A learning health system paves the way for precision medicine
Drugs and medical treatments are typically designed for the “average patient.” Individual patient variability is usually not taken into account, sometimes leading to undesirable clinical outcomes for patients. The goal of precision medicine is to address this unmet medical need by finding the right medication and the right dose for each patient every time. Precision medicine takes into account individual variability when optimizing a patient’s treatment strategy.
The primary goal of precision medicine has not yet been realized. We are still unable to effectively predict treatment regimens that will maximize treatment benefit in different patient populations.
For example, an underweight neonate with poor kidney function does not fall into the spectrum of what is clinically defined as "average", and simply accounting for simple demographic characteristics such as the patient's age and body weight may lead to suboptimal outcomes.
A data-driven framework that creates a mechanism to continuously improve and tailor treatment strategies in real time may be the catalyst to address this knowledge gap and bring precision medicine to the forefront of healthcare.
What is a learning health system?
A learning health care system is one in which science, informatics, and incentives are aligned for continuous improvement and innovation, with best practices seamlessly embedded in the care process. This system would bridge the knowledge gap, so that precision medicine may live up to its promise to simplify and personalize the delivery of complex medications.
Why is a learning health system important?
Clinical trials today are designed to test the efficacy of a certain dose of medication in a group of patients that are deemed “average". Once the drug enters the real world, however, it is often used to treat patients that do not resemble the patient population studied during the clinical trial. Differences in body composition, age, lifestyle, or other environmental and genetic factors may have large effects on how the drug works in the body. Because many clinical trials do not take into account these differences, the true variability in the drug response has not yet been realized.
Understanding drug response variation through clinical trials is like trying to map the night sky with a telescope, looking at only one tiny part of the sky. Yes, you will get some idea of the types of stars and planets that exist in that space, but the sample represents only a tiny fraction of the overall reality. To get a full picture, you would need to look at the night sky several times, in different locations and at different times to get a truly accurate representation of the universe.
Similarly, a clinical trial only gives us a narrow view and understanding of the spectrum of drug response in a specific population. It’s a good start, but by no means a complete picture in terms of fully understanding how drugs may affect patients in broader populations. For a more comprehensive view of drug response, a drug would need to be tested many times in various populations, but this approach would be costly and time-consuming.
A learning system will take advantage of data that is collected within the healthcare system (e.g. electronic health records (EHRs)). Improved data access along with machine learning and other mathematically driven approaches will allow us to learn about drug response characteristics in patient populations that have not been well studied.
For example, we can uncover new predictors of drug response such as a genetic factor in a liver enzyme presumed to influence drug clearance, allowing us to accurately adjust dosing and treatment strategies for patients that carry a mutation for that enzyme.
What makes a learning health system possible?
While the fundamental concept of a continuously learning system in healthcare has been around for some time, only recently has advances in technology and attitudes towards our health system aligned in such a way to make it a reality. Here are a few reasons why a learning health system is now possible today:
Rise of electronic health records
More than 10 years ago, three out of four doctors recorded patient data via manual data entries onto color-coded pieces of paper. Now, approximately 90% of office-based physicians have adopted an EHR system. This allows the consolidation and access to data with relative ease.
The rise of cloud computing
The cloud is by far the best medium for predictive models and algorithms to exist within healthcare. Once models/algorithms are updated, they can be quickly deployed to benefit clinicians and impact patients in different treatment settings around the world.
The shift towards value-based care
Healthcare in this country has been trending towards a value-based system that ties economic benefit for providers/payers/pharma companies to the actual therapeutic benefit for patients. This has introduced a macro level economic force to continuously improve the way we treat patients.
Better algorithms and predictive models
This is in part due to the enormous amount of data that we can generate from the healthcare system, but it is also a reflection of the progress that has been made in the field of machine learning in recent years.
A learning health system in action
The ImproveCareNow (ICN) network presents a great example of a learning health system. ICN began in 2007 with the goal to transform the health, care, and costs for all children and adolescents with Inflammatory Bowel Disease (IBD). With funding from the National Institutes of Health (NIH), ICN has evolved into an enduring learning health system that harnesses the inherent motivation and collective intelligence of patients, families, clinicians, and researchers, accelerating innovation, discovery, and the application of new knowledge.
One success story by the ICN group was the thorough evaluation of anti-tumor necrosis factor-a (anti-TNFa) agents in the management of pediatric Crohn’s disease. Previous studies showing the benefit of the drug had only been conducted in adult patients, and not in children due to time and cost.
The ICN study analyzed 4000 pediatric patients suffering from Crohn’s disease from 35 separate practices and classified each patient as an initiator or a non-initiator for the anti-TNFa therapy. The evaluation showed that patients treated with anti-TNFa had higher rates of clinical remission.
The utilization of ICN’s continuous learning health system allowed generalization of likely benefits to patients. By studying real-time data from a wide variety of cases, ICN was able to avoid strict inclusion criteria and yield relevant results for a patient population not formally observed during the clinical trial. This is a concrete example that shows the value of a learning health system in observational research to uncover new clinical knowledge.
How InsightRX works toward precision medicine
Much of my work at InsightRX has been devoted to creating a learning health system as it relates to treatment accuracy. We deploy predictive models via an EHR integrated framework to help clinicians and clinical pharmacists individualize treatment and dosing. However, treatment accuracy and precision do not end there. Over time, the data that we collect are then used to improve the predictive accuracy of the underlying models and algorithms. Specifically, we investigate drug response characteristics in patient populations that have not been well studied and explore the effect of other clinical factors such as genetics and disease type.
The new knowledge that is gleaned from this “re-learning” step is then deployed back into the cloud-based tool to further improve treatment accuracy for clinicians.
Although we have ways to go when it comes to developing a fully scalable learning system across many therapeutic areas, the promise of precision medicine will be realized through this kind of framework. We hope to continue making strides toward this goal.
(2) Ramsey L, Mizuno T, Vinks A, Margolis PA. Learning health systems as facilitators of precision medicine. Clinical Pharmacology & Therapeutics 2017; 101(3): 359-367