1
00:00:06,547 -->
00:00:08,977
Krishna Gade: good morning,
or good afternoon, wherever
2
00:00:08,977 -->
00:00:09,997
you're joining us from.
3
00:00:11,527 -->
00:00:17,417
Uh, thank you for joining this AI
explained with, uh, Dr. Girish Nadkarni.
4
00:00:17,601 -->
00:00:20,331
So Girish, uh, you know,
thanks for joining.
5
00:00:20,361 -->
00:00:25,131
Uh, just like from, you know, from an like
audience perspective, could you just give
6
00:00:25,131 -->
00:00:31,081
a brief intro about yourself, you know,
what you do at, you know, uh, Mount Sinai
7
00:00:31,081 -->
00:00:32,421
and you know, your role in terms of AI.
8
00:00:33,636 -->
00:00:34,026
Dr. Girish N. Nadkarni: Yeah.
9
00:00:34,026 -->
00:00:36,216
Um, uh, absolutely.
10
00:00:36,216 -->
00:00:38,716
So, uh, my name is Girish Nadkarni.
11
00:00:38,736 -->
00:00:42,906
I'm a clinician, uh, but also
an informaticist/AI scientist.
12
00:00:42,906 -->
00:00:47,076
I've been working in the field of
applied AI for over a decade now.
13
00:00:47,436 -->
00:00:49,896
Uh, my specific interests are a few.
14
00:00:50,346 -->
00:00:54,981
Um, one, um, is the safe, effective
and ethical clinical implementation
15
00:00:54,981 -->
00:00:58,281
and scaling of AI across health
systems and sort of the best
16
00:00:58,281 -->
00:01:00,801
governance strategies to do that.
17
00:01:00,981 -->
00:01:02,391
You know, uh, that's one.
18
00:01:02,451 -->
00:01:05,361
Um, the second is, uh, uh.
19
00:01:05,976 -->
00:01:10,356
Particularly around LLM sort of
assurance testing of large language
20
00:01:10,356 -->
00:01:13,596
models to make sure that they align
with like best clinical practices
21
00:01:13,596 -->
00:01:14,886
and recommendations, right.
22
00:01:15,216 -->
00:01:22,656
Uh, the third is, uh, sort of the agent
take experience of AI as applied to
23
00:01:22,656 -->
00:01:26,256
medicine in the terms of things uh,
and we've talked about this question at
24
00:01:26,256 -->
00:01:30,426
line about things that are, especially
for clinical decisions that are sort of
25
00:01:30,426 -->
00:01:35,796
non-critical, need to be made to be fast
and like are easily reversible, right?
26
00:01:36,006 -->
00:01:42,366
So in that role, um, I, uh, I'm the
chair of the Windreich Department of
27
00:01:42,546 -->
00:01:47,986
AI and Human Health at Mount Sinai,
and I also lead a research/clinical
28
00:01:48,006 -->
00:01:51,996
institute called Hasso Plattner Institute
of Digital Health at Mount Sinai.
29
00:01:52,326 -->
00:01:55,626
And I have a role on the health
system side of things as well,
30
00:01:55,836 -->
00:02:00,096
particularly around governance,
um, and clinical implementation.
31
00:02:00,426 -->
00:02:01,476
Um, so yeah.
32
00:02:01,596 -->
00:02:02,406
Um, and, uh.
33
00:02:03,241 -->
00:02:06,691
Um, I, uh, have in my other life
I've started some companies.
34
00:02:07,321 -->
00:02:07,831
Krishna Gade: That's great.
35
00:02:08,191 -->
00:02:08,401
Yeah.
36
00:02:08,401 -->
00:02:09,241
Thank you so much.
37
00:02:09,241 -->
00:02:12,841
Uh, you know, it's, it's a, it's very
rare to see someone with like, so
38
00:02:12,841 -->
00:02:17,286
much cross disciplinary experience
and background and, um, and so we are
39
00:02:17,286 -->
00:02:21,756
very excited to learn, uh, some of
the AI uh, applications in healthcare.
40
00:02:22,056 -->
00:02:25,836
So, you know, I guess like generative
AI has gained a lot of traction, right?
41
00:02:25,836 -->
00:02:29,016
You know, how is GenAI being
used today in clinical settings?
42
00:02:29,406 -->
00:02:32,886
Are there any specific examples that
you can walk us through, you know?
43
00:02:32,886 -->
00:02:33,096
Yeah.
44
00:02:33,186 -->
00:02:35,196
You know, like AI power
diagnostics or whatnot?
45
00:02:36,306 -->
00:02:36,666
Dr. Girish N. Nadkarni: Yeah.
46
00:02:36,666 -->
00:02:40,206
So, uh, well, so I think let's
just talk about this, right?
47
00:02:41,631 -->
00:02:48,081
Let's, let's differentiate clinical
settings from what actually happens when
48
00:02:48,081 -->
00:02:53,481
the patient interaction and patient facing
stuff and like all the rest of it, right?
49
00:02:53,811 -->
00:02:58,071
All the rest of it is probably like
85% of the whole enterprise, right?
50
00:02:58,301 -->
00:03:05,751
Things like billing, coding, um,
back office tasks, register creation,
51
00:03:06,081 -->
00:03:07,251
et cetera, et cetera, right?
52
00:03:07,731 -->
00:03:19,251
So I think there is an impetus on
generative AI to, um, uh, um, uh, push a
53
00:03:19,251 -->
00:03:24,471
lot of the tasks that initially required
human intervention or human labor,
54
00:03:25,161 -->
00:03:26,871
especially in the back office field.
55
00:03:27,231 -->
00:03:31,311
But then there is also now
slowly a trickle down effect
56
00:03:31,581 -->
00:03:33,451
into the clinical fields, right?
57
00:03:34,491 -->
00:03:35,841
So I'll give you an example.
58
00:03:35,841 -->
00:03:37,221
In the back office fields, right?
59
00:03:38,106 -->
00:03:42,756
So generative AI at Sinai, we
are using it, uh, for automating
60
00:03:42,756 -->
00:03:45,816
a lot of back office tasks, like
scheduling appointments, right?
61
00:03:46,176 -->
00:03:48,906
Bill better billing and coding,
better financial management,
62
00:03:49,176 -->
00:03:50,706
better insurance plan enrollment.
63
00:03:50,706 -->
00:03:52,686
And those things already exist, right?
64
00:03:53,016 -->
00:03:54,996
But slowly moving into
the clinical field, right?
65
00:03:54,996 -->
00:04:03,096
I think where the first big, uh, push will
be is, uh removing the back office, his
66
00:04:03,096 -->
00:04:04,986
task from the practice of medicine, right.
67
00:04:04,986 -->
00:04:08,706
So let, have you been to a
pain uh, doctor recently?
68
00:04:09,066 -->
00:04:10,386
Krishna Gade: Yeah, just like yesterday.
69
00:04:11,586 -->
00:04:12,066
Dr. Girish N. Nadkarni: Where was it?
70
00:04:12,186 -->
00:04:12,456
If I may.
71
00:04:12,456 -->
00:04:14,676
Krishna Gade: Kaiser,
Kaiser in the Bay Area yeah.
72
00:04:14,731 -->
00:04:15,906
Dr. Girish N. Nadkarni:
How, how long was the visit?
73
00:04:16,986 -->
00:04:20,166
Krishna Gade: It, you know, it took
like a, an hour like end-to-end, right?
74
00:04:20,166 -->
00:04:21,306
Like, you know, basically.
75
00:04:21,306 -->
00:04:21,366
Dr. Girish N. Nadkarni: Yeah.
76
00:04:21,366 -->
00:04:22,536
But how much time did the doctor spend?
77
00:04:22,536 -->
00:04:24,006
Krishna Gade: Well, the, the
doctor probably spends like
78
00:04:24,006 -->
00:04:25,206
10, 15 minutes right like.
79
00:04:25,566 -->
00:04:27,846
Dr. Girish N. Nadkarni: In that 10, 15
minutes, how much time did the doctor
80
00:04:27,846 -->
00:04:29,736
spend looking at you versus the computer?
81
00:04:30,681 -->
00:04:32,151
Krishna Gade: Maybe like
less than that right?
82
00:04:32,151 -->
00:04:33,411
Like five minutes, 10 minutes.
83
00:04:33,411 -->
00:04:33,681
Yeah.
84
00:04:34,011 -->
00:04:34,161
Dr. Girish N. Nadkarni: Yeah.
85
00:04:34,161 -->
00:04:37,881
So I mean, is it fair to say that,
you know, in the 15 minutes again
86
00:04:37,881 -->
00:04:41,871
of a hour long thing, which should
have been ideally like 20 minutes?
87
00:04:41,931 -->
00:04:42,201
Krishna Gade: Yeah.
88
00:04:42,651 -->
00:04:44,281
Dr. Girish N. Nadkarni:
Uh, the physician spend...
89
00:04:45,456 -->
00:04:46,436
Krishna Gade: Half of the time.
90
00:04:46,606 -->
00:04:46,696
Yeah.
91
00:04:46,696 -->
00:04:47,766
Dr. Girish N. Nadkarni:
Five minutes looking at you,
92
00:04:48,036 -->
00:04:48,756
Krishna Gade: Right absolutely.
93
00:04:48,756 -->
00:04:49,026
Yeah.
94
00:04:49,296 -->
00:04:51,096
Dr. Girish N. Nadkarni: So I
mean, that's where ambient AI
95
00:04:51,096 -->
00:04:52,386
and generative AI comes in.
96
00:04:52,386 -->
00:04:52,566
Right?
