AI use in healthcare ramps up for app maker Cognoa
Applications of AI in healthcare have been relatively restricted due to regulatory and data challenges, but one startup is finding ways to make AI effective.
The healthcare industry has not traditionally been considered a technology pioneer, but in spite of some built-in obstacles, medical technology companies are making strides, particularly with the use of AI tools.
That's the case with Cognoa, a health IT vendor that develops apps for helping parents and clinicians diagnose autism and other child behavioral health problems. Cognoa's machine learning-based tool was recently recognized by the Food and Drug Administration as a medical device to aid in the diagnosis of autism. The company currently provides an app that is available to parents and caregivers through their employers' benefits and through a network of therapy service providers.
Halim Abbas, Cognoa's chief AI officer, said AI use in healthcare has been challenging in part because privacy regulations limit what you can do with patient data, including making that personal health data less available. But new technology is emerging that's making it easier to for data scientists to get the data they need to run more complicated models and get deeper answers to questions.
"The industry has dramatically improved, and a lot of these tasks that used to be challenging are now commonplace," Abbas said.
Handling data is step one
Cognoa works with a mix of data pulled from clinical patient records and from users of its app. Families can download the app and upload information about their children and take assessments. All of this feeds into machine learning models that identify certain characteristics that are associated with an autism diagnosis.
The data from medical records is typically heavily structured, while the user-generated data tends to be more free-form. Natural language processing was, until fairly recently, exceedingly difficult for most in the health IT world to do, but Abbas said effective NLP is necessary to capture information locked in natural language text to get a fuller picture of traits that could be associated with behavioral disorders.
That's the kind of thing AI is really good at, finding subtle connections in very large data sets.
Halim AbbasCognoa chief AI officer
"That's the kind of thing AI is really good at, finding subtle connections in very large data sets," he said.
Managing access to these sensitive data sets was a big challenge for Cognoa and has been the kind of thing that has slowed AI use in healthcare.
But Abbas worked with a software vendor called Immuta, which has a data governance tool that Cognoa used to define user access data types in accordance with health data privacy regulations.
Abbas' team defines user roles throughout the company and creates templates that define the level of data access roles should have. The tool then automatically applies these access rules to specific data sets, rather than locking users out based on broader access-level restrictions.
Abbas said this data governance process has sped up development of his company's AI model. Time that data scientists once spent requesting data access can now be applied to improving the model.
Don't try to start new
The proliferation of widely available deep learning models has also helped Cognoa advance its models. Rather than building every feature anew, the data science team is able to incorporate pre-built models that help minimize some of the development effort.
For example, one app Cognoa is developing enables parents to upload videos of their children's behavior to contribute to diagnosing attention deficit hyperactivity disorder. The team incorporated pre-written object-detection deep learning models to do basic tasks like identifying the child in the video. This enabled the team to focus on more important functions, such as connecting the child's observed behavior to diagnostic criteria.
Abbas said explainability is critical when developing AI tools for diagnosing child behavioral disorders. But that doesn't mean that every piece of the model needs to be explainable. In the case of the video, the model doesn't need to explain why it thinks a certain grouping of pixels in a video is a child. It needs to explain why it thinks the child's behavior is associated with a certain disorder.
This distinction has enabled Cognoa to benefit from some of the machine learning advances happening outside the healthcare industry and is the kind of approach that could make AI use in healthcare a reality.
"The guidelines in the healthcare industry pose restrictions that make doing data science and AI tricky," Abbas said. "This is how you build a big breakthrough in an industry that really needs one."
