The Future Of AI In Post-Covid Healthcare

Kira Radinsky, co-founder and Chairman of Diagnostic Robotics, wants to make healthcare more affordable and accessible. The lessons learned from initial deployments of the startup’s AI-based digital triage platform in Israel and the U.S. and the valuable experience gained during the Covid-19 pandemic, point to a future of better healthcare: Providing the right treatment at the right time in the most appropriate setting.

At the Mayo Clinic, Diagnostic Robotics’ triage platform suggests possible diagnoses and provides a risk score for each patient based on their answers to questions regarding their medical conditions. The Mayo Clinic’s Dr. John Halamka calls it “Waze for healthcare,” stressing its use as a navigation system, matching patients with the right healthcare resource at the hospital’s emergency room or even before they arrive there. 

The State of Rhode Island has used Diagnostic Robotics’ platform to help manage its response to the Covid-19 pandemic. Answering questions on a Web-based questionnaire, users can make decisions about when to seek care and testing, connecting them to local information and resources. “Diagnostic Robotics is a leader in using innovative technology to ensure people get the right care at the right time – including during the Covid-19 pandemic,” said Rhode Island Governor, Gina M. Raimondo.

“We try to automate the process of getting the patient treated in the right location,” says Radinsky. The average emergency room wait time in the United States is about 40 minutes, according to the CDC, and millions of patients wait at least two hours to see a doctor. “Many patients likely without true emergencies are told to go to the emergency room by physicians out in the community,” say two Boston University researchers, adding that “the Covid-19 pandemic shows just how important a well-functioning emergency medicine system is.”

Getting to see a primary care physician in the first place, however, is a growing problem, even in the U.S. The Association of American Medical Colleges (AAMC) estimates that there will be a shortage of between 21,400 and 55,200 primary care physicians by 2033. Already in September 2019, 35% of respondents to an AAMC survey reported they had trouble finding a doctor in the past two or three years, a 10% increase from when the question was asked in 2015.

“We are trying to build the capacity of primary care physicians to treat more people, to increase the scalability of medical treatment,” says Radinsky. It is the right way to go about using AI in healthcare, “starting with something small and tangible.” This is in contrast to the noble but misplaced ambitions (and failed promises) to use AI in its current state in “moonshots” such as curing cancer. “Everybody is talking about the singularity. Let’s get to a point where AI can differentiate between a banana and a cat,” cautiones Radinsky.

In the same practical—and hype-resisting—vain, she is also adamant that “the computer will not replace the doctor.” Radinsky sees her digital triage platform as a decision-support system, helping better manage load and flow and process, with better outcomes. The system continuously learns, both from clinical results and from the feedback of practitioners. It’s important, she says, not just to make a prediction about the right course of action but also to be able to explain it to the doctors, a key to increasing adoption.

Making accurate predictions based on analysis of large data stores is something that 34-year-old Radinsky has been occupied with for a long time. After completing her Computer Science PhD at the Technion, Israel Institute of Technology, she worked at Microsoft Research, developing algorithms that leverage web knowledge and dynamics (e.g., changes in user search behavior) to predict future events. Based on the analysis of past and present news report, she and her colleagues developed a global warning system, most famously accurately predicting the first Cholera outbreak in Cuba in more than 100 years.

Given this track record, I asked Radinsky why no one has predicted the Covid outbreak. “We knew SARS viruses are getting worse, but knowing when exactly [the next outbreak will happen) was super hard—there was no existing pattern,” she says, adding that “one day we will have enough data to predict all of this.” By this she means having not only large enough number of observations, but finding out correlations between different data points and understanding the cause and effect mechanism behind what is studied. Predicting accurately the Cholera outbreak came after the analysis revealed that the probability of an outbreak of Cholera (a waterborne disease) in the event of floods, increases significantly if two years before there was a drought, specifically in countries with low GDP and low concentration of water resources, as access to clean water reduces the impact of the disease.

So we need not just a lot more data but also a similar casual model to predict accurately the arrival of the next SARS-related outbreak. Radinsky points to deforestation as one place to begin to look for cause and effect in the context of the coronavirus and similar infectious diseases, occurring when the virus is transmitted from animals to humans (see here and here).

There is only one silver lining of a calamity like the Covid pandemic—if you learn anything from it. Radinsky shares a couple of lessons from Diagnostic Robotics’ experience in assisting the Israeli government in monitoring the spread and managing the response to the pandemic. It is important to not just accurately make predictions but also to be prepared with a plan of action following the prediction. And, in contrast to the conventional wisdom, they found that people over 65 are very willing to use of digital tools. “If you just change the interface for older patients, you get numbers similar to younger patients,” says Radinsky.

The Covid experience also provided further insights into treating chronic patients and pinpointing when their condition is deteriorating, requiring intervention. The large data sets mined by Diagnostic Robotics provide a unique opportunity for conducting research into what has worked and what hasn’t worked in the past for a particular patient profile. “If we have a large enough sample,” says Radinsky, “we can mimic a controlled clinical trials.”

This is yet another way by which Diagnostic Robotics demonstrates the future of healthcare: Not only intervening before a patient goes to an emergency room by providing triage at home, but intervening six months or even a year in advance and getting a deteriorating chronic patient to see their primary care physician. It’s all based on what the digital triage platform has learned from numerous similar cases, from the medical literature, and from the experience of many physicians. Eventually, Radinsky believes, this will lead to the standardization of the healthcare system, like having expert doctors everywhere, no matter the location and the level of available resources. 

More of the future of AI in healthcare can be gleaned at Radinsky’s lab at the Technion where she is a visiting professor. She applies similar methods of analysis to medical texts and historical data sets to find out which existing drugs can be re-purposed to treat medical conditions other than the ones for which they were originally developed. For example, she and her colleagues found that proton pump inhibitors or PPIs (aimed at reducing the production of acid to prevent ulcers) and statins (lowering cholesterol) significantly improve the treatment of hypertension.

In a related, drug discovery research, the researchers analyzed the molecular basis for drugs that were developed to treat Tuberculosis up to the 1930s. The analysis produced new combinations of molecules, including 35 drugs that the FDA approved years later. One such drug was introduced in 1936 but was not used because of its side effects. The version of this drug that the AI system created was actually approved in 1952, becoming a successful treatment for Tuberculosis. The reduction of 16 years in the time it took to “develop” this drug points to the enormous potential of using AI to significantly cut the time it takes and resources required to develop new drugs. “This is the ideal that we as researchers would like to get to—automating ourselves,” says Radinsky.

Her multiple research, business, and public service activities have not prevented Radinsky from finding time to enjoy the company of her entrepreneur husband and their two toddlers. Her mother told her that “you can grow your career as much you want and still have a family” and the first time she heard the term “glass ceiling” was when she moved to the U.S. to intern at Microsoft. Today, she and her husband “share responsibilities and get a lot of help.” And she has come up with what she calls “tactical tricks” to balance it all: “I have one-on-one meetings not only with my employees. I schedule one-of-one meetings with my children as well, every day,” during which she is completely focused on them.

“My childhood dream was to push the boundaries of knowledge of humanity,” says Radinsky. Dream achieved, and more: Using AI to make healthcare accessible and affordable, pushing the boundaries of knowledge to improve humanity’s quality of life.