97
00:04:52,566 -->
00:04:55,416
So if you think about it, right,
uh, you could, you, like you
98
00:04:55,416 -->
00:04:56,706
and I are talking now, right?
99
00:04:56,771 -->
00:05:01,896
We could, we could have a conversation
and you know, ambient AI and now it's a
100
00:05:01,896 -->
00:05:07,026
bunch of companies, Abridge, Suki, DAX
could like listen to a conversation.
101
00:05:07,896 -->
00:05:14,331
Commoditized hardware and uh,
uh, uh, um, then, um, generate a
102
00:05:14,331 -->
00:05:15,711
billable note for the physician.
103
00:05:15,711 -->
00:05:21,231
So doctors sort of become doctors and talk
to the patient and make decisions rather
104
00:05:21,231 -->
00:05:23,051
than being a data entry clerk, right?
105
00:05:23,501 -->
00:05:28,941
Um, so I think that's one area where
there's an infection between back
106
00:05:28,941 -->
00:05:33,621
office tasks and, uh, um, uh, and, um.
107
00:05:33,966 -->
00:05:36,756
Clinical medicine, and I think
that's gonna see like broad
108
00:05:36,816 -->
00:05:40,116
adoption because physicians are
also tired of typing, right?
109
00:05:40,516 -->
00:05:41,626
Uh, yeah.
110
00:05:41,646 -->
00:05:47,346
You know, uh, and I think even if,
uh, there is, uh, not an increase in
111
00:05:47,346 -->
00:05:50,676
productivity, there's gonna be increase
in satisfaction amongst physicians
112
00:05:50,676 -->
00:05:55,356
because they no longer having, uh,
they don't, no longer having, uh,
113
00:05:55,746 -->
00:06:02,976
uh, uh, um, uh, uh, they no longer
have, uh, uh, the, the drudgery of.
114
00:06:04,071 -->
00:06:06,591
Still listening to you and then
also typing into a note, right?
115
00:06:06,591 -->
00:06:09,561
So I would argue that some
of their cognitive, uh.
116
00:06:12,216 -->
00:06:15,726
Krishna Gade: And there are also more
important, uh, serious use cases, right?
117
00:06:15,726 -->
00:06:18,786
Like risk prediction and, you
know, sepsis detection and how is
118
00:06:18,786 -->
00:06:20,436
like AI being used in those cases?
119
00:06:20,436 -->
00:06:21,066
Like where, you know,
120
00:06:21,246 -->
00:06:24,396
Dr. Girish N. Nadkarni: I think that's
a whole new area, predictive AI right.
121
00:06:24,396 -->
00:06:27,686
Which been around longer, in
my opinion, than generative AI.
122
00:06:27,686 -->
00:06:31,056
So we have more experience
with it, and that is where I
123
00:06:31,056 -->
00:06:34,086
think, you know, we need, um.
124
00:06:34,161 -->
00:06:38,831
To think a little bit, sort of deeply
around the risks surrounding it.
125
00:06:38,831 -->
00:06:39,071
Right?
126
00:06:39,281 -->
00:06:43,371
So yeah, we deploy a lot of AI into
clinical care, but I think that
127
00:06:43,371 -->
00:06:47,841
like the decision to not deploy
something is equally important as
128
00:06:47,841 -->
00:06:49,231
a decision to deploy it, right?
129
00:06:49,231 -->
00:06:50,661
What do I mean by that?
130
00:06:51,561 -->
00:06:55,701
If I am physician, I make a mistake
just because of the bandwidth
131
00:06:56,031 -->
00:06:58,281
and like the number of patients I
see is gonna be limited to like.
132
00:06:59,376 -->
00:07:01,106
a patient or five
patients the most, right.
133
00:07:01,776 -->
00:07:03,036
The algorithm is inaccurate.
134
00:07:03,456 -->
00:07:11,286
Who, uh, uh, uh, uh, uh, uh, algorithm
is inaccurate then it scales to
135
00:07:11,286 -->
00:07:12,966
like tens of thousands of patients.
136
00:07:12,966 -->
00:07:13,226
Right?
137
00:07:13,546 -->
00:07:18,801
Um, I think that that is where you sort
of need assurance, you need monitoring,
138
00:07:18,801 -->
00:07:20,541
you need privacy, and you need security.
139
00:07:20,781 -->
00:07:25,281
And I really, you need one level about
that, which is sort of randomized
140
00:07:25,341 -->
00:07:29,751
testing or some sort of AI assurance
where you put things through like an
141
00:07:29,751 -->
00:07:34,491
A/B test, obviously, uh, with highest
levels of patient security, et cetera.
142
00:07:34,761 -->
00:07:37,071
And then, uh, you put
that into clinical care.
143
00:07:37,311 -->
00:07:40,621
Now that being said, at Mount Sinai,
we've deployed several of these
144
00:07:40,621 -->
00:07:45,061
predictive algorithm for things
like worsening of kidney function,
145
00:07:45,066 -->
00:07:47,781
prediction of malfunction, prediction
of falls, et cetera, et cetera.
146
00:07:48,371 -->
00:07:53,131
Now, now, uh, you know, a lot of them
are monitored continuously to allow
147
00:07:53,131 -->
00:07:54,841
for drift, et cetera, et cetera.
148
00:07:55,141 -->
00:08:00,811
And, uh, uh, um, uh, we wanna make
sure that those, uh, are anything
149
00:08:00,811 -->
00:08:05,431
that sort of makes this patient as
not patient facing anything that
150
00:08:05,431 -->
00:08:09,421
impacts clinical care as we know it
is safe, effective, and ethical right.
151
00:08:09,421 -->
00:08:10,381
And we can have a long look on.
152
00:08:11,931 -->
00:08:13,701
Krishna Gade: So, I mean, of
course there's classical ML, as
153
00:08:13,701 -->
00:08:16,041
you said, the predictive AI that's
been around for a long time.
154
00:08:16,041 -->
00:08:19,341
And now there's generative AI, which you
talked about some really interesting use
155
00:08:19,341 -->
00:08:21,261
cases like ambient experience and whatnot.
156
00:08:21,531 -->
00:08:24,711
How are you seeing, seeing the
interplay of these two things, you
157
00:08:24,711 -->
00:08:27,891
know, how are they complimenting each
other in healthcare applications?
158
00:08:27,921 -->
00:08:29,571
Any, any interesting insights there?
159
00:08:30,231 -->
00:08:32,211
Dr. Girish N. Nadkarni: The
2, 2, 2 major insights, right.
160
00:08:33,051 -->
00:08:36,081
Um, so the first is generative AI.
161
00:08:36,966 -->
00:08:39,071
Can actually also be
used as predictive AI.
162
00:08:39,561 -->
00:08:44,886
It could be used as predictive AI in a
sense of you know, few short learning
163
00:08:44,886 -->
00:08:47,596
or small amounts of fine-tuning.
164
00:08:47,946 -->
00:08:50,286
That can actually improve prediction.
165
00:08:50,286 -->
00:08:50,586
Right.
166
00:08:51,156 -->
00:08:57,861
And it can be used as predictive
AI, especially in certain tasks
167
00:08:58,191 -->
00:09:00,381
where there's not much tabular data.
168
00:09:00,411 -->
00:09:02,841
My tabular data, I mean
labs or codes right.
169
00:09:03,021 -->
00:09:07,621
And where a lot of the prediction
thing is incorporated in notes.
170
00:09:08,241 -->
00:09:10,881
. Like examples would be like
mental health disorders, right?
171
00:09:11,531 -->
00:09:15,411
The example would be initial presentation
of a patient to the emergency room
172
00:09:15,651 -->
00:09:20,526
because you don't have a lot of like
work tabular information and then
173
00:09:20,526 -->
00:09:21,231
Krishna Gade: Right, right.
174
00:09:21,546 -->
00:09:22,296
Dr. Girish N. Nadkarni: Mostly notes.
175
00:09:22,296 -->
00:09:22,566
Right?
176
00:09:22,566 -->
00:09:23,436
Like what happened to you?
177
00:09:23,436 -->
00:09:23,526
Krishna Gade: Right.
178
00:09:23,646 -->
00:09:23,826
Correct.
179
00:09:24,346 -->
00:09:24,786
Dr. Girish N. Nadkarni: Proof.
180
00:09:25,086 -->
00:09:26,076
Um, right.
181
00:09:26,346 -->
00:09:28,296
So, so that's one thing, right?
182
00:09:28,806 -->
00:09:32,406
That's one thing where generative AI
is actually used as predictive AI.
183
00:09:32,646 -->
00:09:37,116
And then we, uh, uh, the risk of putting
my own harm, we actually have a paper
184
00:09:37,116 -->
00:09:44,511
on this in JAMIA where you actually
show that the, the, the accuracy, like
185
00:09:44,511 -->
00:09:45,851
we did it for emergency room, right?
186
00:09:46,101 -->
00:09:53,781
.
So the accuracy for uh, uh, uh,
uh, um, uh, generative AI when
187
00:09:54,021 -->
00:09:58,941
you are, uh, um, just feeding it
like 10 examples of emergency room
188
00:09:58,941 -->
00:10:01,161
cases can approach predictive AI.
189
00:10:01,161 -->
00:10:04,071
I mean, it's still lower, but I
mean, I'm assuming if we fed in
190
00:10:04,071 -->
00:10:06,531
more than 10 examples it would
be better because it scale.
191
00:10:06,621 -->
00:10:09,501
It can actually give you a reason
reasoning about why it's making
192
00:10:09,501 -->
00:10:14,676
the prediction, which traditional
ML cannot, right, like traditional
193
00:10:14,706 -->
00:10:15,576
ML, if you think about...
