Improvements in how autism is understood and identified have led to a significant increase in the number of children diagnosed with autism spectrum disorder (ASD). The US’s national public health agency now says that one out of every 36 children in the US has autism, up from its estimate of one in 110 in 2006. That has also led to a rise in autism support clinics.
CentralReach is a SaaS provider that was founded in 2010 by a therapist fed up with the disjointed point solutions her clinic had to contend with while caring for children on the autism spectrum. It has grown into a market leader in the electronic medical record (EMR) space – serving more than 4,000 customers who collectively support around 350,000 children with autism, a significant portion of the two million children diagnosed with ASD in the US.
Its SaaS platform combines practice management, front office and clinical capabilities in one user interface and database – and has afforded it a unique perspective into the scale and extent of the autism care gap.
Care shortages
One of the sector’s biggest issues is a shortage of board certified behavior analysts (BCBAs), who are the beating heart of autism care in the US, working with autistic children to execute on a care plan. Some 250,000 BCBAs are required to serve children with autism across the US; there are only 70,000.
“Our customers have huge wait lists and try to do whatever they can to provide services to as many individuals as they can,” says Chris Sullens, CEO of CentralReach. “But BCBA caseloads continue to average 50-60% below the level that the accreditation board for behavior analysts allows because of high turnover and large amounts of time spent on non-billable administrative tasks…”
“Our system automates much of the workflow out of the box, but despite the benefits it provides, things still fall through the cracks or get missed because of the many insurance-specific nuances with how claims are filed, or how session notes and other documentation are required to be written. UnitedHealthcare wants it one way, TRICARE wants it a different way, and so on,” he explains.
The result, providers end up getting bogged down in paperwork, which increases the chance of errors. Any errors, even simple, small ones, can have huge financial and clinical implications for a practice.
Can AI help reduce the bureaucratic toil?
In the offices of CentralReach, ChatGPT’s initial release triggered a period of playful experimentation by the company’s team of curious technicians.
When they noticed they were getting good outputs when models were properly instructed, realization set in that this could be a game-changer.
“The capacity constraints and high turnover results in a loss of knowledge over time if you can't extract it out of all of the qualitative data being collected,” adds David Stevens, head of AI at CentralReach. “We realized, what if we could use AI to bubble up all of this pertinent clinical information to the people that need it via a simple search?”
See also: From the Marines to the “4 Cs”: BigID on code camps, preventing data exposure and working with MongoDB
That can be surfacing details about a child’s care needs and behavior, so if a clinician changes, a new clinician can step in pre-armed with the full history and other vital knowledge for the individual. One vital input is the type of reinforcer that the child responds most favorably to.
As Stevens puts it “So I need to use the gold stars now for reinforcement and not the purple one, because all of a sudden, [if you get this wrong] you lose what's called instructional control, and the child runs out of the room.”
Unlocking the power of AI
CentralReach started to enhance its software with this in mind. It started by building a powerful API layer to provide access to microservices in a more individualized way and, ultimately, unlock the ability to evolve into an AI-powered EMR platform.
(It serves this via AWS, which ensures it has resilience, cybersecurity, IPAA and other compliance and minimizes heavy lifting for its engineers.)
Today, there are two layers to the company's AI solutions. identify errors that may exist across financial and clinical tasks, catching mistakes and data errors that fall through the cracks as therapists juggle the many things they have to do each day. The second layer is the agentic layer, which unlocks significant ROI and improves the day to day experiences for the therapists and administrators, allowing them to automate tasks such as schedule creation, note drafting and claims submission with AI.
See also: How SEGA’s Felix Baker delivered a data transformation
While the company’s current AI solutions are already game changing, it’s the upcoming AI-powered clinical solution, CentralReach’s Care360 care management platform that has the potential to be the most transformational for autism care. Starting out life as a search and indexing tool, the ultimate ambition is for CR Care360 to mature into a human-in-the-loop clinical guidance tool.
“It is one of our most impactful products but also probably the hardest [to build]” says Sullens. “We need to take the four billion data points we have and label, organize, structure and enhance them so that the LLM knows what they are, how they’re related to each other and can start to connect the dots between the things that need to be filled in on a form, or the results from an assessment…”
“If we can get this right, it has a huge opportunity to impact the field. It’s the big ‘one plus one equals three’, because it'll help our customers serve more children with the same number of BCBAs while improving quality and increasing revenue, which ultimately allows us to have a bigger impact on the community.”
The technical foundations
With billions of existing data points and about three million new session notes per month generated in the CentralReach system, the company needed to think carefully about which database provider to work with.
CentralReach decided to use MongoDB Atlas’ document model to aggregate its diverse data, from assessments to clinical data collection, so it could build rich AI-assisted solutions on top of its database. The MongoDB AI Applications Program, meanwhile, provided a helping hand in designing and optimizing multiple layers of the highly comprehensive buildout.
“As we evolved our clinical solution and expanded its functionality through our CR Care360 solution, we saw an opportunity to evolve our infrastructure to both solve the problem at hand but future proof the solution in the long run. It quickly became apparent that leveraging a document database, like MongoDB, was the best to do it,” says Stevens. “I had used MongoDB at a previous company and what I really appreciated about it was that we could essentially iterate as we went with the data model.
Read this: AI "fireworks" are going off. How do you know which rocket to ride?
“Everyone in the world is learning new things every day as we pursue AI initiatives. MongoDB Atlas’ flexible document model allows us the flexibility to continually iterate and enrich the information we know,” he adds.
!For example, we'll go through a clinical note and what we're starting to do now is extract metadata from that note from natural language.
“We need a good place to store that. We can easily add versions of these different structures in MongoDB documents that enrich the document over time for better and more capabilities. We tried a lot of things, threw a lot of things at the wall, and the MongoDB document model is the way to go.”
As Central Reach CEO Sullens adds: “Auditing against standard operating procedure is a very hard thing to manage at scale. A clinical standard operating procedure is how you maintain quality and treatment fidelity and procedural integrity across your service delivery. It’s a very human-driven, hard to audit, tedious thing, making sure everybody's doing the right things and doing them right.
“We're now moving into basically consuming our customers' standard operating procedures (SOP) – these are big books! They're not simple.
“We take a slice of these SOPs and say, ‘Okay, for this type of thing, here's your standard operating procedure, here's your rules.’ Then what we're doing is we're trying to extract the necessary information, enrich that through LLMs, with classification, summarization, vectorization, status… store all those resultant fields and outputs into the document model that co-resides with the transactional information, where we then can tell the LLM ‘hey, did this person do this particular step properly, to the clinical validation that our customers defined?’ It is, he thinks, a game-changer and the winners will be children who need support.”
Delivered in partnership with MongoDB
