source

AI platforms can help optimize study design in clinical trials by mining thousands of related source documents for key actionable information. Study design focuses on trials intended to provide primary evidence of safety and efficacy (“pivotal” trials) and on regulations to permit substantial flexibility (“adequate and well-controlled trials”). Selecting a study design depends on the stated objectives, expectations, and practicalities, such as available population, site requirements, and the potential impact on medical practices.

“AI can simulate the entire control arm of a clinical trial using a previously collected dataset,” says Dr. Amy Abernethy, Chief Medical and Scientific Officer at Flatiron Health, a New York-based company purchased by Roche for $1.9 billion in early 2018. This is a very ambitious goal, and AI may not yet be ready for that to happen dependably, but the fact remains that study design is still a very human-driven process for obvious reasons. Presently, AI can help by aiding in the more labor-intensive part of this process. For example, it can help by mining source documents within the healthcare domain, finding answers to questions that need answering, providing added scope, depth, and breath to the research before a study is designed.

AI can help sponsors and sites to connect, collaborate, and begin conducting trials. AI facilitates enabling sites to create their own accounts and upload their personal data for populating information in study startup documents and speeding up different feasibility processes. Trials can be streamlined by helping study sites and the sponsor maintain awareness of who needs to do what, when, and track milestones, due dates, timelines, as well as any protocol deviations, adverse events, and other events of interest. AI can also learn from prior forms, documents, contracts, and budgets to tailor new materials for each site.

AI is being used actively to process and understand trial-related documents using NLP. Similarities and differences between different parts of the document are identified and used to build a database of reference text that can be employed in writing future protocols. On contract matters, AI can help identify and extract paragraphs or articles that address similar issues, such as liability and insurance. By identifying similar documents or parts of a document, the sponsor can quickly identify what parts of the document are common and what parts of a document need special attention e.g., legal jurisdictions.     

AI can help select subjects who have a greater likelihood of passing screening and a greater propensity to remain in and complete a study. “Worse than a site that doesn’t enroll patients is a site that enrolls a lot of patients and those patients don’t stay in the study,” said Dr. Stephen Wiviott, executive director of the Clinical Trials Office at Partners HealthCare. Building a patient profile based on reported data using machine learning and other predictive analytics techniques can help in this respect. Wiviott estimates that costs of conducting a trial from beginning to end can be cut by 90 percent if AI is applied throughout the clinical trial process, adding, “So much of this money is spent on humans checking other humans’ work.” Whether or not this level of improvement can be obtained today, it is likely that there is great promise of cost reduction for a typical clinical trial by using AI appropriately.

It is easy to extrapolate a trend from the examples provided above. AI can be a very effective tool in cutting through the labor-intensive, manually exhaustive, and often mind-boggling tedious work related to planning, designing, and executing clinical trials. But this is just a start. In the next segment of this blog post, we will explore some more advanced application of AI within the context of building a more robust clinical trial.  

A related area of intense interest where AI might prove to be effective is in building virtual patients. At the recently concluded World Medical Innovation Forum in Boston, panelists agreed that while it may not be possible for clinical trials to go completely virtual with AI, but offered that advances in AI could replace some subset of patients with virtual modeling, saving both time and money. It must be mentioned here that in many cases we still do not understand how side effects are generated, and we don’t even always understand the mechanism of action in the primary effect. But it may still be possible for AI to allow companies to get drugs approved with less medical evidence in humans, supplementing this by collecting evidence when the drug is on the market, according to Joseph Scheeren, senior advisor for R&D at Bayer. Colin Hill, CEO and cofounder of GNS Healthcare, believes that the future is headed towards virtually simulating drugs down to the molecular level, and using randomized clinical trials to confirm results of virtual experiments.

Another area where AI is likely to replace humans in the near future is data cleanup. Data are messy, and finding meaning amongst the noise and chaos requires a lot of manpower and, consequentially, money. AI can help people clean and analyze data in a smart and automated way—potentially speeding up the clinical trials process. Just cleaning the data after a trial takes one-to-two months, but AI can do it in a day, said Joseph Scheeren, senior advisor for R&D at Bayer. While within a broader context, this example is obviously one that is narrowly defined and most probably related to a very specific clinical trial, but the fact remains that AI has already started showing promises to speeding up data cleanup by a significant factor.