194
00:10:15,666 -->
00:10:17,646
Krishna Gade: You have to retrofit
explainability and whatnot,
195
00:10:17,646 -->
00:10:20,526
but in this case, LLMs can
potentially give a reason out.
196
00:10:20,526 -->
00:10:20,736
Yeah.
197
00:10:20,761 -->
00:10:21,001
Yeah.
198
00:10:21,306 -->
00:10:21,756
Dr. Girish N. Nadkarni: That's one.
199
00:10:22,386 -->
00:10:25,816
The second thing is when you can
work things in concert, right?
200
00:10:25,986 -->
00:10:30,206
And that's when you can think about
potential agents in medicine, right?
201
00:10:30,446 -->
00:10:35,646
So you have a predictive approach
that you know, I'm just gonna give
202
00:10:35,651 -->
00:10:36,726
you a clinical example, right?
203
00:10:36,876 -->
00:10:37,176
Krishna Gade: Sure.
204
00:10:37,581 -->
00:10:41,901
Dr. Girish N. Nadkarni: In the hospital,
who are the five, who are the 10%
205
00:10:41,901 -->
00:10:43,761
of people at most at risk of a fall?
206
00:10:44,586 -->
00:10:46,506
I'm just giving you an example, right?
207
00:10:46,806 -->
00:10:52,086
So, you know, you, you can have a
standard, say a random forest or
208
00:10:52,266 -->
00:10:56,346
some other tree-based algorithm uh,
do that and it can say it has an
209
00:10:56,436 -->
00:11:00,756
accuracy of like 95% and it can, I
had a predictive probability, uh,
210
00:11:00,936 -->
00:11:04,686
predictive, uh, uh, sorry, positive
predictive value of like 70%.
211
00:11:04,896 -->
00:11:08,666
So it, 10 patients, the positive
predict will be set differently.
212
00:11:08,946 -->
00:11:13,506
So instead of like having a physician or
a nurse go and examine those 10 patients,
213
00:11:13,746 -->
00:11:16,776
then automatically put a alert in their
chart, right, saying like, you know,
214
00:11:16,776 -->
00:11:21,161
using a generative AI like saying a
note that this patient has been, uh, uh
215
00:11:22,136 -->
00:11:24,476
flagged that high risk of a fall, right?
216
00:11:24,476 -->
00:11:24,536
Yeah.
217
00:11:24,746 -->
00:11:27,576
You know, next to the patient you to
be, be careful or something, right?
218
00:11:27,916 -->
00:11:32,486
So that's an example of where predictive
and generative way I can operate almost.
219
00:11:32,486 -->
00:11:35,061
Krishna Gade: You can take the predictive
outputs and summarize it for, you
220
00:11:35,061 -->
00:11:39,656
know, and, and for, you know, uh, you
know, doctors and you know nurses.
221
00:11:39,656 -->
00:11:40,226
That's amazing.
222
00:11:40,676 -->
00:11:44,246
Well, one of the biggest concerns
with gen AI is the whole concept of
223
00:11:44,276 -->
00:11:48,661
hallucinations, or what they call
faithfulness or in a groundedness, right?
224
00:11:49,061 -->
00:11:52,301
How do you, how do you ensure, how
do we ensure these AI outputs are
225
00:11:52,301 -->
00:11:53,921
accurate and clinically reliable?
226
00:11:55,421 -->
00:11:56,801
Dr. Girish N. Nadkarni: So, I
mean, you know, you are much more
227
00:11:56,801 -->
00:11:57,971
technical than me, Krishna, right?
228
00:11:57,971 -->
00:12:00,161
You tell me whether the
hallucinations can ever go away.
229
00:12:01,451 -->
00:12:04,961
Krishna Gade: I mean, see, I think,
um, models can never be perfect
230
00:12:04,961 -->
00:12:07,121
and that's what we've learned
even through the classical ml.
231
00:12:07,121 -->
00:12:10,331
So I think there will always
be, uh, issues with it.
232
00:12:10,541 -->
00:12:15,006
It's just a, it's just, I think it's like
more like a creative feature in the, in
233
00:12:15,006 -->
00:12:16,696
the whole, the way, the way the LLMs work.
234
00:12:16,716 -->
00:12:16,896
Right.
235
00:12:16,896 -->
00:12:18,816
So I don't think that
they will ever go away.
236
00:12:18,816 -->
00:12:21,816
But you, I think there are more and
more guardrails and more and more
237
00:12:21,816 -->
00:12:25,086
safety checks people are putting in to
ensure that is, that's not happening.
238
00:12:25,836 -->
00:12:26,076
Dr. Girish N. Nadkarni: Exactly.
239
00:12:26,076 -->
00:12:28,026
So, um, um, okay.
240
00:12:28,026 -->
00:12:29,076
So a few things, right?
241
00:12:29,076 -->
00:12:32,526
So like, so we agree that they will
never completely go away, right?
242
00:12:32,526 -->
00:12:35,316
Because almost a feature, not a bug.
243
00:12:35,526 -->
00:12:35,676
Krishna Gade: Correct.
244
00:12:36,216 -->
00:12:37,476
Dr. Girish N. Nadkarni: It's
a feature of the creativity.
245
00:12:37,476 -->
00:12:37,626
Krishna Gade: Correct.
246
00:12:39,381 -->
00:12:42,471
Dr. Girish N. Nadkarni: So you can
put guardrails around it in the form
247
00:12:42,531 -->
00:12:47,261
of like, you know, not hallucinate
about certain things, right?
248
00:12:47,761 -->
00:12:53,691
You can improve hallucinations with
RAG, basically like, uh, ragged to
249
00:12:53,691 -->
00:12:58,221
a vector database of institutional
policies or medical knowledge, uh,
250
00:12:58,281 -->
00:13:01,491
uh, or uh, uh, uh, what have you.
251
00:13:02,091 -->
00:13:06,001
Now we have some work coming around that,
that I would happy to share at some point
252
00:13:06,381 -->
00:13:10,581
where you, even if you do all of these
things, you know, you drag it and these
253
00:13:10,581 -->
00:13:12,141
are actually real world cases, right?
254
00:13:12,141 -->
00:13:16,251
You basically the, the experiment
was this, you basically take a
255
00:13:16,251 -->
00:13:20,451
thousand real world cases and you
just fabricate a disease name or
256
00:13:20,451 -->
00:13:22,671
fabricate a lab value out of this.
257
00:13:22,881 -->
00:13:27,531
And then you watch how the model,
uh, you know, response with hallis,
258
00:13:28,071 -->
00:13:30,861
further hallucination based upon
bad data, which just happens, right?
259
00:13:31,131 -->
00:13:32,421
Data corruption, et cetera.
260
00:13:32,751 -->
00:13:32,811
Um.
261
00:13:33,831 -->
00:13:37,101
So even if you ragged it, the
hallucination rates goes down, but it
262
00:13:37,101 -->
00:13:39,821
still stays fixed below a certain amount.
263
00:13:39,821 -->
00:13:40,031
Right.
264
00:13:40,571 -->
00:13:45,681
So that is why I think figuring out a
risk-based approach of what to automate
265
00:13:45,681 -->
00:13:50,091
and what needs to be oversight is
gonna be critical because, you know.
266
00:13:51,456 -->
00:13:55,096
If you hallucinate like a, like a
for a noncritical decision, right?
267
00:13:55,266 -->
00:13:59,616
You know, patient needs to go from
bed A to bed B. You know, it's like
268
00:14:00,006 -->
00:14:02,406
if you realize it's wrong after
two hours, doesn't really matter.
269
00:14:02,406 -->
00:14:02,676
Right?
270
00:14:03,156 -->
00:14:06,906
But if if you hallucinate that the patient
needs surgery, that's a big deal, right?
271
00:14:07,206 -->
00:14:09,486
That's why I think a
risk-based approach, like.
272
00:14:10,176 -->
00:14:15,906
Focus on this three parameters of like
how critical is a decision that needs
273
00:14:15,906 -->
00:14:20,256
to be made, how reversible it is, and
how fast it needs to be made actually
274
00:14:20,256 -->
00:14:22,366
helps with the AI governance, right?
275
00:14:22,656 -->
00:14:26,526
So I think, you know, if we agree
that hallucinations are never gonna
276
00:14:26,526 -->
00:14:30,336
go away, but we also agree that, you
know, the benefit of using generative
277
00:14:30,336 -->
00:14:32,646
AI in healthcare is, is important.
278
00:14:32,796 -->
00:14:36,396
Then we need to like both put
guardrails, technical guardrails,
279
00:14:36,396 -->
00:14:37,446
but also governance guardrails.
280
00:14:37,926 -->
00:14:38,226
Krishna Gade: Correct.
281
00:14:38,706 -->
00:14:38,886
Yeah.
282
00:14:38,886 -->
00:14:42,066
So you touched upon, uh, you know,
like this whole thing about continuous
283
00:14:42,066 -->
00:14:43,476
risk evaluation in some ways.
284
00:14:43,476 -->
00:14:46,776
You know, essentially you're
scoring these, uh, general apps.
285
00:14:47,226 -->
00:14:49,146
So, and then you touched
upon this governance aspect.
286
00:14:49,146 -->
00:14:52,416
Could you shed some light on like
the governance processes that happen
287
00:14:52,866 -->
00:14:56,226
around AI, you know, like for example,
in your organization or in general
288
00:14:56,226 -->
00:14:57,331
in healthcare space, you know?
289
00:14:57,666 -->
00:14:58,326
Dr. Girish N. Nadkarni: Yeah, absolutely.
290
00:14:58,326 -->
00:15:04,521
So that, uh, um, governance in AI is
evolving field right now, I'll say that.
