WEBVTT

NOTE week08-clip3-irregular-signals
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 --> 00:00:03.646
In the ICU, measurements
arrive at irregular times.

2
00:00:03.646 --> 00:00:07.850
Heart rate is charted about every
hour, with a long gap overnight,

3
00:00:07.850 --> 00:00:12.350
and lactate is drawn only
three times in the whole day.

4
00:00:12.550 --> 00:00:16.087
The first option is to put
everything on an hourly grid.

5
00:00:16.087 --> 00:00:20.226
Empty hours are filled with
the last measured value,

6
00:00:20.226 --> 00:00:26.083
and a mask of ones and zeros records
which values were actually measured.

7
00:00:26.283 --> 00:00:30.136
The mask matters, because
missing values carry information.

8
00:00:30.136 --> 00:00:34.353
Nobody draws a lactate at two
in the morning without a reason.

9
00:00:34.353 --> 00:00:40.283
The mask lets the model use when a
test was ordered, not just its result.

10
00:00:40.483 --> 00:00:45.408
The second option is the one from clip
one: interval tokens between events.

11
00:00:45.408 --> 00:00:50.780
The overnight gap is no longer
missing data. It becomes a token,

12
00:00:50.780 --> 00:00:54.350
two to six hours, that the model can read.

13
00:00:54.550 --> 00:00:56.989
The third option is a time embedding,

14
00:00:56.989 --> 00:01:03.366
like the sine and cosine position waves
from week two, but measured in real hours.

15
00:01:03.366 --> 00:01:11.039
At two and a half hours, the eight-hour
wave gives 0.92 and minus 0.38,

16
00:01:11.039 --> 00:01:17.750
and the 24-hour wave gives 0.61 and 0.79.

17
00:01:17.950 --> 00:01:23.204
Dense signals such as the ECG are cut
into patches, exactly as in week six.

18
00:01:23.204 --> 00:01:29.140
A ten-second ECG at 500 samples per
second becomes 10 one-second patches

19
00:01:29.140 --> 00:01:32.150
of 500 samples each.

20
00:01:32.350 --> 00:01:37.344
So every clinical data type can
become tokens: codes, lab values,

21
00:01:37.344 --> 00:01:39.917
time, missing values, and signals. Each
choice decides what the model can see,

22
00:01:39.917 --> 00:01:45.547
and what it cannot.