1
00:00:07,500 --> 00:00:12,820
I'm Michael Reilly, Chief Investment Officer of TCW's Equities Group.

2
00:00:12,820 --> 00:00:17,081
I'm here with Beau Pfeiffer, who co-heads our Artificial

3
00:00:17,081 --> 00:00:20,893
Intelligence team, and Matthew Lowe, TCW's Medical

4
00:00:20,893 --> 00:00:25,079
Products Analyst, covering biotech, pharma, and medtech.

5
00:00:25,079 --> 00:00:31,059
In this episode, we explore the transformative
power of generative AI in healthcare.

6
00:00:31,059 --> 00:00:36,739
That's a sector that spends about $4.5 trillion annually in the US alone.

7
00:00:36,739 --> 00:00:40,579
We're around 17% of GDP.

8
00:00:40,579 --> 00:00:45,406
We're going to talk about how leveraging this technology

9
00:00:45,406 --> 00:00:50,318
can significantly enhance patient care, reduce costs, and

10
00:00:50,318 --> 00:00:55,060
really optimize operational efficiency within hospitals.

11
00:00:55,060 --> 00:01:05,539
This means everything from disease diagnosis to personalized treatment
plans to even streamlining how you get your next doctor's appointment.

12
00:01:05,539 --> 00:01:08,739
So thanks for joining me, Beau and Matt.

13
00:01:08,740 --> 00:01:12,140
Maybe we can just start broadly, perhaps with you, Beau.

14
00:01:12,140 --> 00:01:16,299
What are some of the major ways you see AI impacting healthcare?

15
00:01:16,299 --> 00:01:17,299
Yeah.

16
00:01:17,299 --> 00:01:18,299
Hi, Michael.

17
00:01:18,299 --> 00:01:20,299
It's good to be here with you guys.

18
00:01:20,299 --> 00:01:25,620
Bear with me for a minute because there's a lot to cover on this topic.

19
00:01:25,620 --> 00:01:30,083
Let's consider a hypothetical situation that probably

20
00:01:30,083 --> 00:01:34,876
won't be hypothetical for much longer and think about the

21
00:01:34,876 --> 00:01:39,339
impact of AI on healthcare from a patient perspective.

22
00:01:39,340 --> 00:01:41,980
So our patient has symptoms.

23
00:01:41,980 --> 00:01:51,700
They log into their provider portal where they interact in
conversational English with an AI chatbot that takes their history.

24
00:01:51,700 --> 00:01:58,506
Today, this is a lengthy and rigid click here kind of process where really you

25
00:01:58,506 --> 00:02:05,398
spend more time on everything that's not wrong with you, but the AI will listen

26
00:02:05,398 --> 00:02:11,860
to your complaints and quickly zero in on questions that matter clinically.

27
00:02:11,860 --> 00:02:18,409
Based on that interaction, the AI will identify what type of doctor you need to

28
00:02:18,409 --> 00:02:24,958
see, schedule the appointment at a time that works for both you and the doctor,

29
00:02:24,958 --> 00:02:29,052
and provide the doctor with the pertinent medical

30
00:02:29,052 --> 00:02:32,900
history in a complete but very concise summary.

31
00:02:32,900 --> 00:02:38,325
The actual appointment will look pretty much the same as it does today,

32
00:02:38,325 --> 00:02:43,298
but the AI is in the background taking notes and already starting

33
00:02:43,298 --> 00:02:47,819
to home in on a list of conditions that match your symptoms.

34
00:02:47,819 --> 00:02:53,258
The doctor might even ask the AI for some clinical trial data or

35
00:02:53,258 --> 00:02:58,697
references to scientific literature, and get real-time responses

36
00:02:58,697 --> 00:03:04,220
kind of like a medical Siri or Alexa would do on our phones today.

37
00:03:04,219 --> 00:03:08,727
And then when the encounter is over, the AI will immediately

38
00:03:08,727 --> 00:03:13,309
generate the after-visit summary in language that the patient

39
00:03:13,309 --> 00:03:17,300
can actually understand and put it up in their portal.

40
00:03:17,300 --> 00:03:22,612
So then the AI presents an evidence-based treatment plan that might

41
00:03:22,612 --> 00:03:28,394
involve lab work or imaging, and if approved by the doctor, automatically

42
00:03:28,394 --> 00:03:34,020
schedules the follow-up appointment for a CT scan or an MRI or whatever.