291
00:15:04,881 -->
00:15:06,976
So I don't think I have
all of the answers.
292
00:15:07,346 -->
00:15:09,521
. But I have thoughts and
some principles, right?
293
00:15:09,841 -->
00:15:13,851
. So the first thing I think governance,
when you're setting governance policies,
294
00:15:14,301 -->
00:15:18,711
you have to be, make sure that you
have a diverse range of perspectives
295
00:15:18,711 -->
00:15:22,761
on board and in healthcare that means,
you know, people who provide the care,
296
00:15:23,241 -->
00:15:27,111
um, um, like nurses, physicians, MAs.
297
00:15:27,531 -->
00:15:32,106
Um, people who enable the care to be
provided, which ensures like takes uh,
298
00:15:32,346 -->
00:15:36,036
the back office and people to whom the
care is provided to basically the patient.
299
00:15:36,036 -->
00:15:38,946
So you need all of those three
perspectives on board 'cause they have
300
00:15:38,946 -->
00:15:40,566
different perspectives on everything.
301
00:15:40,776 -->
00:15:44,076
And then you can have like a almost a
review panel when you can have a majority
302
00:15:44,076 -->
00:15:45,516
ward for anything to pass through.
303
00:15:46,056 -->
00:15:49,686
But even before the things hit that
review panel, there needs to be some
304
00:15:50,586 -->
00:15:52,056
decision points or rules, right?
305
00:15:53,496 -->
00:15:56,736
First easy thing is like, is it
back office or is it patient facing?
306
00:15:56,736 -->
00:16:00,456
If it's back office, then I would argue,
yeah, you need some rigor, but you
307
00:16:00,456 -->
00:16:04,176
need a lower level of like rigor than
if it's actually impact clinical care.
308
00:16:04,176 -->
00:16:04,356
Right?
309
00:16:04,566 -->
00:16:07,146
And for if things that
impact clinical care, right?
310
00:16:07,266 -->
00:16:11,646
Um, you need to figure out if it's
safe, which requires a lot of like
311
00:16:11,646 -->
00:16:13,716
validation testing and assurance testing.
312
00:16:13,926 -->
00:16:14,616
Is it effective?
313
00:16:14,616 -->
00:16:17,586
Which ideally requires some
sort of A/B testing and which
314
00:16:17,976 -->
00:16:22,026
C3 is ethical, which basically
requires like monitoring for bias.
315
00:16:23,421 -->
00:16:24,351
Krishna Gade: Yeah, absolutely.
316
00:16:24,681 -->
00:16:27,651
And so like, so what should
be the best practices then?
317
00:16:27,651 -->
00:16:30,411
You know, let's say, you know, lots of
healthcare companies, we are working with
318
00:16:30,411 -->
00:16:34,791
a few, uh, they're trying to, you know,
integrate genAI into their workflows.
319
00:16:34,791 -->
00:16:37,761
Now, how should they go about
doing this in a responsible manner?
320
00:16:37,851 -->
00:16:40,191
You know, what are some of the
best practices do you recommend?
321
00:16:40,941 -->
00:16:42,946
Dr. Girish N. Nadkarni: Uh,
ag, again, like this honor.
322
00:16:43,446 -->
00:16:46,236
Um, um, these are my
recommendations, right?
323
00:16:46,926 -->
00:16:50,166
I just say that not my institution's
recommendations, you know, because I'm
324
00:16:50,166 -->
00:16:51,696
one part of the whole purpose, right?
325
00:16:51,696 -->
00:16:55,656
So I think like, first thing
that the best practice is like
326
00:16:55,656 -->
00:16:57,006
clarity of purpose, right?
327
00:16:57,996 -->
00:16:59,796
What problem are you
actually trying to solve?
328
00:16:59,796 -->
00:17:06,426
I mean, there is a wish right now to
appear cool or like to appear like that.
329
00:17:06,636 -->
00:17:11,016
We are doing stuff around AI and we
are like this, but like, okay, and that
330
00:17:11,016 -->
00:17:14,736
sounds fine, but if you're actually
implementing AI without a clear
331
00:17:14,946 -->
00:17:17,706
organizational strategy or a clear.
332
00:17:18,381 -->
00:17:21,021
Solving of a clinical or
operational challenge, right?
333
00:17:21,261 -->
00:17:25,971
Like prioritizing emergency rooms,
clinical decision report, patient flow
334
00:17:25,971 -->
00:17:30,181
optimization, where generative AI has
been shown to concretely add value, right?
335
00:17:30,641 -->
00:17:33,831
The second thing is, you know,
a risk-based approach like
336
00:17:33,831 -->
00:17:34,941
we just talked about, right?
337
00:17:34,941 -->
00:17:41,946
What based upon the metrics or
rubrics of um um, um, reversibility,
338
00:17:42,336 -->
00:17:47,406
um, criticality and, uh, uh, uh, how
fast you make to make the decision.
339
00:17:47,586 -->
00:17:50,166
What's like the level of
oversight that you need, right?
340
00:17:50,166 -->
00:17:54,876
Like, you know, for everything, every,
like if you are a higher criticality,
341
00:17:55,146 -->
00:17:57,936
but also lower reversibility, then
you need physicians, maybe two
342
00:17:57,936 -->
00:17:59,186
physicians to look at it, right?
343
00:17:59,896 -->
00:18:03,891
Then, so you need like clear oversight
through well-defined escalation protocols.
344
00:18:03,891 -->
00:18:05,221
And the third thing...
345
00:18:05,221 -->
00:18:08,161
Krishna Gade: So reversibility from an
AI decision or like reversibility of
346
00:18:08,161 -->
00:18:09,811
human interpretation of that decision.
347
00:18:09,811 -->
00:18:10,531
Like what, what is that
348
00:18:10,531 -->
00:18:12,181
Dr. Girish N. Nadkarni: Reversibility
of the decision themselves
349
00:18:12,301 -->
00:18:13,561
Krishna Gade: Reversibility
of decision themselves.
350
00:18:13,566 -->
00:18:14,006
Okay, okay.
351
00:18:14,006 -->
00:18:16,081
Dr. Girish N. Nadkarni: Like,
like is the this a door that you
352
00:18:16,081 -->
00:18:17,741
cannot reverse, right, for example.
353
00:18:18,171 -->
00:18:21,571
If you do a wrong surgery on a patient
that you can not reverse but if you
354
00:18:22,311 -->
00:18:25,556
send a patient from an emergency room to
the floor, you can easily reverse that.
355
00:18:26,636 -->
00:18:28,666
Mean, there's bunch of
examples around that.
356
00:18:28,666 -->
00:18:34,986
And then, um, um, um, the, the,
the third thing is training.
357
00:18:34,986 -->
00:18:40,176
Right, right now, um, you know,
it's, it's, it's not a new field,
358
00:18:40,416 -->
00:18:44,376
but it's also a field that most
people who make these decisions are
359
00:18:44,376 -->
00:18:47,526
not custom or not used to right.
360
00:18:47,526 -->
00:18:52,746
So I think that needs to be sort of large
scale training across health systems
361
00:18:53,016 -->
00:18:58,341
and sort of training to tailor to what
your job is right now, but how your
362
00:18:58,341 -->
00:19:00,711
job could evolve based upon this right.
363
00:19:00,891 -->
00:19:04,041
And that training is gonna
be different for providers.
364
00:19:04,101 -->
00:19:06,681
It's gonna be different
for physician, uh, nurses.
365
00:19:06,741 -->
00:19:09,561
It's gonna be different
for, uh, uh, other people.
366
00:19:09,561 -->
00:19:09,831
Right.
367
00:19:10,101 -->
00:19:14,931
And then finally sort of build, uh,
some sort of a feedback mechanism.
368
00:19:14,931 -->
00:19:18,111
And that's a bigger conversation
that we can have, right?
369
00:19:18,111 -->
00:19:18,171
Yeah.
370
00:19:18,531 -->
00:19:24,801
So the feedback mechanism is, can
involve, again, on a risk-based basis,
371
00:19:25,131 -->
00:19:29,636
things from clinical trials to long term
everything should be monitored long term.
372
00:19:29,636 -->
00:19:29,916
Sorry.
373
00:19:30,656 -->
00:19:35,136
Um, but, you know, basically continually
measure outcomes, refine the models, get
374
00:19:35,136 -->
00:19:38,586
human feedback, and ensure that any errors
are identified and corrected basically.
375
00:19:38,946 -->
00:19:39,666
Krishna Gade: Right, right.
376
00:19:40,026 -->
00:19:43,476
So, yeah, so this comes back to
like assessing the risks upfront.
377
00:19:43,476 -->
00:19:47,016
You know, which use case you wanna
apply it or not, and then evaluating
378
00:19:47,016 -->
00:19:50,136
it and properly setting up all the
continuous oversight and monitoring,
379
00:19:50,256 -->
00:19:51,996
uh, to, to make sure it works.
380
00:19:52,416 -->
00:19:52,956
That's great.
381
00:19:53,166 -->
00:19:54,546
Now switching gears a little bit.
382
00:19:54,546 -->
00:19:54,786
Right.
383
00:19:54,786 -->
00:19:54,996
So.
384
00:19:56,271 -->
00:19:59,781
AI, you know, is, is seems to be
helping quite a bit in this, you
385
00:19:59,781 -->
00:20:03,261
know, discovering new antibiotics
and assisting in drug discovery.
386
00:20:03,711 -->
00:20:07,731
How do you see like genAI shaping the
future of like, say precision medicine
387
00:20:07,731 -->
00:20:09,111
and, you know, treatment innovations?
