WEBVTT

NOTE week13-clip3-teaming
NOTE Narration read by an AI-generated voice.
NOTE The clip draws its own caption along the bottom of the frame. These cues
NOTE sit there too, so that no part of the animation is ever covered.

1
00:00:00.150 --> 00:00:05.578
Put a clinician and a model together
and there are two obvious rules:

2
00:00:05.578 --> 00:00:08.946
flag if either flags,
or flag only if both do.

3
00:00:08.946 --> 00:00:14.696
Say the clinician catches 80%
of cases and the model 75.

4
00:00:14.696 --> 00:00:20.350
Which rule is better depends
on something nobody measures.

5
00:00:20.550 --> 00:00:26.847
Assume their errors are independent.
Of a hundred real cases, both catch 60,

6
00:00:26.847 --> 00:00:33.434
the clinician alone 20, the model
alone 15, and 5 are missed by both.

7
00:00:33.434 --> 00:00:40.283
Either-flags reaches 0.95,
and both-flags falls to 0.60.

8
00:00:40.483 --> 00:00:46.719
Now assume the model misses whatever
the clinician misses. Both catch 75,

9
00:00:46.719 --> 00:00:51.528
the clinician alone 5,
and 20 are missed by both.

10
00:00:51.528 --> 00:00:56.885
Either-flags falls to 0.80,
which is the clinician alone,

11
00:00:56.885 --> 00:01:04.076
so the model adds nothing. Both-flags
rises to 0.75. The same two sensitivities,

12
00:01:04.076 --> 00:01:06.350
opposite conclusions.

13
00:01:06.550 --> 00:01:08.673
And people are not rules.

14
00:01:08.673 --> 00:01:13.241
A confident score can talk a
clinician out of a correct judgement,

15
00:01:13.241 --> 00:01:19.125
and at one true case in every eight
alerts the rational response is

16
00:01:19.125 --> 00:01:23.300
to stop reading them. Goh and
colleagues measured exactly this:

17
00:01:23.300 --> 00:01:30.479
the model alone scored 92%, and
the physicians using it scored 76,

18
00:01:30.479 --> 00:01:33.417
against 74 without.

19
00:01:33.617 --> 00:01:39.176
Ambient intelligence watches the room
instead of waiting for someone to type.

20
00:01:39.176 --> 00:01:44.356
16 depth sensors detected hand
hygiene at 92.1% sensitivity,

21
00:01:44.356 --> 00:01:51.184
better than the human observers at
85.2. But that came from only 88 events,

22
00:01:51.184 --> 00:01:57.027
so read the interval. In an ICU, 7
sensors detected mobility at 87%,

23
00:01:57.027 --> 00:02:02.817
and counted the staff involved
correctly only 69% of the time.

24
00:02:03.017 --> 00:02:06.627
Sensing a care space is a
governance question first.

25
00:02:06.627 --> 00:02:10.935
Depth sensing keeps identifiable
images out of the room,

26
00:02:10.935 --> 00:02:16.014
which helps, but it does not settle
who consents, who sees the data,

27
00:02:16.014 --> 00:02:21.920
or how long it is kept, and the staff
are being observed too. Validate locally,

28
00:02:21.920 --> 00:02:27.713
watch it after go-live, and measure
the team rather than the model.