43
00:03:34,020 --> 00:03:45,900
That MRI, by the way, will be completed in half the time because AI in the
MRI machine can interpret the scanning data faster and more precisely.

44
00:03:45,900 --> 00:03:50,521
It then reads the scan to identify any abnormalities

45
00:03:50,521 --> 00:03:54,881
in the imaging, incorporates any blood tests that

46
00:03:54,881 --> 00:03:58,979
might have been done, and produces a diagnosis.

47
00:03:58,979 --> 00:04:04,509
The treatment plan is then optimized by analyzing the medical records of tens

48
00:04:04,509 --> 00:04:09,826
of thousands of other patients in the payers organization who have had the

49
00:04:09,826 --> 00:04:14,433
same initial diagnosis to find the best treatment algorithm, and

50
00:04:14,433 --> 00:04:18,899
all of this information is presented to the doctor for consent.

51
00:04:18,899 --> 00:04:23,779
We're not taking the human out of the equation here.

52
00:04:23,779 --> 00:04:29,819
And suppose in this case the treatment
consists of a series of weekly drug infusions.

53
00:04:29,819 --> 00:04:34,724
The AI once again handles all of the scheduling and communicates the

54
00:04:34,724 --> 00:04:39,415
information to the patient while in the background contacting the

55
00:04:39,415 --> 00:04:44,319
managed care provider or the payer to negotiate prior authorization,

56
00:04:44,319 --> 00:04:48,300
for example, if the patient's health plan requires that.

57
00:04:48,300 --> 00:04:53,092
And the drug was probably designed with AI tools to bind with

58
00:04:53,092 --> 00:04:57,885
a very high affinity to the target site of the disease cells,

59
00:04:57,885 --> 00:05:02,059
making use of AI's ability to predict protein folding.

60
00:05:02,060 --> 00:05:08,305
And finally, all of the data collected during the entire process, including the

61
00:05:08,305 --> 00:05:14,238
follow-up and the outcomes, is anonymized and fed back into the database of

62
00:05:14,238 --> 00:05:19,860
medical records to continuously improve the results for future patients.

63
00:05:19,860 --> 00:05:24,032
I'm just wondering, Bo, is the patient going to like

64
00:05:24,032 --> 00:05:28,282
and appreciate this experience or is it going to feel

65
00:05:28,282 --> 00:05:32,139
impersonalized with bots and just all technology?

66
00:05:32,139 --> 00:05:35,420
I mean, what are your thoughts there?

67
00:05:35,420 --> 00:05:40,660
I think it will be to a large degree invisible to the patient.

68
00:05:40,660 --> 00:05:45,288
Certainly the experience of logging onto your portal,

69
00:05:45,288 --> 00:05:49,831
answering a bunch of background questions that we do

70
00:05:49,831 --> 00:05:54,460
with a mouse right now, do you have a history of this?

71
00:05:54,459 --> 00:05:55,860
Do you have a history of that?

72
00:05:55,930 --> 00:06:01,435
I think patients will appreciate actually speaking to something that sounds

73
00:06:01,435 --> 00:06:06,651
like a human and just explaining in plain English what their complaints

74
00:06:06,651 --> 00:06:11,939
are, and then not having to answer a bunch of questions that don't apply.

75
00:06:11,939 --> 00:06:17,939
And then everything else is kind of behind the
scenes and not really visible to the patient anyway.

76
00:06:17,939 --> 00:06:24,060
So I think it'll make the experience a lot more pleasant for patients.

77
00:06:24,060 --> 00:06:25,139
That makes sense.

78
00:06:25,139 --> 00:06:30,099
It sounds like AI is really going to bring about some sizable

79
00:06:30,099 --> 00:06:34,499
both efficiency gains for patient interaction and then

80
00:06:34,499 --> 00:06:38,579
also treatment plans that can really be customized.

81
00:06:38,579 --> 00:06:49,620
Maybe if we turn specifically to drug discovery, Matt, can you give us an
idea of how drug companies are actually using AI to design better drugs?

82
00:06:49,620 --> 00:06:50,620
Yeah, sure.

83
00:06:50,620 --> 00:06:52,579
Thank you for having me, Michael.