388
00:20:09,801 -->
00:20:11,331
Dr. Girish N. Nadkarni: So
that's a great question, right?
389
00:20:11,391 -->
00:20:14,611
And I think that it's sort of a full
stack innovation for everything, right?
390
00:20:15,031 -->
00:20:19,791
. Um, like if you think about what
precision medicine is, right?
391
00:20:20,181 -->
00:20:23,661
Is it's proactive rather than reactive.
392
00:20:24,456 -->
00:20:26,496
It's personalized rather than public.
393
00:20:27,186 -->
00:20:30,606
And it's predictive rather
than reductive, right?
394
00:20:30,996 -->
00:20:37,236
So if take those three things right so
you can know which patients are going
395
00:20:37,236 -->
00:20:42,891
to get sick or which patients are gonna
have an incidence of disease based upon
396
00:20:43,131 -->
00:20:48,111
a combination of several data points,
starting off with their biological data,
397
00:20:48,111 -->
00:20:53,481
which genome expose, but, uh, sorry,
genome, um, uh, proteome, et cetera.
398
00:20:53,751 -->
00:20:56,991
But also like the clinical data,
their environmental data, right?
399
00:20:57,261 -->
00:20:59,076
So that's the predictive part of it.
400
00:20:59,096 -->
00:20:59,316
So.
401
00:21:00,366 -->
00:21:04,776
that's the, you can't just do,
and then you proactively try to
402
00:21:04,776 -->
00:21:08,526
prevent that disease from happening
by personalizing therapy, right?
403
00:21:08,766 -->
00:21:16,041
It's not a fast stretch to think about two
things that are happening concurrently.
404
00:21:16,041 -->
00:21:16,281
Right.
405
00:21:16,281 -->
00:21:23,571
One is the, the increasing personalization
in biotech with like mRNA and, uh,
406
00:21:23,631 -->
00:21:26,171
gene therapies and even gene editing.
407
00:21:26,691 -->
00:21:30,861
But also, um, this AI
development of faster approaches
408
00:21:30,861 -->
00:21:31,731
to do these things, right?
409
00:21:31,731 -->
00:21:36,531
So in the future, and again, it's not
possible, it's, it's definitely possible
410
00:21:36,981 -->
00:21:39,021
to have like a Coke machine, right?
411
00:21:39,021 -->
00:21:43,401
Basically you enter in a person's
information and you press whether
412
00:21:43,401 -->
00:21:47,031
you want like regular Coke or Diet
Coke or vanilla, which vanilla Coke
413
00:21:47,031 -->
00:21:49,191
is terrible by the way uh, Coke.
414
00:21:49,401 -->
00:21:54,111
And you get a drug or a
molecule personalized that
415
00:21:54,111 -->
00:21:55,311
individual patient, right?
416
00:21:55,311 -->
00:21:59,451
I mean, it sounds science fiction, but
if you think about the component parts
417
00:21:59,451 -->
00:22:01,881
of it, right, it's not right, right.
418
00:22:01,881 -->
00:22:04,191
You need, uh, you need, uh,
419
00:22:06,291 -->
00:22:07,791
you need accurate prediction.
420
00:22:07,791 -->
00:22:10,561
This person's gonna get sick because you
don't wanna give medications to someone.
421
00:22:11,021 -->
00:22:19,026
. But then you can have tailored like
tailored drugs to particular genes that
422
00:22:19,026 -->
00:22:22,626
like decrease your risk of cardiovascular
disease, et cetera, et cetera.
423
00:22:23,076 -->
00:22:24,666
And then you need a delivery mechanism.
424
00:22:24,666 -->
00:22:27,036
And a lot of these delivery
mechanisms are becoming oral, right?
425
00:22:27,036 -->
00:22:30,606
So you can have, uh, mRNA
medication in an oral form
426
00:22:30,606 -->
00:22:31,926
that's specifically made for you.
427
00:22:32,831 -->
00:22:34,416
You tell me that this is science fiction.
428
00:22:34,836 -->
00:22:39,006
It's not because having
personalized medicines created.
429
00:22:39,846 -->
00:22:43,631
For rare diseases like you know,
with names, the patient on,
430
00:22:44,106 -->
00:22:47,346
and these are, these are, these
are severe diseases, right.
431
00:22:47,376 -->
00:22:50,436
For example, a girl had like
spinal muscular dystrophy
432
00:22:50,586 -->
00:22:51,636
due to a mutation in a gene.
433
00:22:51,636 -->
00:22:54,636
A drug was created specifically
for her, and she does well now.
434
00:22:54,636 -->
00:22:54,846
Right.
435
00:22:55,176 -->
00:22:58,226
I'm just telling you that this
is possible to scale, right?
436
00:22:58,226 -->
00:23:02,616
With AI, because you can have, you can
screen lots of combinations quickly and
437
00:23:02,616 -->
00:23:08,286
you can rapidly generate any trade novel
molecular design for a particular patient.
438
00:23:08,781 -->
00:23:09,351
Krishna Gade: That's amazing.
439
00:23:09,351 -->
00:23:13,701
So what you're saying is personalized,
completely personalized medicine for some
440
00:23:13,701 -->
00:23:19,121
rare diseases based on your DNA, your sort
of molecular structure, you can personally
441
00:23:19,151 -->
00:23:21,291
Dr. Girish N. Nadkarni: Not just rare,
I would argue, common diseases as well.
442
00:23:21,291 -->
00:23:23,231
Krishna Gade: Common diseases
as well and personalized to you.
443
00:23:23,231 -->
00:23:24,156
Dr. Girish N. Nadkarni:
Trying to prevent diseases.
444
00:23:24,156 -->
00:23:24,316
Right.
445
00:23:24,916 -->
00:23:27,771
Personalized medications
for preventing diseases.
446
00:23:27,771 -->
00:23:28,431
Krishna Gade: Preventing decisions.
447
00:23:29,361 -->
00:23:32,691
Dr. Girish N. Nadkarni: Yeah,
like, you know, I right now, like
448
00:23:32,691 -->
00:23:35,721
the healthcare system in the US
is not really healthcare, right?
449
00:23:35,721 -->
00:23:39,771
It's healthcare because it waits for
you to get sick and then takes, but
450
00:23:39,771 -->
00:23:45,681
if you had a, a way of predicting with
perfect accuracy or with like not a
451
00:23:45,741 -->
00:23:49,551
perfect accuracy, with like reasonable,
and then you had a personalized way of
452
00:23:49,551 -->
00:23:51,366
preventing it, wouldn't that be cool?
453
00:23:52,091 -->
00:23:53,631
Krishna Gade: Yeah, that
would be amazing actually.
454
00:23:53,931 -->
00:23:56,391
You know, I studied a bit of
bioinformatics in my grad school.
455
00:23:56,391 -->
00:23:57,471
This is all very interesting.
456
00:23:57,471 -->
00:23:59,631
We studied all the protein
structured predictions and all.
457
00:23:59,841 -->
00:24:03,441
It seems like genAI can automate
a lot of these things and make,
458
00:24:03,441 -->
00:24:04,791
make them a lot better actually.
459
00:24:04,791 -->
00:24:08,211
Dr. Girish N. Nadkarni: It can make
a, you know, you know, I, you know,
460
00:24:08,511 -->
00:24:11,601
you can, you know, say that science
fiction, but I, I anticipate that the
461
00:24:11,721 -->
00:24:15,831
Coke machine, and again, I, uh, the
Coke machine analogy is not mine, right.
462
00:24:15,831 -->
00:24:17,911
It's my, uh, partner Alex Charney.
463
00:24:18,861 -->
00:24:22,006
Coming sooner rather than later
because it's gonna happen.
464
00:24:22,306 -->
00:24:26,836
Uh uh, and it might start off with like
rare genetic diseases, but it might
465
00:24:26,836 -->
00:24:28,726
end up with common complex diseases.
466
00:24:29,596 -->
00:24:30,226
Krishna Gade: That's amazing.
467
00:24:30,226 -->
00:24:30,796
That's amazing.
468
00:24:31,186 -->
00:24:36,646
And, and so basically when you sort
of, uh, think about genAI across like
469
00:24:36,646 -->
00:24:40,906
healthcare, you know, what are some of
like, let's say three promising use cases
470
00:24:40,906 -->
00:24:44,506
that you think in the, let's say in the
next 12 to 18 months, that you would see
471
00:24:44,506 -->
00:24:46,581
massive, you know, traction happening.
472
00:24:47,796 -->
00:24:50,736
Dr. Girish N. Nadkarni: So first
thing I, uh, uh, three things
473
00:24:51,006 -->
00:24:54,846
and this, this is not, uh, um,
um, specifically clinical, right?
474
00:24:55,056 -->
00:24:55,146
Yeah.
475
00:24:55,146 -->
00:24:57,966
The three things are massive
traction happening is like clinical
476
00:24:57,966 -->
00:25:00,336
summarization and documentation.
477
00:25:00,486 -->
00:25:03,846
And this includes, uh, the areas
that we just talk about, about
478
00:25:03,846 -->
00:25:08,676
ambient AI and uh, right, but it also
includes like mundane things like.
479
00:25:09,186 -->
00:25:15,306
Patient registry creation and like
extraction of data from clinical records
480
00:25:15,306 -->
00:25:19,206
for sending to large quality improvement
initiatives across the country, right?
481
00:25:19,206 -->
00:25:24,606
So that's why the second thing is that
I think in diagnostic imaging, uh, like
482
00:25:24,606 -->
00:25:26,496
diagnostic, not just imaging, right?
483
00:25:26,946 -->
00:25:32,166
Things like novel bio prognostics, uh,
for, uh, predicting disease, right?
