Veradigm, formerly known as Allscripts, tried to outrun a series of crises with a rebrand. Now interim CEO Yin Ho has an audacious plan to use the $140 million acquisition of AI startup ScienceIO to tap into its one major competitive advantage over giants like Epic: data.
By Katie Jennings, Forbes Staff
What looks to be a small wine fridge is tucked away inconspicuously in the corner of Veradigm’s 12th floor office in the Flatiron neighborhood of Manhattan. But rather than bottles of burgundy, the tinted glass sides reveal a server rack. Nestled among fans whirring at high speed are three of the most powerful GPU chips on the market: Nvidia H100s. These three chips – plus a couple more in a closet down the hall – power the AI models of ScienceIO, a small startup that has become the cornerstone of a Hail Mary strategy for the struggling electronic health records company formerly known as Allscripts.
After reaching a $145 million settlement with the Department of Justice over receiving kickbacks from an opioid manufacturer, selling off its money-losing hospital business and failing to file financial statements for a year and a half, 38-year-old Veradigm was delisted from Nasdaq in February. The $820 million market cap company is still awaiting the results of an independent audit into a $20 million revenue misstatement, which it said was caused by a glitch in compliance software from an outside vendor.
Veradigm’s $140 million acquisition of five-year-old ScienceIO in March forms the centerpiece of interim CEO Yin Ho’s answer to the company’s woes: using Science IO’s generative AI models to give new life to Veradigm’s trove of medical data. “We moved quickly because we knew that there was going to come a day where we would have generative AI all around us,” Ho, 54, told Forbes. “And we would have no control over how to use those tools.”
ScienceIO’s models can reduce the labor intensive process of cleaning and structuring hastily written, abbreviation-filled doctor’s notes from months to days. Veradigm plans to apply them to three decades worth of data: 200 million patient records and 5 billion notes generated largely by independent doctor’s offices. Unlike other datasets that come from large academic medical centers in major cities, Veradigm’s data includes substantial swaths of the rural U.S. Ho said it represents “most of the footprint of America.” Veradigm plans to sell millions of de-identified records about patients’ health to pharma companies, who want to understand how people react to drugs outside of carefully orchestrated clinical trials.
ScienceIO was cofounded in 2019 by Will Manidis, a former researcher at cancer genomics company Foundation Medicine and a Thiel fellow. The 25-year-old CEO had raised $25 million from Quiet Capital, Section 32 and Lachy Groom touting an idea that predated the ChatGPT frenzy: the large language models that power AI chatbots could also be used to make sense of the all the clinical notes and records created by doctors and nurses.
ScienceIO’s language models, which currently range from 1 to 13 billion parameters, aim to turn that mess of information into usable, structured data. Unlike OpenAI’s models, which are trained on the entirety of the open internet, ScienceIO’s were all trained on healthcare specific datasets. “We realized really quickly that the only thing that matters for model quality is data quality,” Manidis told Forbes. “LLMs are just overwhelmingly a reflection of the data they are trained on.”
ScienceIO’s models paired with Veradigm’s greatest and largely untapped asset — medical record data rights — could prove to be a potent combination. Most medical records companies don’t own the patient data that flows through their software; their health system customers do. While they can often use de-identified patient data for research, one of the big challenges is how to link patient records between different hospitals without using any names or other identifying details. Electronic health records giant Epic confirmed it does not own any of its customers’ medical record data; Cerner, which Oracle bought for $28 billion, declined to comment.
Veradigm’s setup is different because it owns the data rights, said Manidis. Because of these rights, the company can link patients between different doctor’s offices over long periods of time. Once it has linked those records together, then it deidentifies the information to protect people’s privacy. “Our ability to link over 25 years of longitudinal healthcare data is unique,” he said. Veradigm does not currently sell identifiable data to pharma companies but could potentially get into clinical trial recruitment in the future, Manidis said.
Jeff Smith, a vice president and analyst of life sciences industry research at Gartner, told Forbes there’s “merit” to Veradigm’s attempted pivot, but was quick to caveat that “owning the data is only 60% of the challenge.” It also requires a combination of technology and healthcare expertise that’s rare: any company that could accomplish this is “kind of like discovering a leprechaun,” he said.
That data is potentially very valuable, because it could help pharma companies understand how patients react to drugs over time, in the real world. Recently the FDA has started allowing real-world data from patients outside clinical trials to speed up the drug approval process, which also has the potential to significantly reduce costs. The market for this type of data is expected to hit $7.5 billion this year, according to investment advisory firm Beecher and Co. and is growing at a rate of over 20%. Ho hopes ScienceIO’s AI models will be able to structure, de-identify and package Veradigm’s data to sell to pharma companies starting in 2025, projecting 10% growth in revenue by 2026, according to a presentation at the Barclays Healthcare Investor Conference. The plan is to first target three areas: cardiac, metabolic and central nervous system diseases.