84
00:06:52,610 --> 00:06:59,740
So several biopharma companies are using AI to speed
up and more importantly improve drug development.

85
00:06:59,740 --> 00:07:03,434
Machine learning models are trained using information about

86
00:07:03,434 --> 00:07:06,882
the amino acid sequence or the 3D structure of previous

87
00:07:06,882 --> 00:07:09,900
drug candidates and about properties of interest.

88
00:07:09,899 --> 00:07:18,819
And AI techniques can analyze large scale biomedical data to identify
existing drugs that may have therapeutic potential for different diseases.

89
00:07:18,819 --> 00:07:27,420
So by repurposing existing drugs for new indications, AI can
accelerate the drug discovery process and also reduce costs.

90
00:07:27,420 --> 00:07:32,259
How does it, can you explain a little bit about how AI does that?

91
00:07:32,259 --> 00:07:33,379
What makes this possible?

92
00:07:33,379 --> 00:07:36,980
Yeah, I guess different companies are doing it in different ways.

93
00:07:36,980 --> 00:07:44,660
But I mean, Schrodinger, for example, they have a
computational software platform that's powered by physics.

94
00:07:44,660 --> 00:07:48,139
Effectively it gets better protein binding.

95
00:07:48,139 --> 00:07:50,459
Relay Therapeutics is another company.

96
00:07:50,459 --> 00:07:55,740
Effectively we think of proteins are not static, they're constantly moving.

97
00:07:55,740 --> 00:07:58,899
They don't stay in the same structure all the time.

98
00:07:58,899 --> 00:08:03,579
So traditional drug discovery is really
like looking at old black and white photos.

99
00:08:03,579 --> 00:08:12,220
Whereas Relay Therapeutics, their technology essentially shows how the proteins
move and different binding pockets open up and close at different timeframes.

100
00:08:12,240 --> 00:08:19,360
So essentially Relay is aiming to kind of drug
the undruggable using motion based drug design.

101
00:08:19,360 --> 00:08:23,819
So not every technology platform is doing it in the same way.

102
00:08:23,819 --> 00:08:27,899
Those would be two different examples of
how two different companies are doing it.

103
00:08:27,899 --> 00:08:30,420
Wow, super interesting.

104
00:08:30,420 --> 00:08:43,220
Maybe we can switch gears from drugs and how they're designed and focus a minute
on how AI can help with diagnosing disease, early detection and the like.

105
00:08:43,220 --> 00:08:46,059
You see a lot of that out in the news.

106
00:08:46,059 --> 00:08:47,299
Yep, sure.

107
00:08:47,299 --> 00:08:54,100
So AI powered blood tests are being used to
detect early disease and also disease recurrence.

108
00:08:54,100 --> 00:08:57,259
Essentially all cancer is a bioinformatics play.

109
00:08:57,259 --> 00:09:01,559
You're looking for a change, an insertion or a deletion that

110
00:09:01,559 --> 00:09:05,505
could happen anywhere and in any size and it could show

111
00:09:05,505 --> 00:09:09,100
up in three billion base pairs of the human genome.

112
00:09:09,100 --> 00:09:10,100
And obviously it's...

113
00:09:10,100 --> 00:09:11,100
Yeah, I just want to...

114
00:09:11,100 --> 00:09:12,100
Sorry, go on bud.

115
00:09:12,100 --> 00:09:15,179
Yeah, I just want to emphasize that point.

116
00:09:15,179 --> 00:09:23,100
In some ways, we've made incredible strides in the fight
against cancer, but we're still searching for a cure.

117
00:09:23,100 --> 00:09:34,379
What we know is that detecting cancer early dramatically improves
outcomes and provides the best chance to treat with curative intent.

118
00:09:34,379 --> 00:09:38,210
So we're getting pretty good at treating a lot of cancers when we

119
00:09:38,210 --> 00:09:41,983
catch them early enough and these AI driven tests are absolutely

120
00:09:41,983 --> 00:09:45,059
allowing us to catch the disease earlier and earlier.

121
00:09:45,059 --> 00:09:49,919
Yep, and it's hard to imagine a bigger mountain of data than the human genome.

122
00:09:49,919 --> 00:09:53,659
Essentially every genomics diagnostic company is using AI.

