Nobody likes getting a colonoscopy. For the people who catch colon cancer early thanks to that bowel camera, the standard screening—every 5 to 10 years from age 50 to 75—proves invaluable. For the 993 people in 1,000 who don’t test positive following a colonoscopy, the pain (and for the uninsured, the expense) can be enough to make them skip the next one. People who’ve shirked their exams often number among the 50,000 Americans who die from colon cancer each year. “More non-invasive ways of screening are needed,” says Matthew Kalady, co-director of the colorectal cancer program at the Cleveland Clinic. “If you could pick up colon cancer early and noninvasively with a simple blood test, that would be just fantastic.”
An Israeli health-tech company is trying to use machine learning software to do just that. ColonFlag is the first product from Medial EarlySign, and while poorly named, the software predicts colon cancer twice as well as the fecal exam that’s the industry-standard colonoscopy alternative, according to a 2016 study published in the . ColonFlag compares new blood tests against a patient’s previous diagnostics, as well as Medial’s proprietary database of 20 million anony-mized tests spanning three decades and three continents, to evaluate the patient’s likelihood of harboring cancer. Israel’s second-largest health maintenance organization is already using the software, and Medial (a mashup of “medical” and “algorithms”) is working with Kaiser Permanente and two leading U.S. hospitals to develop other uses for its database and analysis tools.
“Our algorithms can automatically scan all the patient parameters and detect subtle changes over time to find correlative patterns for outcomes we want to predict,” Nir Kalkstein, Medial’s co-founder and chief technology officer, says, characteristically clinical. The database allows his team “to find similar events in the past and then identify from the data correlations that can predict these events.”
Other companies are building massive databases with an eye toward predictive medicine, including heavy hitters such as DeepMind Technologies, owned by Google parent Alphabet Inc. In Boulder, Colo., startup SomaLogic Inc. is predicting heart attacks based on combinations of certain proteins in cardiac patients. In Salt Lake City, Myriad Genetics Inc. assesses hereditary cancer risks based on DNA profiles. DeepMind’s public U.K. tests have largely focused on managing understaffed wards; Myriad’s tests cost thousands of dollars; and most of the other companies have had trouble delivering actionable results. With 45 employees, Medial has the first test that’s becoming an unobtrusive, critical part of doctors’ rounds by using the cheap, easy blood tests they routinely conduct.
Kalkstein has some experience with both bureaucracy and big data sets. He served his mandatory years in the Israel Defense Forces (IDF) with an elite cybersecurity research unit, then started a company called Final (short for “financial algorithms”) on the day of his discharge, in 2001. Despite knowing nothing about finance, he proved adept at writing code that could predict stock market activity based on past market reactions to similar events. “We only looked at the data and the story it told us, without the use of any economic models,” Kalkstein says.
Final made Kalkstein a billionaire, according to the , but left him unsatisfied. By 2009, he says, he decided to “invest my time and resources only in things that will make a positive impact on people’s lives.” With some pals from college and the IDF, he started Medial out of a garage near his house in the Tel Aviv suburbs, recruiting tech executive Ori Geva as chief executive officer and consultant Ofer Ariely as chairman.
The team lacked health-care know-how, but learned quickly. In 2011, Medial held an informal contest with intensive-care-unit doctors at Rabin Medical Center, Israel’s largest hospital, to predict which ICU patients would survive. The data scientists trounced the MDs. “To see in many dimensions at the same time is very difficult,” says Varda Shalev, a primary-care physician who also runs the research and development arm of Maccabi Healthcare Services, which now uses ColonFlag. “With machine learning, it becomes easy.”
Shalev helped lead Kalkstein’s team to use colon cancer as a proving ground in 2011, a couple of years after she lost a patient a little too young for the screening. “You always blame yourself,” she says. Maccabi’s 2 million electronic patient records, stripped of identifying information, were among the first to feed Medial’s database, and Maccabi was the first to build ColonFlag into its alerts system. The software consistently identified patients at 10 times the normal risk of harboring cancer and flagged tumors six months to a year ahead of doctors’ diagnoses, while the cases were still beatable, according to a Kaiser study published last year in the journal . All told, studies by Medial, Kaiser, and Oxford University identified 100 Maccabi patients with cancer and another 100 with potentially precancerous adenomas. The HMO now uses ColonFlag on any blood test from an older patient who’s refused a colonoscopy or the chemical test.
The Israeli company will have to tread carefully to avoid the public outcry faced by companies, including DeepMind, that are building large-scale databases from patient records. “It’s critical that de-identification be done in a rigorous and responsible way,” says Deven McGraw, former deputy director for health information privacy at the U.S. Department of Health and Human Services. Some doctors are skeptical because Medial’s studies have looked at the blood tests of patients known to have later acquired colon cancer. (A blind study is under way.) And a computer’s prediction about the usefulness of a colonoscopy might never be as reliable as, well, getting a colonoscopy. Even Medial executives acknowledge that the success of Kalkstein’s past enterprises hardly guarantees success in this new field. “What Final is doing is much easier,” says Ariely, Medial’s chairman. “Their data is clean, and if they’re wrong, no one dies.”
For now, Kalkstein’s team is focusing mainly on R&D. Medial has raised $40 million from investors led by Hong Kong billionaire Li Ka-shing’s Horizon Ventures, and the company says its next product, due in the second quarter, will predict the onset of diabetes. Further away: a wearable device that seeks to predict seizures in epileptics, giving them precious seconds to pull over their cars or clamber out of their bathtubs. It’s also working on analytics to predict conditions including heart failure, acute kidney injury, and sepsis. There’s always more work to be done, Kalkstein says. “No other sector has such a huge potential of being a force multiplier for the type of resources, algorithmic and financial, which I can invest.”