AI has been here for many decades, but the buzz around this latest wave of AI is louder than any I can remember. You can find articles about AI in every publication and medium. These articles deal mostly with broad themes, rarely delving into how AI might be applied in our work lives. There are plenty of stories about self-driving cars and the latest human expert to be outwitted by an algorithm. But what about the unsexy business that most of us find ourselves in? How can we apply AI to the clinical and commercial activities of pharma?
As a leader of the analytics practice at ZS, I have the good fortune of seeing many applications of AI across clients, industries and practice areas. In this series of interviews, I talked to thought leaders at ZS who have implemented and operationalized AI with our clients, so you can hear more specific examples of AI in practice. We’ll discuss how AI can be considered for product launches, clinical trials, analytics consumption, enterprise applications, user experience and real-world data. While not comprehensive, it’s a good start to a deeper examination of the role AI can play.
While pharma may have a reputation for operating behind the innovation curve, this recent wave of AI has many in the industry motivated to dive in. So where in pharma are we seeing the most traction? Where are the greatest opportunities for AI? And for those who are committed to embarking on this journey, where do they begin?
Naturally, I turned to the leader of the pharmaceutical practice at ZS to help me answer these questions, my good friend and co-conspirator on all things AI, Pratap Khedkar.
Q: Pratap, this is the third wave of AI over the past 25 or so years. As you work with AI in pharma companies, what feels different this time around?
A: There are three elements that are different. The first is simply the impact of better-performing AI, what it can solve for, what it can do. It’s noticeably superior. When I compare it to the second or the first wave, the wide range of problems that are falling to AI is markedly different from a broad perspective.
The second piece is that the data itself has exploded thanks to the internet, thanks to the fact that everybody’s got a phone, thanks to the fact that the paper economy is disappearing and everything is digitized—even digital healthcare records. None of that existed 20 years ago. Because the data was not in electronic form, the algorithms were essentially starved.
A third factor is that AI algorithms themselves have been improved. New ones have been invented, starting around 2004. So, over the last 10 or 15 years, we’ve seen a profusion of new techniques as well.
Q: Do you see applications of AI spanning many functions in pharma or favoring a few?
A: You can use AI for everything from producing the product, to designing the product, to manufacturing the product, to selling the product. And then there are underlying enterprise applications—HR, for instance. In these areas, I think what we’re finding is you can get good use cases if you look hard, but I would say that the two that are the most exciting now are R&D and commercial.
There has been a lot more excitement in R&D because there was a lot more data from clinical trials and consumer activity. But because of the focus on R&D, a lot of commercial uses of AI have been under-served. What the industry needs to do is fire up more uses of AI in commercial, where there’s more room for using AI to make good decisions, from micro-decisions to macro-decisions. You just need to gather specific use cases.
One small example: Perhaps I want to use AI to help a rep. It could be that a rep has to consume 40 reports. That’s a lot of information. If I can ask the AI to sift through these reports and give me the top three insights for a doctor for today, then that’s a very specific use case.
It takes years to see the impact from AI in R&D. In commercial, if you solve the problem, you can see the impact in three months. There is a lot of opportunity for AI to work its magic in commercial.
Q: What advice do you have for AI practitioners within pharma companies?
A: Don’t start with, "Oh, I need to hire 15 people and get a $20 million budget.” Start with clear, specific and even small use cases. What are the three, four, 10 use cases I can think of to start with? Maybe two or three good ones in sales, two or three good ones in marketing for a given brand, or maybe there are a couple of use cases for outpatient services. Go into manufacturing and R&D as well. Go pick out these use cases, brainstorm, do feasibility studies, do some “hackathons.” Without these use cases, your AI will not succeed.
Second piece of advice: Use cases cannot be solved without data. Now, in pharma, we’ve fallen into the trap of saying, “Well, data is only useful if it’s 90% complete and 90% accurate.” Pharma needs to get over that mindset because AI works when you have a lot of broad data. The depth and accuracy are not as important. Being quick and dirty is better than trying to be very precise with the data itself because when the data is a little off, the algorithm can compensate by being a little smarter.