123
00:09:53,659 --> 00:09:58,608
DNA sequencing technologies generate a lot of data and AI helps the diagnostics

124
00:09:58,608 --> 00:10:03,185
companies better mine that data to develop signatures that are predictive

125
00:10:03,185 --> 00:10:07,019
of disease and predictive of response to therapies, et cetera.

126
00:10:07,019 --> 00:10:08,419
Can I just ask you on that, Matt?

127
00:10:08,419 --> 00:10:12,779
I mean, what is the magnitude of the improvement?

128
00:10:12,779 --> 00:10:21,899
How much is AI enabling advances or improvements
in this type of mapping or detection?

129
00:10:21,899 --> 00:10:23,220
How would you characterize it?

130
00:10:23,220 --> 00:10:28,679
Yeah, I think it's hard to quantify at this moment
in time because it's still relatively early.

131
00:10:28,690 --> 00:10:31,299
But I think it has huge potential.

132
00:10:31,299 --> 00:10:40,460
For example, diseases like RA, patients can cycle through various different
therapies before they kind of stumble across the one that's working for them.

133
00:10:40,460 --> 00:10:42,379
It's kind of trial and error.

134
00:10:42,379 --> 00:10:51,940
Whereas AI searching through data, finding out reasons why different patients
respond to different therapies could really hugely increase efficiency.

135
00:10:51,940 --> 00:10:59,500
So we kind of get to the drug that's going to work for that specific patient
right away rather than this kind of trial and error process that we have now.

136
00:10:59,500 --> 00:11:00,500
Fascinating.

137
00:11:00,500 --> 00:11:03,980
That's rheumatoid arthritis you were referring to, right?

138
00:11:03,980 --> 00:11:04,980
Yes.

139
00:11:04,980 --> 00:11:05,980
Okay, great.

140
00:11:05,980 --> 00:11:20,220
Maybe it makes sense to talk about just devices themselves and how AI is really
disrupting that whole medical device world, particularly in the radiology space.

141
00:11:20,220 --> 00:11:21,700
What are we seeing there?

142
00:11:22,019 --> 00:11:28,059
Yeah, so eight years ago there was almost no
AI-enabled medical devices being approved by the FDA.

143
00:11:28,059 --> 00:11:35,019
But now around 40% of all the medical devices
approved by the FDA for radiology are driven by AI.

144
00:11:35,019 --> 00:11:36,019
Wow.

145
00:11:36,019 --> 00:11:38,940
And we expect that percentage to go on increasing.

146
00:11:38,940 --> 00:11:46,700
Around 75% of all the AI products that have been approved
are in radiology, in other words, medical imaging.

147
00:11:46,700 --> 00:11:50,620
And just two examples of the AI radiology products that we currently have.

148
00:11:50,620 --> 00:11:52,259
One would be CLEARLY.

149
00:11:52,259 --> 00:11:55,957
CLEARLY uses AI to interpret CCTA scans to predict who is

150
00:11:55,957 --> 00:11:59,463
at risk of heart disease, essentially allowing doctors

151
00:11:59,463 --> 00:12:02,460
to develop plans to help prevent heart attacks.

152
00:12:02,460 --> 00:12:05,995
The traditional method of diagnosing heart disease uses

153
00:12:05,995 --> 00:12:09,593
CT scans to detect when the buildup of fatty deposits in

154
00:12:09,593 --> 00:12:12,940
the arteries restricts the blood supply to the heart.

155
00:12:12,940 --> 00:12:15,740
But at that point, a heart attack is already likely.

156
00:12:15,740 --> 00:12:18,852
A study published in 2020 showed that the best way of

157
00:12:18,852 --> 00:12:22,080
predicting heart attacks in patients without chest pain

158
00:12:22,080 --> 00:12:25,019
is by looking at the buildup of low-density plaque.

159
00:12:25,019 --> 00:12:31,899
Essentially, patients with more than 4% low-density plaque
are around five times more likely to have a heart attack.

160
00:12:31,899 --> 00:12:37,299
As a result, the focus is now on finding
ways to quantify the buildup of the plaque.

161
00:12:37,299 --> 00:12:46,259
The AI used by CLEARLY has been trained with over
10 million CT images from over 40,000 patients.

162
00:12:46,259 --> 00:12:54,580
As a result, it can rapidly interpret CCTA scans, providing
quantified plaque readings and highlight risk features.

