While the retail, security and entertainment sectors dove in early with the adoption of artificial intelligence (AI) and machine learning (ML) technologies, the life sciences sector has “been a bit of a laggard, frankly,” said Pratap Khedkar, a principal at the Evanston, Ill.-based consultant firm ZS Associates.
The apprehension could be due to the mounting risk inherent in drug discovery and development, but a ZS survey of top pharma executives suggest that 70% think AI is a high priority, particularly at a time when the cost to bring a successful drug across the finish line has reached $2.6 billion or more.
Khedkar led a virtual discussion Tuesday during the first day of Ai4 2020 called “AI in life sciences: What’s working, what’s not, and what’s next.”
The fully digital three-day conference, focused on facilitating the adoption of AI and ML, was originally set to take place in Las Vegas this year until plans changed due to the COVID-19 pandemic.
Similarly, organizations employing AI/ML technology needed to pivot as well. At Brussels, Belgium-based UCB SA, a pilot program was on the path for a roll-out.
“We had a beautifully laid out plan. Then, the pandemic hit,” said Anita Moser, head of assets and optimization for U.S. neurology at UCB. “We had to hit pause a little bit and re-evaluate. Do we further invest and roll out or do we pause? How do we pivot? Our field colleagues, so used to interacting with our stakeholders, were no longer able to.”
Nevertheless, the company went ahead with a full-scale roll-out and adapted the algorithms for COVID-19.
Likewise, Omer Hancer, who leads ZS’s customer-centric marketing and digital analytics solution area, witnessed a client adapt a next-best action (NBA) algorithm to fit the pandemic. The algorithm enables organizations to predict with a high degree of accuracy whether a customer would be receptive to marketing efforts, enabling continuous learning, and replacing the 2,800 touchpoints from pharma each year. It is still marketing, but marketing in a smarter way when physicians or patients are ready for the information. The program went live in early 2019 and was 50% more effective than standard organic calls.
“We had no way of knowing that COVID was in the horizon,” Hancer said. “This particular organization was able to dial up or down certain content pieces and drove over 50% improvement.”
Instead of push marketing techniques of the past, reaching out directly to physicians, pharma saw the need to become more customer-centric, as customers have taken control of their health. AI/ML technologies, Khedkar said, help companies find what he calls the four Cs: the right customer, the right channel, the right content and the right cadence.
Aside from marketing, AI/ML can add value to life sciences not only in helping to create medicine, but also in discerning the correct diagnosis and treatment for each patient, as well as to help patients comply with their treatment regimens.
As of now, “it takes seven years and eight misdiagnoses before you get the right diagnosis of the patient,” Khedkar said, adding that “30-50% of the time the patient does not comply.” Pharma is applying AI to predict types of patients and when they will fall off therapy. “We can predict the patient, we can predict the reasons, and AI is giving us insight as to what can we do about it,” he said.
Other applications could involve the development of an intelligent screener to identify nonalcoholic steatohepatitis patients without a liver biopsy, or it could help in deciding whether to conduct a clinical trial at a certain location.
Data and people pitfalls
The potential is enormous, but there are a few things standing in the way of success. An evolving FDA requires the enhancement of human decision-making within the constraints of the technology. Real-world data also is a necessity. In order to find 300 features that can be placed within a model, the data needs to be robust.
“Data is woefully incomplete. It has to be patched together from different data sources and that degrades the data more and more,” Khedkar said.
And, of course, finding the right people to do the work is a tremendous challenge.
“Not all of our people are as advanced as our analytics,” Moser said. While the technologies of Waze or Netflix are relatively simple to understand, it is not the case within life sciences. Introducing all the analytics to someone at once is not a good idea. “You have to do it in a very gated way.”
Khedkar said 46% of pharma executives have noted a shortage of people with the skills to implement AI.
“The pilot is not about figuring out if the model works. It’s whether the people work with it. They actually have to trust it. The trust involves making sure your data sources aren’t biased,” he said. “In life sciences and healthcare, that is doubly important. It’s not just about accuracy. It is about explainability.”
Two of the biggest hurdles for adopting AI/ML are, in fact, the operating model and the people. Renewing the operating model and adjusting it to respond to change, such as with the COVID-19 pandemic, is essential to making the technology work. Both Khedkar and Moser discussed the resistance within an organization.
“We got caught up in the innovation that we forgot about the operationalization and further advancement in the future. It’s something that can be mitigated, but it’s something I wish we had put a little more planning into,” Moser said. Specifically, she would have liked to better understand the level of resistance and the motives for the resistance.
“Elevating your digital backbone is really quite important to scope out in the beginning,” she said.
Khedkar suggested divvying up the work between data scientists, domain experts, AI/ML administrators and AI/ML engineers, in order to bring the technologies to scale.
“If you diversify these people into four or five (areas), it’s easier to have them work together, rather than asking data scientists to do everything,” he said, emphasizing the value of being fast as opposed to being first.