Inside Precision Medicine

Published Online: August 18, 2022

Read the Full Article >>

Artificial intelligence (AI) has promised major breakthroughs in clinical development success. While the hype has sometimes seemed overblown, AI is starting to deliver on those promises through real-world impact. AI is not a silver bullet for all challenges, but in specific applications, AI methods can provide a real boost to ongoing efforts. Here are three examples of AI’s growing momentum in clinical trial design and execution.

Identifying the Right Patients

Challenge: In many cases, identifying the right patients for clinical trial participation remains difficult.

Finding the right patients for a trial can be a complex task. Inclusion and exclusion criteria can be difficult to map and implement using available datasets, especially when they involve parameters not easily accessible through standard coding. Beyond just qualifying for inclusion, patients’ utilization and adherence patterns matter, too – patients can and do drop out of trials for many reasons, some of which can be understood and predicted using data.

At OM1, we’ve developed a tool, OM1 Patient Finder™, that can help address these challenges. OM1 Patient Finder uses longitudinal health history data and predictive analytics to identify patients most likely to qualify for and participate in specific trials, including those who may have relevant but undiagnosed disease. This approach takes into account patients’ natural history of disease, clinical characteristics, and healthcare utilization patterns, all of which can provide insight into the likelihood of patient appropriateness for clinical studies. 

How does this approach work? First, OM1 Patient Finder determines a ‘profile’ of patients with appropriate characteristics for trial participation. Then this ‘profile’ is applied to broader populations, identifying those most likely to be best candidates. For example, if diagnosis is a challenge (often the case in rare disease), OM1 Patient Finder can help find concentrations of potentially undiagnosed patients; or, if ideal candidates are those with a history of poor response to other therapies for their condition, OM1 Patient Finder can use this information as part of its predictive ‘profile’ application as well.

Identifying the Right Sites

Challenge: Investigators often have incomplete awareness of which sites make most sense for trial participation. 

This application complements identifying the right patients for trial enrollment. Just as patients can have characteristics that provide useful information on their appropriateness for specific trials, site characteristics, like having the right set of associated providers and access to pools of eligible patients, can provide the foundation for an AI algorithm searching for the right sites for a specific trial. 

AI can be used as part of the site selection process to amplify site availability, with a similar effect to amplifying patient availability. Crucially, this approach can work synergistically with patient identification, sometimes resulting in both new sites and associated pools of patients that merit evaluation for potential inclusion.

Expanding Endpoints

Challenge: Using a limited set of trial endpoints can impede adherence, enrollment, and follow-through. 

Each disease has ‘gold standard’ clinical endpoints used in clinical development. Trials and candidates are compared using these validated endpoints to understand how different therapies perform. However, these comparisons can be severely limited in real-world contexts due to constraints on endpoint availability – outside the narrow boundaries of trial protocols, relevant endpoints are often simply unavailable. 

The iceberg analogy is a good way to visualize why expanding endpoint availability using AI can be powerful. Think of coded, recorded disease activity metrics as the tip of the iceberg, above water – these are the gold standard metrics most easily observable, like the ​​Expanded Disability Status Scale (EDSS) in multiple sclerosis, or the Patient Health Questionnaire-9 (PHQ-9) score for depression. Below that, under the water, are deeper layers of available data. The first encountered are annotated disease activity metrics that can be extracted directly from ‘harder to access’ sources, like clinician narratives.

Deeper still, we can use AI and clinical data to estimate endpoints when they are not otherwise recorded. In this process, powerful algorithms are calibrated using millions of datapoints to generate synthetic endpoints that correspond well to our ‘above the water’, gold standard baseline. This process can both increase the number of patients for whom data points are available, and amplify visibility into disease trajectories. Following our analogy, if we think about total data availability, most of it is here – beneath the water, accessible using AI methods. If these estimated endpoints perform well under rigorous evaluation, they can be considered for use in outcome studies alongside the ‘gold standard’.  

Looking ahead, AI will see more and more use across the spectrum of clinical development: everything from novel compound design, to in silico testing, to trial design and dynamic optimization will be enhanced using AI. Across these applications, the core principles remain consistent. The technology can be trained using gold standard data inputs, find patterns associated with optimal outcomes, and then apply those patterns prospectively. Through the use of artificial intelligence, there is a real opportunity to accelerate the clinical development process, and to gain insight into patient populations before, while, and after trials are completed.