And then the third piece of advice I would give you is around organizational mindset. It’s about getting the human to adopt, use, understand and trust this AI output. You can solve a problem well, but nobody uses the AI. That’s quite common, so how will you change human minds in your organization to use this AI output? Work on that, and put energy into adoption and change management from the outset.
We always tell our clients to start small with AI. As Pratap Khedkar shared in my first interview, it’s important to start with clear, specific use cases. For some aspects of the business, it may take some time to carefully consider what these specific use cases are and what opportunities AI can create. Clinical trials present no such difficulties. The opportunities to improve efficiencies are numerous and clear, and AI has begun to make an impact. The industry is aware of the inefficiencies that abound in clinical trial design and execution, and is eager to solve them.
To delve into what AI can do for clinical trials, I spoke with my colleague Venkat Sethuraman, who is the global clinical lead for ZS and sits at the forefront of AI adoption in the clinical trials space.
Q: Where do you see the most promise for AI in clinical trials?
A: The greatest potential I see for AI is in clinical development. You can use AI to design trials in an intelligent way and this is where you start to see a potential cure for clinical trial inefficiencies. AI can scan clinical trial data in literature in real time and give you a landscape view of how your trials are performing and evolving, and it can alert you to challenges as the landscape shifts. You can predict the potential impact of design decisions before making them. For example, how should you sequence indications within clinical programs to maximize trial success? Also, having such rich landscape data enables organizations to better advance clinical development plans that have been archaic and mostly static in nature. When there’s new external information and topline results, the view can be refreshed, and design recommendations can be made by AI, which can then be further validated by a researcher.
Q: How has AI changed the way that pharma companies are approaching clinical trials?
A: There are a lot of opportunities for AI in clinical trials, which are typically run and analyzed in old-school ways. For example, companies have access to vast historical clinical trial data. To gain insights, you still manually pull data from multiple clinical trial studies and translate the data so that it meets a single data standard. You may have some fields that say “gender” and others that say “sex,” and you need to resolve these variations when you pull the data together. We are working with clients to automate this process with AI and their efficiency rate has improved tenfold. Once you bring data into one common data model such as CDISC standards, you can gain rich insights. Companies need to trust AI to improve productivity, but it’s hard for them to believe that they don’t need humans to do this work.
Another area where AI can drive impact is in authoring clinical trial protocols or extracting value from a written protocol. Most companies have standard templates and processes when it comes to authoring protocols. However, “human variation” leads to review delays, amendment issues and increased time to train investigative sites. Algorithms can learn protocol writing from existing protocols. You can develop a process to structure 75% of a protocol with AI and thereby save at least four to six weeks from authoring and review. Most pharma companies have not leveraged AI in this space and that’s why it’s due for disruption.
All of this requires adaptation to new technology and a new way of thinking. Clinical trials have nowhere to go but up, so why not embrace this latest technology? Cars can drive themselves and many people have started to trust them. Pharma companies need to jump in and start experimenting with AI rather than questioning it and over-thinking how they may use it. That’s the advice I give most of my clients.
Q: In laying the groundwork for AI in clinical trial operations, what are the critical elements to consider?
A: There are three that come immediately to mind: diverse data, problem design and solution, and an adaptive mindset.
As is the case with most AI, it all begins with data. Most companies have historical clinical trial data, but today there is a wealth of newer sources, both structured and unstructured data, including real-world data (RWD), that is emerging as a valuable.
We have to leverage this data to say, “I know we can complete this trial on time with this global footprint and these investigators.” In addition, we should be able to say, “This particular trial will excite this investigator, and we know that there are patients in this ZIP code who need this particular treatment and, therefore, we can connect these two together.” The critical element here is being more specific and understanding what the trial requirements are rather than working on a one-size-fits-all approach.
And, finally, develop an adaptive mindset. Typically, companies plan for a certain patient population and then launch the trial. As the enrollment data emerge, we have to adapt models as well as our approach. Clearly, having real-time data and making frequent adjustments requires an agile mindset that’s lacking today.
The past few years have seen an explosion of data, and with it, the need to make use of it. Gathering and stitching together internal and external data sources presents one set of challenges, but what about the analysts who struggle to leverage all of this data? How can they keep up with increasing demands for data-driven insights from the entire organization? How can we empower the whole organization to have access to these critical insights? Here’s a hint: AI is part of the solution.
