WEBVTT

NOTE week08-clip1-timeline-tokens
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 --> 00:00:04.186
A hospital record is a timeline
of events at irregular times.

2
00:00:04.186 --> 00:00:08.406
Our synthetic patient arrives
in the emergency department,

3
00:00:08.406 --> 00:00:13.817
has a lactate of 4.1, gets
norepinephrine, is diagnosed with sepsis,

4
00:00:13.817 --> 00:00:18.017
and moves to the ICU three
and a half hours later.

5
00:00:18.217 --> 00:00:23.079
Codes become tokens. Ethos
splits each code into parts:

6
00:00:23.079 --> 00:00:27.604
the sepsis code A41.9 becomes two tokens,

7
00:00:27.604 --> 00:00:31.099
and the drug code for
norepinephrine becomes three,

8
00:00:31.099 --> 00:00:35.017
from the broad group down to the drug.

9
00:00:35.217 --> 00:00:37.538
Numbers become quantile tokens.

10
00:00:37.538 --> 00:00:43.084
The lactate values in the training data
are cut into ten bins with equal counts.

11
00:00:43.084 --> 00:00:50.617
A lactate of 4.1 falls in the top
bin, so it becomes lactate, then Q10.

12
00:00:50.817 --> 00:00:54.824
Time gaps become interval
tokens. Ethos uses 13 of them,

13
00:00:54.824 --> 00:00:57.765
and adds none for gaps under 5 minutes.

14
00:00:57.765 --> 00:01:01.917
Our 3 gaps of under an hour
become 15 minutes to one hour,

15
00:01:01.917 --> 00:01:05.550
and the last gap becomes 1 to 2 hours.

16
00:01:05.750 --> 00:01:10.867
Six static tokens come first, such
as sex and age in five-year bins.

17
00:01:10.867 --> 00:01:13.649
Then the events follow in time order,

18
00:01:13.649 --> 00:01:19.017
so the whole record becomes
one sequence of 19 tokens.

19
00:01:19.217 --> 00:01:23.069
A decoder transformer then
learns to predict the next token,

20
00:01:23.069 --> 00:01:26.126
exactly as in weeks three to five.

21
00:01:26.126 --> 00:01:31.374
Ethos was trained on 241,000
patients from MIMIC-IV,

22
00:01:31.374 --> 00:01:37.613
and it sees the most recent
2,048 tokens of each timeline.