WEBVTT

1
00:00:00.240 --> 00:00:03.570
- Get ready because if your
doctor isn't using AI right now,

2
00:00:03.570 --> 00:00:04.890
they'll be using it soon.

3
00:00:04.890 --> 00:00:06.900
Today we're talking to
one of the researchers

4
00:00:06.900 --> 00:00:09.180
who's working to bring
artificial intelligence

5
00:00:09.180 --> 00:00:10.740
into the exam room.

6
00:00:10.740 --> 00:00:14.220
So how long until I can
stop going to the doctor

7
00:00:14.220 --> 00:00:16.200
because I can just talk to a robot?

8
00:00:16.200 --> 00:00:19.080
- Well, I think probably
most fundamentally,

9
00:00:19.080 --> 00:00:20.610
you won't want to stop going to the doctor

10
00:00:20.610 --> 00:00:21.870
'cause one of the most important things

11
00:00:21.870 --> 00:00:24.840
is that human connection
with your physician,

12
00:00:24.840 --> 00:00:29.190
but also I think in general,
we're a long way off from that.

13
00:00:29.190 --> 00:00:31.170
- [Thomas] NYUTron is
an AI language model.

14
00:00:31.170 --> 00:00:33.570
It reads the notes that
physicians are already taking

15
00:00:33.570 --> 00:00:35.010
and then makes predictions.

16
00:00:35.010 --> 00:00:38.220
A recent study found that
NYUTron can accurately estimate

17
00:00:38.220 --> 00:00:41.070
patient's risk of death,
their length of hospital stay,

18
00:00:41.070 --> 00:00:43.410
and a whole lot of other
important questions.

19
00:00:43.410 --> 00:00:46.320
- The solution we latched
onto was the physician notes.

20
00:00:46.320 --> 00:00:48.180
You know, when every interaction I have

21
00:00:48.180 --> 00:00:50.250
as a physician with a patient,

22
00:00:50.250 --> 00:00:52.470
I take notes on what I'm doing,

23
00:00:52.470 --> 00:00:55.590
and those notes get fed
into the NYUTron system,

24
00:00:55.590 --> 00:00:58.080
which then outputs all of
this useful information

25
00:00:58.080 --> 00:01:00.360
that the AI's determined based
on just reading what I wrote.

26
00:01:00.360 --> 00:01:01.920
So very much is a assistant

27
00:01:01.920 --> 00:01:04.110
that sits there and reads as I write,

28
00:01:04.110 --> 00:01:05.730
and it can tell me a bunch of information

29
00:01:05.730 --> 00:01:07.350
about the patient in front of me

30
00:01:07.350 --> 00:01:08.760
or tell the organization information

31
00:01:08.760 --> 00:01:10.230
about the patient in front of me.

32
00:01:10.230 --> 00:01:12.240
- [Thomas] NYUTron can predict
how likely patients are

33
00:01:12.240 --> 00:01:15.870
to readmit to a hospital
with 80% accuracy.

34
00:01:15.870 --> 00:01:18.510
But reading doctor's notes
is just the beginning

35
00:01:18.510 --> 00:01:20.010
for AI in medicine.

36
00:01:20.010 --> 00:01:22.440
And the funny thing is
when it comes to AI,

37
00:01:22.440 --> 00:01:25.590
Oermann is more excited
about the boring stuff.

38
00:01:25.590 --> 00:01:28.110
- To me as a doctor, the
most exciting uses of AI

39
00:01:28.110 --> 00:01:30.660
isn't as much for the clinical
component of my practice,

40
00:01:30.660 --> 00:01:32.520
but it's all the non-clinical elements,

41
00:01:32.520 --> 00:01:33.930
especially the ones that take me away

42
00:01:33.930 --> 00:01:36.150
from the clinical part of my practice.

43
00:01:36.150 --> 00:01:40.350
You know, it's the documentation,
the operational stuff,

44
00:01:40.350 --> 00:01:42.270
some of the quality controls.

45
00:01:42.270 --> 00:01:43.680
Everything that takes
me away from patients,

46
00:01:43.680 --> 00:01:46.170
to me, those are the
exciting applications of AI.

47
00:01:46.170 --> 00:01:48.870
- But AI isn't just doing
backroom clerical work.

48
00:01:48.870 --> 00:01:50.610
People are experimenting with having AI

49
00:01:50.610 --> 00:01:52.950
handle all kinds of medical tasks,

50
00:01:52.950 --> 00:01:56.280
including some that require
interpersonal communication.

51
00:01:56.280 --> 00:01:59.070
Earlier this year, a
therapy app called Koko

52
00:01:59.070 --> 00:02:02.430
tested a chatbot handling
users' mental health questions.

53
00:02:02.430 --> 00:02:06.540
In one recent study, chatGPT
did as well as doctors

54
00:02:06.540 --> 00:02:10.140
at diagnosing patients
in emergency situations.

55
00:02:10.140 --> 00:02:11.100
And last April,

56
00:02:11.100 --> 00:02:13.800
a study compared doctors
and chatGPT's answers

57
00:02:13.800 --> 00:02:15.480
to patients' medical questions.

58
00:02:15.480 --> 00:02:17.580
A panel rated chatGPT's answers

59
00:02:17.580 --> 00:02:20.730
as more accurate 79% of the time.

60
00:02:20.730 --> 00:02:23.640
But what's really crazy
is the AI's responses

61
00:02:23.640 --> 00:02:27.300
were rated as empathetic
10 times more often

62
00:02:27.300 --> 00:02:29.430
than the answers from real doctors.

63
00:02:29.430 --> 00:02:31.440
Is there any particular concern,

64
00:02:31.440 --> 00:02:33.210
like should people be
worried when they hear

65
00:02:33.210 --> 00:02:34.890
that AI is gonna be helping their doctor?

66
00:02:34.890 --> 00:02:37.920
- I think anyone who's
asked that factual questions

67
00:02:37.920 --> 00:02:40.770
is probably frequently
surprised at how much it knows,

68
00:02:40.770 --> 00:02:43.590
but also you only know
those things are true

69
00:02:43.590 --> 00:02:46.290
because you know it yourself
and are able to verify it.

70
00:02:46.290 --> 00:02:47.250
And so I think that question

71
00:02:47.250 --> 00:02:50.730
of certainly doctors working
with chatGPT in my mind

72
00:02:50.730 --> 00:02:53.580
have maybe a lower level of concern.

73
00:02:53.580 --> 00:02:55.560
'Cause ultimately physicians
they know medicine,

74
00:02:55.560 --> 00:02:56.970
they can understand that,

75
00:02:56.970 --> 00:03:00.030
hey, if it's saying
this, it's probably off.

76
00:03:00.030 --> 00:03:01.020
I think for me, the big concern

77
00:03:01.020 --> 00:03:03.600
is when patients go to
chatGPT and say something

78
00:03:03.600 --> 00:03:06.030
'cause as a patient you're in some sense

79
00:03:06.030 --> 00:03:09.090
a vulnerable population,
especially if you're sick.

80
00:03:09.090 --> 00:03:11.970
And if you don't have the
medical knowledge to verify it,

81
00:03:11.970 --> 00:03:15.120
misinformation's a major
issue with these models.

82
00:03:15.120 --> 00:03:17.130
- Let us know in the comments
how you feel about it.

83
00:03:17.130 --> 00:03:19.200
Do you want your doctor using AI?

84
00:03:19.200 --> 00:03:22.233
Because like it or not, they'll
probably be doing it soon.