baseline sends fewer unneeded notes but notices less; "2 of 4 weeks" notices more but writes to a steady family almost once a year. Hover a dot for its numbers.
Checked on generated conversations
Simulated: Demo data created for this presentation, not a real call.Separately, a synthetic transcript generator wrote four weeks of chats for a fictional person, with pauses that get longer each week (about 600 → 1,400 ms) and stories retold two weeks later. We ran the real marker code on those transcripts.
Average pause length (ms)
Drift Index per chat
Retold stories counted
| Chat | Week | Pause length (ms) | Drift Index | Retold within 7 days | Retold, any earlier chat |
|---|---|---|---|---|---|
| 1 | 1 | 620 | 0.3 | 0 | 0 |
| 2 | 1 | 590 | -0.5 | 0 | 0 |
| 3 | 1 | 624 | 0.2 | 0 | 0 |
| 4 | 2 | 848 | 3.6 | 0 | 0 |
| 5 | 2 | 853 | 4.8 | 0 | 0 |
| 6 | 2 | 860 | 2.7 | 0 | 0 |
| 7 | 3 | 1116 | 4.9 | 0 | 2 |
| 8 | 3 | 1102 | 4.6 | 0 | 2 |
| 9 | 3 | 1099 | 3.5 | 0 | 1 |
| 10 | 4 | 1408 | 3.7 | 0 | 1 |
| 11 | 4 | 1398 | 5.4 | 0 | 1 |
| 12 | 4 | 1377 | 4.1 | 0 | 2 |
Retold, within 7 days (current counter) Retold, any earlier chat
- The marker code measured the planted pauses exactly. Pace barely moves and there are no fillers, because the generator only varies pauses.
- Remi's rule: note in week 3, using a shortened setup for four weeks of data (week 1 as the baseline, 2 of 3 weeks).
- It found a gap to fix: every retelling comes 14 days after the first, but the repeat counter looks back only 7 days, so it counted none.