To delve into the role that AI is playing in analytics consumption, I spoke with my colleague Mahmood Majeed, leader of ZS’s business technology practice, who has extensive experience exploring innovative ways that AI can augment analytics consumption.
Q: As organizations become more data-driven, what’s the response you’re seeing to the increased demand for insights?
A: The increased demand for insights means we’re seeing more analysis shared, and in more forms than ever before. But with so much information coming our way, it’s hard to discern what’s useful and leverage it to drive decision-making. Searching through every data point and report that’s available wastes time and effort. The challenge that organizations face is to deliver the most pertinent insights in a timely manner and in the most suitable format with an amazing user experience.
Can someone converse with their device, asking it questions about performance to unearth the insights most relevant to them? Can they subscribe to insights in the form of a news feed the way we subscribe to information in our social channels? I see organizations leveraging new and exciting ways to do that with conversational analytics, AI and natural language processing (NLP). These technologies allow us to scan through data and extract insights through natural conversation and text, and offer specific and personalized suggestions. It’s getting more traction now because of disruptive NLP Technologies like Amazon’s Alexa, Google Assistant and Apple’s Siri as well as AI to decipher personalized and relevant insights from large volumes of data. If I’m able to integrate conversational tech into my personal life, why can’t I bring that to my work life?
Q: Where should companies be focusing their investments in the next few years?
A: Analysts are saying that in 2020, 50% of analytics queries will be generated via search, natural language or voice queries. This speaks to the user’s need for more consumable analytics. To meet that need, companies should be shifting now toward generating suggestions and insights rather than sending data and reports, but they need to go further and completely disrupt dashboards and reports as they will need to be augmented through contextual insights that are personalized to the user and the role.
Take salespeople, for example. The concepts of call planning, targeting, execution, performance monitoring, sales reporting and other forms of customer insights will all be combined into what I call contextual insights. They’re actionable, near real time and are embedded in the context of sales processes and fueled by relevant data from across the entire sales enablement process. It’ relevant and contextual to what the salesperson is doing at the time. It simplifies the complexity that exists today between 25 to 30 different systems and turns all that data into powerful, easy-to-digest, useful insights.
If routinely asked questions will be answered more effectively by leveraging machines, this will free up analyst capacity, which can be directed towards solving more challenging business questions. We believe that within 18 to 24 months, we’ll see a significant shift in the way information is consumed and decisions are made in commercial pharma. AI will work side by side with humans to bring this vision to light.
Q: If that’s the trend, how should pharma companies adapt?
A: Start small, and do simple experiments, proofs of concept or pilots at first. This is disrupting the way we work, so significant change management is required. The algorithms also take time to learn and to get better, so patience with such initiatives is a must. AI is less about moonshots and more about a collection of ideas that create impact.
The latest wave of AI continues to break across many industries, including retail, self-driving cars, agriculture and manufacturing. While tales of driverless cars are likely to catch your eye, there’s a less newsworthy but no less important demand for leveraging AI behind the scenes in the workplace. How can non-commercial enterprise functions improve efficiency and enable better decision-making?
To learn more about how pharma’s enterprise functions are adapting to this trend, I spoke with my colleague Shankar Viswanathan, who leads ZS’s advanced data science team in India and has been named one of the top 10 data scientists there.
Q: We often hear how AI can be applied in commercial, clinical or R&D settings across pharma. What applications do you see in other enterprise functions?
A: There’s interest and use across manufacturing, finance, HR and even functions like regulatory and compliance. Centralized groups like enterprise IT are looking to build cross-enterprise AI capabilities that can be leveraged by multiple groups within an organization. Some are looking at leveraging natural language processing and text-finding capabilities that can be applied to different types of unstructured data like operator notes in manufacturing and digital profiles of job candidates for HR. Also, you can use AI to analyze quarterly funding reports from finance or scan for adverse event signals for regulatory and compliance purposes.
Q: Is the aim of these AI programs different from what you see in commercial programs?
