How Lumiata wants to scale medicine with machine learning and APIs

Ash Damle’s parents are doctors, so were his grandparents. But Ash isn’t a doctor. He’s just an MIT-trained artificial intelligence expert who wants to mimic their brains in software.

Well, not mimic their brains so much as scale their knowledge. His mom, for example, is a well-educated pediatrician with years of experience treating sick children. Should she really be spending her time dealing with coughs and colds, he asked during a recent interview, or should her energy be focused kids with more serious problems? It’s a question that rings even truer in areas suffering from a shortage of primary care physicians.

Ash Damle
Ash Damle. Source: Lumiata

But putting some of a doctor’s knowledge into the hands of, say, nurse practitioners or registered nurses required Damle to answer one big question: “How does a physician’s brain work?” he asked, rhetorically. “… It’s not alchemy, there is some logical process.”

The answer came to him about four years ago as the “datafication” of health care really took off with broader use of electronic records, medical sensors and devices. He saw the possibility of using graph analysis — think Facebook’s(s fb) social graph or Google(s goog) PageRank — to connect all the pieces of data physicians see about patients on a daily basis with all the knowledge they’ve acquired over the years. Now, his technology is being tested in about 10 hospitals and other health care facilities and his startup, called Lumiata, has just raised $4 million from Khosla Ventures.

(This actually isn’t the first time graph analysis has been applied to health care. We’ve covered cancer research, life sciences startups and even machine learning startups that have applied graph theory, in some form to medical data. Speakers will talk more about how it, and other cutting-edge techniques, can be applied to data across all sorts of industries at our Structure: Data conference in March.)

On the backend, Lumiata is a big data system that has ingested more than 160 million data points from textbooks, journal articles, public data sets and other places in order to build graph representations of how illnesses and patients are connected. The technical term for what Lumiata does is “multi-dimensional probability distribution,” but, Damle explained, it basically comes down to understanding how time, location and behavior come together to affect how a disease develops and progresses.

A very simple example, he said, are symptoms such as pneumonia and confusion. They’re fairly common in younger patients and older patients, but not really in between. Lumiata might help a nurse  — even a doctor — who heard a 35-year-old complaining about these symptoms realize that something abnormal is afoot and then instruct him or her on potential follow-up questions. A serious condition might require hospitalization; a less-serious one might require a follow-up visit.

A demo of using Lumiata to assess a patient.
A demo of using Lumiata to assess a patient.

Because Lumiata’s technology is able to analyze graphs containing tens of thousands of nodes (e.g., symptoms, diseases and patient data points) and weighted edges (e.g., the connections between them all and how strong they are) in well under a second, it can be used for real-time care — like when a patient visits a hospital — or even more proactive care, Damle explained. A hospital, or even an insurance provider, could actually analyze its entire list of patients to figure out, for example, who might be coming up on a critical time in their recovery from a heart attack and send out a text message asking if they’re experiencing certain symptoms.

The delivery method for Lumiata’s knowledge is a REST API, which means it can tie into existing health care systems and software. That’s potentially useful from a usability perspective (people like the tools they like) but also because it means Lumiata can connect with existing patient records to get an even better sense of how a particular disease might be affecting a particular person.

It’s about bringing in all the small weak signals, things adding up over a year, and asking, “How can we figure out what’s being missed?” Damle said. In the case of a medical device is streaming data back to a server, the same readings could mean different things for different people. There, Damle said, the question becomes “How do I analyze that stream of data within the context of you?”

A visual timeline of how the various health data elements of a patient are associated with one another over a year-long period. Source: Lumiata
A visual timeline of how the various health data elements of a patient are associated with one another over a year-long period. Source: Lumiata

People naturally want to compare Lumiata to IBM’s(s ibm) Watson system because of how both ingest lots of external data and try to help doctors make diagnoses. However, Damle noted, the ability to factor in patient-specific data and the effects of time and place make Lumiata’s output more of a clinical model of a patient, against which a nurse of doctor might ask “what if” questions, and less of a recommendation.

“It’s OK to give the wrong movie recommendation. It’s not to OK give the wrong potential diagnosis,” Damle said.

Actually, Damle said, making sure Lumiata’s graph is accurate is one of the reasons the company has taken years to launch publicly after its creation (it has even undergone a name change from MEDgle in that time). Only medical conclusions with a sufficient level of support make it into the graph, and Lumiata’s team of computer scientists and doctors has spent more than 20,000 so far curating it to make sure the connections the system finds are actually medically sound. They’re also working with academic institutions to validate some findings via controlled studies and publish papers.

Part of that focus on accuracy also means being upfront about what Lumiata is and how users should interpret its findings. In the end, the judgment of a medical professional still matters a lot in making a final determination. That’s one reason the company chose to build the backend using graph analysis, which lends itself to human investigation, rather than more black-box-like AI techniques such as neural networks.

“We’re very, very careful,” Damle said, “of trying not to make claims of ‘It’s this and let’s start the amputation.'”