These are heady times for using artificial intelligence to extract insights from healthcare data—in particular, from the tidal wave of information coming out of fields like genomics and medical imaging.
Yet as innovations proliferate, some age-old business questions have come to the fore. How can startups make money in this emerging field? How can healthcare companies use AI to “bend the curve” of increasing healthcare costs? And, ultimately, how can they get buy-in from government regulators, insurers, doctors, and patients? These were some of the issues that emerged this spring when Xconomy brought together some of San Diego’s most-prominent tech and life sciences leaders for a dinner discussion about the risks and opportunities in the convergence of AI and healthcare.
“Being a healthcare investor, I love the fact that there’s interest now on the tech side,” said Kim Kamdar, a partner in the San Diego office of the venture firm Domain Associates. “It opens up a whole new avenue of potential co-investors for our companies.”
The consensus: It’s still early days for applying machine learning and related techniques in healthcare, and it’s hard to foresee how these innovations will play out. As Xconomy senior editor Jeff Engel has reported, questions abound over the impact AI will have on doctors and healthcare institutions. Yet there is little doubt that transformational change is coming, and tech companies ranging in size from small startups to corporate titans like IBM and GE are scrambling to gain a foothold in this emerging field.
If ever there was a sector in need of transformational disruption, it would be healthcare, where spending in the United States amounts to more than $3.2 trillion a year—and accounts for close to 18 percent of the U.S. gross domestic product.
The sector represents a lucrative-but-daunting target for investors—complicated by regulatory issues, a healthcare system that separates the interests of patients, providers, and payers, and an investment timeline that can take 10 years or more to realize.
There may be no better example of the potential opportunities than Grail, the $1 billion-plus startup spun out by Illumina (NASDAQ: ILMN) to advance diagnostic technology sensitive enough to detect fragments of cancer DNA in a routine blood sample. Yet cautionary tales also abound—most notably with Theranos, the venture-backed diagnostic company that was valued at $9 billion in 2015—and plunged last year to less than a tenth of that.
Interest in healthcare AI runs high in San Diego, which has a well-established life sciences cluster and is home to two genome sequencing giants: Illumina and the life sciences solutions group of Thermo Fisher Scientific (NYSE: TMO). San Diego also has some resident expertise in neural networking technologies that accompanied the rise of HNC Software, a developer of analytic software for the financial industry that is now used by FICO (NYSE: FICO) to predict credit card fraud, among other things. (FICO acquired HNC in 2002 in a stock deal valued at $810 million.)
The dinner conversation that Xconomy convened included Kamdar and other local investors, data scientists, healthcare CTOs, startup founders, academic researchers, and digital health executives. The kickoff question: Is there a proven business model for startups that are applying innovations in machine learning in the life sciences?
The model that came to mind for Larry Smarr, director of the California Institute for Telecommunications and Information Technology (Calit2) headquartered at UC San Diego, was Illumina itself. The company is a pioneer in DNA-sequencing technology, and increasingly, for analyzing genomic data—that is, the genetic variations and biological functions embedded in the code.
“Their cloud solution for analyzing the human genome is pretty substantial,” Smarr said. “The data requires the kind of analytics that we’re hearing [about] around the table. It didn’t used to. But the volume now is on an exponential [scale]. So you really can’t get any healthcare insights out of the data without these algorithms, particularly in genomics and the microbiome.”
Illumina found customers for its gene sequencing technology and data services at genomic research centers, clinical research organizations, research institutions, and biotech and pharmaceutical companies. But, is that a model that can be replicated? In other words, if there was a business built around analyzing data from the microbiome, say, what would it look like?
Smarr pointed across the table to Rob Knight, who holds joint appointments in pediatrics and computer science at UC San Diego. Knight is director of the UC San Diego Center for Microbiome Innovation and a co-founder of the American Gut Project, an exercise in “citizen science” that has collected over 16,000 stool samples to help scientists better understand the role that microbes play in human health.
“So remember, I’m running that as a non-profit,” Knight said. “I think it’s going to be really difficult, because in general, companies that have been based around selling access to DNA sequences have not done particularly well. I’m thinking of Celera, for example, which switched from that model to diagnostics.”
