Presenter: Dr. Peter Kecskemethy, Co-founder and CEO, Kheiron Medical Technologies
Category
🤖
TechTranscript
00:00 Thank you very much.
00:00 It's an absolute privilege to be here.
00:03 And I thought about that probably everyone here
00:08 has heard a lot about what AI could do,
00:12 what the risks could be, how it could be bad for humanity,
00:15 how it could be good for humanity.
00:17 But I'm bringing you slightly different things today.
00:20 And that is a very good news story,
00:23 how AI in practice is actually bringing impact right now,
00:27 today.
00:28 It's not a future, it's not a hypothesis.
00:30 It's actually what's happening now.
00:31 And I'm also going to bring you a secret, hopefully,
00:33 that you can take away from today,
00:36 and maybe a bit of motivation.
00:37 So let's get into that.
00:40 So my company, Chiron, has developed an AI called
00:45 Mia that is supporting doctors in early detection,
00:49 specifically breast cancer screening at the moment.
00:52 And that has got quite a bit of news coverage.
00:57 We were on the front page of the New York Times.
00:59 We were also named after our results
01:02 to be one of the seven biggest breakthroughs of 2023,
01:05 next to CRISPR and messenger RNA.
01:07 So in good company.
01:09 And there's a very good reason why that is the case.
01:13 One of the reasons is we are solving a very hard problem.
01:17 People may not know that breast cancer screening is very, very
01:19 complicated as a diagnostic task.
01:22 It's so complicated, some people say
01:24 it's like finding a snowflake in a snowstorm.
01:27 And actually, the practice of medicine
01:29 covers that, pretty much reflects it.
01:32 Most countries, outside the US, two doctors
01:36 are required to look at every single mammogram, so basically
01:38 every single image that is done for breast cancer screening.
01:42 And even with two doctors, we get at least 20%
01:45 of the cancers missed.
01:47 That means at least every fifth cancer gets missed.
01:50 That's like the best we can do across the globe.
01:52 And in some places, over 70% of the cancers get missed.
01:55 This is the best we are doing in screening right now
01:59 as humanity.
01:59 And you can see a couple of examples
02:02 of a hard case on the left and a hard case on the right.
02:05 On the left, a cancer was missed.
02:07 And on the right, a lot of the times
02:09 the doctors look at all the little tiny nodules
02:11 and they're thinking, OK, which one of these
02:13 is potentially a cancer, which one is not?
02:15 It's extremely hard.
02:16 Some people call it magic sometimes.
02:19 So what can AI do?
02:22 I'm going to talk you through this particular case.
02:26 This is an actual case where a woman went for breast cancer
02:30 screening regularly, so did the best she could do,
02:35 multiple years, and she died of breast cancer.
02:39 And she died of breast cancer because the cancer was not
02:41 picked at the last screening.
02:43 It was not picked up two years before.
02:45 And it was also not picked up four years before.
02:47 Every single time, missed by two doctors.
02:50 This is reality.
02:52 And what can the AI do?
02:54 The AI-- we ran the AI on many thousands, hundreds
02:58 of thousands of cases.
02:59 What the AI could do is find it at the last screening,
03:02 find it two years before, and find it
03:05 on the case for four years before.
03:06 So this could have been over four years earlier detection.
03:10 And we know that when breast cancer is detected early,
03:13 it's over 90% chance that a patient survives.
03:17 That could have been practically a life saved with AI.
03:20 So that's the impact what we can have on a single patient.
03:24 So let's see, what does it mean in reality?
03:27 So what you see here is how a doctor
03:30 looks at a particular case.
03:33 This is a medium complexity case.
03:35 It's not like a high complexity case.
03:37 You see a couple of nodules, but also it's relatively clean.
03:41 In this particular case, again, two doctors missed the cancer,
03:45 and the AI found the cancer.
03:47 This is not a very, very hard case.
03:49 This is a medium case, but the doctors
03:51 have to go through hundreds and hundreds of cases every day.
03:53 There's just not enough energy or attention for the doctors
03:56 to do their jobs well unless they get good tooling,
03:59 good help.
04:00 And I believe AI is one of the fundamental ways we can help.
04:06 In fact, that's exactly what we have shown in live use.
04:09 So what does it mean?
04:12 It means we can have extremely strong impact on a patient,
04:17 on the life of a single patient.
04:19 It also means, because we're talking about software,
04:23 this is automatically scalable.
04:24 You bring a piece of software, a box, to a hospital,
04:27 and you can help doctors start saving more lives.
04:30 That's practically what we're talking about.
04:32 And so roughly the impact that we
04:34 could have that Chiron is focusing on
04:36 is scaling it across the globe.
04:38 So we're estimating we can probably
04:40 help doctors save about 2,000 to 3,000 lives
04:43 every single year in the UK.
04:45 40,000 to 60,000 lives in the US, and hundreds of thousands
04:49 across the globe.
04:51 This is a piece of software that you bring across completely
04:54 scalably right now today.
04:56 And this is not a hypothesis.
04:57 This is based on actual evidence, actual results
05:02 that we have right now in multiple clinical trials
05:05 and multiple assessments across the globe.
05:10 So why is this big?
05:13 Why is it different from the past?
