Biotech Showcase

Part of the challenge in how clinical trial protocols are developed is that they are built on historical criteria, which are usually restrictive and/or tailored for specific populations. With a large pharmaceutical company, TriNetX developed an ML-based diversity protocol evaluation model.

Based on three existing protocols, the team trained a diversity optimization model which looks at different combinations of criteria and how they impact diversity and representation. The goal was an optimal set of criteria that did not meaningfully impact the profile of the clinical trial protocol but did consider different epidemiologies, demographics, and diversity inclusion (DI) recommendations.

Building these models into protocol development is a move in the right direction as it provides a data-driven approach to improving design and planning. Biopharmas can find patient populations that meet their diversity enrollment targets while not compromising the scientific criteria of the trial. As regulators increase the requirements for documentation, ML to predict different optimization scenarios will become more commonplace.

This AI modeling project targeted individuals from the general patient population at increased risk of developing pancreatic cancer that are currently being overlooked by traditional screening guidelines. These guidelines primarily rely on genetic or familial risk factors for eligibility, which account for only about 10 percent of the overall pancreatic cancer population. Early detection can lead to survival rates of over 80 percent within five years in some cases.

While many predictive models have been developed, this project was unique due to the large volume of geographically and ethnically diverse patients accessed from TriNetX’s network. Thirty-five thousand patients with pancreatic cancer and about 1.5 million without from 55 different health systems were analyzed. The model identified 87 top predictive features for Pancreatic Ductal Adenocarcinoma (PDAC) development from demographics, labs, medications, and diagnoses from EHR entries.

The next step in the project is currently ongoing and involves prospective model validation through deployment of the model on the data of approximately 6 million patients to tag those predicted to be at risk of pancreatic cancer within the TriNetX network. This project marks not only the development of a compelling predictive model built on a large swath of the U.S. population, but also, it explores how this model can be deployed into the healthcare system with prospective following of patients.

A second line therapy drug from a top pharma company was going into third line therapy for relapsed follicular lymphoma. The TriNetX team pulled a large electronic health record (EHR) data set to study the factors associated with transitioning from one line of therapy to the next.

The model showed that certain labs and abnormalities within those labs were significantly correlated to patient progression. Some patient and treatment pathways were also positively associated with progression. The team categorized these and then created a scoring algorithm for the company that showed the patients in their target population that will progress into the next line of therapy.