163
00:12:54,580 --> 00:13:00,059
It also reduces the number of cases where it's
necessary to perform an invasive angiogram.

164
00:13:00,059 --> 00:13:05,908
So Matt, if we go back to just the radiology and the advances there, is it

165
00:13:05,908 --> 00:13:11,600
threatening the radiologist's job or are radiologists embracing this and

166
00:13:11,600 --> 00:13:16,123
it's actually just making them much more effective, being

167
00:13:16,123 --> 00:13:20,100
able to pinpoint better, detect earlier, et cetera?

168
00:13:20,100 --> 00:13:22,220
How would you characterize that?

169
00:13:22,220 --> 00:13:23,220
That's a good question.

170
00:13:23,220 --> 00:13:29,139
I mean, I guess there's always early adopters and then later
adopters pretty much with any technology that comes along.

171
00:13:29,139 --> 00:13:33,472
I think generally we see younger physicians are more open to trying the

172
00:13:33,472 --> 00:13:37,925
technology, but I think it's a case of once anyone tries it, they see the

173
00:13:37,925 --> 00:13:42,739
efficiencies that it enables, they see the improvements in the standard of care.

174
00:13:42,740 --> 00:13:47,580
And I think it's just a matter of time until
kind of adoption starts coming up the curve.

175
00:13:47,580 --> 00:13:48,580
Yeah.

176
00:13:48,580 --> 00:13:52,620
I mean, it's really fascinating to see how quickly that space is evolving.

177
00:13:52,620 --> 00:13:53,620
Michael.

178
00:13:53,620 --> 00:13:54,620
Yeah.

179
00:13:54,620 --> 00:14:04,620
You know, this gets to a more fundamental question about AI and whether
it will replace humans and wipe out jobs or will it actually create jobs.

180
00:14:04,620 --> 00:14:14,500
And I don't want to interrupt Matt's train of thought here, but I'll add
one more example on the device side that I kind of touched on earlier.

181
00:14:14,500 --> 00:14:24,379
So GE Healthcare is using AI in their MRI machines to
reduce the amount of time it takes to acquire an image.

182
00:14:24,379 --> 00:14:33,860
And their algorithm uses deep learning models to reduce the
noise in the raw data, which improves the signal to noise ratio.

183
00:14:33,860 --> 00:14:40,899
So pictures have higher resolution, but more importantly,
they take about half the time of current machines.

184
00:14:40,899 --> 00:14:45,943
And that's a really important point because spending 15 minutes inside of a

185
00:14:45,943 --> 00:14:51,185
claustrophobic tube instead of a half an hour or longer is a big deal to a lot

186
00:14:51,185 --> 00:14:56,361
of patients like me, but it's a bigger deal for the provider because they can

187
00:14:56,361 --> 00:15:01,735
now serve twice as many patients with that machine, which means twice the return

188
00:15:01,735 --> 00:15:06,580
on their initial capital investment and faster appointments for patients.

189
00:15:06,690 --> 00:15:11,059
It's very complimentary to the existing radiologist business.

190
00:15:11,059 --> 00:15:12,059
Exactly.

191
00:15:12,059 --> 00:15:16,299
And Bo, I mean, what about generative AI?

192
00:15:16,299 --> 00:15:21,279
What we've been talking about so far sounds
like all of the old school type of AI.

193
00:15:21,279 --> 00:15:29,179
So how does this revolution of chat GPT and the
like, how does that fit in into all of this?

194
00:15:29,179 --> 00:15:36,659
Yeah, the launch of chat GPT in late 22 really
captured investors' imagination and attention.

195
00:15:36,659 --> 00:15:47,460
It put a spotlight on generative AI, which we'll define as the
ability for a machine to create unique content, just like magic.

196
00:15:47,460 --> 00:15:53,129
But it did take some attention away from predictive AI or plain old AI or what

197
00:15:53,129 --> 00:15:58,583
we used to just call AI, which let's define as the ability for a machine to

198
00:15:58,583 --> 00:16:02,961
kind of teach itself or extract patterns from huge data sets

199
00:16:02,961 --> 00:16:06,980
that it's never seen before and then predict an outcome.

200
00:16:06,980 --> 00:16:18,860
And that's the brains behind machine vision and image recognition and natural
language processing, big data analysis and all the other AI tools that we use.