Keeping patients engaged is a huge problem for clinical trials. AI can help retain patients’ interests—allowing for real-time immersion and communication—supporting patient-centric trials. AI can be used to monitor and flag issues with patients’ treatment symptoms and managing medication intakes. Other technologies such as wearables can augment this by allowing patients to share information with researchers in real time, reducing or eliminating the need for patients to travel to sites, increasing patient adherence and compliance. AI can then process this huge real-time data and build intervention plans if a patient is at risk of dropping out of a trial. Reducing the frequency and length of clinical visits can lead to lower site costs and improvement in the quality of patient experience. It is also conceivable that once a model is established, it will be feasible to predict an individual participant’s or patient’s behavior without need for a large quantity of related data.

Disseminating outcomes of clinical trials is a big unresolved problem—a patchwork of confusion and wasted efforts—often leading to trials being repeated at huge costs to innovative companies trying to market their drugs or devices or treatments to patients who cannot wait. AI can help in this regard. King’s College London ran a machine learning project called Robot Reviewer, funded by the Medical Research Council (MRC) and others. The aim of this project was to develop a system that will automate bias assessment in systematic reviews. These syntheses will enable decision makers to consider the entirety of the relevant published evidence. The hope is that this will lead to a better overall meta-analysis for finding new drugs and treatments.

These are some of the exciting efforts in the AI realm that have the potential for streamlining the future of clinical trials. In the next segment of the blog post, we will explore some specific examples of these advancements and how they can be used to improve trial outcomes. 

Designing, conducting, executing, and reporting on clinical trials involves many complex issues. These issues include, but are not limited to, the duration of the trial, the burden the trial places on its participants, communication between healthcare providers and patients, scheduling, tracking, and reporting. These issues directly impact how a trial is planned, executed, and tracked. As data are collected, it then becomes important to generate analytics and share insight with sponsors.  It is also important to collect feedback from patients to make future trials better.

One of the keys in designing an effective trial is putting the patient at the heart of the overall plan. Although it seems commonsense, Patient Centricity is a comparatively new development. Not only does it enhance the patient experience, one study reported that patient-centric trials took almost half the time to recruit participants, recruited double the number of patients, and the study drug was 19% more likely to be launched. Improved patient centricity means more efficiency in conducting a trial and reduction of the overall patient burden.

So, when designing a clinical trial, what should we aim for? The Figure above illustrates that. We should have a design that reduces the time needed to plan and execute it and espouses better communication across the board between the different parties—patients, sponsors, and administrators, including the principal investigator and support staff. It should also help in reducing complexity in scheduling, tracking, and reporting.  Artificial intelligence (AI) can help improve performance in each of these areas.

AI is receiving considerable attention in a wide range of current applications. AI is the generic study of how machines can be made to produce behaviors that are associated with intelligence as found in nature. Machine learning, which is a sub-field of AI, focuses on generating systems that can improve their own performance over time, as well as the theoretical foundations and applications of computational aspects of these systems. Some examples of machine learning include reinforcement learning, neural networks, and evolutionary and swarm intelligence algorithms.

Commonly, machine learning algorithms are used to estimate a relationship between a system’s inputs and outputs using a learning data set that is representative of the behavior found in the system. This learning can be either supervised (with labeled examples) or unsupervised (without labeled examples), and it can be static or dynamic.

More specifically, machine learning is the process by which a computer can learn to do something—that something might be as simple as recognizing handwritten checks, or as complex as safely driving a car on its own. At the core of this technology are mathematical models based on foundations of statistical analysis, statistical model building, and probability. Recently, hardware designs have caught up with the intellectual concepts, making it feasible to implement large-scale AI systems. Large neural networks (deep learning, convolutional, recurrent, long/short-term memory) and other equally innovative and computational intensive model structures can be used to estimate complicated multi-dimensional data. For example, with regard to improving clinical trials, it is now possible to ingest data from hundreds of similar trials and generalize as to why some were successful and others were not. Other components of machine learning, such as evolutionary algorithms, swarm intelligence, and reinforcement learning can be applied generally across multiple data structures, including neural networks but also fuzzy systems, random forest classifiers, and others, to yield rapid insights that were not available 5-10 years ago. 

