Presenter: Ling Shao, Founder and CEO, SpectrumAi
Category
🤖
TechTranscript
00:00My name is Ling Hsiao. I'm the founder and CEO of Spectrum AI.
00:04And I'm here to tell you a story of how an autism startup
00:09found and stumbled into a really unique business model
00:13of how to collect labeled human behavior video data.
00:18Now, at our core, we're trying to help improve the quality of care for kids in the spectrum.
00:24But in the process, we managed to find a way where we can get paid
00:28to collect this labeled human behavior data that has implications far beyond autism
00:34and even far beyond healthcare and in other industries, for example, education.
00:39But let me start at the beginning of my story.
00:42So, I'm a mom of four boys who are all in the autism spectrum.
00:46And for those of you who are not familiar with autism,
00:49it is, in essence, a condition where kids have really significant challenges
00:55learning language and how to socialize with others.
00:59And even after they learn language, sometimes our kids are labeled as acting like robots
01:05because their physical, kind of how they portray themselves physically
01:11does not necessarily line up with how they want to communicate.
01:14And sometimes they speak with a really flat intonation.
01:18Now, full disclosure, I went to MIT as a nuclear engineering undergrad.
01:23So, I'm very familiar with being socially awkward, which makes this moment really awkward right now.
01:29So, I thought my kids were going to follow in my footsteps.
01:32But as it turns out, they had a real medical condition.
01:35And because they all started out nonverbal, the only treatment covered by insurance
01:42was called applied behavior analysis.
01:44And what that means at its core is basically you have clinicians using behavior data
01:51of how your kids respond to try to figure out how to teach them, how they learn,
01:55what's the right education protocol to apply to them.
01:59And now, my little guys, they all learned through what I call pattern recognition,
02:06which means that when we were teaching our youngest son the word for ball,
02:10we showed him a picture of an orange plastic ball and said, that's ball.
02:15But he could not generalize ball to a picture of a green ball or a purple ball
02:20or even the 3D actual object of that orange plastic ball.
02:25We had to show him tons of different photos and even 3D images or 3D objects to say,
02:31what are the possible things that could correlate with this word?
02:37Does that sound like something else we train?
02:40Right?
02:41So, I had this light bulb moment of, oh my gosh, the way we train our kids how to use language
02:47and socialize with others has a really close parallel with how we train our AI and ML models.
02:54But let's hold on to that light bulb moment because I had a more pressing need.
03:00I had to figure out how do I find the best clinicians with the best clinical protocols to treat my kids.
03:06And as it turns out, as you heard in the intro, I am not alone.
03:10The incidence of autism diagnosis has increased over 300% in the last 20 years.
03:17In the last five years, we now have mandatory insurance coverage in all 50 states.
03:24So, I started Spectrum AI as a way to pay it forward and help all families like mine
03:29be able to access high-quality clinical care.
03:32The way I was going to do that is to marry the insurance reimbursement to pay higher
03:38for those clinicians who actually deliver better outcomes.
03:41Now, that sounds super obvious and super simple.
03:44And for the healthcare folks in the room, that's called value-based care.
03:47But it is incredibly difficult to pull off because we don't have objective data
03:52that the health insurance companies can audit to prove if these kids are getting any better.
03:58So, the way we solved that at Spectrum AI was we built an electronic medical record,
04:04a clinical documentation portion of it, to enable therapists to collect data
04:09during their sessions.
04:11And more importantly, they were labeling the video of the actual sessions.
04:16So, we've recorded the video of each session, and as they were taking data
04:19on how the kids were behaving, they were actually labeling our video.
04:23Now, this is important for a couple of different reasons.
04:27One, it helps clinicians collect data easier and more time efficient.
04:33So, they're spending more time with the kids and less time creating reports.
04:38But more importantly, it gave health insurance company an objective bank of data sources
04:45that they could use to compare across aggregate physicians.
04:49And then, because they could audit the system, they can actually reimburse
04:53based on the outcomes the clinicians were actually delivering.
04:56So, I'm really happy to say that three years into our startup, this year,
05:00we had the first national health insurance company launch an initiative
05:04where they paid providers who used our software a higher reimbursement rate
05:08because they had access to this objective data.
05:11And in Q1, we're actually going to launch with two large Bay Area employers,
05:15I'm sure you guys have all heard of, to allow their families to access
05:20higher quality providers as part of their employee benefits.
05:25So, in the near term, I think it's a pretty easy jump for you guys to see
05:29how we're going to use this data to actually start to automate labeling
05:33of some of these more common behaviors.
05:36So that, again, saves the clinicians more time.
05:38But in the longer term, as we collect a larger and larger data set
05:44of accurately labeled human behavior data, because remember, it's being labeled
05:48by clinicians who are literally trained to label human behavior data,
05:53the use cases of this data goes far beyond just improving autism care.
05:59So, for example, education.
06:01Imagine a world where we can measure and figure out which kids actually learn
06:05better using which teaching techniques.
06:08I started my career out as an Army nurse,
06:11and I can definitely see a military use case for this.
06:14Imagine the military using this to assess behavior of their troops
06:18as we think about discharging them for battlefield readiness.
06:24Now, I'm going to end with bragging a little bit about our tech,
06:28because here is actually footage of my youngest.
06:31Remember the kid who couldn't generalize the word ball?
06:35This is actually a screenshot of the interface that the supervisors
06:40or the frontline staff use to see the video that's been labeled
06:44during his session.
06:46This allows them to both edit and annotate, review and annotate
06:50what is actually being labeled,
06:52so they can make the data source even more accurate.
06:54And two, this is the data that insurance companies can review and audit
06:58to see what progress the kids are actually making
07:01so they can validate why they're making these higher reimbursements.
07:05And if you zoom in, you can see that while we made this system
07:11to really help label the behaviors of these kids
07:14to figure out which protocols would help them best,
07:16this label data that we have generated and get paid to generate
07:21is really what we think could be the best value of our company,
07:25and it's something that we've stumbled into
07:27but believe have far-reaching applications
07:30beyond just helping these kids.
07:32So we started using this data to help kids who acted like robots
07:39how to communicate and socialize with their peers,
07:43but nothing stops us from using this data to literally teach actual robots
07:48how to interact with each of us.
07:50So for any of you who feel like this data could be used
07:54for something you're working on,
07:56I hope you will join me on my journey to use this data.