484
00:25:33,561 -->
00:25:37,191
AI for like detecting stuff
on radiology scans, right?
485
00:25:37,671 -->
00:25:40,611
Uh, um, so those are the
things that I think will come
486
00:25:40,611 -->
00:25:41,601
out fast and furious years.
487
00:25:41,601 -->
00:25:45,781
Better ways of disease with the
eventual goal of preventing them.
488
00:25:46,431 -->
00:25:49,371
. And the third thing is a little
sort of left field, right?
489
00:25:49,851 -->
00:25:54,651
Um, I think huge inpatient engagement
and education and empowerment, right?
490
00:25:54,841 -->
00:25:57,051
I'll give you a simple example.
491
00:25:57,201 -->
00:26:00,071
So simple that we should have
thought of it like years back, right?
492
00:26:00,531 -->
00:26:02,721
. Um, like Brown University.
493
00:26:03,141 -->
00:26:06,531
So anytime you go to a doctor, you
have to sign a form, a consent form.
494
00:26:06,531 -->
00:26:06,741
Right?
495
00:26:06,741 -->
00:26:07,821
Like, I'm fine with this.
496
00:26:07,821 -->
00:26:07,851
Okay.
497
00:26:07,941 -->
00:26:08,031
Yeah.
498
00:26:08,841 -->
00:26:11,011
And those consent forms
are like full of jargon.
499
00:26:11,411 -->
00:26:14,811
I mean, I, I couldn't understand
them as a trained physician.
500
00:26:14,816 -->
00:26:17,726
I want people, like the lower
literacy laws gonna show, right.
501
00:26:18,086 -->
00:26:19,376
Who also already did a great thing.
502
00:26:19,376 -->
00:26:19,586
Right.
503
00:26:20,036 -->
00:26:22,886
Can you take this consent form
and translate it without losing
504
00:26:22,886 -->
00:26:26,486
any of the information into
a sixth grade reading level?
505
00:26:26,486 -->
00:26:28,646
And you know, it happened, uh, and they.
506
00:26:29,441 -->
00:26:31,721
Did it across the health system
and patients love it because now
507
00:26:31,721 -->
00:26:32,951
they can actually understand stuff.
508
00:26:32,951 -->
00:26:33,071
Right?
509
00:26:33,071 -->
00:26:33,131
Yeah.
510
00:26:33,371 -->
00:26:38,021
There's a huge thing in making
like complex medical com concepts.
511
00:26:38,021 -->
00:26:38,591
More clar.
512
00:26:39,491 -->
00:26:40,721
More clarifier, right?
513
00:26:40,721 -->
00:26:40,781
Yeah.
514
00:26:40,961 -->
00:26:41,171
So
515
00:26:41,981 -->
00:26:43,541
Krishna Gade: Yeah, I think that's great.
516
00:26:43,541 -->
00:26:46,301
I think, I think the third thing
that you mentioned, it's like
517
00:26:46,301 -->
00:26:48,671
actually close to my heart so the
accessibility of medicine, right?
518
00:26:48,671 -->
00:26:53,471
And I remember when we both met at the
Responsible AI Conference in New York,
519
00:26:53,801 -->
00:26:58,271
you were very passionate about how AI
could transform medicine across the globe.
520
00:26:58,331 -->
00:26:58,571
You know, for.
521
00:26:59,231 -->
00:27:03,711
For like, people who don't have access to
like great, great doctors and hospitals.
522
00:27:03,711 -->
00:27:05,651
Can you, what, what,
what are your thoughts?
523
00:27:05,651 -->
00:27:07,961
You know, how are you, you know,
like, can you share some of that
524
00:27:07,991 -->
00:27:08,981
with our audience, you know?
525
00:27:09,341 -->
00:27:09,521
Dr. Girish N. Nadkarni: Yeah.
526
00:27:09,521 -->
00:27:11,081
So let's, let's have a
conversation about that.
527
00:27:11,081 -->
00:27:11,321
Right.
528
00:27:11,681 -->
00:27:17,471
Um, so, um, medicine is sort
of human expertise encoded.
529
00:27:17,936 -->
00:27:20,006
The in heuristic form, right?
530
00:27:20,186 -->
00:27:22,646
Like you take in inputs and
then you have outputs, right?
531
00:27:23,096 -->
00:27:27,506
And the problem with medicine is like
expertise, until now, any quote unquote
532
00:27:27,506 -->
00:27:29,326
intelligence was limited, right?
533
00:27:29,806 -->
00:27:34,136
And as a result of that, you know,
you couldn't, um, scale this.
534
00:27:34,946 -->
00:27:35,876
Forget about worldwide.
535
00:27:35,876 -->
00:27:37,436
You couldn't really
scale this to a system.
536
00:27:37,676 -->
00:27:39,656
And that's the root of
all access issues, right?
537
00:27:39,656 -->
00:27:44,951
If I, today I were to go and try to find a
primary care provider for a cold, I would
538
00:27:44,951 -->
00:27:46,451
guarantee you I can't find one today.
539
00:27:46,841 -->
00:27:48,971
Um, I mean, I can find one
tomorrow or day after, right?
540
00:27:48,971 -->
00:27:50,981
Because there's limited resources.
541
00:27:50,981 -->
00:27:51,161
Right.
542
00:27:51,881 -->
00:27:57,371
But if you theoretically think about
the fact that it's encoded knowledge
543
00:27:57,701 -->
00:28:02,801
and you can scale it across the globe,
then it's a problem of scale, right?
544
00:28:02,801 -->
00:28:04,391
Which we've solved before.
545
00:28:05,006 -->
00:28:08,216
We, as in tech has, uh, like tech
industry has solved before, right?
546
00:28:08,486 -->
00:28:12,216
Because it's a matter of scale
and just like blitzscaling, right?
547
00:28:12,616 -->
00:28:17,366
. It's a little more dangerous because,
uh, you know, it affects patients.
548
00:28:17,366 -->
00:28:19,586
So that needs to be based
guardrails around it.
549
00:28:19,586 -->
00:28:19,646
Yeah.
550
00:28:20,066 -->
00:28:24,356
But at the same time, I would also
think that a bunch of patients are
551
00:28:24,356 -->
00:28:28,346
already putting their data into ChatGPT
or Google Gemini, or similar, right?
552
00:28:28,586 -->
00:28:30,026
So people are already doing it.
553
00:28:30,071 -->
00:28:34,181
I think we just need to make it better
and put more guardrails around it and
554
00:28:34,181 -->
00:28:38,601
like link it with via RAG to some sort
of verifiable medical knowledge, right?
555
00:28:38,991 -->
00:28:40,871
Because this could be huge, right?
556
00:28:40,871 -->
00:28:44,081
Like for example, I'll give
you a simple example, right?
557
00:28:44,141 -->
00:28:44,501
Um.
558
00:28:45,371 -->
00:28:47,291
Um, uh, where I grew up in India, right?
559
00:28:47,621 -->
00:28:50,261
Uh, TB, Tuberculosis is
extremely prominent, right?
560
00:28:50,621 -->
00:28:55,001
Um, and you know, the, the standard
workflow was, you know, for
561
00:28:55,001 -->
00:28:58,816
diagnosis you had to get like, uh,
three tests, et cetera, et cetera.
562
00:28:59,201 -->
00:29:02,441
Now there are things that, that can
have diagnosed TB with like a hundred
563
00:29:02,441 -->
00:29:04,081
percent accuracy in five minutes.
564
00:29:05,291 -->
00:29:07,451
That completely changes
the workflow, right?
565
00:29:07,451 -->
00:29:13,286
Because now you can go from presentation
to not even a doctor, right?
566
00:29:13,676 -->
00:29:17,636
Someone who's, uh, sort
of paramedical, right?
567
00:29:17,636 -->
00:29:20,456
Like someone with some training,
knowing how to recognize the bad
568
00:29:20,456 -->
00:29:24,096
stuff accompanied by AI agent, right?
569
00:29:24,096 -->
00:29:27,386
. You could get to be diagnosed and you
could get the first dose of treatment
570
00:29:27,391 -->
00:29:30,206
all in all under like 15 minutes, right?
571
00:29:30,376 -->
00:29:31,546
Huge workflow issue.
572
00:29:32,036 -->
00:29:32,276
Right?
573
00:29:32,486 -->
00:29:33,656
So I think that, I think.
574
00:29:34,016 -->
00:29:40,196
The mass adoption of AI into
past clinical healthcare over and
575
00:29:40,196 -->
00:29:46,706
beyond what's happening in the back
office tasks is going to be, uh,
576
00:29:47,516 -->
00:29:52,406
uh, in starting off in the low and
middle income countries, right?
577
00:29:52,796 -->
00:29:58,136
But there's a danger there because
by definition, low and middle income
578
00:29:58,136 -->
00:30:00,506
countries won't necessarily have the same.
579
00:30:01,166 -->
00:30:03,986
Rigor of regulations that the U.S. Does.
580
00:30:03,986 -->
00:30:07,136
So, which not and cannot
take advantage of that.
581
00:30:07,676 -->
00:30:08,066
Krishna Gade: That's right.
582
00:30:08,486 -->
00:30:10,436
So, but do you see the
world of virtual agents?
583
00:30:10,436 -->
00:30:12,026
Like in terms of like, oh yeah.
584
00:30:12,181 -->
00:30:12,421
Dr. Girish N. Nadkarni: Absolutely.
585
00:30:12,421 -->
00:30:13,101
Already happening?
586
00:30:13,101 -->
00:30:13,341
Yeah.
587
00:30:13,341 -->
00:30:13,343
Already happening.