05:15 Actually, computer science and computer vision
05:17 has been used a lot, or at least there
05:20 were a lot of attempts to be used for breast cancer
05:23 screening, but not with too much success.
05:26 And there's a couple of reasons for that,
05:30 but I'm going to talk you through how we had
05:34 a slightly different approach.
05:37 We took two things extremely seriously.
05:39 And I think the importance here is
05:40 what can you learn from that?
05:41 How can you translate it to how you use AI
05:44 or how you develop AI?
05:46 And the two components of that is taking AI seriously,
05:50 that it's a very different type of technology
05:52 than humanity had before, and taking the domain seriously.
05:55 We're talking about a domain where any kind of decision
05:59 impacts a life, either way.
06:03 Either you can over-diagnose or you can potentially
06:06 miss a cancer.
06:07 So taking the AI seriously and taking the domain seriously
06:11 means that you need to collect a lot of data.
06:13 You need to have extreme level of validation.
06:16 I just need to make sure that when the software goes
06:19 to health care, that you're actually
06:21 impacting everyday the patients and the doctors.
06:24 It has a positive impact rather than a negative impact.
06:27 So if you want to know the details,
06:29 you can look it up.
06:31 I have a number of papers in Nature published,
06:34 and also a number of other publications.
06:37 But I'm just going to distill for you
06:39 some of the secret sauce that I believe we all
06:41 need to think about.
06:43 So hopefully, you can use that in your everyday practice.
06:48 And the secret here is that AI is not just a model.
06:52 You have a good model that you think is good,
06:55 is not going to necessarily-- or most likely--
06:58 not going to succeed in practice.
07:00 AI is a type of technology that we have never
07:02 had before as humanity.
07:04 Like, if you have a piece of hardware or software,
07:07 you know what it's used for.
07:08 You test it in a lab, and pretty much that's
07:10 how you're going to use it in everyday practice.
07:13 Your software, AI, specifically, it's
07:16 designed to be able to be intelligent and good on input
07:21 and in circumstances that it has never seen before.
07:24 This is a completely different-- that's
07:26 the difference between AI and other technologies.
07:29 So what does it need?
07:31 Of course, the models need to be good.
07:33 It needs to be good in the lab.
07:34 But also, it needs to be good wherever else you want to use
07:37 it.
07:38 So the algorithms have to be good.
07:39 OK, fine.
07:40 Whenever-- but we need to validate that, again,
07:44 both in the lab as well as in real world use.
07:47 However, we never know, because it's a black box,
07:49 whether it still generalizes.
07:50 Does it still work if I bring it to a specific hospital?
07:54 You need to validate that locally in deployment.
07:56 So that's the deployment stage.
07:58 And also, because it's a black box, it might work one day,
08:00 but it might not work next week, or might not
08:02 work a month or a year later.
08:04 You need to keep monitoring that it is still
08:06 providing the same performance.
08:08 So you need four things for AI to work in practice.
08:10 You need to get the algorithms right.
08:12 You need to validate it.
08:13 You need to make sure it works in practice,
08:17 both upon deployment as well as over time.
08:20 So please, please, if you are impacting lives
08:22 or you're having some kind of very, very important impact
08:26 with AI, do you make sure all of those work?
08:28 Because all of those are covered.
08:29 Otherwise, it's not going to be very good.
08:32 So after that, I'm actually going
08:34 to leave you with the words of one doctor and one patient who
08:39 can tell you about what it means for them
08:43 having used the AI in the UK.
08:46 [VIDEO PLAYBACK]
08:48 - I remember when I was a child, they
08:49 were saying they would eventually
08:50 find a cure for cancer, but it will take a long time.
08:54 And this is-- it's all just progression
08:56 down the right road.
08:57 - We read 5,000 mammograms a year.
09:03 There is a lot of pressure on the service.
09:06 I have starved shortages.
09:08 People are burnt out.
09:10 We have to make our decision first.
09:12 Then we are allowed to see the opinion of me.
09:15 So we press the little button.
09:16 It opens up and shows us yellow circles.
09:19 It says, can you have a second look here?
09:21 - I was quite surprised to hear back from them.
09:24 The penny didn't drop immediately
09:26 until they said they had found a very small amount of cancer.
09:31 And they said it wouldn't have been picked up by the human eye.
09:34 It was too small.
09:35 But the computer earmarked it, and that's how they found it.
09:40 - At the moment, we are using the AI in an evaluation process.
09:43 So we haven't integrated it fully into the system.
09:47 We have modeled that if we did have
09:49 the AI in our workflow, that would be a big difference
09:52 because it would make our turnaround time go from probably
09:54 14 days at the moment down to three days.
09:57 The technology helps move us forward.
10:00 This is the beginning.
10:01 Large scale, we can make a difference.
10:04 - It's a lifesaver, a life changer.
10:08 - So I will just leave you with that.
10:11 I think when we have enough evidence that AI works,
10:13 it's actually our moral obligation to use it.
10:16 Sometimes we need to be very careful.
10:18 Sometimes we need to go through all the steps to validate it.
10:21 But when we do, and it can have a very strong impact,
10:24 we need to actually use it.
10:25 And if you want to learn more, feel free to contact me.
10:28 [APPLAUSE]
10:29 [MUSIC PLAYING]
10:33 [BLANK_AUDIO]