“What I think we need to do is to somehow move that into real time, and discover how to develop a user interface for the microbiome, so you can tell like that”—Knight snapped his fingers—“whether that piece of bread you just ate is having a positive or negative effect for you.”
One company pursuing a strategy along these lines is Tel Aviv-based Nutrino, which has been developing a mobile app and data platform to help users determine how the food they eat affects their personal biochemistry.
“They’re providing you real-time guidance on the impact of your “foodprint” as they call it, on your glucose behavior,” said Annika Jimenez, senior vice president for data at San Diego’s DexCom (NASDAQ: DXCM), which specializes in continuous glucose monitoring technology for managing diabetes.
“It’s like a premium model, but over time they’ll drive to new commercial models targeting enterprises and possibly payers,” Jimenez said.
The key advantage that A.I. provides in healthcare is the ability to extract meaningful insights from exabytes and zettabytes of data—that is, data on a scale far beyond human comprehension.
“That seems to me to be the end goal, longer term,” said Qualcomm Life president Rick Valencia, who seemed skeptical of current strategies in the field for generating revenue. “In the short term, I think the answer to your question is ‘No.’ I just don’t know of anyone at scale who has proven out a business model…[It’s still] early days.”
Navid Alipour, a co-founder and managing partner of San Diego’s Analytics Ventures, said his firm’s portfolio company CureMatch is taking a direct-to-consumer approach, in which cancer patients pay CureMatch to recommend the top three combinations of chemotherapy drugs for each patient’s cancer. The recommendations, based on information in a patient’s own medical record, is intended to help cancer specialists choose a treatment regimen. CureMatch says it uses supercomputer processing to sort through millions of possible three-drug combinations, assessing each combination for factors like unwanted drug-drug interactions, and correlating genomic data to rank the best drug combinations for a specific patient.
CureMetrix, another company in Analytics Ventures’ portfolio, uses machine learning to analyze mammography images for breast cancer—and must still get FDA approval before it can be used in the United States, Alipour said. “It will be a [software as a service] model,” Alipour said. “But we have an institutional investor in Mexico that’s taking us into the top levels of the government there. Breast cancer is a huge problem in Mexico, and there are not many radiologists with a mammography expertise in the country. We’re licensing to the entire country because they have a national healthcare system. So that’s something to think about if you’re outside the U.S. and our insurance system.”
CureMetrix is one of many companies, big and small, that have been applying machine learning to identify anomalies in diagnostic imaging, and image-based pattern recognition seems like “the ultimate use” of the technology, Jimenez said. “But all you have to do is go to [the Strata Data Conference], which is kind of ‘the event’ for big data and data science for the tech community, and the keynote speakers talk about how difficult that use case really is. So you know, it’s maybe not for 10 years…maybe a little bit longer.”
So, when might AI systems supplant radiologists?
Smarr said he was doubtful that artificial intelligence would replace radiologists altogether. Rather, he believes the technology will be used to augment human capabilities—making the worst radiologist more accurate than the best human radiologists could be on their own. “So what you are doing is bringing up the human talent level by augmenting it with vast amounts of data they could never have experienced themselves,” Smarr said. “And I really think that could be more productive in the short term, meaning in the next several decades.”
For companies like DexCom that are focused on the diabetes epidemic, Jimenez said the holy grail is modifying patients’ behavior. That would mean combining the stream of data from glucose monitoring, insulin measurements, patient activity and meals, and applying machine learning to derive insights so the software can send alerts and recommendations back to patients and their doctors, she said.
“But where we are in our maturity as an industry is just publishing numbers,” Jimenez explained. “So we’re just telling people what their glucose number is, which is critical for a type 1 diabetic. But a type 2 diabetic needs to engage with an app, and be compelled to interact with the insights. It’s really all about the development of the app.”
The ultimate goal, perhaps, would be to develop a user interface that uses the insights gained from machine learning to actually prompt diabetic patients to change their behavior.
This point was echoed by Jean Balgrosky, an investor who spent 20 years as the CIO of large, complex healthcare organizations such as San Diego’s Scripps Health. “At the end of the day,” she said, “all this machine learning has to be absorbed and consumed by humans—to take care of humans in healthcare.”