201
00:16:18,860 --> 00:16:24,059
And it's important not to lose sight of the potential for predictive AI.

202
00:16:24,059 --> 00:16:28,109
But at the same time, we just need to be really focused

203
00:16:28,109 --> 00:16:31,869
and aware of the potential impacts of generative AI

204
00:16:31,869 --> 00:16:35,340
because it's a hugely, enormously powerful tool.

205
00:16:35,340 --> 00:16:43,100
We walked through an example of the doctor-patient interaction
and some of the ways generative AI comes into play there.

206
00:16:43,100 --> 00:16:47,311
And if you've ever seen a doctor, and most of us have, you are familiar with

207
00:16:47,311 --> 00:16:51,631
the after visit summary that describes the interaction and any recommendations

208
00:16:51,631 --> 00:16:55,460
from the doctor and do's and don'ts and the follow-up that's required.

209
00:16:55,460 --> 00:17:01,100
And that has to be written, or at least transcribed by the doctor today.

210
00:17:01,100 --> 00:17:06,740
With generative AI, that document is done
before you even leave the examination room.

211
00:17:06,740 --> 00:17:14,579
And the doctor just has to quickly review and sign off on
it while the appointment is still fresh in his or her mind.

212
00:17:14,579 --> 00:17:19,954
So not only does that promise more accurate reporting, because the doctor is not

213
00:17:19,954 --> 00:17:25,195
trying to remember what happened the day before in one of his 25 meetings with

214
00:17:25,195 --> 00:17:30,570
patients, but it can literally save hours per clinic day, which is time that the

215
00:17:30,570 --> 00:17:35,214
doctor can use to see more patients, which means less waiting for you

216
00:17:35,214 --> 00:17:39,659
and me for an appointment and more money for the doctor's practice.

217
00:17:39,660 --> 00:17:41,619
And this is there today.

218
00:17:41,619 --> 00:17:48,180
Microsoft has a health care division that has
already built this generative AI copilot.

219
00:17:48,180 --> 00:17:54,940
And I think it's just a matter of time before we see
it work its way into the health care system broadly.

220
00:17:54,940 --> 00:17:56,259
Sure.

221
00:17:56,259 --> 00:17:58,859
Well, thank you, Matt.

222
00:17:58,859 --> 00:18:03,140
And thank you, Bo, for really a great conversation.

223
00:18:03,140 --> 00:18:09,980
I mean, clearly, AI is pretty dramatically transforming the health care sector.

224
00:18:10,980 --> 00:18:12,700
Yeah, thanks for having me.

225
00:18:12,700 --> 00:18:14,539
Thanks, everybody, for listening.

226
00:18:14,539 --> 00:18:15,539
Thank you.

227
00:18:16,750 --> 00:18:20,750
Thank you for joining us today on TCW Investment Insights.

228
00:18:20,750 --> 00:18:26,349
For more insights from TCW, please visit tcw.com/insights.

229
00:18:26,349 --> 00:18:33,509
This material is for general information purposes only and does not
constitute an offer to sell or solicitation of an offer to buy any security.

230
00:18:33,509 --> 00:18:36,893
TCW, its officers, directors, employees, or clients may have

231
00:18:36,893 --> 00:18:40,054
positions in securities or investments mentioned in this

232
00:18:40,054 --> 00:18:43,769
publication, which positions may change at any time without notice.

233
00:18:43,769 --> 00:18:48,025
While the information and statistical data contained herein are based on sources

234
00:18:48,025 --> 00:18:52,017
believed to be reliable, we do not represent that it is accurate and should

235
00:18:52,017 --> 00:18:55,589
not be relied on as such or be the basis for an investment decision.

236
00:18:55,589 --> 00:18:59,217
The information contained herein may include preliminary information

237
00:18:59,217 --> 00:19:02,476
and/or "forward-looking statements." Due to numerous factors,

238
00:19:02,476 --> 00:19:05,630
actual events may differ substantially from those presented.

239
00:19:05,630 --> 00:19:11,650
TCW assumes no duty to update any forward-looking
statements or opinions in this document.

240
00:19:11,650 --> 00:19:17,750
Any opinions expressed herein are current only as of
the time made and are subject to change without notice.

241
00:19:17,750 --> 00:19:20,589
Past performance is no guarantee of future results.