A: Well, there’s a stronger bottom-line focus in the key aims of AI in these situations. For example, optimizing operations, especially at scale, is important. In manufacturing, there’s a focus on using AI and AI-based systems to predict and mitigate lower-yield batches. Analyzing large volumes of data and sensor data, and combining it with unstructured notes from operators is a priority in manufacturing.
Another priority for AI in these functions is driving agility, again at scale. Finance teams across industries are looking to forecast short-term demand for stock. They’re often dealing with thousands of SKUs across many geographies, countries and regions in the face of both internal and external business changes. This, in turn, feeds downstream decisions such as inventory optimization and site-level manufacturing planning. This is why agility is key.
Another key aim is to derive insights efficiently to enable better decision-making, again at scale. In HR, AI-based systems pre-screen candidate profiles to predict their likelihood of success at an organization. Performance review systems are being augmented to mine feedback notes for talent across the organization and synthesize personalized insights for coaches to leverage with their direct reports. Intelligent automation systems are scouring through reports and prioritizing signals for human review.
In these enterprise functions, there is, therefore, a stronger emphasis on optimization, automation, efficiency and agility, all with a more bottom-line focus. That’s why enterprise IT groups are actively exploring AI algorithm-as-a-service options that can be configured for different contexts across the enterprise.
Q: Are there any key considerations for organizations as they build these types of AI capabilities across the enterprise?
A: Yes. Keep the focus on the human in the loop—the financial analysts, the HR recruiters and the manufacturing operators. Make the predictions and outputs of the AI systems clear and transparent to the humans involved. That’s critical for adoption and impact. People have a hard time trusting recommendations from AI without a clear explanation behind each recommendation. That’s why “explainable AI,” or XAI, is a hot area for research with significant focus and funding from organizations like DARPA. The future of human/AI collaboration requires transparent and clear explanations in order to engender trust.
User experience goes beyond simply developing a relevant and usable interface for software. UX is a discipline that requires a thorough understanding of users’ needs and the context in which they use technology. Whatever solution you may roll out to users, good UX is about meeting those needs.
With the proliferation of AI-driven solutions and proofs of concept, it’s easy to focus on the data science and forget that what you’re developing eventually needs to serve a person on the job. Within pharma companies, the finished product is often the visualization of complex data that appears in software on a laptop or mobile device. However, if users can’t understand these insights or the insights aren’t valuable to them, then the data science was a wasted effort. That’s where UX comes in. More specifically, that’s why UX should have come in a long time ago.
To better understand the role of UX in developing AI solutions, I spoke with ZS Principal Natalie Hanson, leader of our UX practice, whose team has helped develop multiple AI-driven solutions for clients.
Q: Can you share some best practices and lessons learned from visualizing data in AI-driven solutions?
A: It’s not that different from user experience work in any other complex domain. However, when you’re working with AI or machine learning, it’s important to work closely with subject matter experts and data scientists. In other words, experts can look at data and see something exciting in what they’ve found because they deeply understand the domain and the data. To design a compelling visual for data, we really need to deeply understand the story we’re trying to tell.
What’s more challenging about this kind of work is that the data visualization tends to be more dimensional. The data might be better viewed in 3-D or represented across time, for example. In these cases, we may find inspiration from a non-adjacent field, like a video game or the way weather is represented to consumers.
Q: When you think of AI-driven tools such as a recommendation engine, what are some of the key UX components you should keep in mind and how do these components differ from work on other solutions?
A: Again, I’d say there are more similarities than differences. On a recent AI project for a sales audience, my team helped understand how reps wanted to see and interact with the data. The people who were working on the solution were focused on extracting value from the data we had, and that was a hard problem. But then we had to figure out how to share those valuable insights in a way that would be useful to this audience.
The data scientists and machine learning engineers didn’t know anything about what a salesperson’s life is like. We can’t expect them to make that last-mile jump between all this amazing data and how it should be visualized and organized, so we looked at where the reps were in their daily lives when consuming data. Having a good understanding of the context was key because that affects how the data should be represented.