Unlike many de-listed companies, Veradigm is profitable and has cash-on-hand – unaudited 2024 guidance is $620 to $635 million in revenue and $140 million in cash following the acquisition. But some analysts are concerned about the feasibility of Ho’s plan. “While we are encouraged by management’s positive outlook, we believe some will be skeptical of the achievability of these growth targets given performance in recent years,” TD Cowen analysts wrote in a note in March. Veradigm’s stock was flat as of Wednesday’s close since the news of the acquisition, and Ho and Manidis have been on a roadshow to try and convince bankers of the company’s AI-powered future. Veradigm, Manidis told the Barclays Health Conference in March, “is at a re-founding moment.”
“Our ability to link over 25 years of longitudinal healthcare data is unique.”
This isn’t the first time Veradigm has tried to reinvent itself. The company has been a Frankenstein of M&A activity over nearly four decades. Founded in 1986 as Allscripts, the original business involved selling pre-packaged medications to doctors to prescribe in their offices. A decade later, Allscripts was on track to lose $3 million on $75 million of sales. That’s when Glen Tullman, now a serial healthcare entrepreneur of Livongo fame (which sold to Teladoc for $18.5 billion) and investor at 7wire Ventures, took over.
Tullman sold off the profitable parts of Allscript’s business, including a pharmacy benefits manager, and doubled down on software for electronic prescriptions and associated data. He believed there was tremendous potential value in reselling patient data related to diagnosis, treatment and outcome. “The company that is able to collect this data will have the ‘holy grail’ of healthcare information to resell [to insurers],” Tullman recalled in a 2003 Harvard Business School case study.
In 1999, Allscripts went public and Tullman continued to buy and build additional functionality to make it into the electronic health records company that it’s known as today. But his last mega-deal, a $1.3 billion merger with hospital records company Eclipsys in 2010, caused a rift with the board. “The board started to fracture because I wanted to take the EHR to the next level,” Tullman told Forbes. “If you're a doctor and you see somebody who has a particular condition, the EHR ought to make you smarter — it shouldn't just be record keeping.”
Tullman was optimistic about the company’s new direction. “Veradigm’s [current leadership] is smart enough to say, let's pivot and use all this information,” he said. “That's what I wanted to do 15 years ago.”
Manidis is excited to train ScienceIO’s models on Veradigm’s data because it will include patients in rural and low-income geographies from small doctor’s practices that aren’t normally captured. He gives the theoretical example of a pharma company manufacturing the new weight loss and diabetes drugs known as GLP-1s, who wants to know how patients of different ethnic backgrounds in low-resource settings respond better to different drugs. “You can't buy that data set, it just doesn't exist,” said Manidis. But soon, with Veradigm, it will.
The decision to buy ScienceIO, rather than use them as a vendor, was a way to mitigate risk, said Ho. “Because we brought it in house, we're not allowing it to be learning off of any data that we do not control,” she said. In addition to Veradigm’s data, they have also incorporated data from disease registry collaborations with the American College of Cardiology and American Diabetes Association. The acquisition was structured as $96 million upfront and then $44 million paid out in installments over three years as an incentive to keep Manidis and ScienceIO cofounder Gaurav Kaushik at Veradigm for at least that amount of time.
There’s undoubtedly a need for more robust data in the fragmented American health system, where your primary care doctor might not be using the same medical record software as your hospital or specialist. That’s an issue for using real-world data for FDA submissions as well. For example, if a pharma company wants to know if a patient took a certain drug but the dataset doesn’t include key parts of their medical history, they’re out of luck. “It’s very easy to do it badly,” said Shirley Wang, an associate professor at Harvard Medical School who studies the practice. Without all the relevant information, then the data is “not really fit for purpose,” said Wang.
Privacy experts have raised concerns about how AI will impact health privacy. Once data has been de-identified in these massive datasets, it’s no longer covered by the federal healthcare privacy law HIPAA, said Deven McGraw, the former deputy director of health information privacy with the federal Department of Health. There’s a chance that AI models could re-link sensitive health data with people’s identities, and at that point HIPAA would not protect their privacy. Manidis said Veradigm does not train its models “on any data that is re-identifiable.”
For now, Manidis and Ho said they have started pre-selling data products to pharma customers and expect to have pilots up and running by the end of the year. “You will start to see real revenue and margin expansion in the 2025 timeframe,” said Ho. As for 2024? “It’s a year of investing and reinvesting in our company.”
MORE FROM FORBES