588
00:30:13,346 -->
00:30:13,356
Yeah,
589
00:30:13,466 -->
00:30:14,066
Krishna Gade: Already happening.
590
00:30:14,906 -->
00:30:17,036
Dr. Girish N. Nadkarni: I think
it just needs to happen in a more
591
00:30:17,036 -->
00:30:21,596
rigorous and more reproducible and a
more, um, dare I say, ethical fashion.
592
00:30:21,596 -->
00:30:24,236
But it's already, I mean, you know.
593
00:30:25,211 -->
00:30:27,641
The question of whether it's
happening or not is done.
594
00:30:27,911 -->
00:30:28,871
It's how to be happening.
595
00:30:29,231 -->
00:30:31,781
The question is how do we make
sure that it's safe and effect?
596
00:30:32,351 -->
00:30:32,621
Krishna Gade: That's right.
597
00:30:32,831 -->
00:30:35,081
So I think as a related audience
question on that, you know, what
598
00:30:35,081 -->
00:30:38,991
are those particular clinical,
clinical scenarios where genAI
599
00:30:39,011 -->
00:30:41,051
should not be used or too risky?
600
00:30:41,051 -->
00:30:42,521
So this is kind of a counter question.
601
00:30:43,451 -->
00:30:44,651
Dr. Girish N. Nadkarni: Well,
that's a really good question.
602
00:30:44,651 -->
00:30:44,951
Right.
603
00:30:45,131 -->
00:30:53,651
Um, um, I don't think it should be used
when there's a question of capacity.
604
00:30:54,101 -->
00:30:58,481
And by capacity I mean it in a
specific medical and a legal sense.
605
00:30:58,481 -->
00:31:04,391
Capacity basically means that
you cannot, because of severe
606
00:31:04,391 -->
00:31:08,981
mental or physical issues, um, you
cannot, don't have the capability
607
00:31:08,981 -->
00:31:11,111
to take your own decisions, right?
608
00:31:11,741 -->
00:31:16,511
And there it should not
be used because, uh, of.
609
00:31:17,081 -->
00:31:20,501
The fact that, you know, patients
don't really have the ability to
610
00:31:20,741 -->
00:31:22,301
distinguish right from wrong, right?
611
00:31:22,301 -->
00:31:26,261
And that's, I don't mean to sound
paternalistic, but the specific
612
00:31:26,291 -->
00:31:29,561
medicolegal definition of capacity,
which says that, you know, then you
613
00:31:29,561 -->
00:31:32,471
need like two or three physicians
or like ethics board to come in.
614
00:31:32,741 -->
00:31:36,641
So I think that's one example, um,
of where it should not be used.
615
00:31:36,911 -->
00:31:39,221
The second one is, uh,
616
00:31:41,951 -->
00:31:42,131
well.
617
00:31:42,836 -->
00:31:43,886
This is starting off with nothing.
618
00:31:43,886 -->
00:31:46,136
Nothing should be used unless
it's safe, effective, and ethical.
619
00:31:46,136 -->
00:31:46,286
Right?
620
00:31:46,286 -->
00:31:46,346
Yeah.
621
00:31:46,376 -->
00:31:50,696
But like, you know, where it absolutely
should not be used on, on the
622
00:31:50,696 -->
00:31:52,316
pediatric and the child health fund.
623
00:31:52,436 -->
00:31:54,596
I'm a little conflicted, right?
624
00:31:54,926 -->
00:31:58,946
I don't know the answer to that,
but that's something which has, um.
625
00:32:00,326 -->
00:32:02,456
Uh, which has tricky, right?
626
00:32:02,456 -->
00:32:06,146
Because of consent and
because of, uh, issues, right?
627
00:32:06,146 -->
00:32:10,616
And that's why even if you see it in the
current marketplace, uh, in the current,
628
00:32:10,616 -->
00:32:19,916
uh, world, uh, um, you know, um, uh,
uh, it's not being used particularly in
629
00:32:20,546 -->
00:32:22,196
not a lot of products in kids right now.
630
00:32:22,196 -->
00:32:22,496
Right.
631
00:32:22,766 -->
00:32:25,706
Which is an issue bio in of itself, right?
632
00:32:26,096 -->
00:32:28,916
Um, and the third thing
is sort of specific.
633
00:32:29,711 -->
00:32:31,061
Protected populations, right?
634
00:32:31,621 -->
00:32:33,641
. That the, those are definitions, right?
635
00:32:33,641 -->
00:32:38,951
Like prisoners and like, uh, um, you
know, people, uh, with like significant,
636
00:32:39,311 -->
00:32:42,671
um, medical and or mental health issues.
637
00:32:42,671 -->
00:32:45,521
So those are, that, that,
that's tricky ethical situation.
638
00:32:45,551 -->
00:32:46,571
That's a really good question.
639
00:32:46,571 -->
00:32:47,796
I need to think a little bit more about.
640
00:32:48,106 -->
00:32:48,396
Krishna Gade: Yeah.
641
00:32:48,676 -->
00:32:50,036
I mean there's this whole sort of, uh.
642
00:32:50,811 -->
00:32:53,361
I, I don't know if it's a
dystopian scenario, but like, you
643
00:32:53,361 -->
00:32:56,241
know, you think about like Elon
Musk's, Optimus Robot, right?
644
00:32:56,241 -->
00:32:59,181
Like that is, you know, running,
you know, walking uphills, you
645
00:32:59,181 -->
00:33:00,861
know, catching tennis balls.
646
00:33:00,861 -->
00:33:04,431
Now if you think about like generative
AI systems connected to robotics now,
647
00:33:04,431 -->
00:33:09,321
do you see a world where, you know, AI
driven robots could outperform human
648
00:33:09,321 -->
00:33:14,511
surgeons are, you know, the sort of, you
know, pretty high risk, uh, situation?
649
00:33:14,781 -->
00:33:18,836
Dr. Girish N. Nadkarni: Well,
so, so, so, uh, uh, uh, um, so.
650
00:33:19,976 -->
00:33:20,576
Yes.
651
00:33:21,086 -->
00:33:28,886
No, the question is when, uh,
uh, uh, um, the question is when
652
00:33:29,336 -->
00:33:32,106
and, um, I don't know the answer.
653
00:33:33,126 -->
00:33:35,426
Could, I mean, if you talk
to different people, right.
654
00:33:35,486 -->
00:33:35,996
Again, yeah.
655
00:33:36,866 -->
00:33:39,626
Because the understanding, the real world.
656
00:33:40,061 -->
00:33:44,051
Is harder than, much, much harder
than understanding language, right?
657
00:33:44,571 -->
00:33:51,911
. That's why, um, the hardware is there,
but the world models are, and the LLMs
658
00:33:51,971 -->
00:33:53,501
are not particularly there, right?
659
00:33:54,011 -->
00:33:57,701
. Uh, um, uh, you know, um, um,
660
00:34:00,041 -->
00:34:05,231
um, I think it's not going
to be easy to do this.
661
00:34:05,831 -->
00:34:06,371
Um.
662
00:34:08,641 -->
00:34:10,106
Krishna Gade: It's too
ambiguous and you need more,
663
00:34:10,371 -->
00:34:11,996
Dr. Girish N. Nadkarni: I, I don't
know, I honestly don't know answer.
664
00:34:12,056 -->
00:34:13,106
Krishna Gade: Yeah, yeah.
665
00:34:13,106 -->
00:34:14,486
Maybe, maybe in a hundred years.
666
00:34:15,691 -->
00:34:17,211
Dr. Girish N. Nadkarni: I think
a bit sooner than that, but
667
00:34:17,211 -->
00:34:21,386
like, I don't know if like, it'll
require the, the development of
668
00:34:21,386 -->
00:34:24,116
like robust world models, right?
669
00:34:24,116 -->
00:34:27,446
Because understanding the physical
world is much harder and understanding
670
00:34:27,776 -->
00:34:28,526
the digital world, right?
671
00:34:28,526 -->
00:34:31,256
So it requires development
of robust world models.
672
00:34:31,706 -->
00:34:32,006
Krishna Gade: Right.
673
00:34:32,096 -->
00:34:32,636
Right, right.
674
00:34:32,996 -->
00:34:33,566
Makes sense.
675
00:34:33,776 -->
00:34:37,406
So, and then there's, uh, something
related to the, uh, so whole sort of,
676
00:34:37,406 -->
00:34:39,956
uh, administrative efficiency, right.
677
00:34:39,956 -->
00:34:41,486
Which we, you should
talked about at length.
678
00:34:42,056 -->
00:34:44,636
Uh, but you know, there's an
audience question as well on that.
679
00:34:45,026 -->
00:34:47,456
You know, there are some people
who absorb information better
680
00:34:47,456 -->
00:34:49,016
when taking notes or writing.
681
00:34:49,256 -->
00:34:50,606
You know, AI is taking notes.
682
00:34:50,956 -->
00:34:53,626
You know, how will we include
these, those doctor styles, right?
683
00:34:53,626 -->
00:34:57,166
Like, do you see a world that that's
gonna be a hindrance or it's a, it's,
684
00:34:57,166 -->
00:34:58,461
it's kind of a template of everything.
685
00:34:58,461 -->
00:35:00,766
Dr. Girish N. Nadkarni: I mean, they, they
wanna like notes, type notes, all right?
686
00:35:00,766 -->
00:35:01,306
They can do that.
687
00:35:01,306 -->
00:35:05,316
And then you can just add in to the then
addendums of notes to the end, right?
688
00:35:05,316 -->
00:35:10,006
I mean, like, I would argue that lots
of doctors want to have a conversation.