For example, we had a client who had all these amazing, robust, sales-related data points and they had pushed it to iPads for their reps. But the tool, when teams logged into it, started them at the national level, so they had to drill down to a region, drill down to a territory and then drill down to a physician. They had this massive amount of data to wade through, and they’re on iPads, sitting in a parking lot. Suddenly, it becomes onerous to make use of that data in that context.
We saw all kinds of crazy work-arounds for that. Reps were doing things like opening the physician view and grabbing a screenshot and saving it to photos or printing it out and putting it in a binder. Basically, what they were saying is, “The data’s good, but it needs to be more accessible at this point in my day.” Our job was making sure that we were providing the right information at the right time in the right way.
We also look at what we call information architecture or taxonomy. How do the reps think about these kinds of insights? How can we organize the information according to their mental models? And we also consider what we call progressive disclosure. That’s about how much information is really needed at the outset before someone wants to dig deeper.
Q: What are your thoughts on new innovations like conversational AI, which lets you hear spoken insights, or augmented reality (AR) and virtual reality (VR), which let you consume information in 3-D?
A: A lot of this has to do with users, and context is key. What we saw with reps, for example, is that they don’t want to talk to a chat bot because of privacy issues. They can’t be in a public place with their phone saying, “Hey, tell me about physician so-and-so,” and having it talk back to them. That would be disclosing their physician’s information. Also, many reps have security protocols on their devices, so we thought about helping to make optimal use of their time in the car, but many of the reps can’t use their devices while driving.
I think we’re just in the infancy of exploring what AR and VR might do. For example, if you’re a rep serving a large hospital and you’re trying to build new relationships, how do you find a physician, a head of purchasing, a head of surgery? Could there be an AR solution that helps them locate these people?
What’s exciting and challenging about creating solutions in augmented or virtual reality is we have to design for multiple senses in a way that we haven’t done in the enterprise context before. We’re drawing on design that may be familiar to gamers. For example, sound plays a more critical role, as do sensory cues like vibrations. The use of these different elements has to be done in a thoughtful way that informs and guides the user without overwhelming them, and in a way that’s appropriate for the context of use.
But it will be a long time before users will want the next generation in data visualization. When you think of the era of mobile reporting that started around 2013, we’re just now at the point where users have a strong point of view about how they like to see data and what helps them and what doesn’t. I don’t know if there’s any kind of guidance on how long that evolution takes, but I would say that if you want to think about bringing AR, VR or other innovations, be practical and realistic about your audience’s readiness and how quickly they will adapt to it. What might make perfect sense for a young patient may not work for a seasoned sales rep operating in a hospital setting.
Real-world data (RWD) is as old as data itself. It was relatively useless, however, until it was digitized. And then once it was digitized, we struggled to analyze so much data at scale with any kind of efficiency.
Enter our ability to store and analyze large volumes of data and advancements in data science, algorithms and technology. With the help of AI, it’s now possible to truly leverage RWD and begin to understand patient populations in holistic ways that have so far eluded pharmaceutical companies. It’s accelerating the use of evidence across healthcare to support the focus on patient health outcomes.
To better understand the relationship between AI and RWD and its potential, I spoke with Principal Asheesh Shukla, who leads ZS’s real-world data and patient insights practice, and has worked extensively with AI.
Q: What’s your definition of RWD, and what’s its current state across pharma? How do you see AI playing a role?
A: RWD is any information that’s generated by the healthcare ecosystem as healthcare is delivered to patients. That includes data from outpatient settings, hospitalizations, clinics, pharmacies, labs, and increasingly wearables and other devices. Regulatory and service authorities like the FDA or CMS have adapted well to the digitization of healthcare, and they acknowledge the many opportunities that RWD has opened up for healthcare delivery.
I think the overhype behind RWD has died down now and we’re shifting from post-disillusionment to becoming more informed. We’re aware of the problems we may encounter with RWD. For example, we lack universal identifiers to help us distinguish patients so we can match them between data sets. We need to improve data capture accuracy and consistency. We need better and more consistent population coverage. And that’s where the story converges with what’s happening on the AI side. As the storage and computing costs continue to shrink, many of these AI methods that have been around for a while are making a resurgence. We’re working around the known issues with data quality and still generating relevant insights. We’re predicting events with increasing precision and relevance.