AI also involves natural language processing (NLP), a technology that allows a computer to convey an understanding of text and process that text for human consumption. For example, when details of a new trial are submitted to ClinicalTrials.gov—a repository of most clinical trials in the US—some of those details are in structured format, but others are not, such as eligibility criteria. If an eligibility criterion indicates that a pediatric study should include patients between 13 to 18 years of age, there are numerous ways of stating that information. Similarly, adult onset diabetes may be listed as Type 2 diabetes, Type II diabetes, diabetes mellitus 2, T2DM, or other options—each of which describes the same set of disease. Computers have to be programmed to handle the inherent variability of language to accurately model the variables and parameters of interest and NLP tools offer a solution to this complex problem.

Just as NLP is a way to model and understand human language, and ML offers tools to help learn models of complex data, predictive analytics combines these technologies so that we can predict how likely a particular event is to happen based on past data. For example, by looking at certain personal data, we may be able to predict if a trial participant will more or less likely to continue with the trial or drop out—a vital piece of information that can make the difference between success and failure.  We can also estimate and quantify the burden that a patient is experiencing in a trial and then use predictive analytics to address the “what if” scenarios that could be expected to lower that burden. In turn, predictive analytics can also be applied from the sponsor’s perspective to assist with study site selection, trial management, funding allocation, and other matters.

How can these emerging technologies have a positive impact on clinical trials? We will address some specifics and some practical examples in the next post.

Retention is a critical issue: 85% of clinical trials fail to retain enough patients. Timeliness is problematic: 80% of clinical trials do not end on time. Under-enrollment is common: 50% of sites cannot enroll enough patients to be viable. Furthermore, 30% of patients drop out of trials, and that’s just on average. In certain cases, the percentage is much higher. The entire arena of clinical trial preparation, execution, and reporting faces many problems, and we need to do better.

We leave today with another sobering statistic. Last year, over 1.7 million people in the United States were diagnosed with cancer for the first time. Around 10,000 clinical trials were run to test potentially life-saving cancer drugs, which ended up enrolling less than 5% of these patients. We must do better than this, but how? The $65B clinical trials market is screaming for a makeover. The question to ponder is: Can the adoption of latest technological advancements in artificial intelligence help alleviate these issues and help design, conduct, and report more effective clinical trials? That will be the topic of this blog in the next few weeks and months. Stay tuned!

Nex Cubed is a frontier tech investment firm that, among other things, operates an incubator for early stage technology companies. Kelsey shared the mission and vision for theNex Cubed incubator and explained the thesis around frontier technology and AI. The current cohort is nothing short of amazing. The companies are  are sequencing human movement (BioVirtua), accelerating clinical trials (Trials.ai), connecting brands with influencers (Endorsify), unifying data management (Mesh Candy), and automating retail pricing (RapidMathmatix).

For Trials.ai, tonight was an important coming out party. Throughout our incubation, guided by our expert mentors Tom Giles and Iqubal Hussain, we laid a plan that would enable Trials.ai to disrupt the clinical trials market. More important, however, is what we accomplished during our incubation. 

1. Customers. We closed our biggest customer yet and in 3 short months we booked more revenue than we had in the prior year.
2. We built an amazing team. In order to tackle the technical challenge we added four full time resources  to our engineering/product team.
3. We raised money. Prior to joining Nex3, Trials.ai was totally bootstrapped by customer and founder money. We closed a pre-seed convertible note round of $750,000 to enable us to execute on our plan. 

Kim, as always, delivered an excellent pitch. The crowd connected with our mission and approach to the market. Needless to say the next few hours were validating as many in attendance had experience with clinical trials and resonated with our mission and vision for the future. 

The Nex Cubed incubator was transformative for Trials.ai and we certainly couldn’t have gotten this far without them. Big thanks to the whole team at Nex Cubed as well as our fellow cohort.