689
00:35:11,256 -->
00:35:11,476
Krishna Gade: Yeah.
690
00:35:11,816 -->
00:35:12,316
Makes sense.
691
00:35:12,736 -->
00:35:13,216
Makes sense.
692
00:35:13,426 -->
00:35:14,206
So I guess, you know.
693
00:35:14,861 -->
00:35:19,451
Finally, you know, where do you see the
biggest bottlenecks in scaling gen AI
694
00:35:19,451 -->
00:35:21,311
solutions in, in, in, in healthcare?
695
00:35:21,341 -->
00:35:26,861
You know, is it coming from, uh, let's
say like we talked about the reliability,
696
00:35:26,861 -->
00:35:29,651
accuracy, this whole, uh, risk issues.
697
00:35:29,771 -->
00:35:33,671
Uh, is it coming from pro, you know,
governance side of things, process
698
00:35:33,671 -->
00:35:36,551
side of things, you know, where,
where, where are the bottlenecks?
699
00:35:38,861 -->
00:35:40,181
Dr. Girish N. Nadkarni:
Lots of bottlenecks, right?
700
00:35:40,301 -->
00:35:43,991
I mean, uh, there is a
standing maxim that, um.
701
00:35:44,426 -->
00:35:54,956
Uh, uh, uh, um, uh, uh, uh, that,
that is standing maxim that, uh, um,
702
00:35:55,886 -->
00:35:58,406
workflow eats technology for breakfast.
703
00:35:58,406 -->
00:36:01,556
So I think if, okay.
704
00:36:02,666 -->
00:36:05,906
Do we agree that AI in its
current form is a massive.
705
00:36:06,746 -->
00:36:08,606
Transformational technology probably.
706
00:36:08,606 -->
00:36:08,816
Right?
707
00:36:08,816 -->
00:36:10,586
I mean, you found a company
around it, so you've probably,
708
00:36:12,596 -->
00:36:17,636
so, but any transformational
technology takes time because there
709
00:36:17,636 -->
00:36:20,846
needs the, the technology term might
be transformational, but society
710
00:36:20,846 -->
00:36:22,556
needs to transform around it, right?
711
00:36:22,951 -->
00:36:26,036
For example, I'll give you
a clear example, right?
712
00:36:26,306 -->
00:36:27,116
Electricity.
713
00:36:27,746 -->
00:36:31,856
The became invented, forget what
invented, became cheaply, easy to
714
00:36:31,856 -->
00:36:36,506
produce cheaply, but the time it took
for electricity to replace steam was
715
00:36:36,506 -->
00:36:38,556
approximately 50 to 60 years, right?
716
00:36:38,836 -->
00:36:42,836
. That's because, uh, um, you
know, factories had to be
717
00:36:42,836 -->
00:36:44,066
reconfigured for electricity.
718
00:36:44,166 -->
00:36:48,536
Um, people had to be trained,
infrastructure had to be laid.
719
00:36:48,926 -->
00:36:51,986
All of those things, and it
society had to evolve, right?
720
00:36:52,196 -->
00:36:52,436
Krishna Gade: Yeah.
721
00:36:52,796 -->
00:36:54,866
Be comfortable with it and trust it.
722
00:36:54,866 -->
00:36:54,986
Yeah.
723
00:36:56,036 -->
00:36:59,996
Dr. Girish N. Nadkarni: Now the, and here
I would like to hear your opinion, right?
724
00:37:00,656 -->
00:37:00,866
Yeah.
725
00:37:00,956 -->
00:37:03,746
Just in thinking about healthcare, right?
726
00:37:03,751 -->
00:37:03,871
Yeah.
727
00:37:04,346 -->
00:37:07,766
Um, right now, hospitals and
health systems physically,
728
00:37:08,396 -->
00:37:13,736
logistically, organizationally, are
configured in a certain way, right?
729
00:37:14,216 -->
00:37:14,516
Um.
730
00:37:16,046 -->
00:37:19,736
To reconfigure them is gonna
take time, effort, and energy.
731
00:37:19,946 -->
00:37:20,156
Krishna Gade: Yeah.
732
00:37:20,306 -->
00:37:22,346
Dr. Girish N. Nadkarni: And
if to reconfigure them early,
733
00:37:22,556 -->
00:37:26,186
then the tipping point is gonna
be a massive financial risk.
734
00:37:26,186 -->
00:37:26,396
Yeah.
735
00:37:26,731 -->
00:37:27,021
Yeah.
736
00:37:27,566 -->
00:37:28,166
In this case.
737
00:37:28,166 -->
00:37:29,096
And tell you what you think, right?
738
00:37:29,096 -->
00:37:29,156
Yeah.
739
00:37:29,186 -->
00:37:31,406
Being early is almost
the same as being wrong.
740
00:37:31,556 -->
00:37:32,306
Krishna Gade: Yeah, absolutely.
741
00:37:32,306 -->
00:37:35,861
I think, you know, and in, in this
why in many industries and healthcare
742
00:37:35,881 -->
00:37:37,286
is obviously under industry where.
743
00:37:37,796 -->
00:37:40,976
What we are seeing is AI is being used
to drive more efficiencies, right?
744
00:37:40,976 -->
00:37:45,206
Like you talked about things like, you
know, assessing readmission risk or you
745
00:37:45,206 -->
00:37:49,796
know, getting an alert if this patient
is, you know, likely to fall or something.
746
00:37:49,796 -->
00:37:53,306
Or like maybe like a sepsis detection
system that that can give you an early
747
00:37:53,306 -->
00:37:57,386
warning or you know, you know, the
whole ambient experience you talked
748
00:37:57,386 -->
00:37:59,126
about where even if you are wrong.
749
00:37:59,661 -->
00:38:03,261
It's not like gonna be so bad,
like the false positives are not
750
00:38:03,261 -->
00:38:07,401
bad, but, but getting that insight
quickly can save a lot of time and
751
00:38:07,401 -->
00:38:11,391
energy and human resources, you
know, being spent on those problems.
752
00:38:11,391 -->
00:38:14,511
I think that's where we are seeing
like the early adoption happen,
753
00:38:14,661 -->
00:38:19,086
of course, like having an AI robot
doing this ophthalm, ophthalm, uh,
754
00:38:19,086 -->
00:38:22,131
like eye surgery that's probably
like, uh, tens of years away, right?
755
00:38:22,131 -->
00:38:22,606
Like, you know, uh,
756
00:38:23,421 -->
00:38:26,691
Dr. Girish N. Nadkarni: I think
society society broadly, but
757
00:38:26,691 -->
00:38:28,221
also specifically healthcare.
758
00:38:28,646 -->
00:38:32,856
Society needs to transform
around it, right?
759
00:38:33,046 -->
00:38:33,336
Yeah.
760
00:38:33,516 -->
00:38:35,651
We agree that the technology,
transformational, it just needs
761
00:38:35,651 -->
00:38:39,431
the remaining world to conform
around it, which will happen.
762
00:38:39,521 -->
00:38:39,941
Trust me.
763
00:38:40,301 -->
00:38:41,051
Krishna Gade: Yeah, absolutely.
764
00:38:41,261 -->
00:38:43,181
I think the personalized
medicine that you talked about
765
00:38:43,181 -->
00:38:44,351
is very, very interesting, right?
766
00:38:44,351 -->
00:38:47,711
And that's both innovative and
could we get game changing actually?
767
00:38:48,131 -->
00:38:51,281
Uh, so I guess, you know, we are
gonna wrap up this conversation,
768
00:38:51,461 -->
00:38:52,811
you know, two, three minutes now.
769
00:38:52,956 -->
00:38:56,536
If you had one final takeaway for
a healthcare leader like you, you
770
00:38:56,536 -->
00:38:59,816
know, trying to drive genAI in their
organization, what would it be?
771
00:39:00,911 -->
00:39:04,631
Dr. Girish N. Nadkarni: I mean this time
this might soundy and like, uh, cheesy,
772
00:39:04,871 -->
00:39:06,521
but keep the patient at the center right.
773
00:39:06,521 -->
00:39:07,751
Do what's best for the patient.
774
00:39:08,831 -->
00:39:14,801
If your focus and your effort and
your energy and your purpose starts
775
00:39:14,801 -->
00:39:17,411
from that, then even if you're
wrong, you'll probably be right.
776
00:39:17,981 -->
00:39:19,091
Krishna Gade: Yeah, absolutely.
777
00:39:19,601 -->
00:39:19,811
Yeah.
778
00:39:20,351 -->
00:39:22,281
Well, thank you so much, Girish.
779
00:39:22,301 -->
00:39:24,971
That's a great way to end
the, uh, conversation today.
780
00:39:24,971 -->
00:39:28,321
I. And thanks for sharing your
valuable insights and time with us.
781
00:39:28,351 -->
00:39:33,211
Um, uh, that's it for this, you
know, this week's AI explain.
782
00:39:33,421 -->
00:39:36,841
Uh, thank you so much for everybody
for joining us, and, uh, we'll
783
00:39:36,841 -->
00:39:38,071
see you in another session.
784
00:39:39,151 -->
00:39:39,421
Dr. Girish N. Nadkarni: Thank you.
785
00:39:39,421 -->
00:39:39,763
Thank you Krishna.
786
00:39:39,763 -->
00:39:42,151
And thank you everyone to
the, uh, for having me on.
787
00:39:42,151 -->
00:39:43,261
It was, it was a blast.
788
00:39:43,681 -->
00:39:43,921
Krishna Gade: Awesome.
789
00:39:44,071 -->
00:39:44,401
Thank you.
790
00:39:44,581 -->
00:39:44,731
Bye-bye.