AI methodologies are hungry for data. Machine learning requires us to deal with a large representation of what happens in real life, and there are very few examples of such data sets in the marketplace, so machine learning has to discover connections across data sets and corelate events in patients’ treatment journeys, continually making refinements. That’s where AI is most effective and RWD offers the best fodder for AI methods.
Q: How can companies leverage AI in real-world data?
A: I think it’s worthwhile to first look at where RWD can be leveraged. Historically, real-world data was used to develop what we call “potential value.” For example, which disease areas should we invest in? What kind of product profile should we target? It was also used to validate and prove the value of a product to various stakeholders such as regulatory authorities, payers and providers.
Now we see how AI can be used across the pharma value chain. We started with potential value in research, then we moved on to proving the value during product launches. Now we’re seeing value prediction in clinical development. We’re seeing AI in trial design, feasibility and execution. Everything we’re doing on the clinical R&D side incorporates more real-world data, like site selection, enrollment predictions, protocol design and so on.
The most exciting use of RWD and AI is in providing value to patients. In pharma, we’ve been hearing about value-based contracting, targeted patient services, and looking at identifying untreated and undertreated patients with slow disease progressions. You can use RWD and AI to make better patient event predictions and provide more timely, targeted patient assistance. The patient journey is analyzed to make next best action and next best customer suggestions to the field so that reps’ conversations with physicians stay relevant and value oriented.
Where AI really proves its worth is with prediction and providing value across healthcare, including with clinical decision support. Prediction problems work nicely with AI methods and techniques. AI is also used for tracking value over a longer period in longitudinal studies, looking at longer duration and demonstrating therapy outcomes to the payer population, providers and regulatory authorities. That typically requires tracking a large volume of data over a long period of time. And those are the two areas where AI is playing a critical role because the traditional techniques for data analysis just don’t work as well in terms of the base they can operate from, as well as the cost and the time it takes to deliver results.
Q: What sorts of success or progress should companies expect with AI and RWD? Should they anticipate quick ROI or a long learning curve?
A: I think the answer is both. Pharma has always run campaigns for physicians and detailing for physicians. We’ve had campaigns for patients. We’ve run patient services. There are these existing processes, services and functional areas that have been around for a long time. We believe you can improve and reimagine these areas from multiple dimensions by infusing AI and real-world data at the same time.
For example, in clinical trial design, we’re not only seeing RWD and AI methods that process and mine patient data, but we’re also seeing clinical trial feasibility predictions based on historical and current patient data available around selected sites, recruitment viability, coverage and other parameters.
On the commercial side, next best actions and next best customer recommendations for the field are becoming prioritized programs where we use AI and RWD. We take real-world data and use AI methods to track a patient’s journey and make predictions about potential events and confidence associated with the predictions informing sales and marketing decisions. In the pilots of these solutions, we’re seeing companies have a lot of success with very good, even double-digit ROI, both in terms of sales as well as significant cost savings from targeted efforts of up to 30%.
I also believe we should look at ways that real-world data and AI can create disruption outside the boundaries of existing processes. For example, historically, post commercialization label expansion triggered a post-marketing clinical trial. These are expensive and difficult to execute. Now we have trials generating real-world evidence that can complement or in some cases replace them.
Reimagining existing processes and disrupting old processes should happen in parallel. We always recommend that clients focus on existing processes first because that’s what fuels and keeps the momentum going in the organization and gets the funding required for them to experiment on the disruption side.
Q: What does the future hold for AI and RWD?
A: The future, as I see it, is in the ongoing expansion in healthcare from disease management to wellness management. RWD is population-focused right now, but can it also be more personalized? So far, there has been an either-or mentality, partly driven by privacy concerns, lagging regulations and data veracity. As we get precise and consistent in our ability to leverage real-world insights for key decision-making, the trust in RWD and AI methods will increase. As we’re able to combine determinants of health beyond medical data and be inclusive of genomics, social determinants, environmental and other data sets, our ability to personalize prediction will improve. Can we shift to a disease management model where we’re not responding to illness but predicting it and prescribing therapies to avoid the onset of disease? To me, that’s always been the true promise of AI and RWD.